University of Florida Digital Collections Home
UFDC Home myUFDC Home  |   Help 
<%BANNER%>

Statistical Methods in Data Mining of Brain Dynamics

Permanent Link: http://ufdc.ufl.edu/UFE0022141/00001

Material Information

Title: Statistical Methods in Data Mining of Brain Dynamics
Physical Description: 1 online resource (83 p.)
Language: english
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2008

Subjects

Subjects / Keywords: autoregression, cointegrated, eeg, electroencephalogram, phase, synchronization, var
Industrial and Systems Engineering -- Dissertations, Academic -- UF
Genre: Industrial and Systems Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: This study discusses statistical approaches used for data mining of multichannel electroencephalogram recordings. Such recordings represent massive data sets that contain hidden patterns of complex dynamical processes in the brain. Formally, multichannel EEG can be viewed as a multiple time series, and therefore, a natural idea for summarizing such data is to utilize autoregressive modeling of multivariate stochastic processes. In particular, we thoroughly discuss various concepts and approaches related vector autoregressive processes, including stable stationary VAR models of order p and nonstationary systems with integrated and cointegrated variables, as well as procedures for estimating parameters of the systems (e.g., order, lag, or cointegration rank). The work highlights some stability issues that may arise in the application of vector autoregression to mining EEG data, and questions the applicability of Granger causality in the frequency domain to multichannel EEG. Synchronization has been found to be an important characteristic of the abnormal brain dynamics manifested by epilepsy and Parkinson's disease. We review two approaches for extracting the instantaneous phase from time series. In this study, we generalize the concept of the phase synchronization, and propose a novel approach based on multivariate analysis via modeling cointegrated VAR(p) processes.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Thesis: Thesis (Ph.D.)--University of Florida, 2008.
Local: Adviser: Pardalos, Panagote M.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2008
System ID: UFE0022141:00001

Permanent Link: http://ufdc.ufl.edu/UFE0022141/00001

Material Information

Title: Statistical Methods in Data Mining of Brain Dynamics
Physical Description: 1 online resource (83 p.)
Language: english
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2008

Subjects

Subjects / Keywords: autoregression, cointegrated, eeg, electroencephalogram, phase, synchronization, var
Industrial and Systems Engineering -- Dissertations, Academic -- UF
Genre: Industrial and Systems Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: This study discusses statistical approaches used for data mining of multichannel electroencephalogram recordings. Such recordings represent massive data sets that contain hidden patterns of complex dynamical processes in the brain. Formally, multichannel EEG can be viewed as a multiple time series, and therefore, a natural idea for summarizing such data is to utilize autoregressive modeling of multivariate stochastic processes. In particular, we thoroughly discuss various concepts and approaches related vector autoregressive processes, including stable stationary VAR models of order p and nonstationary systems with integrated and cointegrated variables, as well as procedures for estimating parameters of the systems (e.g., order, lag, or cointegration rank). The work highlights some stability issues that may arise in the application of vector autoregression to mining EEG data, and questions the applicability of Granger causality in the frequency domain to multichannel EEG. Synchronization has been found to be an important characteristic of the abnormal brain dynamics manifested by epilepsy and Parkinson's disease. We review two approaches for extracting the instantaneous phase from time series. In this study, we generalize the concept of the phase synchronization, and propose a novel approach based on multivariate analysis via modeling cointegrated VAR(p) processes.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Thesis: Thesis (Ph.D.)--University of Florida, 2008.
Local: Adviser: Pardalos, Panagote M.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2008
System ID: UFE0022141:00001


This item has the following downloads:


Full Text
xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E20101109_AAAABX INGEST_TIME 2010-11-09T20:40:03Z PACKAGE UFE0022141_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
FILE SIZE 6951 DFID F20101109_AABJUV ORIGIN DEPOSITOR PATH kammerdiner_a_Page_39thm.jpg GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5
74ce947e51710e6571503a29119378f9
SHA-1
d6ae64827887abd1199604a366925e8c5796188d
25271604 F20101109_AABJPX kammerdiner_a_Page_57.tif
0b85c66d9117759a745c04b4d562a385
8bba2ac44352fa4cb378e7a031201b2540c34beb
28189 F20101109_AABJIC kammerdiner_a_Page_42.pro
911dcce953fa0986137e9c6e063bd5ce
abeb9a18a75c9778a1bc6c93db9d544c89c95294
31332 F20101109_AABJDF kammerdiner_a_Page_29.QC.jpg
91c613bc3b4feae1fa9e23c54b67d546
11b5710f55e270bc0d2245d7a00d682cdb8768a4
5286 F20101109_AABJUW kammerdiner_a_Page_42thm.jpg
2e484a471c2dc39a0df818ad3dc44bc4
abf00547657909c7ca30a750043590d126abf67f
F20101109_AABJPY kammerdiner_a_Page_60.tif
75474e8f5b9677e75aa756a168daf8ab
e6ec6769cf74788abc096bfe1c557b3250101a8f
56973 F20101109_AABJID kammerdiner_a_Page_61.pro
426acb2f62335a83e9920be00159531a
4b13e166843438a186ba5c0372fee762e62d3a85
42570 F20101109_AABJDG kammerdiner_a_Page_83.jpg
1d9b14ef2610292d6bd596b5213d1824
d7f3de165f1365400e620da91b71959006633902
115231 F20101109_AABJNA kammerdiner_a_Page_54.jpg
27d009818dff24ebe77c17a4cab84c77
0f7d64bf9b3ba7638f65958fbd93ca095e5ec1e8
6187 F20101109_AABJUX kammerdiner_a_Page_45thm.jpg
0441cd52b1b59506158aa951ffbc8aea
a80c38c86ef66ad37f3d66c39fc57cd0650d37cd
F20101109_AABJPZ kammerdiner_a_Page_64.tif
89a0a4fd68b3359d9be4c8c5e1444949
42d8f77c365ba4eb7ea49528c8a81ae2b0323c27
31711 F20101109_AABJIE kammerdiner_a_Page_22.QC.jpg
006786edd0e0def424a698029c93d6ba
05d2b2a618ee1ca5f95e0ab0ff0efb9b56abafee
25399 F20101109_AABJDH kammerdiner_a_Page_07.pro
083209cef46c6405a63a18541f94bd22
4c9d449f898dceb5efff496928516ae04da99b3e
98709 F20101109_AABJNB kammerdiner_a_Page_55.jpg
6428a97c4e3ea28cb3d02cb9c6b3c281
b0ca8d64163b52af4c759d490e565dd0e70d96ce
4821 F20101109_AABJUY kammerdiner_a_Page_47thm.jpg
b3bc156ce9c2b219ddb73a64fe9e6fbf
eb2f2ecde4217dcdd8165d8c5d0d923cdfa79a8b
7952 F20101109_AABJIF kammerdiner_a_Page_28thm.jpg
8f7f39c6f5ed86b6515bd97d5cc61231
17ebd493ce46efaa64e970a2dcf8ebc57fc287b0
85852 F20101109_AABJNC kammerdiner_a_Page_56.jpg
b7360afb4c22c21b5074973f26031bc4
85c51e1a91a901ecd43f4570d3d9cff9f584f90d
2286 F20101109_AABJSA kammerdiner_a_Page_24.txt
d183ba8c5c1c58f10f6ec54f262c748b
1b88979f1e66e2714aed59289f63d1ab2a1f8090
F20101109_AABJIG kammerdiner_a_Page_40.tif
6614166abfc605980e6db419f476614a
0b5e515c295e496e9e4d7eca178c6615ccdbcc12
8483 F20101109_AABJDI kammerdiner_a_Page_79thm.jpg
cacbc33125a67a75fa622c9f23984f10
dc3cef15c2444e17b9ca07651894eb29edc7b52f
107847 F20101109_AABJND kammerdiner_a_Page_58.jpg
9d9c860c3c4859b83bc30c51a6770de9
b0c375c40e4152ac40c4fe132cdc1a78c0cf08fd
20210 F20101109_AABJUZ kammerdiner_a_Page_49.QC.jpg
6f7969605669edb39e941c6987822b55
aa9b7d94fc31ee451faad58e5cefdd76cf399572
2414 F20101109_AABJSB kammerdiner_a_Page_25.txt
f3c983873d49225e9ff7687402e7c73b
f905df5161014fa34c766e2ee7fd62b44354259f
847978 F20101109_AABJIH kammerdiner_a_Page_48.jp2
961804920ac79eb47f41e90eeefad998
7cb4dce7bd4ecc4345cbc19144e35953b2e9cb9d
F20101109_AABJDJ kammerdiner_a_Page_68.tif
b43cd7f09a7aa75ac67be35ab349d6f8
dc1e639f526ffd860a42cda6e7c38160d933e6c9
33730 F20101109_AABJNE kammerdiner_a_Page_59.jpg
06b1bc4b4ebeb6475c5e88e0f75d88a9
83f16e463df653962144db5ef6f80a5395c7647c
1361 F20101109_AABJSC kammerdiner_a_Page_26.txt
0c2b014e37b931ae4ba6e34277304372
11c6fabafd3f0ae4d971bba373af8b61df9fb19a
2301 F20101109_AABJII kammerdiner_a_Page_16.txt
658dbd0c8ededf839941a7b71976e91c
6ccba1987c035bce96c45f21c4c2910f8a2e9560
2253 F20101109_AABJDK kammerdiner_a_Page_12.txt
3906b7666dcd6e9fcb38db5f5ef9e23f
c3d42450f8b6a369cb6478ca09de1aef58f2dcc2
84126 F20101109_AABJNF kammerdiner_a_Page_65.jpg
4e24b5331a37abce2a7db773f7d6bb38
0fd2c9f7124b687fb99abe80056a88e6d0b97a9a
208 F20101109_AABJSD kammerdiner_a_Page_27.txt
858b614bc1e2854c495dcc38bcf49f6b
3a5eee5665f84257724d62ae35175f025e3fe960
15895 F20101109_AABJIJ kammerdiner_a_Page_07.QC.jpg
a359909bf41fa5e54ce689ef1b1c3d4d
34ea6cf7076760f22f298a71ce6f1e315a3068e4
1053954 F20101109_AABJDL kammerdiner_a_Page_67.tif
fc5d36ea3287026812c2a90ddb102cd1
0502cf8fa60a35042b5de79d942555d60070d8ed
95670 F20101109_AABJNG kammerdiner_a_Page_66.jpg
62157bb45f6b5518b32660547dfc5880
4f6c617dcbefc2abf6b722a2a42a652014fa221d
2026 F20101109_AABJSE kammerdiner_a_Page_29.txt
072e59fea2a8a54df7df06df899f6e88
e95c3172138827306d20206da3b2973caf0fe462
130733 F20101109_AABJIK kammerdiner_a_Page_79.jpg
f2ff59d366b19ee1df43894ff9055cbd
56b38ca7cbee8e43ea816eb3f244fc9b678076b8
1051980 F20101109_AABJDM kammerdiner_a_Page_36.jp2
04a8347f1f24d3c514455c4d61d97a69
dde4e6987e7377f2a3438249b61ddff9daad4879
128138 F20101109_AABJNH kammerdiner_a_Page_70.jpg
485c9ce52c8cacb0f53c4c49e5ea1e4b
f00a5f2323d1fa74565dd6d2bffdbb090469be9e
2069 F20101109_AABJSF kammerdiner_a_Page_30.txt
0d0746123094f03a123564a006d9cdf9
2f22232e798aa719c7ec6fd6c0db027c3f22ec8e
1031417 F20101109_AABJDN kammerdiner_a_Page_57.jp2
b9a8fd28d3fff276a0e7aa1cb17f94cc
f5b645b379823bb7b5803ecea60d77472d2a87ee
113510 F20101109_AABJNI kammerdiner_a_Page_75.jpg
cf4c6b8534a1ac97de2176802b9d18d3
1d6050b39fc110948a9aafacaf619cf1b466a010
F20101109_AABJIL kammerdiner_a_Page_30.tif
eff1d79afebffda328983b6ea0e32e5d
c20986fef0915109219e65acc20c275ca5525b8f
1877 F20101109_AABJSG kammerdiner_a_Page_32.txt
e9b0ebe135ecd3df05deeb2e1f4fa7e9
180cff7b7bfb55ffcd17d47a4c6509aa7dbb2a16
F20101109_AABJDO kammerdiner_a_Page_37.tif
c48a704436d60bb64063e432b2feedcf
6eeb4710bd2d3e6608379946eae52d325ec1c59b
121530 F20101109_AABJNJ kammerdiner_a_Page_82.jpg
8d7699b6e877bfdb31a1cb460d20af3a
1046c6e5afbf1753dd1ee1f7598b9e5b2a9a6213
424 F20101109_AABJIM kammerdiner_a_Page_01.txt
c16c6fff3f1f606675cb5f65d8f4e6f2
71dcf66a9eaa4bf9d5e97e9bf6bbf12880d99f09
1809 F20101109_AABJSH kammerdiner_a_Page_33.txt
86f047d78c6c711e8bbdeed2b96731e9
e672c7bfd329ffc6f0e261df8c308a6203938eba
103216 F20101109_AABJDP kammerdiner_a_Page_51.jpg
d0c504a8a44c8a1c860ba0518b7be079
8d325d9b8533180187d7230de191145d9fb5d4dd
23025 F20101109_AABJNK kammerdiner_a_Page_01.jp2
c713d32a0c772eea48c6e758f7e58fd7
6335d15de37c71d815844440ad5a9e3f348e19b2
1880 F20101109_AABJSI kammerdiner_a_Page_34.txt
c86379572757c5931721d6411af03acf
81429ad651890d1b9110f2f59a4ec975f05814c3
2335 F20101109_AABJDQ kammerdiner_a_Page_21.txt
ce107807c767c14776f2f59bfe2880b2
a58a7c84ae2adec9309b61b2d5c669194cd4eb61
5616 F20101109_AABJNL kammerdiner_a_Page_02.jp2
c0c1798ffa494185347fd122da353165
bca41f8c750da6b6c54b73ce4e45cfea8103c580
7523 F20101109_AABJIN kammerdiner_a_Page_19thm.jpg
aae48c41b5f6fcdcc052680f6b6880b0
dc6ae55d4accc3abbf8aad8104257c0e232bbd76
2108 F20101109_AABJSJ kammerdiner_a_Page_35.txt
a2b5bc82169a26cd16ff1d74734f7f63
e61ac555080cefb8f8fb9a0ef2a675e3a20a46d5
F20101109_AABJDR kammerdiner_a_Page_56.tif
50d529087b04143ea5c67a51db508675
25be5c87296226b4c518a4f640a21dddc5c3bf9a
368136 F20101109_AABJNM kammerdiner_a_Page_08.jp2
bbecf37849463862c111f5164d4312cf
1b1d52f0ad505b983940a8eafd5fe811e218603a
6955 F20101109_AABJIO kammerdiner_a_Page_48thm.jpg
59137fa975874a32bef305b60775e6d9
ef669bd107bccfd4198f768e8e76b07a84e877ad
2214 F20101109_AABJSK kammerdiner_a_Page_36.txt
1e471cb4fc55c03c24404c19971458a7
e990d86b07fc70276cb817699148d7a190715651
2141 F20101109_AABJDS kammerdiner_a_Page_44.txt
cc72f4138fcf4834b51361529db3e45f
6a49c76c1d3513842347fea86a374621eda34437
F20101109_AABJNN kammerdiner_a_Page_14.jp2
0fa7107ea13d9a9686aeb231ed9aa537
030021febb47c9f1cd6a0ff56494fea3e3650ee7
52781 F20101109_AABJIP kammerdiner_a_Page_38.pro
4219d270427d18b5741b59940df67a97
ec612ea81cc94b6309528e61b5ae8fafd5a35dc5
2265 F20101109_AABJSL kammerdiner_a_Page_37.txt
fc1d58b61b02f039c60ad187d36240a3
1cc66335fe24bb42d978d827d1acd4b9c2a9e42d
101980 F20101109_AABJDT kammerdiner_a_Page_44.jpg
228b7836dff3366750cf2a5d6fd46968
3f886bfc7fcaade0dfe8e2daab52839989879ec0
F20101109_AABJNO kammerdiner_a_Page_16.jp2
2b91e915a2d09cf7f294bf39fcdcf33c
e057dcc6b722592b1f8c0b183776a46d482f85aa
96616 F20101109_AABJIQ kammerdiner_a_Page_63.jpg
d145bb953508984a15207f2c5d3212fd
54bcc5a871076f79b1057d1cb2bc23b9d0720400
2208 F20101109_AABJSM kammerdiner_a_Page_38.txt
ce4d15422d6a7549942b058072a56cce
44a8dc67a4e2654bebd3b164592acdb81ca117a7
7709 F20101109_AABJDU kammerdiner_a_Page_67thm.jpg
6e66e6c1a4719834b9c736d84706956a
2a298b6607cf309c8dcf0b596398b0ce1d9b775f
1051985 F20101109_AABJNP kammerdiner_a_Page_19.jp2
b6e2a620f87c1b3fb76d2211eb189763
cb8b71f18bcd19cccba4bc7f9569270b442f392f
20643 F20101109_AABJIR kammerdiner_a_Page_46.QC.jpg
1bcad6e1d2bbb54e62c2a2a429c122a8
d5fa4a9b0fb6cd3c38f532e625e3f68e8622281b
1955 F20101109_AABJSN kammerdiner_a_Page_39.txt
f4bfad50d2deb34068c9d4ec13284ad8
b3b42c0f0fcc855227abe6708e024acad0bcfcbb
F20101109_AABJDV kammerdiner_a_Page_22.jp2
76641ce2cb2ad3243aac46f0f1dc5f62
add9b4e30318d524458a6e642fe15fe5e92b0a27
1051981 F20101109_AABJNQ kammerdiner_a_Page_23.jp2
5461f87a89ec345a38fe6137f2b14ddc
91eeb2f3d007f99a085eb19850e1dc76b5f06895
816 F20101109_AABJIS kammerdiner_a_Page_74thm.jpg
2ff44667248923d76332d7c697e71a3b
9b1efcdbf094ca9678c8dac2aaad96ef6030452c
1953 F20101109_AABJSO kammerdiner_a_Page_40.txt
caa02b5950b19a5f23ad5171856ed861
0e02a48162ae6be064592e936ffd4be74646243d
20782 F20101109_AABJDW kammerdiner_a_Page_09.pro
ad5588b47c79dd64df3ab81584022278
8da4175283d5068aa419b9aaa81b6e4edac00099
1051964 F20101109_AABJNR kammerdiner_a_Page_24.jp2
6edb78bbddcf73038d7049d1fbc3f944
1d26b5f78e81d2daec11594fd05578744411d84d
F20101109_AABJIT kammerdiner_a_Page_65.tif
e73af691ccb2913e99ff07304f263486
301879c0cae6c83e99bd9a92efea1257cbe2bd49
2071 F20101109_AABJSP kammerdiner_a_Page_41.txt
024f9f9c9dd9eb35e7de21f71fc56ce5
066883dbb81985c02747c6ae9e0ec2076333546b
590 F20101109_AABJDX kammerdiner_a_Page_59.txt
6e207d9ba25ba737db5c1f364788fb47
500db08fef7469da381178e8849a9d09a1e0e1b3
1034704 F20101109_AABJIU kammerdiner_a_Page_20.jp2
0dab597065f0c6df4c47510af1e83337
cf22fc01d1f41c21a573b0d8c5c66c1461002a98
1963 F20101109_AABJSQ kammerdiner_a_Page_45.txt
7afbe3ea68d6f6b984a85904c14fbcc2
36ef6fbc74dc1ec540b53558c26fc875d8f72166
907777 F20101109_AABJDY kammerdiner_a_Page_32.jp2
fe6a643c7ecdfa7848af3f3d6d02f28b
44bce4ff8834804081e0b3dc05967f6631538406
756651 F20101109_AABJNS kammerdiner_a_Page_26.jp2
b3b56301f7c041191d0e1572fca3b66a
f0e1fa2412060896675696892c03b7a04993044e
36294 F20101109_AABJIV kammerdiner_a_Page_79.QC.jpg
5cd11d3e03421eaea6ac9eca4f0e19d6
01279c4e46d346a2ce2298ea8c9332bbe011f157
133 F20101109_AABJBA kammerdiner_a_Page_74.txt
201c2381506fc12e5347210184b74ad2
05a0f6ed50798fd481ebdd5571e79600fa194498
1656 F20101109_AABJSR kammerdiner_a_Page_46.txt
04da69e0f22527aff78805c92c6a1107
04ba0f5d947bdd229e1905311a8e58fb97077df8
134982 F20101109_AABJDZ kammerdiner_a_Page_78.jpg
d72bbb001d5a553221132e27ba1ad9bf
0fbbb93a405ea367a2ae5ef8ebf7ba93030c2e19
489532 F20101109_AABJNT kammerdiner_a_Page_27.jp2
60fa30baf8f577a1c0b92e11c3da8f8e
482eb47f133a491bc89fe5ea6fbbdcca6953b5af
57773 F20101109_AABJIW kammerdiner_a_Page_23.pro
743f9c0016c16b8e71b1f830e6eb3678
12e8e57ec3a55e902ee20badb18db5aa26c7c045
7628 F20101109_AABJBB kammerdiner_a_Page_31thm.jpg
01563f6717fc41901685175425fd9824
a6ad5e4d3a2ab876095b57bf58f3e33dd54b9f4b
366 F20101109_AABJSS kammerdiner_a_Page_50.txt
18b4dcc49f949bdf0636f18984e4b74a
de4d2eba036964e555f20ec06b0b647e0c6e18cd
1051975 F20101109_AABJNU kammerdiner_a_Page_29.jp2
c7c0f90e67bbbe54ee537adf526e23cc
04f61e833f8d0dd91709ba9974841bdce25d3584
8014 F20101109_AABJIX kammerdiner_a_Page_80thm.jpg
905ef58c92541e1263c6ae35ad32e634
e7519c53ba9acf79210fef7c52b6b6d73822a327
35175 F20101109_AABJBC kammerdiner_a_Page_81.QC.jpg
ecbdbcf6798e641193c125daa25ab735
133c8ac4ed0d7704cff0061236aa560fec4f2539
2033 F20101109_AABJST kammerdiner_a_Page_51.txt
09bc82aee5fe80fe0ebf70dad23d06aa
b879ac817d83220e5c71323ada4c128c35757617
789781 F20101109_AABJNV kammerdiner_a_Page_33.jp2
51df392449cedffdd7331523e53e2154
302abe361c6b4ecdc8c7468e4bd0dd71a9fdf474
35681 F20101109_AABJGA kammerdiner_a_Page_23.QC.jpg
fa347363f1dc92087876adde06e311a3
5e3dc8ea3f3fcbee21a7a08a0b72f856b686ed08
149900 F20101109_AABJIY kammerdiner_a_Page_50.jp2
f9eb82592eef17707afe5df8882d0fbb
ab6349efcd2afa77b00dff32d2483ce47daafaf3
43515 F20101109_AABJBD kammerdiner_a_Page_09.jpg
947b9c7043cf966bd043285e990d1cdf
62eb8367f863457862134625e0d3b84bc89ecd85
2313 F20101109_AABJSU kammerdiner_a_Page_53.txt
7aa268a6fab11507a1de0ff8d2dbe440
a6d33016320113e2fc4f6705f7a164496f2a423c
948289 F20101109_AABJNW kammerdiner_a_Page_34.jp2
9e99a2991843be289d53571acda89a20
770747859650fa349af845abf5e397da24222521
36391 F20101109_AABJGB kammerdiner_a_Page_78.QC.jpg
f9628b5f28c3eb1fad3042d4aad9c95c
0f08b86cf3bab3f2c4f2809cb3c34702d784918d
1051982 F20101109_AABJIZ kammerdiner_a_Page_05.jp2
bc39663111d9feb94c75cce126418a8a
0d3f920fc28466f637ebe2cec130a999f2ecf267
5662 F20101109_AABJBE kammerdiner_a_Page_50.QC.jpg
5ae25f5924bab76688c2321c393da3b2
d8329d7f1ba43ccf55586d1aafa5483dcf5e8305
2332 F20101109_AABJSV kammerdiner_a_Page_54.txt
fb42c28c700b08da1c1aa5ef15c55101
af258052f246aceaf8631a715f6c1a4856053326
1051956 F20101109_AABJNX kammerdiner_a_Page_35.jp2
8bf87388b1a10b3648b5b2308cc68c44
f2d932dd9cd886d950e9e66afbc33b693a25f299
2285 F20101109_AABJGC kammerdiner_a_Page_60.txt
382796bad4bd6501d5ef0c27fbf9af68
1c9dde8dc2e718a3c4a6360b5b85242d642d4bf2
33430 F20101109_AABJBF kammerdiner_a_Page_58.QC.jpg
917bb646157b92cfa8c5f9821e14e1c2
4979d50881abbf9bdb643cfc05f92634e9f67225
2070 F20101109_AABJSW kammerdiner_a_Page_55.txt
a0ec34fd46316d04cfd7258ec46c4d30
d842fcccd3a6f7dfa11e8ad86a5457d2cfba88ae
119357 F20101109_AABJLA kammerdiner_a_Page_80.jpg
144727177eae1a103c3981dca6b9ded5
518d35ae5a9979a2ac120f1f3d4a9f48e5eb96b6
1051978 F20101109_AABJNY kammerdiner_a_Page_37.jp2
7ac1b426e497e5509f3f0a7787b7b78f
a449afa10acc87ebeca39bf48a84b98d44ecc472
57233 F20101109_AABJGD kammerdiner_a_Page_67.pro
9f131cc86d6f796d119c380aa96b2e9d
9d69f17d3f331bb97e8635b40521b8feb80e4312
1051986 F20101109_AABJLB kammerdiner_a_Page_44.jp2
69a824a8d35adf8176b8d3a1db6163e5
4b2aa56b06083821404373e13f53677e34873a6f
988574 F20101109_AABJNZ kammerdiner_a_Page_39.jp2
433f3454cbb964b552165b2f072504c2
31b379425a3fc4537959a827f468931dc5a6037a
24775 F20101109_AABJGE kammerdiner_a_Page_48.QC.jpg
e444fa330d6d0f1e7c4dd50b2f257d96
3ff743b9987c69bd020f3e224227efec2ddafe6f
27709 F20101109_AABJBG kammerdiner_a_Page_11.QC.jpg
056c1993ee985da35f469ffde6b3f03d
5f0eca53fd38b0615470e7a462d04d6b2cbd3ce4
1986 F20101109_AABJSX kammerdiner_a_Page_57.txt
692f4ecb173080df256db7e0bbe9c542
1fa65e5c6e8009c53fa28433b265b75c46dd63c7
32459 F20101109_AABJLC kammerdiner_a_Page_43.QC.jpg
3fc973de76c7a6f611d9b4876df0cf71
c914793958a6e6eda0ef134f08f1c2ae42d9786f
77177 F20101109_AABJGF kammerdiner_a_Page_30.jp2
a7f1e45c0de39449d291ae88bad655fa
fa537a22420c5a9941ddcbe3ed80eb0c1a89d50e
2463 F20101109_AABJBH kammerdiner_a_Page_80.txt
e232016dfe388626e998993fda53073b
7e936ca1b51f38c8136407b2e11e28d6d600d8cd
1948 F20101109_AABJSY kammerdiner_a_Page_65.txt
8aebacacacf9d2afd6956356b488f6ca
c13e007c6cf298977a5f578e91ada9b1e054a066
F20101109_AABJQA kammerdiner_a_Page_69.tif
2dce3301e4f578ad2a069e868fd6a3b3
c2a1fa06fd9f3421b36034fff4d3d2cd225080e4
110897 F20101109_AABJLD kammerdiner_a_Page_68.jpg
6c40df32b63545d9a6299672055ce1b9
6bc0e34d75fd17d59df02a686eec612cf2ac5b22
17460 F20101109_AABJGG kammerdiner_a_Page_10.jp2
9d0cd8197da5a518473af1f05d5d6865
f69325e51b77c1d4d889b6fcaf4f5d4f294e077d
57876 F20101109_AABJBI kammerdiner_a_Page_14.pro
cff7dc2201ffe3bdbc440cdbf55bd963
173547ca9e0a5f5d4bc0a4bb80d557e4fae084e5
2036 F20101109_AABJSZ kammerdiner_a_Page_66.txt
6c53c733aecf380a5d92dd20ae238328
beab256946118c81f75e74af822fc633f1dd8dca
F20101109_AABJQB kammerdiner_a_Page_72.tif
88e62067af3388e62fa52ba979f0093d
c04901f7bd3ada8c69ed9409b5645ffd2f7100a0
118631 F20101109_AABJLE kammerdiner_a_Page_67.jp2
40fae2a367ec64419de51548447de59e
9b00d2f67d4882229ff35ffb69bd75c0d29b7dc9
123508 F20101109_AABJGH kammerdiner_a_Page_76.jpg
136486b1a1d60d4df21f5f05b96543ba
62aa0dcfee3307204eeab69ad357fc4a1c6c7c21
1827 F20101109_AABJBJ kammerdiner_a_Page_48.txt
a812840515572e44fb90f047639b9dab
5ea14a31c030e767944f0aa6c99abfc2b0b8ac7d
F20101109_AABJQC kammerdiner_a_Page_75.tif
5746d51405015ad37c959a261aa9b9aa
8cf480e4fb26c70016ebb9226aa1360c1463942f
7632 F20101109_AABJLF kammerdiner_a_Page_01.pro
e3b3a12933e97b7d8fa464b324c0364e
71f604cba5f171ef812d69c66bbd8db637241575
3465 F20101109_AABJGI kammerdiner_a_Page_06thm.jpg
beb60ce9632307c810d48a765f990ed6
5cf1187982973a34c0975aac3210eb4efacb7ce0
32549 F20101109_AABJBK kammerdiner_a_Page_44.QC.jpg
cc3bc46a5c55abaafff7b2c3a849f3aa
a3d5ce822cec0aac0078646dfabac8591bbb14f9
5377 F20101109_AABJVA kammerdiner_a_Page_49thm.jpg
64de83a7619a96ed26d008902a4c4870
8a72334872b3442fa4fa69ea11246f1c766d4f42
F20101109_AABJQD kammerdiner_a_Page_78.tif
22a5f777ccea17a5a27c2704ed6e83ef
c8d170d3e15c34c4e17253e80f8255f3c136fc20
6494 F20101109_AABJLG kammerdiner_a_Page_10.pro
fb78ac17683455f821c84fa22c38d0aa
daa428c9807aafe3a45e121965dfe827ce8160a6
7940 F20101109_AABJGJ kammerdiner_a_Page_21thm.jpg
b49154348f862f63e744d7258dad63a2
4b437361d52461c8859e7f0e0a3331ba2ed16b38
F20101109_AABJBL kammerdiner_a_Page_80.tif
11582216565cc1f6bdfc1681e2496cff
876902796fdd67577ea969179706f0532fdca18b
1716 F20101109_AABJVB kammerdiner_a_Page_50thm.jpg
95acd8eea66b304da9e07fd22af13007
fd9e122f01dda3ece65010fc4926dfc1905d759a
F20101109_AABJQE kammerdiner_a_Page_79.tif
da9641fd7bd3e2a92dddaee4081cf8e1
6e501fee0df8e77c6592f22d99b37f648b3ac06e
83515 F20101109_AABJGK kammerdiner_a_Page_32.jpg
42ba23cd0388477d2008bb11e3df6dad
c7c979faf508e1dc4a07471503eeed8c34977d1b
912595 F20101109_AABJBM kammerdiner_a_Page_65.jp2
1204690aeae2a88ea0327824a3019c90
ca18eed72e25903bb7e60c484692d2ca51d95bdf
5075 F20101109_AABJLH kammerdiner_a_Page_03.jp2
b14a671fdd22efc8542760c429c9792c
4000fe99838d3595b6360f21dd2d741bd2477556
7685 F20101109_AABJVC kammerdiner_a_Page_51thm.jpg
338a4d914a2746dff9a17684866a94ac
fb822bd47367753467a7af90693ebf943b48d8a4
F20101109_AABJQF kammerdiner_a_Page_81.tif
eacaeaafff3f52739d9ac7a999b62c0f
7a489295737d19e9d7080bef352e6f97e5ab8002
138062 F20101109_AABJBN kammerdiner_a_Page_81.jp2
ab7d6ed5bd97753ef800ab8d03c37118
1ce2af967aef161913e3412418a8ed0994f4b68b
329 F20101109_AABJLI kammerdiner_a_Page_10.txt
e56e1d2a7ff98d496979ddf87fff6be5
d9a2d1d4f184c4a57bb8e626ada5adf71c5721f6
35554 F20101109_AABJVD kammerdiner_a_Page_53.QC.jpg
45596559415fed1835d94e1153053417
a6bd76215fb34cc4984b9198656de31eea841cac
F20101109_AABJQG kammerdiner_a_Page_82.tif
48f67ef11eec953ff6d6ce2b6663e438
9c06ba8098948590c64435543bb6f51a6480d13a
390 F20101109_AABJGL kammerdiner_a_Page_72.txt
a54ace4b7f15de06849b25d21519b620
77df821d8bcade56466aa5383b0e825f62d7641e
33157 F20101109_AABIZO kammerdiner_a_Page_49.pro
7661249e7107acf7087922fa47602365
d93f14c55ed6ede61139912dcf137d7a4ed2c7fe
34309 F20101109_AABJBO kammerdiner_a_Page_82.QC.jpg
a9333e89101f1139e8a6191418e0b60f
e1dfb89dd5ffe637676f8bbbbbf89be9ac91a4e5
16906 F20101109_AABJLJ kammerdiner_a_Page_47.QC.jpg
0c873b21b5961048550776b0a172406b
d248e92f4cd361c40729f5966a0aa1ccf1b3b864
8135 F20101109_AABJVE kammerdiner_a_Page_53thm.jpg
f0edfa22138c4e80153d1bd569d5a60d
8dfa1bbb2bf3f1a7f3848efe70de3ce028cbc230
F20101109_AABJQH kammerdiner_a_Page_83.tif
8c30aadf95b91cdbd3ba1230ba69d345
f37e5b60584d39bdea6fad8ac2aa90f7fd5e725e
1495 F20101109_AABJGM kammerdiner_a_Page_10thm.jpg
140a4d92fa75b61bac46c0e5c1a49f5f
b1224bb930d2ecde696650f437484c21220ddce8
2116 F20101109_AABIZP kammerdiner_a_Page_19.txt
72dceefd241c89a850471354cc24e1b8
f1a76f653abbf7739b105bf016fd57312591f519
1051910 F20101109_AABJBP kammerdiner_a_Page_38.jp2
e6269d8d051797c356bbab6dc06c2187
fcb464a5fbf940606ed6b0fcb6963c0153427523
F20101109_AABJLK kammerdiner_a_Page_25.jp2
4d69f5e7c4ce83debf0dfea1170044cc
3dbcca5bcd35477246b9a33ea8fda55b5e831f65
30868 F20101109_AABJVF kammerdiner_a_Page_55.QC.jpg
064ee45f28f31eccaf1c4e6640b520da
111957e02c745fa24248fbdd1fb11759374569b0
56349 F20101109_AABJQI kammerdiner_a_Page_05.pro
a309fe2e0151f50b92192b0b9a1d7fba
77616fc84d18ef040a256d6fd1bb2bbf0199e655
735544 F20101109_AABJGN kammerdiner_a_Page_49.jp2
5eb5d2181dd38224fa65773c4bd057fd
14dbe9a0c0f221009078cbfe8161c718205a5933
113461 F20101109_AABIZQ kammerdiner_a_Page_61.jpg
4b5b0fbccec82b2a27418e5f387bcd92
d5e672e864293841fdd72229cc012004929cd041
7812 F20101109_AABJBQ kammerdiner_a_Page_74.jpg
b9ddff27297b3c51079e76e5c40b3f96
b182d9b60f405e4bad3311a4f66ed264d55c6e4a
5909 F20101109_AABJLL kammerdiner_a_Page_46thm.jpg
40ef1127f2cb2b4862cdde67eb64ed17
bd3759b942c7ea0b4c536f4ed35d3bb0e7616623
7325 F20101109_AABJVG kammerdiner_a_Page_55thm.jpg
6d9e3736b1af911299cef8604f8391d4
9f09a1335de18bbd130fbf794a2696008e84f5e3
11154 F20101109_AABJQJ kammerdiner_a_Page_08.pro
53d61dd98ce837399319b29a10b00a29
28807be82d7f7e6f62c8d1dcbd1cb0c9e74f4165
F20101109_AABJGO kammerdiner_a_Page_51.tif
44df86b6b243474f2e82a7ca0a618a56
dae99818444d6176d3d991261a702358aaad7595
816603 F20101109_AABIZR kammerdiner_a_Page_06.jp2
4e9fc3b27f2a8619059743dc1d95d1c5
721060d7ecb6a25f8dbeb9aec94ddbc65a88453c
1287 F20101109_AABJBR kammerdiner_a_Page_09.txt
050531dec248ea705f2514aae89ec7c3
d090085e7978de209eb3efdb2d96bd7eb9a7218d
484 F20101109_AABJLM kammerdiner_a_Page_02thm.jpg
55039f823f640ed7e6e441ccccba5312
670297444f6fe60cdf8d7f6d9b4ca1e55014617b
30667 F20101109_AABJVH kammerdiner_a_Page_57.QC.jpg
b00d56816f234adceddfbac14a23918a
208a45692eb3276ed9a0cbe7e504f3682031df3e
45541 F20101109_AABJQK kammerdiner_a_Page_11.pro
6b2bad54e9c56f09891f8f8e61b89abd
89edfc3a7b06a823969b4b94bdcdc26b704bfcb5
2725 F20101109_AABJGP kammerdiner_a_Page_81.txt
eb5a214de8126cae8f194d0b7af75d11
1adf26633e574c84d1368714e723c1560db5d80a
61646 F20101109_AABIZS kammerdiner_a_Page_80.pro
3d31cc4547c10c6c8e00079547be8085
6ed9ac2bbedad35d10c9058d97f533feb0172b67
1051951 F20101109_AABJBS kammerdiner_a_Page_21.jp2
c1342b5708703ab33f82892202bb30b6
319268ef96f9cb48b6ae8e92e6355d0357d32570
F20101109_AABJLN kammerdiner_a_Page_05.tif
6ce609eaa37e4ed6b9a5ea9648a8dffd
d38e498bc4db92299fef434e5406c18d296f6c5d
7467 F20101109_AABJVI kammerdiner_a_Page_57thm.jpg
e6a4e3fa42088d2018fdf0b9d534b25d
a5aaae04efaf0540f0a1c55a7719cef037088357
54874 F20101109_AABJQL kammerdiner_a_Page_12.pro
f1575d4a0ea3d027d1b9227d9e7d6da6
18a455c117856fa4f06aee4979a6b5b4cc93a64e
53038 F20101109_AABJGQ kammerdiner_a_Page_58.pro
e7c766a71c625ddf302ca51b08418a42
c1b316427fa0b00dc08504e59a2e0f58cf2a2775
F20101109_AABIZT kammerdiner_a_Page_73.tif
89d79ec9025d3e913064e6a1c5764f0d
b55a9516bb3717004a0ee07c32634058148af376
F20101109_AABJBT kammerdiner_a_Page_26.tif
94c9695fa6649067299d8e47f97ce070
8efd5f0e52a747b35b015f9b65f93a07cdf0e0e6
99449 F20101109_AABJLO kammerdiner_a_Page_11.jp2
42cce0d63f6e40920db65588e3b5f22a
58b92968bfd7a7bb64471e09c5712af19252a3d5
10443 F20101109_AABJVJ kammerdiner_a_Page_59.QC.jpg
07723b4b524fe1011a5a80891eb56a3d
91393a5b8f451a2224670fd40fb956c9439e57da
57736 F20101109_AABJQM kammerdiner_a_Page_13.pro
70d3c4e91d4a28eacf44ff3e4b2c725e
2d2aae0265aaef077d484512bb4edbcb5ff7e269
141433 F20101109_AABJGR kammerdiner_a_Page_77.jp2
c760836da524ce556780f050a6beb2fd
c46b5fb74ae237a80ddd47a754adb3109b7c8e0f
7930 F20101109_AABIZU kammerdiner_a_Page_43thm.jpg
558b745e2b335b31d98acf6d793efeee
9f3961f6dc9fc61c22105cc8cac62a5000157b70
26820 F20101109_AABJBU kammerdiner_a_Page_56.QC.jpg
35ede2d108590965fe2106ec2a05157f
4235e3c9dc3ed2e0a17abcd4551e6a6950e0e228
1051965 F20101109_AABJLP kammerdiner_a_Page_17.jp2
ee2e6bf18055206d6e693249765d0e6e
59ce393a80195aa4eb3d3781c2cb9820ac8e1cc7
2551 F20101109_AABJVK kammerdiner_a_Page_59thm.jpg
a3fcd0ac9a49a659db68a7b9b6bf6aac
6bb48f6999ae7d4e9d71ff471997c86c3448a8f5
56980 F20101109_AABJQN kammerdiner_a_Page_15.pro
81ddecb54cfffae8a528914e0f2c4d8c
83fca1f97ab92077e1b4cd057b8a2d11dfe5f372
32233 F20101109_AABJGS kammerdiner_a_Page_51.QC.jpg
5a059f24f8b6ddc195e798a4d2bb3e24
bce5b6b761e8a08c31aea2080ce9f9142fb92d5c
44362 F20101109_AABIZV kammerdiner_a_Page_83.jp2
67b706993fd7446bc0cb51a4953a950f
79aa682f8bef01721734efa266fd40ed662940a9
125801 F20101109_AABJBV kammerdiner_a_Page_81.jpg
34cd58c06844cc7d0062f59cf86a2054
9d2532bb1e6b3cf939810f6334d45d7dba642b0d
36298 F20101109_AABJVL kammerdiner_a_Page_60.QC.jpg
d2fe271b9c2a82f8af6f0c80061a416c
619a9ee519ebaf353981dcb0ef26cadbd869137c
57980 F20101109_AABJQO kammerdiner_a_Page_16.pro
b10393742ea96d741184afe3cdfb70f3
adb2457c76f6941ff2c7211802b3556efa7d9295
32320 F20101109_AABJGT kammerdiner_a_Page_46.pro
dfd758e47766125819a5953a16119a6a
ffa3cc194ae6a909c2c5c62a9dc7a3b52e8f3014
2221 F20101109_AABIZW kammerdiner_a_Page_28.txt
11c4906cf2027f09952a666b75f7ca8c
5de8a552187cf570f8928758e11eff439db65063
35460 F20101109_AABJBW kammerdiner_a_Page_54.QC.jpg
644569b96fcffb4abf1c652c2fa250cb
cd693398990c9b1ec92575020711e896a7489908
F20101109_AABJLQ kammerdiner_a_Page_23.tif
74318a79c7c43fa70957b56a423fb580
1bc3e7aea8713872d0752cde1b74b09912ce1127
8325 F20101109_AABJVM kammerdiner_a_Page_60thm.jpg
c4c3ce1eecaa2fa8c690114795013b45
b7c8d891afeee125a11f2ec8a0d94ce2ef511e82
59465 F20101109_AABJQP kammerdiner_a_Page_17.pro
6c5aa80a36cd8f9980445ae59c8e9f64
3be6b67a3286e54305efe67d776fb0a0367cc854
33869 F20101109_AABJGU kammerdiner_a_Page_12.QC.jpg
35b5899d4945a57713fc1d249d1df8f7
1a60d30f2c6fcd2c09162867d0fc7a2e290f98cc
55600 F20101109_AABIZX kammerdiner_a_Page_36.pro
e2ede60cb160bf8003da4be5c18fe93c
fff511fb44eba06c6c706fd9e291ebd57f0c2b49
7822 F20101109_AABJBX kammerdiner_a_Page_44thm.jpg
dbeb09c8d1fd45d34d42c3723c83fdea
38a3f8bef538a588eb8cc867fd9de8447143c3c4
7336 F20101109_AABJLR kammerdiner_a_Page_66thm.jpg
d84fd740c68f6ef4e396fc37705c097d
5bf96ece8173771104a5c86265387c0b75b91026
36248 F20101109_AABJVN kammerdiner_a_Page_62.QC.jpg
df58f6f75468ff73fdf1c71b46ff62dc
13805597f9581f3a4a05984c3fc6d7a0c40c7fe9
60708 F20101109_AABJQQ kammerdiner_a_Page_18.pro
768d33713b8e42ac54958da61614fcd6
fe7fd6005cfd9bbb1603b0faa46568e767eb7059
28510 F20101109_AABJGV kammerdiner_a_Page_39.QC.jpg
ae10176e1a0cedd3db95dc25a813b766
131c1c955706f7f6daa51f0f83140573eb43e624
37184 F20101109_AABIZY kammerdiner_a_Page_18.QC.jpg
368e674adb802f550323203f00874dbb
911e2db7516b8a0aa49e08b7d55afc69414f3d61
38238 F20101109_AABJBY kammerdiner_a_Page_04.jp2
7c67fbb62ea7f69286ccdf7cedcc919c
ade38a6a0222f241fd38dd3a2334f0f878f81b58
107659 F20101109_AABJLS kammerdiner_a_Page_67.jpg
ca431965cce76c7a0e61c85448c6a045
8804f09f84d8bbe01bcd4a52017347b22a45e361
8335 F20101109_AABJVO kammerdiner_a_Page_62thm.jpg
4d7c3a5024c91e7e4b1a5fe5e53b1374
6558cd201548dbb7c90d4adbe8eca0a4940603ae
47836 F20101109_AABJQR kammerdiner_a_Page_20.pro
6e81a58922d3493757e65bc67814f965
56ed05e31482676a53855457f5510d5e219576d8
F20101109_AABJGW kammerdiner_a_Page_38.jpg
8ed54fe6bd3ebcf9f4f1d926f530662c
0166d02ccf6007712e61633072bc6130c5791ed3
68631 F20101109_AABIZZ kammerdiner_a_Page_70.pro
56bc044d09e99925e32f5d6bf58c7d4a
75b68bcd4e2cf0f55ba471f0a5689f9fa7b8e820
F20101109_AABJBZ kammerdiner_a_Page_06.tif
aee2e172af1fc8d8fd63cdbdf8b20cb8
1cf01cff45a299fed930030d61df195d6417f350
1051915 F20101109_AABJLT kammerdiner_a_Page_15.jp2
bbcdda77e509c37028a61b05fedfb9ce
03dd28954ee3a7ffe7a1be2d4bba56bac75251d3
7242 F20101109_AABJVP kammerdiner_a_Page_63thm.jpg
e3991fff8f41b3ddb407099b0d2355c5
2673b3b575ee7c43152d2470e51620eedb418b2c
57277 F20101109_AABJQS kammerdiner_a_Page_21.pro
dcd2e1f8b8cc0de7c2ccbb89716f3f18
8879121c35593d6cf394d3299e682113ee9d3956
47956 F20101109_AABJGX kammerdiner_a_Page_64.pro
17690ee63ad3233978d6100e6f231dc2
71fbdc8b20083e35ababbedcd102fc2c7fb787de
F20101109_AABJLU kammerdiner_a_Page_33.tif
9189aa5181560d7e48f899508d94b3cf
d908ad0bc5a627de57ab3d28c8fd3436992100f2
7387 F20101109_AABJVQ kammerdiner_a_Page_64thm.jpg
947c78f8a426de93943d1f4e76d285c0
2271a54abc47b27e6c562a412f1d9db96a1daa14
51288 F20101109_AABJQT kammerdiner_a_Page_22.pro
3b0da8b9a73a4ae0c0dbd66ea3e9535f
69e9a95383471d04064f14f5f06f865a868c2860
113785 F20101109_AABJGY kammerdiner_a_Page_13.jpg
3e37bef396f058b0ef5d4311f405613f
8d834021c00aa0db5f16e96d71dc07984d3d756a
99193 F20101109_AABJLV UFE0022141_00001.mets FULL
fbb0f1675e3b20c8a0e4bb810cbab858
0681139baff8bb9829035f7a83b398105a030c59
58690 F20101109_AABJEA kammerdiner_a_Page_75.pro
cc74a30e3a57f3ba32b458b1f7fd2b9d
cb80a1a26ca00730c3df584cfe1d874cfb0a42e9
26632 F20101109_AABJVR kammerdiner_a_Page_65.QC.jpg
9cd915ba368263653bbbd7c1bcc7d9fa
a0296c23f9f4e02bdcdf0999097a7b3b8427b2d5
57936 F20101109_AABJQU kammerdiner_a_Page_24.pro
1b73962c36440c581b6d1905268a2bf4
adaf9183f0ac46df924197b3e9f02856be1ea8a6
8543 F20101109_AABJGZ kammerdiner_a_Page_25thm.jpg
92a031d74dd4222407b71d4e13698bdd
358f77780e085f4c20b62b7da7d6459f8c5f4fe1
94492 F20101109_AABJEB kammerdiner_a_Page_64.jpg
62f1c55af0fa606918dc75b914635002
ae859370e4a395d5f164a8b0fe28ba6e9e2b0a27
13720 F20101109_AABJVS kammerdiner_a_Page_71.QC.jpg
3984c4b6fcb124c640deffbafab2c7be
623ce4fe7649078f99b89a10fa828d901f7837b5
36468 F20101109_AABJEC kammerdiner_a_Page_17.QC.jpg
06621c5a93924f16a2e27a6dce9f51ea
915b304e2138ff481c08d93e5191f3536a47c82b
3816 F20101109_AABJVT kammerdiner_a_Page_71thm.jpg
0d83ec8383ffdf8f6801dd1ab748e9a5
0a5c9b8ed0de8fb19d1088a08ba6e8280f628551
61260 F20101109_AABJQV kammerdiner_a_Page_25.pro
8e1a9c5b5fd2a7a02eee66fb6997e991
8d1bafa5ac05068452a4d94241a7bd0b2b48b29e
35181 F20101109_AABJJA kammerdiner_a_Page_28.QC.jpg
5fd89adfd88484ecc98172d336a6dc2f
d60851c7349ab201e65ab4ee9d3aada427c0a640
25027 F20101109_AABJLY kammerdiner_a_Page_01.jpg
74780b7947224890aa64bba84ca02c84
252c73728450a2c7e4314a618313f78d0a942fbf
8545 F20101109_AABJED kammerdiner_a_Page_18thm.jpg
ed44366ec7ee9a92e13f006a80130db7
894905abc101c17b65be3559bd90c6866d03fc4f
5628 F20101109_AABJVU kammerdiner_a_Page_72thm.jpg
a9ca14b480287ec4c50eeb2b29105407
6e894f16846887901962cf68ffaf815f5be1b09b
33915 F20101109_AABJQW kammerdiner_a_Page_26.pro
4e787a3560d0897bc6cae11030df26d9
79a99db754ea30c6f8eebe10800066a00706d130
119714 F20101109_AABJJB kammerdiner_a_Page_25.jpg
6511ecd19960ba56a5d6eeaab3a3fa4a
5501b0248b4d54eb761184a000f318b40c93c8fa
4258 F20101109_AABJLZ kammerdiner_a_Page_02.jpg
f3efd36f8c64d320c12ba6fb26ff1ba5
cdbb1f111a5b605894f6037594a5afbc285e4b0c
1051977 F20101109_AABJEE kammerdiner_a_Page_13.jp2
a63633e8d1d5a56d8270d447726a7615
931d4cfa9e4759feaa335aa60831b6b0a1eacb35
31669 F20101109_AABJVV kammerdiner_a_Page_73.QC.jpg
f6c3459777220b7b3e2f07a39eb4d4c3
822e0463e23aa305cba916439fc3793217a98e42
3102 F20101109_AABJQX kammerdiner_a_Page_27.pro
cd81071a6a2addeda9316d16e1450497
551aee219c1bbb844ea71c11a5879a54064903fe
49136 F20101109_AABJJC kammerdiner_a_Page_35.pro
d96c811aee9e272e79a81ce9697e181e
3a956c9ebdd0f394822488ba1fa9fadc7fe9dcfe
115507 F20101109_AABJEF kammerdiner_a_Page_62.jpg
fe2addb32beb844c1812459c9aeae0bf
4f6364cd0da7115c041af8d427de1d6352e239b3
7124 F20101109_AABJVW kammerdiner_a_Page_73thm.jpg
9822eee014eb436c9e965678475364e5
1703c4ebe44fe5e807eef14173608a5d969ec69f
55010 F20101109_AABJQY kammerdiner_a_Page_28.pro
274b6c37cd50528fdec14c8e7cdb602c
88e0317982c00599e465ce1ae4a256191ff65d5b
1020386 F20101109_AABJOA kammerdiner_a_Page_41.jp2
3c3bf9cc038c29f209c916ad361eb791
6692d0dd4e3e8713dcf6e8c51ef3d8e6bdfd31d7
2243 F20101109_AABJJD kammerdiner_a_Page_61.txt
a6b6fbc21be6a70a271f40a2a60f8448
ce6f96b260bfd807a6869478298349e81a733c12
620821 F20101109_AABJEG kammerdiner_a_Page_71.jp2
fb9136f50e54cb266e34009871333610
b914a3889c5e857de49bcaa60138d8f9f3dac4ca
2695 F20101109_AABJVX kammerdiner_a_Page_74.QC.jpg
25dcbfd8b4706497d4a7f2fdfc01ed69
ba737f85359de22cc93c273a269508c2bedcd73d
48757 F20101109_AABJQZ kammerdiner_a_Page_29.pro
62f6c03005d6d16e7eca0557848fe97d
d60cf170baf204e5531af91115a355b78df60def
582570 F20101109_AABJOB kammerdiner_a_Page_42.jp2
6f8149e0d535bfd1bf0bd75363990684
cb1abfce312210957433731cee0e94b8684781dc
50492 F20101109_AABJJE kammerdiner_a_Page_55.pro
61cc75b149e75bf22aa85b6b4ec3c5f3
0c5656bad76fd0c507f5b98710b3da7d4d709e1a
7759 F20101109_AABJEH kammerdiner_a_Page_38thm.jpg
163fe6566be3a9ae5e3bf57d49dc7479
c5977b07775a3c84762f4d43b99480c5c034678e
31921 F20101109_AABJVY kammerdiner_a_Page_75.QC.jpg
f266a9b22084d99bd16b7ce4adc1db14
f2a4a7380094ca57d86a230c326db7a930df5060
669065 F20101109_AABJOC kammerdiner_a_Page_46.jp2
572f9634744d807ddd8762118fc6ae6c
5bd6eb88eabbe6a91ee8681a7b85a8a4ac3c6710
48705 F20101109_AABJJF kammerdiner_a_Page_66.pro
6881c9a410564d50dbed2259f2f5149a
a1dacdcb59283bce83020314655fb4864f0d0357
102908 F20101109_AABJEI kammerdiner_a_Page_73.jpg
5acde59c148faa999808a884ffad6c45
a2d2e56a3c544304f1a9cfc90217de5da500a238
34820 F20101109_AABJVZ kammerdiner_a_Page_76.QC.jpg
0135ec5a03ec9317b6efc52781c95db8
6ebe5dc8c88258378ca9cec925fe85c444d8e22e
2248 F20101109_AABJTA kammerdiner_a_Page_67.txt
c7bf34d1181897e16f1471ed057f8cb7
7824f6950c5e4b30750acda1db19f0b9511bf504
133791 F20101109_AABJJG kammerdiner_a_Page_76.jp2
7713e38d6db4bf0972a01a19aa424969
ac44c95bcf6299955669395e0932aceb581b9211
543671 F20101109_AABJOD kammerdiner_a_Page_47.jp2
66a93401c75ac5900074750f8c5a6b8d
b3056ecd80ad4d5a3d956db188f215a58e0c03a2
2225 F20101109_AABJTB kammerdiner_a_Page_69.txt
49bc428f638ad1dc04addce00612ef1b
2c0d1aab640f3b4e875c34bd779661763a3c9f80
114335 F20101109_AABJJH kammerdiner_a_Page_16.jpg
4645cf1d33b19addb2d5cf84425457f9
b13b50bc05f344664e26ef8fddfbe24406b9476f
2337 F20101109_AABJEJ kammerdiner_a_Page_52.txt
3ee308aebcd6288ff4e5c5696592d01d
0c575fda33592dc99eb0f5ba632bbf94cace889d
1051900 F20101109_AABJOE kammerdiner_a_Page_51.jp2
1e4d95d56a6769766704a08396b46b79
3a68e3868963ae39fc52aee9e69ecd3549fcd9ac
3013 F20101109_AABJTC kammerdiner_a_Page_70.txt
f84b2bf0fefaf1cf3dff78dbf865662a
2e9d0010e07c876b33c778fd018fe108a73978f4
F20101109_AABJJI kammerdiner_a_Page_36.tif
e741fe202216fd88ae9c828856ccf880
e3373f2ddb695108b0c1fb33ae622b54dd89b5ae
99071 F20101109_AABJEK kammerdiner_a_Page_22.jpg
6c12309fb10ab708e658b970c7552a13
7d3895cae35ce61c914048ed445194d14e8f1ca9
F20101109_AABJOF kammerdiner_a_Page_52.jp2
467512f75911e87f88c901b3f468a1a5
d977653cc510b8f055d919587cd6cb6b183586ba
580 F20101109_AABJTD kammerdiner_a_Page_71.txt
43d6e6f8fe197687fda0adbf24bd2f7d
1c542b150333344dacdca91e338df53ae892dc83
8305 F20101109_AABJJJ kammerdiner_a_Page_61thm.jpg
946967ba81bdfa0ca6341c6a29f891cc
d824038bc51ce25324ab27d02d45450b748536ff
1398 F20101109_AABJEL kammerdiner_a_Page_49.txt
61eda346cdd2ea0261a6cbe4356eb0f6
66e5665bbe82b8eff12b9b0bc8a1a9c7f85821c3
F20101109_AABJOG kammerdiner_a_Page_53.jp2
b1b7dc824151a29ecfea432200b1706d
b0bd5404addc5d09dd8cc4dbbdb8e6518ab956bb
2554 F20101109_AABJTE kammerdiner_a_Page_76.txt
55e4a7658da5e9ceb4fd8eb99be9ee08
0f5b5cb293c85ea32c5f8aa53980cb67e196248d
41698 F20101109_AABJJK kammerdiner_a_Page_32.pro
02249fc7cf486b77743851c06e4df6ce
5587b64ea55798ff3edabc009c698dc2fe70e362
36491 F20101109_AABJEM kammerdiner_a_Page_70.QC.jpg
82a8145e3e25a7f5539852a240ee070d
512b8229653ef97c1071f24d37344061934cee9f
1051950 F20101109_AABJOH kammerdiner_a_Page_54.jp2
2950c34ab68ca65e73e6cee108f20f9c
cda6c46b09dfa232c2b877139742803035c3da8e
1825 F20101109_AABJTF kammerdiner_a_Page_01thm.jpg
f7c5c8103bcaf766e5eb49e9b0536103
e24ffafe7b4ec4299ef26e60b8c18f9ceb1065ff
1037777 F20101109_AABJJL kammerdiner_a_Page_40.jp2
9fe006baf5f774b539815af13398863a
f909799b4c99e97de8e04b03d0e1bba5b8ba52b0
2540 F20101109_AABJEN kammerdiner_a_Page_82.txt
d789fe6ba97dd4f430f4419fdfe38b78
cbd06c6a6d0e7e2550389d8c3a0312b35e7de6ba
1051892 F20101109_AABJOI kammerdiner_a_Page_55.jp2
05068ce6783ec6a2bf0c579e3f2cb0b3
257534fa25ab8278f85f0d2f077735584c65cdcd
935336 F20101109_AABJTG kammerdiner_a.pdf
82899352ace162cdc3aca1d10f8e0790
d7763c0aabfb316dea8553d992bff69622d694be
33706 F20101109_AABJJM kammerdiner_a_Page_80.QC.jpg
b9920bd606627b051acf77e118b11708
7323087491b002149e8defd0bda8bb3d31e44f74
97875 F20101109_AABJEO kammerdiner_a_Page_31.jpg
cf7554acdd70f91e823965d72e36b679
7b3170d7c063a833bd503b6fd753523185ab2a53
922796 F20101109_AABJOJ kammerdiner_a_Page_56.jp2
f6a0f7b117d382439e82fec3e0f4e477
e2e0fdecca4adb49de8f3d58aeac7c1991c41893
35477 F20101109_AABJTH kammerdiner_a_Page_16.QC.jpg
69a071edc155e387505348ce4181c7d7
8737a8833d20eca66172998271843c46f8160252
6787 F20101109_AABJJN kammerdiner_a_Page_50.pro
53fd5b45a266ed05845338452dd302b8
4c67e44faf402490130cbc5a99bac1b533d1a1ef
7971 F20101109_AABJEP kammerdiner_a_Page_37thm.jpg
18f17575100bef57d101cff1b04bada6
c7b1d0a667e2d728542d40b4a922bb521f3af3c2
F20101109_AABJOK kammerdiner_a_Page_58.jp2
2f7e41577e90c12857d394277678b4b0
e52f1b03d974a9fe0330c7a1d3d3c7cc2cd3ea43
1521 F20101109_AABJTI kammerdiner_a_Page_02.QC.jpg
81a8d49be3b4331abd2e56a025dde64a
9180ded9ecc8a17d23e7cc3150aa1f30370ab968
888 F20101109_AABJEQ kammerdiner_a_Page_03.pro
951db765d5db61c391d0be7135673bad
4abe56586bbdd664bac38cc1e58ed35a6f802e98
333520 F20101109_AABJOL kammerdiner_a_Page_59.jp2
c38bb258fd92a59d8e8c3c35f22777af
f2b376982c2367432a4b8e36286ed38a76bb2829
35851 F20101109_AABJTJ kammerdiner_a_Page_52.QC.jpg
ac4317d67db39e386419aa342134af6d
8ef4398a1bf38b4d15a34d2580433b469ccafff9
23701 F20101109_AABJJO kammerdiner_a_Page_06.pro
ab265d6866940b986d299fdbb3bb9de3
e7dff403065593039045e3d9fe2a9b0c4fc6a2c3
2744 F20101109_AABJER kammerdiner_a_Page_79.txt
e56de5d33d88180e55ccd8307decd876
5d71dac014c2c853f50abad01dd2abb1cb332953
1051969 F20101109_AABJOM kammerdiner_a_Page_60.jp2
b29baa9c3540b904ff1cd66da4d52d52
890b22bd9ec35601a98e84f9e7d4c28cca406333
29858 F20101109_AABJTK kammerdiner_a_Page_66.QC.jpg
fe6a0c1c81a27e23fd57ea20f63594a0
30bd111658c908d483f3885f0593a48d2ecdbbf8
F20101109_AABJJP kammerdiner_a_Page_17.tif
a9265bd7c4934c67977015ef73cd1126
cbfa9419836167b866a7715102e5ce09f35ba470
28502 F20101109_AABJES kammerdiner_a_Page_34.QC.jpg
47c67002f56dc07ccdadbff63d738f37
14c649f051aea895acd36bb4e3b6f850a92adfc6
F20101109_AABJON kammerdiner_a_Page_63.jp2
76a7557eb89c6f6b66016bb6838f14d6
22d478c2addc127efa4b1f42caa314eaae0dc5b8
8444 F20101109_AABJTL kammerdiner_a_Page_78thm.jpg
4459fc1033250ac8b7fa155a6255e933
1e47a9974a3219359fdf5f2b0b0cfadae0af3e0c
2279 F20101109_AABJJQ kammerdiner_a_Page_13.txt
e84039bf67efabe27722beebfee51cce
88a0eb4f5c9a86f53adf4476ec70c4113d8bb816
21673 F20101109_AABJET kammerdiner_a_Page_26.QC.jpg
95819edafbc49a1ba32fd26742b9a743
3c84107d9ed1683a7ea249f49bd8a264d2209e59
1051928 F20101109_AABJOO kammerdiner_a_Page_68.jp2
1adcaa624626cb49d9544790525eed3d
c61da96b83266b0f8bcadb8094809cd2c2d7ef4a
8133 F20101109_AABJTM kammerdiner_a_Page_15thm.jpg
b169fddd12b0b180fa32bf7bebf09d2d
18fd3b091e03e8781caf9cbafd42c37bf5489d0c
7402 F20101109_AABJJR kammerdiner_a_Page_41thm.jpg
666d2196fd4fd46488b07ed889b480dd
390405ec150aed7ca0b26350c767e574f6faaa74
8370 F20101109_AABJEU kammerdiner_a_Page_81thm.jpg
33a76ab1417de88fcae3f731e83d1784
b9d57e27e495d7071d89a6b66170196779360325
130732 F20101109_AABJOP kammerdiner_a_Page_70.jp2
90564176eee87ddbbc82b564b71e439d
c36fec8c4ef53617ca561d2df3defb6eafa4d401
34672 F20101109_AABJTN kammerdiner_a_Page_69.QC.jpg
087166a7fca7c5f6db1afbb125c77198
d417ab8837c92758fe6b151105f5fb189140e22d
852012 F20101109_AABJJS kammerdiner_a_Page_07.jp2
043422038b1f1904ab67e711935a876b
709557b8381359213701ddbd0055fb2622704ccc
8423998 F20101109_AABJEV kammerdiner_a_Page_27.tif
bee9e23fd5cb87d42679458bfcc6a974
102240f846624d6c232f0221f4c28fd78529a67c
1051954 F20101109_AABJOQ kammerdiner_a_Page_72.jp2
f34b7f07e6d9baf0b343e567351f8dce
58382bf3a9cd01da82227e71e977bf91b4e919ed
128346 F20101109_AABJTO UFE0022141_00001.xml
11a176271cac60056408c0a0ac17eb0e
faf69e0ecd5ef679afce2538630b8952065f248f
F20101109_AABJJT kammerdiner_a_Page_61.tif
7bdcda3f0d4c912c972836b147f98dd2
7e80e3b4b8b782bceadf92202c6955b3c9fed2df
F20101109_AABJEW kammerdiner_a_Page_74.tif
071fff89824bde2f9b6a71f25c9cb2b5
322c0bdceba44152131b42fe02216ef5f41467da
10669 F20101109_AABJOR kammerdiner_a_Page_74.jp2
54f9931cd9972d8602f33c69da31fc4b
c38ec4dc2808517a715f88a6cc0c6c390fc9929e
7560 F20101109_AABJTP kammerdiner_a_Page_01.QC.jpg
8ded2073345fb35c17e7b5b42005827f
919ba0511888ee5aded90e2bee3c90b807452af0
F20101109_AABJJU kammerdiner_a_Page_76.tif
8e0be7f45d1cc0db3732c8820fcecb94
810342038a02e1ea48706084021e70e4d684ccbe
F20101109_AABJEX kammerdiner_a_Page_69.jp2
2d2a901a8983fd50e2b2bb59bb94b56a
dc2f203de524b403cf8d05575065ef876f9e2f5f
124375 F20101109_AABJOS kammerdiner_a_Page_75.jp2
2c1434612c4b5ac6c60985c0946e1f38
787b7f7a9e9233b334b6f8d63bb0f16968808abb
1269 F20101109_AABJTQ kammerdiner_a_Page_03.QC.jpg
704b3c0db7fec774c7cefa5eadd1673c
31b836485f77b04552a82e4731b1559c6029e205
1059 F20101109_AABJJV kammerdiner_a_Page_07.txt
9fe9362871e3ed44ade8a19e7a1069ef
c9107c43e8fe05027279a4b4604d0f77325a4848
2192 F20101109_AABJCA kammerdiner_a_Page_43.txt
7ced62d4cdb524b2e3332b483b3f5456
a89c85cb3fe0561a4de3a74ac0a846110534ffa9
1048053 F20101109_AABJEY kammerdiner_a_Page_66.jp2
fad34e56781219d0c53bd926c29ea6d6
f312b52a404aae1283edd0561cb24ab680be4986
2678 F20101109_AABJTR kammerdiner_a_Page_04thm.jpg
29a6f16283e80352d805a23ba08cb297
4d7edf236c5a476a70da0244fca141b2ad264214
42472 F20101109_AABJJW kammerdiner_a_Page_56.pro
f271a28106e092fbe69d780e26c43b7b
9394af33b6629f6b47630a825ae7be05acbba5f7
46556 F20101109_AABJCB kammerdiner_a_Page_39.pro
526bbd186d372f92ed0d86e95e4600e0
ad784f80037bea56b29c56605f747b5d6a8568c1
51229 F20101109_AABJEZ kammerdiner_a_Page_19.pro
1b01d228b0cf7d3cd8be483b21d355d9
82b0cf8aa6ed4b889fdf2af59d149637021b3a7e
146525 F20101109_AABJOT kammerdiner_a_Page_78.jp2
bf48380428b000165ce460b58d6e62eb
2eac4d634db8a1de10bfe4d1ad0f8cc1feee01c0
7027 F20101109_AABJTS kammerdiner_a_Page_05thm.jpg
241a6f2664771685ef91d8ffe385b7be
929c3b77807fe934a93facc65545cc7eeec18ecf
114356 F20101109_AABJJX kammerdiner_a_Page_24.jpg
c15515a2b7feb079b05ececdeacce48f
d3ec5898c508851ede7dcd782f82bb4f3d57ca59
6826 F20101109_AABJCC kammerdiner_a_Page_32thm.jpg
27a59171ad7072f7570cc1cdb908719e
1b91881be21e11277818f67cbe5b1e5d85b34d50
138330 F20101109_AABJOU kammerdiner_a_Page_79.jp2
6995635f392377ae302b96237e5b5f18
9909f1f98e715b08d3b80a08c67e7644b026d05b
3790 F20101109_AABJTT kammerdiner_a_Page_07thm.jpg
3456f2767eb0a15e07d2049c1a96d6c9
38ed39ccbcfb42541f76f4ce5f3ee15ef32e0586
8039 F20101109_AABJJY kammerdiner_a_Page_36thm.jpg
2f8ecf1519638e82655092d2c4afb44f
ae54f8e2b2f2e4d194859f68d0bfa8d3ffb77dd0
574 F20101109_AABJCD kammerdiner_a_Page_03thm.jpg
c1628fa7f2b24eb6951c476d5853cd5e
254e6fcd03b2f2112a577ed1e1a1b6f4ea1b2471
127755 F20101109_AABJOV kammerdiner_a_Page_80.jp2
641433db5e7aa56f6ccba953694f95b8
d810f26075cb08e602b25beabcfeacd79a6eb803
1981 F20101109_AABJHA kammerdiner_a_Page_11.txt
8667418c707c58d3ca3d9718e88432d8
5ec8bf7201ae18590ea2e79fe09ce1b3f41ce08f
7989 F20101109_AABJTU kammerdiner_a_Page_08.QC.jpg
ce152db10d2e184bd40f9f4d52532608
605a05b804ad2875480942f0ca46456687f8e466
38759 F20101109_AABJJZ kammerdiner_a_Page_48.pro
732fcea0d35d85a599846faa41857696
74aba750f5a28cf23df64a8f718f45d9800697db
32292 F20101109_AABJCE kammerdiner_a_Page_38.QC.jpg
c63f54d39937832b5883d8121a3383aa
02bc78ef4ac670de952c615f9f66e0ce88deae8c
129971 F20101109_AABJOW kammerdiner_a_Page_82.jp2
138ff8344f767d4b4d1b04ca537ee2dc
279a0e62e61b6a11be8bb63ef54d6ebed8bddfb6
18719 F20101109_AABJHB kammerdiner_a_Page_83.pro
663e65c763dd462ac72a0a33bf24c0be
ff5008e7b3063c3b1601a169d538901ff16171cc
2245 F20101109_AABJTV kammerdiner_a_Page_08thm.jpg
32b8d0b2b33fcd7dc0cf27c99d32a90e
9a757fca77933b63dddc2383fe6e5ad68636b04c
F20101109_AABJCF kammerdiner_a_Page_71.tif
6767fb4da904dc35ab09c66ff6f3f0ea
11ca88411915ffb95648f5e546906704aa7f716c
F20101109_AABJOX kammerdiner_a_Page_01.tif
75674852da4c1c12178c5edc78226d5a
ab262adfac88cc1a3bf7d5507e0a93fe119c5ce2
8164 F20101109_AABJHC kammerdiner_a_Page_76thm.jpg
e3fdea3b79c95a984de312b465cafd7e
b741ee3dd99131e6b82d5f13a504cd55468056ca
3621 F20101109_AABJTW kammerdiner_a_Page_09thm.jpg
b89604f0244b54604e3d7cd71b28e96b
d9c61eac79edabe08c11ab6edd8eb30fa0798222
F20101109_AABJCG kammerdiner_a_Page_21.tif
fcace90c3d9b9bab15ccd3b9ee3f9929
dcd5969f852fb909348ae8e91fd207bf6d0a3f45
36750 F20101109_AABJMA kammerdiner_a_Page_04.jpg
829a890a2cc0ca61833e46d2b08b0856
e9c8868565a3ec4639c38a42361f8fcac93466d5
F20101109_AABJOY kammerdiner_a_Page_04.tif
ff621ef72393753b32a05494ed298e9b
132a5bf3a8cdca46e9f209341d629a61888d8def
13978 F20101109_AABJHD kammerdiner_a_Page_59.pro
77af0671936c98f2627760fef587fb94
e1d3e0f62ba79f8f59ffb647ac4f6bd4bc7f7554
6647 F20101109_AABJTX kammerdiner_a_Page_11thm.jpg
e2218f61d0c5aabf0d37680b12c021a8
1ad79cc476f40bd4f9e736e3ee447f2b72c7db28
120601 F20101109_AABJMB kammerdiner_a_Page_05.jpg
bbd1ec0272c5263e3f375908dc7675e8
43806eedc5090edbe57761dbe089e41df71dd6ae
F20101109_AABJOZ kammerdiner_a_Page_07.tif
2eef4e470d835f0e9db6b1b68c465fc0
ab586d46b91c25c03ea65e5b9cb5a085100bbbff
89610 F20101109_AABJHE kammerdiner_a_Page_11.jpg
48aa12271f0c56462b3707597e83f589
f4434c071ff5f6f7a7fd7bb7be6cee79352d12a9
25978 F20101109_AABJCH kammerdiner_a_Page_08.jpg
0f1654e50d9d7e45bcaa0fd88d6e68d6
cddea5b81418643777d61958d71f0cf79947266a
50450 F20101109_AABJMC kammerdiner_a_Page_06.jpg
e146f55c6e8a425d1eaa72962f475f78
0ea4b556c63a44cb5995d922611cde99b57cc89f
47888 F20101109_AABJHF kammerdiner_a_Page_40.pro
5071089602455755a2e9b926d39a4f4c
f77e592f871167538a41e8944cb926a14bac8687
34667 F20101109_AABJTY kammerdiner_a_Page_13.QC.jpg
d033ae256e9cd3439286b7f9126fc49b
bb4c0d14ea38f987bfcd271e8e69c66ebf43b6a6
37710 F20101109_AABJRA kammerdiner_a_Page_33.pro
7ffdd0b81feb17f346f5a563a4f47b80
e80249ccc49d254ef4d256a237fded11334da8c3
56137 F20101109_AABJCI kammerdiner_a_Page_60.pro
060eb4eb93d650850220b6e948d21436
1ede4dc21d35b91f28746cb8f005dc8f3822d8b5
110403 F20101109_AABJMD kammerdiner_a_Page_12.jpg
1d2ba16a0670a944f07d8f1476be87fe
6b1c8d514256c30f2bb17b46cf0fc5f8b4a8673d
30113 F20101109_AABJHG kammerdiner_a_Page_64.QC.jpg
2f1694785e6d36b6788fe822e61d20a4
45da7520e13063dc16f8ea1d49caccb039399c22
35325 F20101109_AABJTZ kammerdiner_a_Page_14.QC.jpg
a98ad7407a0f7dd1b60ed8f1a89dc7f9
6bebd8a6eaceff213ccacbbae441b3c875ff77e6
F20101109_AABJRB kammerdiner_a_Page_34.pro
486c558f47d8967678c3392c3c4a137a
d1f8ae1a6ccdab7f3d742f60e34ebe1ad3a9cbbc
F20101109_AABJCJ kammerdiner_a_Page_24.tif
14969e40a8bd9a89d11e895e208bf354
9a8dac75533dc67fc7a768791549bde539d07e10
112556 F20101109_AABJME kammerdiner_a_Page_15.jpg
7f8fb4c7051e86bb023614b786793ce0
c3d458bcc63f402a2d11cb85277593dcbe96f0a2
62452 F20101109_AABJHH kammerdiner_a_Page_46.jpg
68c6a09570d833bf9ca72e3305427b91
5d5721f704336c7b106e18e8887d5012c26274a0
56595 F20101109_AABJRC kammerdiner_a_Page_37.pro
9a5c3ceeac9df03e5dc33bfb62af83fc
4f90d1ad47a6c404f22da23b8a65d17aec055379
7931 F20101109_AABJCK kammerdiner_a_Page_58thm.jpg
22c05593c24198a489540048a6e9fd6e
1e445d81a9fb6bf690f5ec4c3e631a966b2789a4
101120 F20101109_AABJMF kammerdiner_a_Page_19.jpg
f7df07882af0ccda0045f2b410678a85
3b36b47f662df891cb368b11229ef3f7f8b71775
F20101109_AABJHI kammerdiner_a_Page_41.tif
06635375e22d3b272b69e8a25ef8db0e
89e87029fadf0598063e5689d77789df9a4d13f3
8187 F20101109_AABJWA kammerdiner_a_Page_82thm.jpg
13a8cdf38f2d9b43cce9457975565f28
909d46ec2c5a6fc61ef6caf72bef8653e09fe4aa
48816 F20101109_AABJRD kammerdiner_a_Page_41.pro
8b9c7ac20154467ea64c95e0762b7a7b
fac3ae3506abf3f6876d063a02b6e70c1a3ded21
8327 F20101109_AABJCL kammerdiner_a_Page_52thm.jpg
39c76eb6004e408d31fb2d275996b235
3dbe8d1de2867a1eeb553e719c5cfe5d45e424b1
112732 F20101109_AABJMG kammerdiner_a_Page_21.jpg
2455a7b49c2eccc06cbed6a2d9d9deca
6765a5a14c6ca6e8418dcf82e2f5d1d4504593dd
F20101109_AABJHJ kammerdiner_a_Page_58.txt
a8807162676c404c407a477b76610387
f3496c87030f8aef74bac115631448e33baf5c4f
3100 F20101109_AABJWB kammerdiner_a_Page_83thm.jpg
264c98600d0b036f854fb7625f909a83
7b179489a4d38d2b1f93d169f929403053ccc663
53524 F20101109_AABJRE kammerdiner_a_Page_43.pro
629dc30cc59de07d7973fb1a6a58c42b
2ccb64657c57ba20f58e8d87bc10cd69b5b8935b
8411 F20101109_AABJCM kammerdiner_a_Page_70thm.jpg
7082ddf5af0b4df79f5611bf1848520d
ba184fffe111237d1556c4b671f8cce5c435b454
116167 F20101109_AABJMH kammerdiner_a_Page_23.jpg
2430128aef4be3afac45a08c2ff12e54
0f87b80b0f8b39b2068ccd1c96c311da7a916535
48690 F20101109_AABJHK kammerdiner_a_Page_09.jp2
c12244a70c513f678ad670c26cd27ee0
c7ca40e9b91cfcb2c125502a39f0b20f3b6bc70b
52774 F20101109_AABJRF kammerdiner_a_Page_44.pro
ca3bc56553dbe017d428a2509785615c
7b5c98f3f9e33d1e0df13b44b47d5515cf58780c
105285 F20101109_AABJCN kammerdiner_a_Page_43.jpg
acc99961c90bb20aa7b5d8a210578e38
50b033c3c535e8adcca370eb99051fb885854d54
68952 F20101109_AABJMI kammerdiner_a_Page_26.jpg
457f51cb69c31d94c2bf005c172746b8
cb35ca00b77b7c67fe69650bbbf6cf760f26adca
38346 F20101109_AABJHL kammerdiner_a_Page_30.pro
c374abb9b0e3311a08df566a48bd9661
8870e7ee03755c60ae7542c3bc195271febff409
35848 F20101109_AABJRG kammerdiner_a_Page_45.pro
1830648fa3360614ae77a5ae8cb9c346
ae4aecd479d2f23a8a3e28be1927659a0efb9a10
34163 F20101109_AABJCO kammerdiner_a_Page_37.QC.jpg
a01a512bc8df72edb5f88763ad82f019
6b5ccec0305d361e3140a2e66cdbf344a0a4aa54
49573 F20101109_AABJMJ kammerdiner_a_Page_27.jpg
f6d57da51a59e1b4963093775880bf08
8e9f21fced3c1c45228c78096673d3c079d1cf7b
59254 F20101109_AABJRH kammerdiner_a_Page_52.pro
e231c7dd61f82c618bfe6d22c162f439
b70f1ab6e6dbbb8f92338294d89c1a54066cea59
968 F20101109_AABJCP kammerdiner_a_Page_02.pro
c22e4b52da5f6557e9058d1b699c5c67
dcd494eb3cba9b7b7c1ba7e5041438f2ea1634ec
113318 F20101109_AABJMK kammerdiner_a_Page_28.jpg
ed691001be2ee81854ad7ce9ab91253c
c021552c61898ffa7a40e6f2421f3fb2f4021b44
114798 F20101109_AABJHM kammerdiner_a_Page_14.jpg
c25e5017d1e78e9f0b173d051420ec23
5323e49847c5d15b3740887bb5194eda0e70f419
57944 F20101109_AABJRI kammerdiner_a_Page_62.pro
56c4199177ddd015e04470017e2478cf
71effe58c54e360dd0189eeb31aa6f97b0592772
130142 F20101109_AABJCQ kammerdiner_a_Page_77.jpg
cd0ce5beb2cdaaee81ebaa3ed0c0e165
b6af101610b91a95ae800fd75cfd6c9654fbaeac
96220 F20101109_AABJML kammerdiner_a_Page_29.jpg
fa1f7013ec62cfed34ee23c0867d1d67
99c588a2fbe8224fdab495dd4b696b0200ab4d05
F20101109_AABJHN kammerdiner_a_Page_62.tif
3774335e191c5adc7f56a13e29b548eb
9e8b4d26cd6a7ecf9f4b1efd6db2b44d47d7173e
49298 F20101109_AABJRJ kammerdiner_a_Page_63.pro
684aee71d8b5c5c8a3b113ff4a08780c
6621dde2f40742ea44cd0a2d090cc8ce57ad4f15
F20101109_AABJCR kammerdiner_a_Page_31.tif
95a28995e01a65b8c64ead75872c2d1a
3ad6bb0b39d8197f09f2d568c7af00d58003fb60
73987 F20101109_AABJMM kammerdiner_a_Page_30.jpg
094f1aab5f25bd973895fdc3bb5390fa
73eb0fc97ad6f4b81913d75b8c9d455356e6bc50
F20101109_AABJHO kammerdiner_a_Page_66.tif
dd12933cc4ee6ecd2f37d4d5df634cfe
524b7c641dd348caee5dc75983b6a4c8cdf3ad71
43982 F20101109_AABJRK kammerdiner_a_Page_65.pro
143019c563d647faaaa226af7e060ccd
9d9c8084c5d7e97abde3ac382f6a1666eed576a1
88783 F20101109_AABJMN kammerdiner_a_Page_34.jpg
cb7720df5285127ead755e2d579df8b8
aec299fb10c70aa82ffd40efdf60937370fa611f
8024 F20101109_AABJHP kammerdiner_a_Page_13thm.jpg
104a88d224cc0b7519aa265ed0ce680c
bfad6f7590e05bd79bef1da47620687c0b8ca386
1950 F20101109_AABJCS kammerdiner_a_Page_56.txt
9879b4728fe09097c513d3ba4b9a858b
ad3d1160fa94962bb4dd1c2987d91aac98aeac94
53437 F20101109_AABJRL kammerdiner_a_Page_68.pro
9dbfb8371ddeb24216c78e94273dfd58
0bce2bc0d505ee419ab45ae68e313b89ea5cb131
96931 F20101109_AABJMO kammerdiner_a_Page_35.jpg
67dd6a05d2cadd64b51c89e8899b3fd6
e7ba5c981f204bb0f69c4e9f4dcb8df1e24694ca
F20101109_AABJHQ kammerdiner_a_Page_35.tif
9f80a516f823eb8b4228c2592ee03c2a
ac0d65f3ff08e9d4b5f35e645ad92760517410b4
58810 F20101109_AABJCT kammerdiner_a_Page_53.pro
1c5f639fee0fb42574891d4be56ae775
aacf547835afab56bf95d79260874d523c4f3fb3
52729 F20101109_AABJRM kammerdiner_a_Page_73.pro
42f39d1c6172a128a0e2baa30f1fed79
156629f1a9684d83e435951ef171503f42cd97c5
109258 F20101109_AABJMP kammerdiner_a_Page_37.jpg
3727390ee667e7d9043c615641879d57
fa946cd4f6d55e2b5910d7515b171b58b2b37e17
5100 F20101109_AABJHR kammerdiner_a_Page_10.QC.jpg
7c5b73378d6e687bf632ffd2e151b655
6877da9014918cd01981eda9aebb991ab4e85fdd
116100 F20101109_AABJCU kammerdiner_a_Page_17.jpg
f75aef14c9392db70b8f56ca8a1e401f
239c9229027cdd3490be5832fbaba4e2168d522e
3243 F20101109_AABJRN kammerdiner_a_Page_74.pro
dd643b57ea3fb70dea3efbee304bd9a6
86cb22caa5881587fd1866b197b814533fe9a593
89626 F20101109_AABJMQ kammerdiner_a_Page_39.jpg
5eaf2b65c990cce47898fb24f41dda9f
343e602ee1aa23ae868af937763c030a2088451a
30563 F20101109_AABJHS kammerdiner_a_Page_40.QC.jpg
f5c25b58bfb36b94c65801f9751a9263
522d58a76e65b464460b4d40f54ebfcbe6518b98
791 F20101109_AABJCV kammerdiner_a_Page_83.txt
41d49e9b179278f1dc56ab1cfdf458d0
b443603cc96e5a54d181517e7f00d5938ba6ee42
68951 F20101109_AABJRO kammerdiner_a_Page_77.pro
99d0ebf4bd5c5a0940ef14ce86981ddf
3f796055fa0cc27707eb90bc8b0c7ecd12d84ba5
63881 F20101109_AABJHT kammerdiner_a_Page_76.pro
347fc249bfb827511b1ee642acca0e21
e9859b4e096dbe799d6fa3369ade7c795aab8920
97045 F20101109_AABJCW kammerdiner_a_Page_57.jpg
b924536d03568f1c1dc2e8f4316899b1
c6ed21245eb33f3cecb12ccf49c92137b22a0360
68151 F20101109_AABJRP kammerdiner_a_Page_81.pro
9ee7ce4937a32e40d183d25ebf8f27ca
f1170fb8ba76c94265ce67b30e83309a28ee7a09
97721 F20101109_AABJMR kammerdiner_a_Page_40.jpg
dc2a278d08a2fa8984d1bfe925e4d3d8
4398c14cd62b55e6e7d5df409dc248d64b496425
F20101109_AABJHU kammerdiner_a_Page_20.tif
fb75c57a90ac01837ade528e65c9f748
6df9b7af9d586819c8d2b4db908683b30543ee37
8257 F20101109_AABJCX kammerdiner_a_Page_17thm.jpg
1c9a2051fc5654e25caf1f10e5376273
5af92b670e53dd2069037f5efe1b94fcdca6ab4b
62844 F20101109_AABJRQ kammerdiner_a_Page_82.pro
cb868bc9eb81edd5c262ffd89c9e0f2a
733ad3789418793b16011f4fb67767ed1a956c0c
93120 F20101109_AABJMS kammerdiner_a_Page_41.jpg
8e0fe3e91fca55743e6d7f8be4cf40d1
f1398305f35748ed9528e39bd46700e9a6beb541
35492 F20101109_AABJHV kammerdiner_a_Page_77.QC.jpg
9fa5b9be1c9400cfa6e6841dee6b3301
8af8b778d43d99095d6529582dc120a66f359c7f
2102 F20101109_AABJAA kammerdiner_a_Page_42.txt
5bb79e858288f32e6326ed5709fd7cb7
65200fa8564c63ee73682c7d4e13657f384de241
33269 F20101109_AABJCY kammerdiner_a_Page_05.QC.jpg
09e28557f92923f87ebac8c6060f0d42
d313c76554b6d179af2c758a11496817095e04f5
95 F20101109_AABJRR kammerdiner_a_Page_02.txt
665b1bcb715511a54eb1900b89518873
38074449f2d817b804e3466ba4e8feec9b895073
54219 F20101109_AABJMT kammerdiner_a_Page_42.jpg
60db4b077535265e2db9cf08eedfafd9
971e3d2f42c7c37b132beaebecbb5f04ca136eb4
24001 F20101109_AABJHW kammerdiner_a_Page_72.QC.jpg
11dade1af2e37cfeddbf641cf9f4abfa
5c13671aa75e2871075bc4486b7158ded48e1f12
14655 F20101109_AABJAB kammerdiner_a_Page_09.QC.jpg
30f398f566c0247459b3dd09a898b761
805191646fb321fc517ced8413fcc26ee6a66ad2
F20101109_AABJCZ kammerdiner_a_Page_02.tif
a70cf8f12b14f600dccb875f95ca4104
8aa54be51ce09c8d46b322619f1bcf38d2f9a401
686 F20101109_AABJRS kammerdiner_a_Page_04.txt
d831ac1939df75d5a166ff13d88b4226
c638a0a946593f4ae606139d21cc7d0a9f93cf23
51906 F20101109_AABJMU kammerdiner_a_Page_47.jpg
3cc17c2a17c05befc1fb80b6d5a2625f
64c2f7fdc5b43b56aa7e07d8a0c8093c8bf511f5
2848 F20101109_AABJHX kammerdiner_a_Page_78.txt
f81297f23412f5ba32751f13050f2322
44280d05816a53a707888c74e6d6875d0beb4581
49534 F20101109_AABJAC kammerdiner_a_Page_51.pro
dd9c685d5e0f5e7b06b7e552488b0ceb
55af0a7d302a5b38f62a9cc6cae2076c1d2ef7ef
2360 F20101109_AABJRT kammerdiner_a_Page_05.txt
12680e6fe2d4e44cb8d4f1425ed7f388
800380e9d94ea02c0bdbae902994ae244da38b06
81436 F20101109_AABJMV kammerdiner_a_Page_48.jpg
97363c819ff6b3da9b5337526865a4cc
b5d57b31f26ec001c58312c24c33d23685463c70
1051983 F20101109_AABJFA kammerdiner_a_Page_62.jp2
1eacb037a353406360ff647310e9d5f4
aa664fb80f21818b7401658ebcec169202a0f26c
8328 F20101109_AABJHY kammerdiner_a_Page_16thm.jpg
5a125caa08e406a06505903f6504b9cc
32a4078598a8ab7c0b8302e05952eba8986b9748
2292 F20101109_AABJAD kammerdiner_a_Page_62.txt
d819f05534e498dc084db4329bfa2fd6
1ac3941b299aa90cd49ee7272badc2bf56d97717
480 F20101109_AABJRU kammerdiner_a_Page_08.txt
49729683a492777787aade9c8ff2d461
362435cc5ad31cf6a848b3b570ec75555be33d4f
68868 F20101109_AABJMW kammerdiner_a_Page_49.jpg
bb12c35461238cf066002b56eace796b
74c6754a76f2e2539fb9816c17fb6e11c8f7fe8e
1998 F20101109_AABJFB kammerdiner_a_Page_63.txt
5813be9bbbc20c8da8d67a195704ef8a
ee1696263e2bed18727f8c6f25b1c7b137e346e2
112575 F20101109_AABJHZ kammerdiner_a_Page_73.jp2
8c9e36220449f83cd9c1648e7266f651
bcc6e9888c3c84895d52d117e39b5259ff6a541b
6845 F20101109_AABJAE kammerdiner_a_Page_65thm.jpg
3359a066ca28a67a80b45d0248df53fd
b493a2acfb593c4e827438f602deb360ffa2dfdc
F20101109_AABJRV kammerdiner_a_Page_15.txt
eb7c5e8bc50043de342fcdb677c078fe
9aadf447d33b1ecabd335565940ffbecec552ce6
17915 F20101109_AABJMX kammerdiner_a_Page_50.jpg
608e880d1bc3c1362c066d30ef5e21c2
10bffc9031247f32def1c765fdd39cd0ce62778f
23412 F20101109_AABJFC kammerdiner_a_Page_47.pro
6ed86943bd4348641a5f18114dd10f2e
958fbcc42b5c66100ccbf7e284472c1e9990dbd5
115976 F20101109_AABJMY kammerdiner_a_Page_52.jpg
c9c65d8f42a681827d66636b1caffedb
3749ea7e642ed4731e8bab50a9cedcce563eeb9f
F20101109_AABJFD kammerdiner_a_Page_63.tif
dd2d9b3915c6a0a5259f153907719ca9
8251e2a0fb41dfc6e8432f7c37ac628571f4cf80
8203 F20101109_AABJAF kammerdiner_a_Page_77thm.jpg
7c2ac5a27b224fea69f4663f7a0798b3
759daeeeb250d4b7ebc4d681a693363364bf9e80
F20101109_AABJKA kammerdiner_a_Page_20.txt
f33adbfbe4838bc7f6cec6b6bba1956f
a517d230c4297cee47aad62f57e4d15430f43d0b
2334 F20101109_AABJRW kammerdiner_a_Page_17.txt
eec2a2029e7c763b3b792808ba651f2e
c3f7650a7f5fad771d7b7d3d64189e90bd656d9a
114431 F20101109_AABJMZ kammerdiner_a_Page_53.jpg
712cfa934f248f8c991a451350c98aae
a6f7465c545c6f43e34c3fe7c0a420c87d20dc7f
34316 F20101109_AABJFE kammerdiner_a_Page_68.QC.jpg
edf40ff91dc044ad95cbb69e37dad6dc
60e218125bcba02dab6ccaf1f016ca18e4ebf2bc
13405 F20101109_AABJAG kammerdiner_a_Page_71.pro
ac81d67cfd6c50c9239898cf459325a9
9cbe58df29e1f49e9c52349266ab6388b906dc37
88 F20101109_AABJKB kammerdiner_a_Page_03.txt
f647853c02f69091de7ee4281a7336d6
ba822324dfad209aad698cedbd87271f3af96cee
2383 F20101109_AABJRX kammerdiner_a_Page_18.txt
e8e0904845eef14897b6c82a2677a3d1
7b630309baef8b0b25a8028345caf087c9376cbe
1051970 F20101109_AABJFF kammerdiner_a_Page_31.jp2
4b13a6a3d3181726760c33c1e9f0c6b1
b8ba2d3948ef5fe62f77aa4730aedac086558783
F20101109_AABJAH kammerdiner_a_Page_11.tif
d799debc52e0e1b586df05045385b05c
0573e29fc2db20084c69df0e2c097c847241126d
8122 F20101109_AABJKC kammerdiner_a_Page_68thm.jpg
21121019d1a4318a2397fac762d0fb78
f13c964bd5874cb3746cae6f2de41b1d06a427fb
2171 F20101109_AABJRY kammerdiner_a_Page_22.txt
326fb6347d06a4b66594ca394f1b6d58
839440aca9c1423e956f84ebfc8b914a59cb7381
F20101109_AABJPA kammerdiner_a_Page_09.tif
0506026812a0079cd94a75fc7cf119fc
4e473b3f7244d9c4dd898bacd307ca8707d5a993
F20101109_AABJFG kammerdiner_a_Page_59.tif
284044fd5201b2f3ed77efd454468d9b
c86765b2fe9a5c0d0febc8babae8ea787561eac5
5988 F20101109_AABJAI kammerdiner_a_Page_30thm.jpg
132d630067155a65a02bc4ba86f138bb
5e7eafeac1273051065934589e2a9d714a104014
F20101109_AABJKD kammerdiner_a_Page_77.tif
205e2fd64769a82bded51be672910563
929f40c9de542fd954cfd304b82782d7da0896ab
2310 F20101109_AABJRZ kammerdiner_a_Page_23.txt
1d5f3dd989cff8291f954e050b493fc4
15b324e48ea4cd76c8adc7cbedc71db5cacccf32
F20101109_AABJPB kammerdiner_a_Page_10.tif
dd79e55db5ceab10f648707cca358439
bb7481118f09fcde2a1e067b9c6f4ca7d9ea86a2
F20101109_AABJFH kammerdiner_a_Page_52.tif
7fcad3f21314f12ce7951b14b923bf88
bcf41b53782003b5ba5dde40c14e5a01e141d522
1276 F20101109_AABJAJ kammerdiner_a_Page_47.txt
b2e6f6d4065032b24c844c084b94ca54
83f8b189ee641068d9f9a4f2660f62b1875c779e
F20101109_AABJKE kammerdiner_a_Page_03.tif
be8880b688a3b498add70906d26ba840
7db2d840037b918e7a5eb8242cf37ab3df51bcb9
F20101109_AABJPC kammerdiner_a_Page_12.tif
19b04e593b2f06aa50bafa7380a7341d
6ef2d2b3b43e82635438b8287972036ce56e6ff7
68637 F20101109_AABJFI kammerdiner_a_Page_79.pro
c19d1f7cb24f9bf54560ca81b61de9e6
c5e51e2738af66b2f086dc8c0b76a2fda6b1b86d
F20101109_AABJAK kammerdiner_a_Page_18.jp2
d1ed36b95b89ed3baee4e89e96781b33
58ea442dd3c6c1bedd2257363bbe35681d977790
F20101109_AABJKF kammerdiner_a_Page_34.tif
13ecb41b0478cad184ca859812d9123a
f8237a943939b67ca54f02a21fdb4b8a2c3eb37b
7963 F20101109_AABJUA kammerdiner_a_Page_14thm.jpg
0eeee16f540968fbf724fbf46444d95f
2c57ceec73a03526ae8e40280f0c20ab2e55e542
F20101109_AABJPD kammerdiner_a_Page_13.tif
19512fe1da73d7e17c14019f338577ee
07ad10aee2d58af5a3dbd355cfe59677802ef1d7
71025 F20101109_AABJFJ kammerdiner_a_Page_45.jpg
025c7b98f870ef348a46e739a99c26ee
bb69e10bd69481fd1826424f1f9f14fe38707106
112119 F20101109_AABJAL kammerdiner_a_Page_69.jpg
69576f80165405146f6af7c68092a417
32a3dab0770bf6a6ea29bbb286e36ead65a8c109
17273 F20101109_AABJKG kammerdiner_a_Page_42.QC.jpg
8324ef39514cee3b73568ee988de2712
47fc4ba6fecddfa56ce9d415adafafee5bae2e9d
34892 F20101109_AABJUB kammerdiner_a_Page_15.QC.jpg
19c6a4e633241e81eeac0167a42851d8
12cb6fff3056cd1837df806b093666f1b3647420
F20101109_AABJPE kammerdiner_a_Page_14.tif
5dd3cab1c17d11a808accaae6a31cfce
016f5a667eab35dfb93a9c917e2bec2f308c6df7
2028 F20101109_AABJAM kammerdiner_a_Page_31.txt
aeab8bbb84ca4cb96d25d9052d620862
0200acf87ce33e1168f9567a1fada152255b9595
1051923 F20101109_AABJKH kammerdiner_a_Page_43.jp2
f8ddb59397aa4bdc8358c05f5dbb43a1
7656cc0ff9e3f913731ba6e0969f102c12b5eb9c
31844 F20101109_AABJUC kammerdiner_a_Page_19.QC.jpg
608f2b20caa4620fb84bd2d7718f35a2
6e463ab7e1f1ab4d0a59d43f4c085d0c4cce146d
F20101109_AABJPF kammerdiner_a_Page_15.tif
4c265270a7fc9fad89108baf3fae73fe
37c731b950829e5b3db878eea2f2870a2ce8526b
114986 F20101109_AABJFK kammerdiner_a_Page_60.jpg
56ec0e3e8ba35496aa45e4b709636df0
65d7c5ecaed634c328def637f864e6cc8b064f10
F20101109_AABJAN kammerdiner_a_Page_70.tif
e75198005a7209939b368448d7ba09d6
9a40f81572e61289331bfe1934abca30c3c3001e
30125 F20101109_AABJKI kammerdiner_a_Page_41.QC.jpg
417b9d0e5db40364adeb12745954201a
ab0c72386bb63ee3742390cdac77836fbeb456eb
29434 F20101109_AABJUD kammerdiner_a_Page_20.QC.jpg
3436e18248ff730624c79caae17b3332
64e333fd9c0dbf1d6dfb87b03eac971ffb09b57c
F20101109_AABJPG kammerdiner_a_Page_18.tif
e9feb7ddcf382e635ac4819eb7e995dc
870ae2abce4c9b67ab019733839e69f059f109e4
2371 F20101109_AABJFL kammerdiner_a_Page_75.txt
babc37c854be98800bded1f05ecb0c33
64cda95316a5a90b86503b32412a6d504563b743
11756 F20101109_AABJAO kammerdiner_a_Page_04.QC.jpg
4924f3d9a93ea2283fb8e3dae0853ec7
de14a68713a5d5f99b9d6ae007e90f747b0d3fe2
F20101109_AABJKJ kammerdiner_a_Page_08.tif
f68ae0dd3aa43d62fb67f59b4db0f2e9
21ac495f1b99228bdd93ffab230893b9451d6512
7548 F20101109_AABJUE kammerdiner_a_Page_20thm.jpg
993a5060f236eb5d711103b3d7675e63
c8701d47ab173c7eaf99d6028d5b20eeb49460cc
F20101109_AABJPH kammerdiner_a_Page_19.tif
eeb372235f24207b9ea0e2ef465a2eee
7009cdfaf85bb6cccb7cba771358dd0217418b34
45806 F20101109_AABJFM kammerdiner_a_Page_57.pro
384cc911407998a2cb85a895e4d63aac
0a56378072f78d532ceafaf0060e2d26bb0861c3
32805 F20101109_AABJAP kammerdiner_a_Page_36.QC.jpg
36d6d54ad3373e2bf1947af63ceb4a65
795e65dbc259033dbcc13bee0124287ee336b0cd
35479 F20101109_AABJKK kammerdiner_a_Page_61.QC.jpg
66427081be40634c9c55a1be9607db27
24110c0e53e1d3ad06d511e42510c08038d74ae4
34782 F20101109_AABJUF kammerdiner_a_Page_21.QC.jpg
7dd74ff1bbed20604cc5b0ef59ca95fb
6088bda1d858434bf1e0bc4a620b7aa2b3bc3be9
F20101109_AABJPI kammerdiner_a_Page_22.tif
680110b7e941a3ed6f8ff6b88bd3a3c4
e52194de6488fbed3f54da0aabeead82e9415430
107949 F20101109_AABJFN kammerdiner_a_Page_36.jpg
72bc5504c233304bcc99c95935cb8d8b
7cfb6741c6b090a8871891e33911ecdd1a435db5
2124 F20101109_AABJAQ kammerdiner_a_Page_68.txt
04cc39567a99278bfc1aac6a047618b7
a9e38719cd6a9e28bdd21fbfc1a6ec4404d270d4
7761 F20101109_AABJKL kammerdiner_a_Page_75thm.jpg
469791f1d973ab34a59d8f8642ae31d0
f6c6700b091706162fb75f00fb8f1fc1def8dfb2
7587 F20101109_AABJUG kammerdiner_a_Page_22thm.jpg
5de9f016f8179443c3d0b4c7eb582592
72c57a66ef59d7d8d9a70d140629babc443d7d96
F20101109_AABJPJ kammerdiner_a_Page_25.tif
20420c682f3dbbcaa0ef8224e85e4c8a
837fcc38f7370b044c238d74f1f7eb0f0679e109
1051953 F20101109_AABJAR kammerdiner_a_Page_28.jp2
f8b37aa30e9a43e0bc8d961f291ed03f
fde0b19cf7b4631c3335f7bb900fb4910c02c7e1
737412 F20101109_AABJKM kammerdiner_a_Page_45.jp2
4f017e1b22ca0534bf511852a56d8b57
85ee57d29285cce394c956470ffb4519e4ac1d83
120318 F20101109_AABJFO kammerdiner_a_Page_18.jpg
5b0b26bad1f671b14d0394e39e609004
46ab667bfedf7a8e6ad66a2d417024932a9f284a
8372 F20101109_AABJUH kammerdiner_a_Page_23thm.jpg
b89c200c668027ada4d0e2f81a482f9e
8cfafb6154994fd110443ffa3914a7031d5ae383
F20101109_AABJPK kammerdiner_a_Page_29.tif
a83824fb0b003730301b9d76948201af
364b56f86db496449ec60cb9d0b7746fc9c03e1b
8098 F20101109_AABJAS kammerdiner_a_Page_54thm.jpg
53804741e081b8896a7eb82d03168fed
844fd26135b82aac3f127f965cff66dc6d98f7f1
15890 F20101109_AABJKN kammerdiner_a_Page_10.jpg
b0f43c2eae8d5b301808bc58b4448285
cc5848d7e074a821e9175818bd3794e67e271e07
4707 F20101109_AABJFP kammerdiner_a_Page_72.pro
9eefbc3e5d5cde0632c32454dcedbe26
cf8aa03704cf9bde670674b6f88d5494e8a8c3ab
F20101109_AABJUI kammerdiner_a_Page_24thm.jpg
269b85cb3263906043207dd688844941
f2f4f61a2e5a561105dd9972a75482b893876b57
F20101109_AABJPL kammerdiner_a_Page_32.tif
68eeed91b00382edfdc13b07a9d69431
d44167b929723d2df7ee9dcbab6f1399fb039179
F20101109_AABJAT kammerdiner_a_Page_64.txt
1df0b1b85dd22fb957439c762a709030
d1e9fafad1789c4558de1b650b7cd440a2ad2552
94728 F20101109_AABJKO kammerdiner_a_Page_20.jpg
7e8960f4304c1d6db89cf0785a2a4e15
0f164420b45fa1d4ca9704f9695e56a6fa45b23c
1049 F20101109_AABJFQ kammerdiner_a_Page_06.txt
05f37c58d0d049449b99e966ebb334bc
73b878c1c8663d2f40c5dbb194ab48842d902e92
37111 F20101109_AABJUJ kammerdiner_a_Page_25.QC.jpg
e54ad4a77951a6800f548f690159ba51
879e7d2530436f68aebd49aad57bb7f78d0b6f89
F20101109_AABJPM kammerdiner_a_Page_38.tif
89b06a2097321c5f4de840d773b4377a
a8d5387eab649f353cac48f1d4e53dcf4bd38553
1032216 F20101109_AABJAU kammerdiner_a_Page_64.jp2
2f7f502bb5cc40fef275688272711adf
358ccdd646b93644011c3096ef91c0bd451b52cd
87902 F20101109_AABJFR kammerdiner_a_Page_72.jpg
9bc29d8892a06cfae0317485faf960cf
a8e7791948513fae6196c096a6658783e7bb4df4
5355 F20101109_AABJUK kammerdiner_a_Page_26thm.jpg
f6a546da86b2541c0553b880ee063d3d
9e0ad6b56517b83f9d8d20ce96bbdb7c184159a2
F20101109_AABJPN kammerdiner_a_Page_39.tif
d8c7adc36fcddfe2ad029dad460ab7a5
5c85110c77b34df48fd5f90cab6a1ab50cbecb90
7118 F20101109_AABJAV kammerdiner_a_Page_40thm.jpg
f98d615382c0f4622d97dd66827ae83b
dc0a766e858b7e870a21ee1b784db5e99a7d10c4
50246 F20101109_AABJKP kammerdiner_a_Page_31.pro
710daf665da44d269c0b5950903de81b
7a408eadb4a1d40f911ec253263173c286da74e0
2274 F20101109_AABJFS kammerdiner_a_Page_14.txt
76ea86811e6fc1d5397189a83400d626
c7dcf2e19d5fedfa82f8ef02e134d62356401740
16544 F20101109_AABJUL kammerdiner_a_Page_27.QC.jpg
157bcfa5db3e5ed80deaa91388f25b73
de24cfb78cc92c408d61bb83be71f3fbfe20325f
F20101109_AABJPO kammerdiner_a_Page_43.tif
224a452d6d0bdf1d7179300acf2f434e
9b77e4fa632ff4fdf8c567df2edc9d5f8cfe42a1
46826 F20101109_AABJAW kammerdiner_a_Page_71.jpg
869d7e5b0f17745181888e77921bc616
61442abe1e53722350c50cfdf58da102d7cecc36
50394 F20101109_AABJKQ kammerdiner_a_Page_07.jpg
6f6e10934e805cbc529721eeb859c39c
2c1811e9d3e1668aaae713f1f4be356a0661cff9
13830 F20101109_AABJFT kammerdiner_a_Page_83.QC.jpg
13a9e1b2349f6c84c7ec1abacb856aec
a4d1f13be83af711c8ff485bf47378bd64a8329a
5085 F20101109_AABJUM kammerdiner_a_Page_27thm.jpg
3307a36e6845e0938654f448af1fed4a
f99781fd5535c10c5835c30ad20e74030a95b4dd
F20101109_AABJPP kammerdiner_a_Page_44.tif
e7e6a6fd00f1f419f1d38d7c15a2ea77
0f88df86dd6cec0d398ad8e404752236493fc1c1
F20101109_AABJAX kammerdiner_a_Page_58.tif
85576aa3a8d90152b3c517a36f92e1a1
f49342ee61b380582f7a19cc0666b72f3f3cb666
71372 F20101109_AABJKR kammerdiner_a_Page_78.pro
a50cd83c1671a97fa889f4f116ddb332
f655976fff21860a041cc0d7d8cf576e3fb25a90
16296 F20101109_AABJFU kammerdiner_a_Page_04.pro
acde82aa699593a8ed24891874231a56
3ff3d6e45747918669c806594522c72266b90428
7903 F20101109_AABJUN kammerdiner_a_Page_29thm.jpg
90cea5588caf12730ef1e526e34a347a
7dceeb4d1a812926baa7dabe470a3ca352a93471
F20101109_AABJPQ kammerdiner_a_Page_45.tif
0d9d3be1dc3cfd4b13a3ededfd0b49ce
cd6325bcd4ec11200251f95460d9100955e844be
29791 F20101109_AABJAY kammerdiner_a_Page_63.QC.jpg
3e639555b885cc51ab5f1e918e83c203
e307657281a5ed8740ba6a58404f4dd8a5da3291
34036 F20101109_AABJKS kammerdiner_a_Page_67.QC.jpg
401592a5dabdeb3efbbaaef795605b26
d41a22bceffd4e1bc4684711cc3b31b77195f317
23484 F20101109_AABJFV kammerdiner_a_Page_45.QC.jpg
74fce7f8f734df2b27141965b509ca90
81c7bca450f0b69df57c0b71012796800974839f
22632 F20101109_AABJUO kammerdiner_a_Page_30.QC.jpg
202b84ac47dcfd5ff19b7502aff07bec
c93077ea4c6493024809844bca7e03b39aad8ba8
F20101109_AABJPR kammerdiner_a_Page_46.tif
d20fe978615943d987a09cccb469f436
cb35b03d9599d325275a68301d97f66dd896db14
F20101109_AABJAZ kammerdiner_a_Page_16.tif
44e0da8720c6f3c3d6589a7e6584e377
94fccdf6db3299bf3c0ed4d3a99ef6114b332f48
2159 F20101109_AABJKT kammerdiner_a_Page_73.txt
2149276c8692b5799f3ef25881e7f4d9
b0bae38b0c0194d4e5001db98f58bcea39b1dbc5
15144 F20101109_AABJFW kammerdiner_a_Page_06.QC.jpg
665a6e6bf75c8719d34bc81931c61860
53af5d68b1563f8ddce11cf6f6302c9b107e309e
31177 F20101109_AABJUP kammerdiner_a_Page_31.QC.jpg
5ff0ec4158907f7df29766213ce79586
6b88ec1756bcd1796c56b865a4a6a2ad8b55b8ec
F20101109_AABJPS kammerdiner_a_Page_47.tif
64cf4e24bb5d643bfbd3337478d2e8fe
1ed9feb78b9c7407039dfc615e18a54054054094
8037 F20101109_AABJKU kammerdiner_a_Page_69thm.jpg
36a271ea8c22c9bbf975e4ca2539ced0
ee6e15245a64a1d5e8c4b0984280bbda23de5de7
F20101109_AABJFX kammerdiner_a_Page_50.tif
4609753acdba30275e71246cba2b62cc
9a0e1937fa43a06763033c59dad0103c4eead90b
26652 F20101109_AABJUQ kammerdiner_a_Page_32.QC.jpg
e01d7bb6baf8a8c465c348d62dff8431
19ee3262985e5de793e944b286c8aa39ecf1f377
F20101109_AABJPT kammerdiner_a_Page_48.tif
8ee75d5fd9d015943c41dda204e10f64
65e1def1d58017574d6b837985ff84325fb0b86c
F20101109_AABJKV kammerdiner_a_Page_28.tif
32f3396b778390e7968f32519e2e5eae
315a60a173972b61134ba791d4648f3f7403e180
F20101109_AABJDA kammerdiner_a_Page_53.tif
239e789684229ea69759f7774a8ddeea
204ba581f75177b900d4e0fa246e380c9fd864f8
7858 F20101109_AABJFY kammerdiner_a_Page_12thm.jpg
0be538f9603de9742b1a1b2fda1ad8f5
ff39831d5b39d5012436291e53186629ca1a0abb
24893 F20101109_AABJUR kammerdiner_a_Page_33.QC.jpg
0787f5d31f40f450e8ad1ca36aa72c87
a0041fdd4b8f93a5944b8f93db0321e255bb646e
6922 F20101109_AABJKW kammerdiner_a_Page_56thm.jpg
1450dcf90e82ba96113081e0d5c4fe0d
f16732e2c97a867de3dfdd7a353c077351a5744b
6367 F20101109_AABJDB kammerdiner_a_Page_33thm.jpg
7f73e35e4aacc432ddd725f9ae0ef8a4
01657ebee10ee2d1249b1613b78c9f4c9e4b7f33
F20101109_AABJFZ kammerdiner_a_Page_12.jp2
88b956f0b0b6c22d98fe0fb449c50057
6816543ac9a38cd41782fc176d5119693dab7440
6834 F20101109_AABJUS kammerdiner_a_Page_34thm.jpg
d0bc5fe90e20b9da4e6602cad7b6a672
aba023d5dcc71cecb52e1cd77fb87ebd01cf3eff
F20101109_AABJPU kammerdiner_a_Page_49.tif
9b713b7098ec6c491e0ced51544c64c2
bc235287085d4a9d78869d66ac35b503f750a80b
2757 F20101109_AABJKX kammerdiner_a_Page_77.txt
8b7b2dec2a8304a9052b3c4b94a87183
729048358729f7c3d270922387928e4986823b4b
56483 F20101109_AABJDC kammerdiner_a_Page_69.pro
dba2df7ed1c6d6cc269ca4f0d4e54003
cd3f8b48b2304fe91bbbb327a380f9f8d1dd9c51
30703 F20101109_AABJUT kammerdiner_a_Page_35.QC.jpg
989080c088a2905e801f160b395bf2cf
1fce03449b87e87d0188710ac5f9ef405edc6bad
F20101109_AABJPV kammerdiner_a_Page_54.tif
8db9496d29475dae819b117050f666b7
076349d12eee72c774bddf478a0fe27f7b1c2f9a
35285 F20101109_AABJIA kammerdiner_a_Page_24.QC.jpg
e26f6c5e8b0a40b07d7461a7af0f510d
49093f7457db3e278ac7a9ad76437786f210fc79
3636 F20101109_AABJKY kammerdiner_a_Page_03.jpg
fca511999548654585d9c902bdda70cd
e895ab9e0006c6aa3854392f341d022b2eeac530
76044 F20101109_AABJDD kammerdiner_a_Page_33.jpg
2db0943722be38943c0941bf4226ff06
0ebf9d5003db93fc71811c871d35cdaac9e2d994
7718 F20101109_AABJUU kammerdiner_a_Page_35thm.jpg
1b3073dec4ca31fc0611678d82cc7123
df2de8cd5ddd559c5274512bcb11eb3ee4bceb98
F20101109_AABJPW kammerdiner_a_Page_55.tif
c74b38eef9a3d99122d2976854bbf46d
7ac4a9bf7bca85ae76da747dcd8e0b2384ae7ed5
1051966 F20101109_AABJIB kammerdiner_a_Page_61.jp2
4b4fe711062820d34944fac1a933cf2e
f25d52ad9dab7800bf1a5b3cf17af311b8de548f
F20101109_AABJKZ kammerdiner_a_Page_42.tif
b6fc1d15d6ee9a893866d1849fbc29f1
27e30003cf2c5ceecd5d8db09b46db6e7d09d02b
58356 F20101109_AABJDE kammerdiner_a_Page_54.pro
800fa431c9e9747283e7d70f674d970a
3732bcb2d7ae530ac84dc6fca2f27e2fba6083ab







STATISTICAL METHODS IN DATA MINING OF BRAIN DYNAMICS


By

ALLA R. KAMMERDINER



















A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA

2008


































2008 Alla R. Kammerdiner

































To my wonderful family.









ACKNOWLEDGMENTS

I would like to show my heartfelt appreciation to my advisor Dr. Panos M.

Pardalos for his support and mentoring. Working with him has helped me grow not

only professionally, but also as a person. I am also very grateful to other members

who served on my supervisory committee, J. Cole Smith, William W. Hager, Vladimir

L. Boginski, and H. Edwin Romeijn, for their valuable comments on my research for

this dissertation. Last but not least, I thank my husband Jason, my parents Olga and

Aleksandr, my brother Mikhail, and all the rest of my great family for their unconditional

love and support.









TABLE OF CONTENTS


page


ACKNOW LEDGMENTS .................................

LIST O F TABLES . . . . . . . . . .

LIST OF FIGURES . . . . . . . . .

LIST OF SYMBOLS ................... ................

A B ST R A C T . . . . . . . . . .

CHAPTER


1 INTRODUCTION ..................................

1.1 Statistical Methods for Data Mining .....................
1.2 Electroencephalographic Recordings ......................
1.3 Feature extraction . . . . . . . .
1.4 Contribution Summ ary .............................
1.4.1 Testing Applicability of Frequency Domain Estimates of Granger
Causality for EEG time series .. .................
1.4.2 Generalization of Phase Synchronization via Cointegrated VAR .

2 AUTOREGRESSIVE MODELING OF MULTIPLE TIME SERIES .......

2.1 Multivariate Autoregressive Modeling in EEG Data Mining .........
2.2 Tests of Granger Causality ........................
2.3 Vector Autoregressive Models (VAR) .. .................
2.3.1 Methods for VAR Parameter Estimation .. ............
2.3.2 VAR Order Selection Criteria .. .................
2.3.3 Stability Condition and Other Assumptions of VAR .........
2.4 Inegrated and Cointegrated VAR .. ...................
2.4.1 Augmented Dickey-Fuller Test for Testing the Null Hypothesis of
the Presence of a Unit Root .. ..................
2.4.2 Phillips-Ouliaris Cointegration Test ................
2.4.3 Estimation of Cointegrated VAR(p) Processes .. .........
2.4.4 Testing for the Rank of Cointegration .................

3 PHASE SYNCHRONY IN BRAIN DYNAMICS ..................


The Role of Phase Synchronization in Neural Dynamics .......
Phase Estimation using Hilbert Transform .. ...........
Phase Estimation via Wavelet Transform .. ............
Comparison between Two Approaches to Phase Extraction ......
Measures of Phase Synchrony .. .................









4 APPLICATION OF VECTOR AUTOREGRESSION TO MINING BRAIN DY-
N A M ICS . . . . . . . . . 60

4.1 Numerical Issues in Estimating Parameters of Vector Autoregression from
EEG ................... .. .................. 60
4.2 Multivariate Approach to Phase Synchrony via Cointegrated VAR . 62
4.2.1 Cointegration Rank as a Measure of Synchronization among Different
EEG C('!i ,. . ... .... ...... ..... .. .. 64
4.2.2 Absence Seizures ..... . . .... .. ....... 67
4.2.3 Numerical Study of Synchrony in Multichannel EEG Recordings
from Patients with Absence Epilepsy ................ .. 68

5 CONCLUSION .................. ................. 73

REFERENCES .............................. .. ..... 75

BIOGRAPHICAL SKETCH .................. ............ 83









LIST OF TABLES


Table page

2-1 ADF test: Critical values of the T(c- 1) and i statistics . ..... 49

2-2 Phillips-Ouliaris demeaned: Critical values of the Z, statistic . ... 49

2-3 Johansen test: Critical values of the ALR(r, K) statistic . . ..... 49

2-4 Johansen test: Critical values of the ALR(r, r + 1) statistic . .... 50

4-1 Seizure 1: Results of the ADF unit root tests .................. 70

4-2 Seizure 2: Results of the ADF unit root tests ......... . ....... 70

4-3 Seizure 3: Results of the ADF unit root tests .................. 70

4-4 Seizure 1: Results of the Johansen cointegration rank tests . .... 70

4-5 Seizure 2: Results of the Johansen cointegration rank tests . .... 70

4-6 Seizure 3: Results of the Johansen cointegration rank tests . .... 70









LIST OF FIGURES


Figure page

1-1 The International 10-20 system for placement of EEG electrodes . ... 27

4-1 Raw data: Numbers of unstable roots for different T and p . ..... 71

4-2 Filtered data: Numbers of unstable roots for different T and p . ... 71

4-3 EEG segment with absence seizure .................. ..... .. 72









LIST OF SYMBOLS, NOMENCLATURE, OR ABBREVIATIONS

Most of the notation is unambiguously defined in the text where it is introduced. To

provide some general guidelines, we include the following list of commonly used symbols.
ARIMA(p, d, q) autoregressive integrated moving average process

AR(p) autoregressive process of order p

CPV C 1 r!v principal value

X complex conjugate of X

converges to

Sconverges in distribution to

Sconvolution of functions

det determinant

e element of

equals

: equals by definition

Equivalent to

x estimator of x

3 exists

E expectation

exp exponential function

V for all

I identity matrix

IK (K x K) identity matrix

S imaginary part of a complex value

oo infinity

I(d) integrated process of order d

S is distributed as

lim limit










in natural logarithm

Pr probability of a random event

rank rank of a matrix

R real part of a complex value

Y sum of terms
o superposition

trace trace

X' transpose of X

Var variance

VAR(p) vector autoregressive process of order p









Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

STATISTICAL METHODS IN DATA MINING OF BRAIN DYNAMICS

By

Alla R. Kammerdiner

May 2008

C'!I i': Panos M. Pardalos
Major: Industrial and Systems Engineering

This study discusses statistical approaches used for data mining of multichannel

electroencephalogram recordings. Such recordings represent massive data sets that contain

hidden patterns of complex dynamical processes in the brain. Formally, multichannel EEG

can be viewed as a multiple time series, and therefore, a natural idea for summarizing such

data is to utilize autoregressive modeling of multivariate stochastic processes.

In particular, we thoroughly discuss various concepts and approaches related

vector autoregressive processes, including stable stationary VAR models of order p and

nonstationary systems with integrated and cointegrated variables, as well as procedures for

estimating parameters of the systems (e.g. order, lag, or cointegration rank).

The work highlights some stability issues that may arise in the application of vector

autoregression to mining EEG data, and questions the applicability of Granger causality in

the frequency domain to multichannel EEG.

Synchronization has been found to be an important characteristic of the abnormal

brain dynamics manifested by epilepsy and Parkinson's disease. We review two approaches

for extracting the instantaneous phase from time series. In this study, we generalize the

concept of the phase synchronization, and propose a novel approach based on multivariate

analysis via modeling cointegrated VAR(p) processes.









CHAPTER 1
INTRODUCTION

1.1 Statistical Methods for Data Mining

With significant improvements in obtaining, processing and storing the information

electronically in the last decades, it became routine to accumulate large amounts of data

in various fields of research and business. However, any information is only useful if it

can be analyzed to draw some meaningful conclusions. A drastic increase in information

loads makes the task of interpreting collected data especially challenging. Not only

visual inspection and analysis of such massive data become extremely time-consuming

and often ineffective, but also the limitations of the traditional numerical data analysis

techniques in application to the massive collections of data necessitate the development

of new approaches. As a result of continuous attempts of scientific community to extract

useful information from large data sets, a multi-discipline field of data mining has been

developed.

Data mining together with data preprocessing constitute the central part of a more

general process of knowledge discovery in databases (KDD). The KKD process can

be described as a sequence of actions, which selects the raw data in data warehouses

and transforms the selected data in order to discover valid, understandable, novel and

potentially useful knowledge from the data. Data preprocessing that is applied to raw

data to improve the quality of the data often influences the selection and facilitates

the application of data mining techniques. Proper preprocessing of raw data leads to a

decrease in the time needed to mine the data, and boosts the overall mining efficiency.

The techniques used for data preprocessing can be roughly subdivided into data

(1. ,iii,: data integration and data reduction [64]. Data cleaning techniques handle the

problem of incomplete, inconsistent and erroneous data, remove the noise inherently

present in the raw data, minimize redundancy in the data, etc. Data integration is

concerned with combining heterogeneous data collected from different sources to form a









consistent data set. The process of data reduction amounts to identifying useful features,

which are capable of adequately representing the data, and it is usually performed using

dimensionality reduction and feature extraction methods.

The two fundamental tasks assigned to data mining are a descriptive task of

discovering hidden patterns and relationships in given data, and a predictive task

of forecasting or classifying the model's behavior from available data. Data mining

includes regression, classification, (1 l-1. i-i.- image restoration, learning association

rules and extracting functional dependencies, data summarization, etc. Data mining

is closely connected to other research areas such as statistics, machine learning and

artificial intelligence, optimization, visualization and databases. Data mining utilizes many

important results from the related fields, while keeping the main focus on the algorithms

and architectures, scalability of the number of features and instances, and automated

managing of massive quantities of diverse data.

t ii,: areas of data mining employ various approaches developed in the field of

optimization. In particular, it is shown in [7] that many basic problems in data mining,

including classification and (< -1. ii11 .- can be formulated as mathematical programming

problems and solved using optimization techniques.

In fact, Bradley at el. [7] demonstrated that a problem of minimizing the number

of misclassified points in two-class classification can be viewed as a linear program

with equilibrium constraints (LPEC). LPEC is a linear program (LP) with a single

complementarity constraint. Such constraint imposes a condition of orthogonality between

two linear functions. LPEC formulation arises in the instances when the constraints of the

problem include another LP problem.

In addition, the problem of feature selection in two-class classification by finding a

separating plane that utilizes minimum number of features can be given a mathematical

programming formulation as a parametric problem. Furthermore, the classification

via support vector machines (SVMs that find the separating plane maximizing the









margin between two classes while minimizing misclassification errors) can be stated

as a quadratic programming problem [6]. Moreover, as indicated in [7], the above

mathematical programming formulations have been extended to be effectively emploi,- .1

by other data mining approaches, including neural networks training, calculation of

nonlinear discriminants, and building decision trees. The clustering problem has a

complex formulation as a minimization problem with the objective given by a sum of

the minimums of a set of convex functions [5]. In general, this objective function is

neither convex nor concave. See review by Bradley at el. [7] for additional information

about mathematical programming formulations for various problems in data mining, the

application of optimization techniques, as well as the challenges that the field of data

mining offers to optimization.

i in!;: data mining approaches, such as classification,
segmentation, have been applied to time series a", -ii-- [45]. Many traditional statistical

approaches are also applied to mining time series. For instance, regression is one of the

most commonly used techniques for modeling and forecasting time series. Among the

statistical models applied to regression in time series are linear autoregression (AR),

autoregressive moving average process (ARMA), autoregressive integrated moving average

process (ARIMA), as well as their multivariate analogs (i.e. vector autoregression, etc.)

Time series arise in various applied areas, including economics and finance, meteorology,

biomedicine, etc. For instance, the study of seismic activity related to earthquakes

produces two-dimensional time series, where each measurement consists of the time

and the magnitude of a registered seismic event. Many biomedical signals, such as

electrocardiogram (ECG), electroencephalogram (EEG) and electrooculogram (EOG)

represent time series that can be interpreted via application of regression, segmentation,

neural networks, and other data mining methodologies. Sound signals are another example

of time series that are effectively analyzed using different data mining techniques.









Time series, which originate in different fields, are generated by diverse underlying

processes, and as a result they are often characterized by very distinct properties. Indeed,

as indicated in [14], although the normal time scale is a very natural choice of parameter

for the time series describing physical processes, the regular time looses its natural

meaning when dealing with many financial time series. Because most of financial time

series are irregularly spaced in physical time, the concept of "business time" or ,ll ,ii-,

time" is introduced to represent a new time parameter with respect to which time series

are regularly scaled. This procedure of time deformation allows the relabeled time series to

be viewed as stationary on a new time scale. Financial time series also often exhibit clear

seasonal trends, which obviously cannot be found when examining time series produced by

speech.

Statistical testing of several multivariate time series determined that the time series

from AUSLAN and BCI data sets can be considered stationary, whereas BCI MPI and

EEG contained non-stationary time series [108]. As a result of inherent differences in time

series data from diverse sources, some data mining methods that are successfully applied

to time series in one research area may not necessarily be applicable to analysis of time

series that stem from another applied field.

Time series obtained from electroencephalogram (EEG) recordings have several

interesting properties that distinguish them from other time series. Although some studies

apply one-dimensional modeling by considering one channel in EEG recording at a time,

in general, EEG data should be treated as multivariate time series. The multivariate

approach becomes especially important in view of its ability to investigate spatio-temporal

dependencies in the EEG data in contrast to being limited to only temporal relations in

a one-dimensional case. There is a disagreement among researchers studying EEG data

on whether the series should be modeled by a non-linear stochastic process or they can be

better described by a deterministic chaotic dynamical system.









Storage of continuous EEG recordings sampled from multiple channels at high

frequency during several hours (or even di -) from multiple subjects, wether it is for a

sleep study or diagnosis of neurological disorder, may take gigabytes of memory. The

application of data mining approaches to EEG time series allows automatic handling and

analysis of such large data sets. With new technological advances in collecting EEG data,

it becomes particularly important to develop new efficient data mining methods designed

specifically for mining EEG data.

1.2 Electroencephalographic Recordings

Electroencephalography is one of the most commonly applied methods of extracting

neurophysiological signals. It originated in 1875, when an English physician Richard Caton

measured the electrical activity from the exposed brains of monkeys and rabbits [100].

Generally -p1' i1:;i EEG represents a digital or a graphic record of the electrical

activity in the brain, and can be measured by either non-invasive or invasive methods.

EEG (obtained during a non-invasive procedure) is defined as a record of electrical activity

of an alternating type measured from the scalp surface after being picked up by metal

electrodes and conductive media [67]. There are two types of EEG produced by invasive

procedures, the electrocorticogram, which measures the brain's electrical activity directly

from the cortical surface, and the electrogram, which is an EEG obtained using deep

probes.

EEG estimates and records the relative change in electric potentials produced by

a large number of electric dipoles during a period of neural excitations. The activation

of neurons (brain cells) generates local current flows in the brain. EEG records mostly

the electrical currents that flow during synaptic excitations of the dendrites of numerous

pyramidal neurons in the cerebral cortex. EEG recorded from the scalp surface can only

detect the electrical activity produced by massive populations of active neurons. On the

other hand, EEG recorded using deep probe electrodes implanted into the brain can pick

up a signal from a small group of neurons, which can be further filtered out to obtain the









electric potentials generated by individual neurons. EEG has become an effective device

in the area of neurological research as well as clinical neurology, because of its capacity

to reveal both abnormal and normal brain activity. It is believed that by birth, human

brain has already developed the full number of neural cells, which is approximately 1011

neurons [68]. This gives an average density of about 104 neural cells per cubic millimeter

of the brain. Neural cells are interconnected through synaptic connections in the brain

into neural nets. The brain of an average adult contains approximately 500 trillion

synapses. The total number of neurons decreases with age. As a result the total number

of synaptic connections declines with .I-,: even though the number of synapses per one

neuron increases with age.

To ensure the consistency in referencing locations of electrodes in EEG experiments,

the International 10-20 system for EEG electrode placement was developed [37]. The 10-20

EEG system is used to describe the respective locations of scalp electrodes during EEG

recording in relation to the underlying area of cerebral cortex.

According to the 10-20 system, anatomical landmarks of a skull, such nasion, inion

and preauricular points, are identified for consecutive placement of the electrodes at fixed

distances from these points in steps of either 10 or 20 percent. This approach is devised

to take into account possible variations of head size. In addition, the method is easily

applicable in practical use. As a result, the 10-20 EEG system became very widely used

for positioning electrodes.

In the 10-20 system, the points are denoted with one or two letters, and can be also

followed by a number (as shown on Figure 1-1). The letters roughly represent the lobe

location (with exception of letters C and Z), whereas the numbers serve for identifying

the corresponding hemisphere. More specifically, the points located on the left hemisphere

of the brain are represented by odd numbers (1, 3, 5, and 7), and the sites on the right

hemisphere are marked with even numbers (2, 4, 6, and 8). The sites located on the

frontal, temporal, parietal and occipital lobes are denoted by the corresponding initials









F, T, P, and 0. The letters C and Z refer to the points placed in the central area. In

particular, Z represents a point on the midline, and C refers to the line parallel to the

midline. Note that the central area is not a lobe.

EEG signal resembles a collection of sinusoids of various amplitude and frequency.

Power spectrum is extracted from the raw EEG data using Fourier transform to obtain the

information about the contribution of sinusoidal waves of different frequency. The power

spectrum of EEG is continuous, ranging from 0 Hz up to a half of the sampling frequency.

Depending on the state of the brain, certain frequencies appear to be more prevalent.

There are four in i r frequency bands, alpha, beta, delta and theta, which presence in

EEG during various states of consciousness has been extensively studied. These bands

represent sine waves of relatively low frequency, with delta ranging from 0.5 to 4 Hz, theta

4 8 Hz, alpha 8 13Hz, and beta over 13 Hz. Alpha waves discovered by Adrian and

Matthews in 1934 are the best-known and the most studied among the four frequency

bands [100]. They are induced by closing eyes and by relaxation, and terminated with eyes

opening or due to thinking, calculating, and other analytical activities. In particular, in

most people, eye closing produces rapid changes in brain activity manifesting themselves

in EEG as an adjustment of the dominant frequency band from beta to alpha. EEG is

capable of discriminating between different states, such as i1 ii.- alertness, stress state,

various sleep stages, hypnosis, etc. Presence of beta band is dominant during the state

of alertness with eyes open. Drowsiness or the resting are usually characterized by the

rise in alpha activity. During the sleep, presence of lower frequency waves becomes more

apparent. A higher proportion of delta band frequencies is observed during stages III and

IV of the non-rapid eye movement sleep (NREM). EEG recorded from distinct regions in

the brain exhibits different spectrum of wave frequencies. In addition, the brain patterns

are unique for every individual.

Practical applications of EEG include epilepsy research and localization of the focus

of epileptic seizures, testing of epilepsy drug effects; determining areas of damage due









to stroke, head injury, etc.; monitoring alertness, coma and brain death; testing afferent

pathr- --, by evoked potentials; research in sleep physiology and sleep disorder; controlling

anaesthesia depth, etc. [3].

1.3 Feature extraction

Formally, EEG signal can be described as a deterministic multidimensional nonlinear

non-stationary time series [94]. In order to properly reflect the spatio-temporal properties

of brain dynamics, the i, i1 ,-i of EEG data must involve a simultaneous investigation of

the dependencies across channels with respect to time.

Different features have been proposed for analysis of EEG time series, including

Fourier transform, wavelets, cross-correlation, coherence, Granger causality and partial

directed coherence, mutual information and transfer entropy, global and phase synchronization,

Lyapunov exponents and correlation dimension, etc.

Since EEG can be viewed as a collection of sine waves, EEG series are often Jin 1v. 1

in a frequency domain. In addition, some frequency bands have shown to p1 iv specific

roles in various states of consciousness, and so the frequency information in EEG can be

particularly important. Subsequently, the Fourier transform with a running time window,

also known as short time Fourier transform (STFT), became one of the most widely used

methods for extracting features from EEG. STFT is obtained from Fourier transform by

applying a time window function g with a time shift r. Mathematically, STFT is given by

the following formula:


S(r, f) x(t)g(t -r) exp{-217r f t}dt, (1-1)
-oo

where S(r, f) denotes the STFT with time window g located at time 7, corresponding to

frequency f; and x(t) is a signal at time t. In other words, STFT S(r, f) represents the

power spectrum of the signal estimated around time 7. The drawback of STFT is that

there is a trade off between time accuracy and frequency precision. By making the window









g smaller, the resolution of the time parameter is improved, unfortunately, at the expense

of the resolution in frequency.

An alternative to Fourier transform is wavelet transform (WT), which is a transformation

of the signal based on a special function, called mother wavelet (M\W). The mother

wavelet is shifted in time by a location parameter 7, and then adjusted by a scale

parameter a. More precisely, the wavelet transform is defined by the following formula:

1 +0 t T
W(r,a) = x(t)( )dt, (1-2)


where ) is a mother wavelet, a is scale parameter, is a time location parameter, and x is

a signal.

The scale parameter a in WT is analogous to the frequency parameter f in STFT.

In particular, the large values of parameter a (a > 1) stretch the wavelet, and so they

represent low frequencies, whereas the small values of a (a < 1) shrink the wavelet

function, which corresponds to higher frequencies. An advantage of using wavelets is that

the high frequency components can be analyzed with a higher time accuracy than the

lower frequency components of the signal.

As follows from (1-2), W(r, a) can be interpreted as the projection of the signal onto

the appropriately shifted and scaled wavelet W, i.e. W(r, a) is a contribution of the wavelet

to the signal x(t).

While Fourier and wavelet transforms are usually applied to study each channel of

EEG signal individually, the cross-correlation, coherence and Granger causality measure

the interdependency between different channels. The cross-correlation function quantifies

the linear correlation between two processes. Given two normalized signals x(t) and

y(t) with zero means and unit variances, the cross-correlation between these signals is

estimated as:
N-T
C (r) xr+tt)y(t), (1-3)
t=i









where N denotes the total number of sampled points, and r is a time delay parameter

between two signals. The cross-correlation estimate ranges between -1 and 1. The positive

values of cross-correlation indicate the direct correlation between the signals (i.e. x and y

tend to be similar in both their absolute value and have the same sign), and the negative

values correspond to inversely correlated signals (i.e. signals have similar absolute values,

but different signs). Although zero cross-correlation value shows that the signals are not

linearly correlated, it does not necessarily imply that two signals are not interrelated in a

non-linear fashion.

The coherence function is a frequency domain analog of the cross-correlation measure.

Coherence is obtained from cross-correlation by applying Fourier transform to (1-3). The

estimate of the coherence spectrum of two signals is called periodogram. The periodogram

is calculated by subdividing the signals into a number of epochs of the same length, and

then applying the following formula:

| 2S1 (f) 12
P)2 (f) = (1-4)

where Sx(f), Sxx(f), Syy(f) denote the average values of the cross-spectral density

between x and y, and the individual auto-spectral densities of x and y, respectively. The

periodogram values range between 0 and 1, where zero coherence indicates that the signals

are linearly independent at a given frequency, and the maximum periodogram value of 1

shows that two signals are completely linearly dependent at a chosen frequency.

While the cross-correlation and coherence are features, which reflect the linear

dependency between two channels in the data, the concept of Granger causality is capable

of not only establishing the linear dependency, but also specifying the direction of such

dependency. In other words, by applying Granger causality, it becomes possible to

identify causal relationship among the channels of EEG. Granger causality is based on

the multivariate autoregressive modeling of time series. It has also received an alternative

reformulation in the frequency domain via spectral decomposition for stochastic processes.









Several studies have shown that EEG can be effectively modeled using chaos

theory [35, 89]. In chaotic systems, trajectories originating from very close initial

conditions diverge exponentially. The system dynamics are characterized by the rate

of the divergence of the trajectories, which is measured by Lyapunov exponents and

dynamical phase.

Short term largest Lyapunov exponent (denoted STLmax), which is an estimate of

the maximum Lyapunov exponent for non-stationary data, is a dynamical measure of the

chaoticity in the brain. Next, the method for estimating STLmax is summarized.

First, using the method of d. 1 ,i' [69], the embedding phase space is constructed from

a data segment x(t) with t E [0, T], so that the vector Xi of the phase space is given by

Xi (x(t),x(t + 7),..., x(ti + (p- 1))), (1-5)

where ti E [1, T (p 1)r], p is a chosen dimension of the embedding phase space, and T

denotes the time d. 1 iv between the components of each phase space vector.

Next, the estimate L of the short term largest Lyapunov exponent STLmax is

computed as follows:

1 -2 X(t + At) X(t + At) (
L 0 log2 (- ( -' (1-6)
i-N X(ti) X(ti)
where Na is the total number of local maximum Lyapunov exponents that are estimated

during the time interval [0, T]; At is the evolution time for the displacement vector

X(ti) X(tj); X(ti) represents the point of the fiducial trajectory such that t = ti,

X(to) = (x(to),x(to + -),... ,x(to + (p 1)r)), and X(tj) is an appropriately selected

vector that is .,.i i,:ent to X(ti) in the embedding phase space. In [34], lasemedis at

el. si-- -1- .1 a method for estimating STLmax in the EEG data based on the Wolf's

algorithm for time series [107].

The short term largest Lyapunov exponent STLmax is proved to be an especially

useful EEG feature for studying the dynamics of the epileptic brain [35, 89]. In particular,









spatio-temporal transitions during interictal, preictal, ictal, and postictal states can be

characterized by the changes in STLmax profiles [90].

1.4 Contribution Summary

1.4.1 Testing Applicability of Frequency Domain Estimates of Granger
Causality for EEG time series

The rules of interaction between various parts of the brain are one of the key

problems in theoretical and applied neuroscience. One of the most commonly used

v-,- to record the neural information in multiple areas of the brain is via a multichannel

electroencephalogram recording. Quantitatively, such recording is a multivariate time

series. In the last decade, the investigation of causal relationships exhibited by the

electroencephalographic time series became a very active area of research in neuroscience

and related fields. Several approaches for studying causality in EEG are proposed, most

of which are based on the definition of Granger causality via the spectral representation of

time series (or processes), which was introduced by Geweke [25] for analysis of econometric

series. Because frequency domain contains valuable information about the brain processes,

Geweke's definition of causality seems to be particularly useful for analysis of EEG data.

Both definitions of Granger causality, the original one given by Granger and the

frequency domain definition, are introduced via vector autoregressive modeling of multiple

time series. Precisely, the linear vector autoregression is used to fit the data. Based on this

model, two competing hypothesis about the data are considered and tested statistically to

determine, which of these two assumptions is supported by the data. In other words, to

test the causality, the hypothesis of data being modeled as linear autoregression (that does

not include another series) is compared to the alternative of the data best described by

including information from the other series.

Since Granger causality is defined based on linear vector autoregressive model (VAR),

the applicability of Granger causality depends upon the underlying assumptions of the

model. The fundamental assumptions of the VAR are stability, stationarity, and gaussian









distribution of the error term. If such conditions are violated, the model is not suitable for

the data. In this case, one needs to find a model that is more appropriate for the given

data. For example, many financial time series data exhibit periodic or seasonal behavior.

To account for such cases, various modifications of vector autoregression (including those

relaxing the stability condition) are constructed for analysis of econometric time series.

Although Bernasconi and Konig [2] tested the stationarity of different EEG data, and

concluded that a time interval of 1 second is the interval on which the EEG time series

can be considered stationary, the underlying assumptions of the VAR modeling such as

stationarity and stability are rarely statistically tested in applications to EEG data.

In particular, the stability assumption means that the reverse characteristic

polynomial of the model does not have roots inside a unit circle. In order to highlight

the importance of the stability condition for VAR, it is necessary to point out that when

stability is violated, the model may simply follow a random walk, or it may even exhibit

explosive behavior.

To the best of our knowledge, the stability condition of the VAR estimated from the

EEG time series was not investigated before our study [43]. In many studies that utilize

the vector autoregression to examine Granger causality among series, the verification of

the conditions assumed by the VAR model is often omitted.

The results of our numerical experiments indicate that the stability condition of

vector autoregressive model is often violated in application to the EEG data. More

specifically, we found that the stability assumption imposed on the linear VAR models

may be violated even in the case when the sample size parameter T is much larger than

the lag parameter p of the estimated model. In addition, we showed that despite the

fact that it is common in practice to filter the EEG data within a certain frequency

band, filtering the EEG time series within some restricted frequency band often results

in significant reduction of the (T,p) domain, where the estimated VAR(p) models remain

stable.









Based on our numerical studies, we concluded that suitable extensions of multivariate

autoregression to unstable processes may fit the data better, and so, be more appropriate

for the EEG time series analysis than linear vector autoregressive modeling [43].

Comprehensive statistical testing is necessary in order to make conclusions on what

multivariate models are the most appropriate for extracting the directional dependencies

between channels in a frequency domain from multichannel EEG data.

1.4.2 Generalization of Phase Synchronization via Cointegrated VAR

The temporal integration of various functional areas in different parts of the brain is

believed to be essential for normal cognitive processes. This results in constant interaction

among the brain regions. Many studies highlight the importance of neural synchrony in

such large-scale integration [15, 101, 104, 105]. Actually, it was found that oscillation

of various neuronal groups in given frequency bands leads to temporary phase-locking

between such groups of neurons. This observation has stimulated the development of

robust approaches that allow one to measure the phase-synchrony in a given frequency

band from experimentally recorded biomedical signals such as EEG.

In particular, the importance of synchronization of neuronal discharges has been

shown by a variety of animal studies using microelectrode recordings of brain activity [83,

96], and even at coarser levels of resolution by other studies in animals and humans [21].

The phase synchronization in the brain extracted from EEG data using Hilbert or

wavelet transforms has recently been shown to be an especially promising tool in ain 1, -;

of EEG data recorded from patients with various types of epilepsy [86].

In our recent study [44], we introduce a novel concept of generalized phase synchronization,

which is based on vector autoregressive modeling. This new notion of phase synchronization

is constructed as an extension of the classical definition of phase synchronization between

two systems. In fact, the phase synchronization is usually defined as the condition that

some integer combination of the instantaneous phases of two signals is constant. Often this

condition is relaxed by allowing for a bounded linear combination of two phases, in order









to account for noise in the measurements. This classical approach is clearly bivariate. But

what if we are interested in studying a synchrony among several parts of a system? Is

there such a notion?

To construct a more general multivariate concept of phase synchronization, we

extended the classical definition by considering a linear combination of phases for a

finite number of signals that represents a stationary process. All the individual signals

together form a common system described by some multivariate process. We note that a

vector process, such that a linear combination of its individual components is a stationary

process, can be modeled as a cointegrated vector autoregressive time series.

Furthermore, it is easy to see (as shown in Section 4.2.1) that the cointegrated rank of

the regression determines how restricted the behavior of such system is. This means that

the rank r of cointegrated autoregressive model, estimated from the multiple time series

of the instantaneous phases, measures how large the vector subspace, which generates the

changes in the phase values, is.

This new measure of cointegration was applied to absence epilepsy EEG data in [44].

The data sets collected from the patients with other types of epilepsy are currently being

investigated.































A top view


B profile view

Figure 1-1. The International 10-20 system for placement of EEG electrodes









CHAPTER 2
AUTOREGRESSIVE MODELING OF MULTIPLE TIME SERIES

2.1 Multivariate Autoregressive Modeling in EEG Data Mining

Several methods for joint spatio-temporal analysis of multichannel EEG recordings

based on the idea of Granger causality were presented in the last decade. The concept

of Granger causality was first introduced by Clive Granger [26, 27] for measuring

linear dependence and feedback in economic time series. Later, this idea was further

extended by John Geweke [25], who proposed an equivalent measure based on the spectral

representation of time series. Both Granger's and Geweke's approaches employ the vector

autoregressive modeling to derive estimates of underlying causal relations in the data.

However, the latter approach is found particularly useful for analysis of EEG time series,

since it investigates the causal relation in the frequency domain instead of the time

domain as in the former approach. In particular, the spectral measure of Granger causality

proposed by John Geweke was employ, ,1 on intracortical local field potentials recorded

from 8 electrodes during go/no-go trials of cat's visual responses [2]. Another study [57]

utilized a similar method of directed transfer function (which is equivalent to the spectral

measure of Granger causality) to examine causal influences in the primate visual cortex

during the task of visual pattern recognition. The direct transfer function approach to

Granger causality was also applied to analyzing brain connectivity patterns on human

EEG data recorded during stage 2 sleep [42].

Michael Eichler proposed a graphical approach for modeling Granger-causal

relationships in multivariate time series [17] and later applied this method to studying

connectivity in neural systems [18, 19]. Luiz Baccala and Koichi Sameshima introduced

a concept of the partial directed coherence for inference of Granger causality in the

frequency domain based on the linear vector autoregressive modeling, and applied it to

investigating the functional interactions among different brain structures [1, 92].









Vector autoregressive modeling is a common basis in various approaches to estimating

Granger causality. Next, we review the details of the multivariate autoregressive modeling

and investigate some limitations in its application to the EEG data.

2.2 Tests of Granger Causality

w1 i: different tests of Granger Causality are developed. Some versions of the test are

based on a vector autoregressive model, others are based on a multivariate moving average

representation. For simplicity, we present the alternative definition based on the bivariate

regression. For a detailed review of Granger causality tests, see the book on time series by

Hamilton [30].

Let X(t) and Y(t), t E Z denote two time series (or discrete time stochastic processes)

with the corresponding realizations xt and yt, t E Z. Suppose that Qx,t and Qy,t denote

all the information about the realizations of processes X and Y, respectively, up to time t.

Then, the relationship of Granger causality between such series can be formally defined as

follows:

A time series X(t) is said to Gri,,n,. -cause Y(t) if there exists p = 1, 2,... such that

the mean squared error (lMSl) of the p-step forecast of Y(t) based on the information

Qx,t and QY,t is smaller than the MSE of the p-step forecast of Y(t) based on Qyt alone,
i.e.

p = 1,2,... : Ey(plx| t, y, t) < Ey(plQY,t), (2-1)

where Ey(p|fQ) is the MSE of the p-step forecast of Y(t) based on information f.

Using the above definition, we now present the test for Granger causality based on the

bivariate autoregressive model. Suppose that for some integer lag parameter p > 0, the

realizations of time series Y(t) are given by the model


yt = v + aiy-i + Oixt-i + Et, (2-2)

where Et is a standard white noise (or innovation process, i.e. Et has zero mean and zero

autocorrelation).









We test the hypothesis


Ho: i 0 Vi ,2,...,p (2-3)


against the alternative

H1 : 3j, 1 < j < p s.t. 0j 0 (2-4)

Note that if the null hypothesis is accepted, then a time series Y(t) is believed to be NOT

Granger-caused by X(t). Meanwhile, rejecting the null hypothesis (i.e. accepting the

alternative) means that X(t) is believed to cause Y(t) in Granger's sense.

Let T be the sample size parameter. The model parameters for the null hypothesis Ho

and the parameters for the alternative HI are estimated from the sample data using the

ordinary least squares method (or other methods) to obtain the estimates of the forecast

errors ot and Eit, respectively, t = 1, 2,..., T. Then the sum of squared residuals RSSo

under the assumption of null hypothesis Ho is

T
RSSo0 = t, (2-5)
t=1

and the sum of squared residuals RSSI under the alternative H1 is

T
RSS, = t. (2-6)
t=1

By conducting the F-test of the null hypothesis, one can find the test statistic

(RSSo RSSi)/p
1 RSS ~ Fp, T-2p-1. (2-7)
RSSl/(T 2p 1)

If the test statistic Si exceeds the specified critical value, then the null hypothesis that

X(t) does not Granger-causes Y(t) is rejected. Otherwise, Ho is accepted.

An .,-vmptotically equivalent test of Granger causality is given by the following

statistic
T(RSSo RSS)
2 (RSSp) (28)
RSS1









It is noteworthy to point out that the tests of Granger causality are very sensitive to

the choice of the lag length parameter p, and to the methods utilized for handling any

non-stationarity in the time series.

The bivariate approach to testing Granger causality can be naturally extended to

the multivariate case by partitioning the vector autoregressive process Z(t) into two

components X(t) and Y(t), so that Z(t) = (X(t),Y(t)), and then testing the suitable

zero constraints on the coefficients of vector autoregression. For the derivation of the Wald

statistic and the F-statistic for testing Granger causality in the multivariate case, see the

book on multiple time series by Liitkepohl [59].

2.3 Vector Autoregressive Models (VAR)

2.3.1 Methods for VAR Parameter Estimation

The vector autoregressive (VAR) model of finite order serves as a foundation for

establishing Granger causal relations in multidimensional time series.

Let p denote a positive integer, and let yt denote the K-variate time series (i.e.

realizations of K-dimensional process Y(t)). A vector autoregressive model of order p,

denoted VAR(p), is formally defined as follows:


yt =v+Ayt+...+A tp t, t 1, ... yt.., (2-9)

where yt = (lt, ... Kt)' is a (K x 1) random vector, v (vl,..., VK)' is a fixed

(K x 1) vector representing a non-zero mean EY(t), the Aj, i = ,... ,p are fixed

(K x K)-dimensional coefficient matrices, and Et = ( Et,..., Kt)' is a K-dimensional white

noise process (i.e. E [Et] = 0, E [E:'] = 0, for s / t, and E [FsE] = EY). It is assumed

that the covariance matrix Es is nonsingular. In addition, three important conditions are

usually imposed on the time series in the VAR model. The first condition is stability of

the process Y(t), the second is stationarity of Y(t), while the third one supposes that the

underlying white noise process Et is Gaussian.









Suppose that the the lag length parameter p is specified. Although, in the above

definition of a VAR(p) model, the process mean v, the coefficient matrices Ai, and the

covariance matrix Es are assumed to be known, in practice, these parameters must

be derived from the sample data. There are three main approaches to estimating the

parameters of a VAR(p) time series, namely, the multivariate least squares method,

the Yule-Walker estimator, and the maximum likelihood estimation [59]. Under the

assumptions of stability and Gaussian distribution, these approaches lead to estimators

with the same .,-vmptotic properties. However, the .,-vmptotic results should be used

cautiously in inference from small samples. As a result, different approaches may

sometimes lead to different results when estimating the model parameters using small

samples.

Let us now briefly present the multivariate least squares estimation, which is a higher

dimensional extension of the well-known method of ordinary least squares. For more

detailed discussion, refer to [59].

Suppose that the available data include (T + p) successive realizations of estimated

multiple time series represented by K-dimensional vectors


Y-p+ ... ,o, Y 1, l .. YT

where p is the fixed lag length, and T is the sample size parameter. For convenience, we

partition the data into the pre-sample y-p+,... yo and the sample yl,..., yr values. In

addition, the following notation is introduced:


Y:= (y,,...,yr) (K x T),

1

Z:= yt ((Kp+ 1) x 1), (2-10)


Ut-p+1
Z:= (Zo,...,ZT-) ((Kp + ) x T),









B:= (v,A,,...,Ap) (K x (Kp+ 1)),

E:= (1E,..., T) (K xT),


where t = 1,..., T. Then using this notation in (2-9), the vector autoregressive model of

order p can be represented in the compact form:


Y = BZ + E, (2-11)


and the coefficients B of the model are given by the least squares estimator:


B = YZ'(ZZ')-\. (2-12)


The covariance matrix can be estimated in various v--iv. Since ZE = E [EtE'], the

estimator
T
E f E ti't = (Y BZ)(Y BZ)' (2-13)
t=1
is consistent. However, this estimator of the covariance matrix Es is not unbiased.

Therefore, it is often replaced by the following unbiased estimator


E~ t' (Y BZ)(Y BZ)' (214)
T- Kp- 1 t T-Kp 1

Obviously, both estimators are consistent estimators of the covariance matrix, and they

are .,- mptotically equivalent.

When estimating the coefficients of the vector autoregressive model from data, we

assumed the order p of the VAR(p) to be known. In practice, however, it is unknown, and

therefore, needs to be derived from the data. Since zero coefficient matrices are allowed,

one could simply set p to some upper bound on the VAR order. On the other hand,

selecting an unnecessary large p would affect the forecast precision of the estimated model.

Therefore, it is advantageous to apply some suitable criteria for optimal selection of the

lag length parameter p.









2.3.2 VAR Order Selection Criteria

Various criteria for choosing the optimal model order are developed. Some of the

most commonly used are the final prediction error (FPE) criterion, Akaike's information

criterion (AIC), Hannan-Quinn criterion (HQ), and Schwarz or B ,-i ,- information

criterion (SC) [59].

Let E,(m) denote the maximum likelihood estimator of ~, computed by fitting the

VAR model of order m. The FPE criterion proposed by Akaike in 1969 is based on the

idea that minimizing the mean square error improves the forecast of the model. For a

VAR(p) time series, the FPE criterion is defined as


FPE(m) ( ) K det (1 (m)) (2-15)

Using the FPE criterion, the estimate PFPE of the model order p is selected so that

FPE (PFPE) min FPE(m), (2-16)
m=1,...,M

where M denotes some upper boundary on the model order. In other words, first, for each

m = 1,... M, the vector autoregressive model of order m is estimated from the data, and

the respective values of the FPE(m) are calculated using (2-15); then the order producing

the smallest value of FPE(m) is chosen among the possible orders m 1,..., M.

AIC is another popular order selection criteria that was also introduced by Akaike.

Given a VAR(m) model, the Akaike information criteria is defined as follows:

2m inK2 I
AIC(m) =2 +n (I1,(m) (2-17)

Similarly to the FPE criterion, the VAR(m) models are estimated for different m =

1,..., M to obtain the corresponding AIC(m) values for each order. Then the estimate

PAIC of the model order p with the smallest AIC(m) is selected.









The Hannan-Quinn order selection criterion, HQ, is given by:


HQ(m) mK2 nT + In (m)) (2-18)
T

As before, among the model parameters m = 1,..., M, the parameter m having the

smallest value of HQ(m) is chosen as the estimator PHQ of the true model order p.

Last, but not least, we present Schwarz criterion, which was derived using B li-, -i i

arguments. The SC is formulated as:

mK21n T 1
SC(m) = nT +ln (,(m ) (2-19)

and the order minimizing SC(m) is chosen among m = 1,..., M as the estimator psc of

the model order p.

Some interesting statistical properties of the above criteria are proved in [59].

In particular, it is shown that AIC and FPE criteria for VAR order selection are

.,-i-,,il! 11 ically equivalent, although these estimators of the model order are not consistent.

On the other hand, the other two criteria provide consistent estimators of the order

parameter p. More precisely, in the univariate case (K = 1), the Hannan-Quinn criterion

is consistent (i.e. limTr,+ Pr{p = p} = 1). In addition, the HQ criterion is strongly

consistent for K > 2 (i.e. Pr {limT,+,- p } = 1). The SC is shown to be strongly

consistent for any dimension K.

It is important to keep in mind that even though FPE and AIC do not provide

consistent estimators, they are not necessarily inferior to HQ and SC. Actually, in small

samples, and even in larger samples, FPE and AIC may produce better forecast, although

they may not estimate the model order correctly.

2.3.3 Stability Condition and Other Assumptions of VAR

As mentioned above, the conditions of stability, stationarity and Gaussian distribution

are usually imposed on time series when dealing with the VAR models. Below we define

these conditions, and discuss their role. A K-dimensional VAR(p) time series (2-9) are









called stable, if


det(IK Alz ... Apz) / 0 for complex z : zl < 1. (2-20)


In other words, the VAR(p) process (2-9) satisfies the -Il,.1ii:,1 condition when its reverse

characteristic p ..1;,,;. ..;;,,.l (given by det(IK Alz ... ApzP)) has no roots on and inside

the complex unit circle.

The stability condition guarantees that there exists a moving average (1 A)

representation for the VAR(p) process. Also stability ensures that the process is a

well-defined stochastic process with the distributions of its univariate components and

joint distribution of the process yt uniquely determined by the innovation process Et. For

a stable VAR(p)process, both the process mean and the autocovariance are time-invariant

(which, according to the definition below, implies stationarity).

When the stability condition is violated, the process variance is increasing with time

and unbounded. Specifically, if the reverse characteristic polynomial of the time series has

a single unit root, and all the other roots are outside the complex unit circle, then the

time series behavior is similar to a random walk. In this special case, the variance increases

linearly with time, the correlation between yt and yt+h approaches 1, and the process mean

E [Y(t)] exhibits a linear trend for v / 0. In addition, if one of the roots of the reverse

characteristic polynomial lies strictly inside the complex unit circle, then such process is

explosive, i.e. the process variance grows exponentially. Various approaches are developed

in the time series literature to address the time series with the unit roots. For example,

the unit roots can be removed by taking differences. However, the explosive time series are

not as well-studied, because it is believed that an exponential increase in the variance of

the economic time series is not well founded. As one can see the stability assumption pi' "-

an important role in VAR(p).

A wide-sense stationarity for stochastic processes is imposed on the VAR time series

as follows. A stochastic process Y(t) is considered stat.:..., *i ;, if









1. E [Y(t)] = v for all t;

2. E [(Y(t + h) v) (Y(t) -)'] = F(h) = F(-h) for all t and = 0,1,....

In other words, the stationarity condition supposes that the first and the second moments

are time invariant. Also note that the process mean v and the autocovariance matrix

Fy(h) are finite. It is shown (see Proposition 2.1 in [59]) that

A stable VAR(p) time series yt, t 0, 1, 2,... is stationary.

Since stability of a time series implies that the series is stationary, the stability condition (2-20)

is sometimes cited in the literature as the stationarity condition. However, it is important

to remember that these two conditions are not equivalent. In fact, although a stable vector

autoregressive series is alv-- stationary, the converse is not true, i.e. an unstable time

series is not necessary non-stationary.

The Gaussian distribution assumption is introduced into the VAR(p) model through

Et. Specifically, given representation (2-9) of the VAR (p), the innovation process Et is

assumed to be Gaussian white noise. This condition implies that yt is a Gaussian process,

i.e. any subcollection t,... t+h follows a multivariate normal distribution for all possible

values of t and h.

2.4 Inegrated and Cointegrated VAR

In previous sections of this chapter, we considered VAR processes, for which the

stationarity and stability assumptions are satisfied. However, in practice, many time series

data are fit better by unstable non-stationary processes. In this chapter, we introduce

integrated and cointegrated processes, which are found especially useful in econometric

studies, and for which the stability and stationarity conditions are violated.

Recall that the VAR(p) process (2-9) satisfies the -/.i..:1/; condition when its reverse

characteristic 1y. Jl;;.i,,.:/.l det(IK A1z ... ApzP) has no roots on and inside a complex

unit circle. If an unstable process has a single unit root and all the other roots outside

of the complex unit circle, then such process exhibits a behavior similar to that of a

random walk. In other words, the variance of such process increases linearly to infinity,









and the correlation between the variables Y(t) and Y(t h) tends to 1 as t o0. On

the other hand, when the root of reverse characteristic polynomial lies inside the unit

circle, the process becomes explosive, i.e. its variance increases exponentially. In real-life

applications, the former case is of the most practical interest.

This renders the following definition of an integrated process.

A one-dimensional process with d roots on the unit circle is said to be integrated of

order d (denoted as I(d)).

It can be shown [59] that the integrated I(d) process Y(t) of order d with all roots of

its reverse characteristic polynomial being equal to 1 can be made stable by differencing

the original process d times. For example, the integrated I(1) process Y(t) becomes stable

after taking the first differences (1 L)Y(t) = Y(t) Y(t 1), where L represents the

lag operator. More generally, for the I(d) process Y(t), its transformation (1 L)dY(t) is

stable.

An example of an integrated I(d) process in the univariate case is an autoregressive

integrated moving average process ARIMA(p, d, q), which is sometimes called fractionally

difference autoregressive moving average process for d c (-0.5, 0.5). The one-dimensional

process Y(t) is said to be ARIMA(p, d, q), if Z(t) := (1 L)dY(t) is a stationary

autoregressive moving average ARMA(p,q) process, i.e.
P q
aZ(t i) = bjt-j, (2-21)
i=0 j=0

where Et-js are independent normally distributed random variables with mean 0 and

variance a2, and L is the differencing operator introduced above.

It is noteworthy to point out that taking differences may distort the relationship

among the variables (i.e. one-dimensional components) in some VAR(p) models. In

particular, this is the case for systems with cointegrated variables. It turns out that fitting

VAR(p) model after differencing the original cointegrated process produces inadequate

results. Next, we discuss such processes.









Cointegrated processes were first introduced by Clive Granger in 1981, and gained

a great deal of popularity in both theoretical and applied econometrics. Indeed, many

economic variables are expected to be in equilibrium relationship, for example, household

income and expenditures, or prices of a given commodity in different markets.

Suppose that sampled values yit of K different variables of interest i(t) are combined

into the K-dimensional vectors yt = lt,..., YKt)'. In addition, suppose that the variables

are in a long-run equilibrium relation


c Y(t) := ci Yi(t) +... + CK YK(t) = 0, (2-22)

where c = (c1,..., CK) is a K-dimensional real vector. During any given time interval, the

relation (2-22) may not necessarily be satisfied precisely by the sample yt, instead we may

have:

c yt := ci yt + C. + CK YKt = Et, (2-23)

where Et is a stochastic process that denotes the deviation from the equilibrium relation at

time t. If our assumption about the long-run equilibrium among individual variables I(t),

i = 1,..., K is valid then it is reasonable to expect that the variables Y(t) move together,

i.e. the stochastic process Et is stable. On the other hand, this does not contradict the

possibility that the variables deviate substantially as a group. Therefore, it is possible that

although each individual component i(t) is integrated, there is a linear combination of

Y(t), i = 1,..., K, which is stationary. Integrated processes with such property are called

cointegrated.

Without loss of generality, we assume that all individual one-dimensional components

Yi(t) (i = 1,..., K) are either I(1) or I(0) processes. Then the combined K-dimensional

VAR(p) process

Y(t) = v+ AIY(t- 1) +... + ApY(t- p) + Et (2-24)









is said to be cointegrated of rank r, when the correspondent matrix


I IK A1 ... Ap (2-25)


has rank r.

Since some one-dimensional components of the cointegrated VAR(p) process are

integrated processes, one may be interested in testing the presence of a unit root in the

univariate series. In the following section, we present a commonly used unit root test,

which was derived by Dickey and Fuller [16].

2.4.1 Augmented Dickey-Fuller Test for Testing the Null Hypothesis of the
Presence of a Unit Root

The augmented Dickey-Fuller (or ADF) test is a widely used statistical test for

detecting the existence of a unit-root of the reverse characteristic polynomial in a

univariate time series. By fitting an autoregressive AR(k) model, this test investigates

the null hypothesis of an autoregressive integrated moving average ARIMA(p, 1, 0)

process against the alternative of a stationary ARIMA(p + 1, 0, 0) process. The limiting

distribution of the ADF test for p < k 1 was derived by Dickey and Fuller [16], and it

can be shown that this distribution is the same for k > 1 and for k = 1. Fuller tabulated

the approximate critical values for the ADF test with k > 1 and p < k 1 for -/ .. '.:

sample sizes.

Finite-sample critical values for the ADF test for i,1, sample size were obtained

by means of response surface analysis by MacKinnon [60], who also showed that an

approximate .,-,iii!,ll ic distribution function for the test can be derived via response

surface estimation of quantiles [61].

Although the ..-, ,.ii!ll ic distribution of the ADF test statistic does not depend on

the lag order, it is noted by C'!. uii et al. [13] that empirical applications must deal with

finite samples, in which case the distribution of the ADF test statistic can be sensitive to









the lag order. Taking this into account, they closely examined the roles of the sample size

and the lag order in finding the finite-sample critical values of the ADF test.

As we noted above, the limiting distribution of the ADF test statistic is the same for

k > 1 and k = 1. Hence, for simplicity, we consider the case of k = 1. In fact, let Y denote

the autoregressive AR(1) model


Y(t) c Y(t- 1)+ Et, t 1,2,... (2-26)


where Y(0) = 0, c is a real number, and Et ~ N(0, a2) (i.e. Et is normally distributed with

zero mean and variance a2 for all t = 1, 2,...).

Note that when |c| < 1, the process Y(t) converges to a stationary process as t oo;

whereas, in the case of c| = 1, the process Y(t) is not stationary with variance ta2.

Furthermore, when |cl > 1, not only the process is not stationary, but the variance of Y(t)

grows exponentially with time t.

From the AR(1) model (2-26), one can see that in the case when c = 1, in order to

make the process stationary, the series can be appropriately transformed by differencing.

Furthermore, notice that the condition c = 1 in (2-26) is clearly equivalent to the

requirement that the reverse characteristic polynomial det(1 cz) = 1 z of AR(1) has a

unit root. In other words, to determine whether an autoregressive time series AR(1) has a

unit root, we must test the null hypothesis Ho : c = 1.

Let y y2, .. yT denote a sample of T consecutive observations of the AR(1) process

Y(t), then the maximum likelihood estimator of c is the least squares estimator

C t= tt- (2 27)
T y2 1

Note that 'c is a consistent estimator of the regression coefficient c.

Since each yt, t 1,..., T is a realization of an AR(1) process, it follows from (2-26)

that yt = c yt-1 + Et holds, and so by plugging this last condition into Equation (2-27), the









estimator c of the regression coefficient can also be written as:

(zc =tt I 1+t) 1 (2-28)
T
S z2T 1
c t 1 -1 i+ t1 t-1it
zT 12
Et 1 2t t
Yt 1 Lt- Et
c+
Y, E I-itL
3t 1u/ 1

Subtracting c from both sides of Equation 2-28 and multiplying each side by T lead

to the ADF statistic

T(Z- c) = i t-I t (2-29)
2 t2 1 t[1t- 1
Dickey and Fuller [16] derived the following representation of the limiting distribution for

statistic T(c- c):

T(c-c)= i- (W2 -1), as T oo (2-30)
2

where
OO
r F= df Xf, (2-31)
i=1

W = v2 dXj, (2-32)
i=1
2(-1)i+1
d = )' (2-33)


and random variables X, i = 1, 2,..., are independent and identically distributed

according to the normal distribution with zero mean and variance a2. Note that a symbol

= denotes convergence in distribution.

In [16], Dickey and Fuller considered the following "Stu. li ii. i i statistic based on

the likelihood ratio test of the hypothesis Ho : c = 1.



^ S y) (2-34)
t=2-2









where

^S2 7 _1,- i12 (2-35)
T 2

and is computed from (2-27)

Tables of the critical values for the .i-i1!,l .1' ic distributions of the ADF test statistic

T(c 1) and the statistic ? can be found in Fuller [22]. We summarize some of the

information in Table 2-1, which lists the p-values for .,-vmptotic distributions of T(c- 1)

and ? corresponding to percentiles of 90, 95, and 99 percent.

2.4.2 Phillips-Ouliaris Cointegration Test

The unit root tests based on analysis of residuals were introduced by Phillips [75]. In

particular, in his study Phillips first considered two statistics Z, and Zt for testing the

null of no cointegration in time series.

Because many unit root tests, constructed before 1987, were founded on the

assumption that the errors in the regression are independent with common variance

(which is rarely met in practice), Phillips wanted to relax the rather strict condition that

the time series are driven by independent identically distributed innovations. In other

words, he wanted to develop the testing procedures based on the least squares regression

estimation and the associated regression t statistic, which would allow for rather general

weakly dependent and heterogeneously distributed sequence of error terms.

The properties of .-i- i!! i i.. i distributions of residual based tests for the presence

of cointegration in multiple time series were thoroughly investigated by Phillips and

Ouliaris [76]. The characteristic feature of these tests is that they utilize the residuals

computed from regressions among the univariate components of multivariate series. The

residual based procedures developed by Phillips and Ouliaris are designed to test the null

of no cointegration by means of testing the null hypothesis of the unit root presence in

the residuals against the alternative of a root that lies inside the complex unit circle. The

hypothesis Ho of the absence of cointegration is rejected, if the null of a unit root in the









residuals is rejected. In the nutshell, the procedures are simply residual based unit root

tests.

As noted in [76], the residual based unit root tests are .,-vi! ,tically similar, and can

be represented via the standard Brownian motion. Moreover, the ADF and Zt tests are

proved to be .,-imptotically equivalent. However, these two tests are not as powerful as

the test based on statistic Zo, because it was shown by Phillips and Ouliaris [76] that the

rate of divergence under cointegration assumption is slower for the ADF and Zt than other

tests, such as the Z,-statistic test. The later test (i.e. the cointegration test based on Z,)

is also widely known as the Phillips-Ouliaris cointegration test.

It is noteworthy that the null hypothesis for the Phillips-Ouliaris test is that of no

cointegration (instead of cointegration). This formulation is chosen because of some 1n i i"

pitfalls found in procedures that are designed to test the null of cointegration in multiple

time series. These defects (discussed in more detail in [76]) are significant enough to be a

strong argument against the indiscriminate use of the test formulations based on the null

of cointegration, and to support the continuing use of residual based unit root tests.

Consider the K-dimensional vector autoregressive process Y(t). Let us partition

Y(t) = (Ut, V )' into the univariate component Ut = Y1(t) and the (K 1)-dimensional

Vt Y( t),..., YK(t))'.

The residuals are determined by fitting linear cointegrating regression


U(t)= cV(t) + t, t 1,2,... (2-36)


Residual based tests are formulated to test the null hypothesis that the multiple time

series Y(t) are not cointegrated using the scalar unit root tests, such as the ADF test,

which are applied to the residuals t, t = 1, 2,... in (2-36).

In [76], the ADF test as well as two additional tests Z, and Zt, developed earlier by

Phillips [75], were applied to check for the presence of a unit root in the residuals t. In









order to perform the unit root test, we fit an AR(1) model to t, t = 1, 2,... according to


at = _- + Qt, t 1,2,... (2-37)

Then the statistic Z, in Phillips-Ouliaris test is defined as follows:

1 2 2
Z = T( 1 1) (2-38)
T2 t=-2 2t-I

whereas the Zt statistic is given by the following formula:

(t 2 1) 1 Si -S 2
Z Y ) = _2 l 1 (2 39)
where I1 T

where


s 2 2 Q (2-40)

T T T
4Si -T Yj Qt +r7"; T L s, (241)
t=1 s=1 t=s+1
,.. 1- (2-42)
1+1

Note that s2 and STi are consistent estimators for the variance a2 of Qt and the partial sum

variance a2 =limT, E (E 52), where ST z= t1 is the partial sum of the error terms

in (2-37).

The critical values for Z, and Zt statistics can be found in [76] (Tables I and II).

Phillips and Ouliaris tabulated the values for cointegrating regressions with at most 5

explanatory variables. Some estimates of the critical values for the Phillips-Ouliaris test

(Zn) are listed in Table 2-2.

2.4.3 Estimation of Cointegrated VAR(p) Processes

Several methods can be employ, ,1 to estimate the parameters of a Cointegrated

VAR(p) model, including modifications of the approaches used for estimation of the

standard VAR(p) processes.









In this section we present the maximum likelihood approach to estimating a Gaussian

cointegrated VAR(p) process. Suppose yt is a realization of a K-dimensional VAR(p)

process with cointegration rank r, such that 0 < r < K. Without loss of generality, we

assume that Y(t) has zero mean, i.e. the intercept v = 0 in (2-24).

Given a realization yt, t = 2,..., of Y(t), one seeks to determine the coefficients of

the following model:


Yt= Alyt-1 + ... + Apyt+p+ Et, t= 1,2,..., (2-43)

subject to the constraint


rank(I) = rank(IK A -...- A) = r. (2-44)


Note that Et is assumed to be a Gaussian white noise with a nonsingular covariance matrix

Es. Furthermore, the initial conditions y-p+l,... ,yo are supposed to be fixed.

In order to impose the cointegration constraint, the model (2-43) is reparameterized

in the following fashion [59]:


Ayt DAyt-_ +... + Dp_,Ayt-p+l + Fyt-p + t, t 1,2,... (2-45)


where Ay = yt yt-1, and matrix II can be represented as a product II = HC of matrices

of rank r, i.e. H is (K x r) and C is (r x K).

Consider


AY :- [AY1,...,Ayr],

Ayt

AXt : (2-46)

AXYt-p+2
AX := [AXo,..., AX,_-] ,

D := [D1,...,Dp-1],

Y-p := [y-p, ...-, YT-p]









Then the log-likelihood function for a sample of size T can be written as:

KT T
In 1 ln[2r] TIn [det S,]
2 2
trace ((AY DAX + HCYp)' E (AY DAX + HCYp)) (2 47)
2

The proof of the following theorem on the maximum likelihood estimators of a

cointegrated VAR process can be found in [59] (Proposition 11.1).

Theorem 1. (reproduced from [59])



M : I AX'(AXAX')-AX,

Ro := AYM,

R1: Y_, -,
1
Sj RiR j, =0, 1.

Let G be the lower ', :,,,,,il,1ir matrix with positive .':,i..'.1i such that GS G' = IK

Denote A1 > ... > AK to be the .i; ,; ,;,,,. of GSloSoo So1G',

and

vi,..., v2 be the corresponding orthonormal eigenvectors.

Then the 1 A-1.:. O 7, ,.od function in (2-47) is maximized for

C := [v ,... ,]'G,

H -AYMY'IC' (CY, _IY'C')1

-So1C' (CS11C')-

D : (AY + HCYp)AX (AXAX')-,

Z := (AY DAX + HCY_) (AY DAX + HCY_)'.
T









The maximum is


KT T + ln KT(24
max[lnl] K2 ln[27]- 2 In [det oo] + ( A) (2-48)
i= 1

2.4.4 Testing for the Rank of Cointegration

Based on Theorem 1, one can easily derive the likelihood ratio statistic for testing

a candidate value ro of the cointegration rank r of a VAR(p) process against a larger

cointegration rank rl.

Given a VAR(p) process y(t) defined by (2-24), suppose we wish to test a hypothesis

Ho against an alternative H1, where


Ho : r= ro against H1 : ro < r < rl. (2-49)


Under assumption that the noise Et is a Gaussian process, the maximum of the likelihood

function for a cointegrated VAR(p) model with cointegration rank r is computed in

Theorem 1. From that result, the value of the LR statistic for testing (2-49) can be

determined in the following manner:


ALR(ro,r1) 2[lnLmax(ri)- lnLmax(ro)] (2-50)
r1 TO



-T ln( A),
i=ro+1

where Lmax(ri), i = 0, 1, denotes the maximum of the Gaussian likelihood function

for cointegration rank ri. The advantage of this test is in the simplicity with which the

LR statistic can be computed. On the other hand, the .,-i-,iii!l' tic distribution of the

LR statistic (2-50) is nonstandard. Specifically, the LR statistic is not ..i-','i!iil1 l ly

distributed according to x2-distribution. Nevertheless, the ..i-mptotic distribution of the

cointegration rank test statistic ALR depends only on two factors:









the difference K r between the process dimensionality and the cointegration rank;

and

the alternative hypothesis.

As a result, the selected percentage points of the .i-:-.1,l' I '1ic distribution of the test

statistic ALR were tabulated by Johansen and Juselius in [38]. The percentage points of

.I-i-,iIl ,-l ic distribution of ALR are given in Tables 2-3 and 2-4.

Table 2-1. Critical values of the .,-imptotic distributions of the T(c- 1) and r for
performing unit root check by the ADF test (reproduced from [22])
Statistic 911 95' 9' '.
c 0.93 1.28 2.03
S0.89 1.28 2.01


Table 2-2.


Critical values of the ..i-mptotic distributions of the Z, statistic for testing the
null of no cointegration (Phillips-Ouliaris demeaned, reproduced from [76]).
Parameter n (n = K 1) represents the number of explanatory variables


911' .
-17.0390
-22.1948
-27.5846
-32.7382
-37.0074


Table 2-3.


95',
-20.4935
-26.0943
-32.0615
-37.1508
-41.9388


9-r
-28.3218
-34.1686
-41.1348
-47.5118
-52.1723


Percentage points of the .,-i-ii! l 'i ic distributions of the ALR(r, K) for testing
the cointegration rank (reproduced from [38])
91 r' 95' 9' '
6.69 8.08 11.58
15.58 17.84 21.96
28.44 31.26 37.29































Table 2-4. Percentage points of the .,-i~,ll, '1 ic distributions of the ALR(r, r + 1) for testing
the cointegration rank (reproduced from [38])
K -r 91 r', 95' I
1 6.69 8.08 11.58
2 12.78 14.60 18.78
3 18.96 21.28 26.15









CHAPTER 3
PHASE SYNCHRONY IN BRAIN DYNAMICS

In this chapter we introduce a concept of phase synchronization, and consider two

methods for estimating the phase of a signal, specifically using the Hilbert transform and

via the wavelet transform.

3.1 The Role of Phase Synchronization in Neural Dynamics

The word "synchrony" originates from a combination of two Greek words cva (syn,

meaning common) and Xpovoc (chronos, meaning time), and it can be translated as

!' iplening at the same time". A concept of synchronization can be defined as a process

of active adjustment between the rhythms of different oscillating systems due to some

kind of interaction or coupling between them [78]. Synchronization phenomena were

discovered in the seventeenth century by C. Huygens who first observed synchronization

between two pendulum clocks hanging from a common support [33]. Since then, the

study of synchronization between dynamical systems became an active field of research

in many scientific and technical disciplines, including solid state physics [74], plasma

physics [84], communication [11, 48], electronics [72, 77], laser dynamics [20, 87, 98], and

control [80, 88].

Synchronization phenomena can also be found in physiological systems, such as heart

and brain. Synchronization processes in physiological systems were discovered by B. van

der Pol in the beginning of the twentieth century. In particular, van der Pol was the first

to apply oscillation theory to the human heart [103].

One of the important research areas in neuroscience explores the role of synchronization

in neural dynamics. Much effort is given to investigation of synchronization phenomena on

all different levels of organization of brain tissue, starting with pairs of individual neurons

to larger scales, such as within a given area of the brain or between distinct parts of the

brain.









M ini: studies emphasize that normal cognitive processes call for the transient

integration of numerous functional areas in various regions of the brain, and as a result,

the brain regions are in constant interaction with each other [15, 101, 104, 105]. Neural

synchrony p1 ii- a vital role in such large-scale integration. In fact, various neuronal

groups oscillate in specific frequency bands and become phase-locked over a limited period

of time. This observation has stimulated the development of robust approaches that

allow one to measure the phase-synchrony in a given frequency band from experimentally

recorded biomedical signals, such as EEG.

In particular, the importance of synchronization of neuronal discharges has been

shown by a variety of animal studies using microelectrode recordings of brain activity [83,

96]. The findings in the microelectrode-recording studies are also supported at coarser

levels of resolution by other studies in animals and humans [21].

An electrophysiological signal is recorded via a low impedance extracellular

microelectrode by placing the microelectrode sufficiently far from individual local neurons

in order to prevent any particular cell from dominating the signal. Next, to obtain the

local field potential (LFP), the signal is low-pass filtered, with a cut off at approximately

300 Hz. Due to the low impedance and positioning of the micro electrode, the recorded

signal is predominantly induced by the activity of a large number of neurons. The

unfiltered signal reflects the sum of action potentials from cells within approximately

50-350 micrometers from the tip of the electrode [53] and slower ionic events from within

0.5-3 millimeters from the tip of the electrode [39]. The spike component of the signal is

removed by low-pass filter, whereas the lower frequency signal, the LFP, is preserved in

the signal. It is assumed that the local field potential characterizes the synchronized input

into the observed area, in the contrast to the spike data, which represents the output from

the area.

Local field potentials (LFPs) of various degrees of spatial resolution can be recorded

by scalp EEG or MEG. In fact, studies have shown that the presence of gamma and beta









band responses can be detected during visual discrimination tasks on the human scalp [97]

and in subdural electrocorticograms [50, 54]. In addition, some recent findings -~i--. -1

that long-range synchronization analogous to the one found in microelectrode studies in

animals can also be detected between surface recordings [82].

It has been shown that synchronization is a significant attribute of the signal recorded

from the patients affected by several neurological disorders. In particular, researchers have

found that epilepsy [65] and Parkinsons disease [99] manifest as a pathological form of the

synchronization process.

It is noted in [56] that although the cross-correlograms between spike discharges may

be adequate for microelectrode studies, the quantification of phase synchrony between

meso- or macro-electrodes (i.e. EEG/\! I G, intracranial recordings) calls for entirely

different methods. Therefore, they emphasize an importance of clearly distinguishing

between synchrony as an appropriate estimate of phase relation, and the classical measures

of coherence or spectral covariance that have been extensively used in neuroscience [8, 10,

62]. Le Van Quyen et al. discuss two important limitations of coherence [56].

The first limitation arises because the standard approaches for measuring coherence [12]

based on Fourier analysis are known to be highly dependent on the stationarity of the

measured signal, whereas the signals recorded from the brain, such as EEG, appear

to be clearly non-stationary. Applying the timefrequency estimation method, which is

not founded on the assumption of stationarity, could improve this limitation towards

estimating a stable, instantaneous coherence as well as synchrony between two concurrent

brain signals.

The second limitation stems from the fact that classical coherence is a measure of

spectral covariance. Hence, it is not able to separate the effects of amplitude and phase

in the relations between two signals. Because we are concerned with examining the

specific hypothesis that phase-locking synchrony is the pertinent biological mechanism of

transient integration in the brain, coherence serves only as an indirect measure. In order









to investigate the phase relations between different areas in the brain directly, the phase

component should be extracted separately from the amplitude component for a given

frequency or frequency band, which can be quite unstable or even chaotic. In a nutshell,

coherence gives only an indirect and approximate indication of phase synchrony. There has

been a general increase of interest in understanding bivariate data by studying their phase

synchronization over time not only in neuroscience, but also in other research fields [86].

In other words, even though our discussion of the phase synchrony is focused on EEG

data, its applications can also be extended to the fields other than neuroscience.

Classical concept of the synchronization of two oscillators is described as an active

adjustment of their rhythmicity that manifests in phase-locking between the synchronized

oscillators. Specifically, given two signals X1(t) and X2(t), and their corresponding

instantaneous phases 01(t) and 02(t), the basic definition of the phase locking states that


ni(t) m2(t) = C cost, (3-1)

where integers n and m specify the phase locking ratio.

When investigating phase synchrony in neurophysiological signals, one must assume

that the constant phase locking ratio is valid within a limited time interval T, which

usually means a few hundreds of milliseconds. When examining neural signals, one

must keep in mind that discovering the presence of the phase locking between EEG

recordings from two distant parts of the brain is not straightforward. The detection of

phase synchrony in neural signals is problematic because of several factors particularly

when working not on the level of a single neuron, but rather with large neuronal

populations, whose activity is picked up by macroscopic or meso- electrodes. As noted

in [56], as a consequence of volume conduction effects in brain tissues, the activity of a

single neuronal population can be recorded by two distant electrodes, which results in

spurious phase-locking between their signals. Furthermore, in non-invasive EEG, the

true synchronies are hidden in a significant background noise, and so, in the synchronous









state, the phase shifts back and forth around some constant value. Hence, the signals

can be viewed as synchronous or not synchronous only in a statistical sense. Therefore,

this necessitates the development of novel approaches that are capable of extracting the

true phase synchronies from noisy data, and so, the condition (3-1) must be adjusted to

account for the noise as follows:


|ni(t)- m0r2(t)|
where C denotes a positive constant.

The following two key steps are instrumental in investigating the phase synchrony:

1. estimate instantaneous phase of each signal;

2. provide a statistical criteria to quantify the degree of phase-locking.

Two methods for detecting phase-locking applied to neuronal signals have recently

been considered in the literature. Tass and colleagues [99] extracted the instantaneous

phases from original signals by means of the Hilbert transform, and then applied to

magnetoencephalographic (\ll;G) motor signals in patients affected by Parkinson's

disease [99], and also to the synchronization between cardiovascular and respiratory

rhythms [93]. On the other hand, Lachaux et al. [50] estimated the phases from the

original signals by means of convolution with a complex wavelet, and then applied it to

EEG and intracranial data during cognitive tasks [51, 82].

The first step in quantifying phase synchronization between two time series X and Y

is the determination of their instantaneous phases Qx(t) and Qy(t). This is achieved either

via the Hilbert transform or via the wavelet transform. These approaches are presented in

the next two sections.

3.2 Phase Estimation using Hilbert Transform

The first method used to extract the instantaneous phase from the time series is

based on the ,1:.~i/.:l .-..1..,l approach, which was first introduced by D. Gabor [23] and

later extended for model systems and experimental data [70, 86].









The Hilbert ',. .", irm of a given real-valued function f(t) with domain T is defined as

a real-valued function f(t) on T as follows:


f(t) = CPV f(T)g(t r)dTr CPV g(T)h(t T)dr, (3-3)

where
1
g(t) := t eT,
7tT
and symbol CPV signifies that the integral is taken in the sense of C',,. 1I,; principal value.

Notice that f(t) can be viewed as a convolution g(t) f(t) of the original function

f(t) with the function g(t). This means that the Hilbert transform can be performed by

applying an ideal filter, whose amplitude response equals to 1, and phase response is a

constant r/2 lag at all frequencies.

Given an arbitrary continuous real-valued time series X(t), the corresponding I,.,l.l.'

/.::,,l is defined as the following complex-valued function:


x(t) = X(t) + 2 X(t) = ax(t) exp {z. x(t)}, (3-4)

where t denotes time, z is a unit on the complex axis, X(t) denotes the Hilbert transform

of the time series X(t), ax(t) is the corresponding instantaneous amplitude, and Ox(t)

represents the instantaneous phase of the signal via Hilbert convolution.

It follows from Equation 3-4 that the instantaneous phase Ox(t) of X(t) can be

computed as:

x (t) arctan () (3-5)
X (t)
A key advantage of the analytic approach is that the phase can be easily computed

for an arbitrary broad-band signal. On the other hand, instantaneous amplitude and phase

have a clear physical meaning only if X(t) is a narrow-band signal. Therefore, filtration

is required in order to separate the frequency band of interest from the background brain

activity.









3.3 Phase Estimation via Wavelet Transform

An alternative approach to determining the instantaneous phase of the signal is

based on the wavelet transform. This method of phase estimation was proposed by

Lachaux and colleagues [50, 52], and is similar to the Hilbert transform method presented

above. In their approach, Lachaux et al. extract the instantaneous phase by applying the

convolution of the original signal with a complex Morlet wavelet. They consider the Morlet

wavelet (also known as Gabor function) at time t and frequency w given by the following

formula:

t,' (e) = exp {z 27 ( t)} exp- -2t} (3-6)

Notice that t,,(-r) is the product of a sinusoidal wave at frequency w, and a Gaussian

function centered at time t, with a standard deviation a proportional to the inverse of w.

It depends solely on o, which sets the number of cycles of the wavelet to 6wJa.

According to [56], given the time series X(t), the coefficient of the Morlet transform

as a function of time t and frequency w is defined as follows:

Wx(t,) Xr)(-).r,,,)dr, (37)

where t,, (r) denotes the complex conjugate of the Morlet wavelet t,(-(r).

The following slight modification of the Morlet wavelet is introduced in [81]:


(t) exp 1- } exp {z. ,t} exp -2 ) (3-8)

where parameters wo and a represent the center frequency and the rate of decay of the

wavelet function, respectively. This is proportional to the number of cycles and related to

the frequency span by the uncertainty principle.

Similarly to the above, a complex time series of wavelet coefficients is obtained via the

convolution of X(t) with Q(t):


Wx() = o X)(t) = J r)X(t 7)dr iix(t) exp I () (3-9)
J-+









where ax(t) and Qx(t), respectively, are the instantaneous amplitude and the phase of the

signal X(t) extracted via the Morlet wavelet.

As in the case of the Hilbert transform, the phases can be determined from

Equation (3-9) as

x(t) arctan { [Wx( }, (3 10)

where Z [Wx(t)] and S [Wx(t)] denote the real and imaginary parts of the complex

transformed time series Wx(t), respectively.

3.4 Comparison between Two Approaches to Phase Extraction.

The above two definitions of the instantaneous phase are closely related, despite

the fact that they are based on very different transformations. In particular, the

connection between the phases obtained via the Hilbert and wavelet transforms was

demonstrated experimentally in [56] and also explained theoretically by Quian Quiroga

and colleagues [81].

In a nutshell, the phase Qx(t) extracted from the signal using the wavelet transform

corresponds approximately to the phase Qx(t) determined via the Hilbert convolution,

which would be performed after band pass filtering the time series. Furthermore, if the

phase estimation based on wavelet transform were performed by a convolution with an

analytic wavelet, and if this wavelet were applied to do the band pass filtering in the

Hilbert approach, then such approaches would, in fact, be equivalent.

It is easy to see that in the method based on the wavelet convolution, the center

frequency u and the rate of decay a of the wavelet can serve as parameters that allow us

to modify the frequency range of interest. On the other hand, the actual phase extraction

via the Hilbert transform is free of parameters, and so the correspondent phase preserves

information from the entire power spectrum and not just the main frequency band as in

the case of the wavelet convolution. As a result, it is possible to achieve a comparison

of narrow band and broad band synchronization simply by using both methods of phase

extraction without performing any additional filtering.









3.5 Measures of Phase Synchrony

Various measures of phase synchrony between two signal are proposed based on the

phases extracted via the Hilbert and the wavelet transforms, including standard deviation,

mutual information and Shannon entropy [49, 56]. However, most of the currently used

measures of phase synchronization are based on bivariate indexes.

In C'! lpter 4, we propose a novel multivariate approach to detecting phase synchronization

in the phases extracted from multiple time series, such as multichannel EEG.









CHAPTER 4
APPLICATION OF VECTOR AUTOREGRESSION TO MINING BRAIN DYNAMICS

4.1 Numerical Issues in Estimating Parameters of Vector Autoregression
from EEG

In the applications of Granger causality and related measures to the EEG, the

directional dependencies in neural data are analyzed based on autoregressive modeling.

Although, Bernasconi and Konig [2] applied statistical testing to verify the stationarity

of the data, and established the duration of the stationary interval for EEG to be

approximately 1 second. The statistical testing of underlying assumptions of the VAR

(which was thoroughly discussed in C'! lpter 4) is often omitted in the later studies.

As shown above, the stability condition is a very important assumption of vector

autoregression. In this study, we examine how the parameters of the model order and

sample size influence the stability of the derived VAR model.

In order to estimate VAR(p) model from data and investigate the properties of the

derived model, the rodent intracranial EEG data were used. The data set consisted of

the electroencephalographic recordings from 6 electrodes (implanted in left frontal, right

frontal, two left hippocampal and two right hippocampal parts of the rodent's brain)

sampled at 200 Hz.

To examine the applicability of the vector autoregressive modeling to EEG data, we

estimated the VAR(p) model parameters for different values of lag order p and different

sample sizes T. The sample sizes T E [1, 300] were used, and the lag length parameter

p varied between 1 and 30. In addition, we filtered data using a Rectangular band pass

Hamming window with 100 coefficients into the frequency bands of 0 -30 Hz, 30 -60 Hz,

60 -90 Hz, 90 120 Hz, and 120 150 Hz. The raw data and the five differently filtered

data represented separate data sets. For each data set, we ran the model estimation

procedure with T = 1,2,..., 300 and p = 1,2,..., 30. The procedure for estimating

coefficients of the model was implemented in the MATLAB environment based on the

multivariate least squares method presented above. For every VAR model derived from









the data, the stability condition was checked, and the number of the roots of the reverse

characteristic polynomial (RCP) on and inside the complex unit circle was stored. For

each data set, a 3-dimensional surface plot was produced by graphing (T, p, n), where

n denotes the number of the RCP roots on and inside the unit circle, whereas T and p

represent the sample size and the model order (or lag length), respectively. The obtained

surface plots are di- l i' '1 in Figures 4-1 and 4-2.

The surface plot in Figure 4-1 supports the fact [59] that the sample size parameter

should significantly exceed the lag length p, i.e. T > p. On the other hand, it can be

seen from Figure 4-1 that even for T > p, the stability assumptions may still be violated.

Indeed, for T = 132 considerably larger than p = 2, the estimated VAR has two RCP roots

on or inside the unit circle (n = 2).

Figure 4-2 shows that for the filtered data, the (T,p) region, where the stability

condition of the estimated VAR(p) model is violated, covers almost the whole domain.

Whereas the (T,p) region that corresponds to stable VAR(p) models is much smaller

than the stable region in Figure 4-1, and characterized by large T and very small p. Very

similar results were obtained for the number of the RCP roots inside the unit circle, when

estimating parameters of VAR(p) with different p using filters in the 30-60 Hz, 60-90 Hz,

90-120 Hz, and 120-150 Hz bands.

The experiment was repeated with consistent results on various samples from the six

data sets. The results of our experiments clearly show that the stability condition imposed

on the linear VAR model is often violated even for the parameters T > p. Furthermore,

filtering the data within some restricted frequency band often leads to reduction of the

(T,p) domain where the estimated VAR(p) models remain stable. In practice, the EEG

data are often filtered within a certain frequency band. In many studies that utilize the

vector autoregression to investigate Granger causality in the biomedical time series, the

verification of the conditions assumed by the VAR model is often omitted. From our point









of view, the relevant statistical tests ju- I iVii:-; the suitability of the model should alvb-7

be performed when estimating the model from data.

Various modifications of vector autoregression (which relax the stability condition of

no roots inside unit circle) are developed for analysis of economic time series. Although

Bernasconi and Konig [2] examined the stationarity of different EEG data, and concluded

that a time interval of approximately one second is the interval during which the EEG

time series can be considered stationary, the underlying assumptions of the VAR modeling

such as stationarity and stability are rarely statistically tested in applications to EEG

data. The results of our experiments indicate that the stability assumptions of vector

autoregressive model are often violated in application to the EEG data. This observation

-i --.- -i; that suitable extensions of multivariate autoregression to unstable processes may

fit the data better, and as a result, such extensions of VAR may be more appropriate for

the EEG time series analysis than the standard linear vector autoregressive modeling.

Additional statistical testing is required in order to make conclusions on what

multivariate models are the most suitable for extracting the directional dependencies

between channels in a frequency domain from multichannel EEG data.

4.2 Multivariate Approach to Phase Synchrony via Cointegrated VAR

We propose a new approach to measuring the synchrony among the instantaneous

phases extracted from multivariate time series. This approach is based on the Cointegrated

VAR modeling of time series.

Given the signal represented formally as a multiple time series X(t), one can extract

the instantaneous phases Qx(t) from each one-dimensional component Xi(t) of the signal

as shown in C'! lpter 3 (either via a convolution with the Morlet wavelet or by applying

the Hilbert transform). The phase extraction procedure produces a new multiple time

series Qx(t) of the correspondent phases.

Next, we derive new measures of phase synchrony of the signal based on the concepts

introduced in C'!i pter 3. Let us observe that the left-hand side of Equation (3-1)









represents the linear combination of the respective phases Ox, (t) and x,2 (t) with integer

coefficients. Also recall that condition (3-1), which defines phase-locking between two

signals XI(t) and X2(t), needs to be modified in practice to account for the noise in the

signal. Taking into account presence of the stochastic noise in the phase series, let us

introduce a i,. ',/fi. ,1 concept of the phase synchrony between two signals by rI' 1, i.,:. the

'i, ,i,,d:l;/ condition on the coefficients in the linear combination as follows.

Two signals XI(t) and X2(t) are considered to be gener /ll/; phase-synchronized, if the

correspondent instantaneous phases Ox, (t) and x,2 (t) satisfy the condition below:


3 c1, C2 : C1, () + 2x (t) = Zt, (4 1)

where zt ~ N(C, a') is a stochastic variable that represents the deviation from the

constant level C as a result of the noise. Notice that in the contrast to condition (3-1)

in the classic definition of phase synchronization, the coefficients cl and c2 in the

definition (4-1) of generalized phase-synchrony do not need to be integer.

Furthermore, it is straightforward that the new condition (4-1) means that a

two-dimensional process X(t) = (XI(t),X2(t))' is cointegrated. Based on this observation,

we can extend our modified concept of phase synchronization between two signals to the

multivariate case in the following manner.

The multichannel signal X(t) = (Xi(t),... ,XK(t)) is considered to be phase-

synchronized of rank r, if the process Ox(t) composed of the correspondent instantaneous

phases Ox,(t), i 1,..., K is cointegrated of rank r.

In the subsequent subsections, we first discuss the role of the cointegration rank in

the framework of multivariate phase-synchronization, and then apply this approach to

multichannel EEG data collected from the patients with absence epilepsy.









4.2.1 Cointegration Rank as a Measure of Synchronization among Different
EEG Channels

Note that integrated autoregressive processes I(d) are shown to exhibit behavior

similar to that of a random walk. In a short paper [66], Michael Murray used an example

of drunkard and her dog to illustrate the concept of the cointegration. To explain

our reasoning behind the rank of cointegration as a measure of synchrony, we briefly

summarize and then further extend his analogy.

The nonstationary processes (such as a random walk) are often introduced by teachers

of statistics by comparing them (it) with the drunkard's walk. The drunkard wonders

aimlessly, so that the direction of each step is random and completely independent of her

previous steps. In other words, the meandering of the drunkard is described by a random

walk:

Xt Xt-1= Et, t= 1,2,..., (4-2)

where Xt represents the position of the drunk at time t, and Et is a stationary white-noise,

which models the drunk's step at time t).

As Murray noticed [66], an unleashed puppy is another creature, whose behavior

reminds a random walk. Indeed, each new scent that puppy's nose comes upon dictates

a direction for the pup's next step so strongly that the last scent along with its direction

is forgotten as soon as the new scent appears. Having shown that the puppies follow the

random walk yt, t = 1, 2,..., let us represent the puppy's walk as:


yt Yt-1 = t, t = 2,..., (4-3)

where ct is a stationary white-noise (i.e. puppy's step at time t).

The well-known feature of a random walk is that the best predictor of the future

value is the most recently observed one. In other words, the longer it has been since we

had seen the drunk, or the dog, the further away from the initial place, on average, they









are at the moment. As a result, even if the drunk and the dog crossed their walks at some

location, as the time goes on, they tend to wander further away from each other.

However, if the puppy belongs to the drunkard, then they will remain relatively close

to each other at all the time, similarly to the individual integrated processes that together

form a cointegrated process. Indeed, the drunk would still wonder aimlessly in a random

walk fashion, as would her puppy. However, from time to time she would remember about

her dog and call for it, the puppy would recognize her voice, and bark. They would hear

each other and make their next step in each other's direction.

The paths of the drunk and her dog are still nonstationary, but they are no longer

independent from each other. As a matter of fact, at each time, the puppy and its master

are likely to be found not far from each other. If this is true, then the distance between

two paths is stationary, and the walks of the drunk xt and her dog yt are said to be

cointegrated, i.e. xt and yt are integrated 1(1), and there is a linear combination of xt and

yt (with non-zero weights) that is I(0), i.e. stationary.
Mathematically, the cointegrating relationship between a lady and her puppy can be

written as:


Xt- Xt_- =t + C(Yt-1 Xt), (4-4)

yt yt-1 = t + d(xt-1 yt-1), (4-5)

at time t = 1, 2,.... Note that Et and et, as before, represent the stationary white-noise

steps of the drunk and her dog.

Since Equation 4-4 can be easily rewritten in form of (2-45) as follows:


Xt Et C -C Xt-1
A (4-6)
t t -d d y/t-1









then


c -c
-d d


and so, rank(II) = 1. This shows that the cointegrating relationship between the drunk

lady and her puppy has the cointegration rank 1.

Note that rank(I) = 0, if and only if c = d = 0. In such case, (4-4) becomes simply a

system of equations (4-2) and (4-3), which models two independent random walks driven

by independent white noise processes E and e. On the other hand, when at least one of the

coefficients c and d is non-zero, then by multiplying system (4-6) by a vector [d, c]', we

have:

dAxt + cAyt = dEt + et, t = t, 2,..., (4-7)

which means that the model is driven by a single common stochastic trend dEt + ct.

Although the example described by Murray is clearly a bivariate cointegrated

VAR(1), it can be extended to an illustration of the multivariate cointegrated process.

Consider, for example, a heard of sheep guarded by two dogs, where the sheep wonder

aimlessly in the field, while the dogs run around and bring the sheep that have -11 i, l- too

far back into the flock. f -, for example, a faster dog guards sheep from the east, south,

and west, whereas a slower dog from the north, then the cointegrated process appear to

have the cointegration rank of 2. Clearly, two dogs are able to keep a flock of sheep closer

together, than a single dog can. In other words, the higher cointegration rank the more

restrictive it is.

In fact, let us consider a K-dimensional cointegrated vector autoregressive process,

and let r denote the cointegration rank of the process. Similarly to the bivariate example

above, we can see that when the rank is zero (r = 0), the univariate components of the

process are independent, and the model is driven by K independent white noise processes

(i.e. there is no cointegration). In the case of r = 1, we can decompose the multivariate









process onto K 2 independent components, and two dependent components that form a

common stochastic trend. Hence, in the case r = 1, the cointegrated model is driven by

(K 2) + 1 = K 1 independent stochastic processes. By induction, we can show that for

a cointegrated VAR process with the cointegration rank r, 0 < r < K 1, the VAR model

is generated by K r independent stochastic trends.

Therefore, the smaller is the cointegration rank r, the larger is the number K r

of the underlying independent stochastic trends, and so (the larger) is the vector space

in which our cointegrated model can travel. And the other way around, increasing the

cointegration rank of the model shrinks the underlying domain of the process, i.e. makes it

bounded to a smaller hyperplane. For r = K, the VAR(p) is a stable process, which clearly

has the most constrained domain. For r = 0, the VAR process is not cointegrated and

unrestricted.

Thus, in the framework of generalized phase-synchronization introduced above, the

cointegration rank represents a fundamental measure of synchrony in the multi-channel

signal, such as EEG. In particular, we z v that the signal is ../,ipl. /. ;/ r;,. ;. ronous,

if the cointegration rank r is zero. On the other hand, when the multivariate process is

stable (i.e. the rank coincides with the dimension of the process, r = K), the signal is said

to be perf,. I/,/ synchronous.

4.2.2 Absence Seizures

Absence seizures (or petit mal seizures) are known to occur in several forms of

epilepsy, whereas absence epilepsy refers to a type of epilepsy in which only the absence

seizures occur. Absence epilepsy is usually characterized by age of onset, and often affects

teenage population. Absence seizures usually begin in childhood or adolescence, and often

run in families, which may -ir--.- -1 a genetic predisposition. Absence seizures are marked

by momentary lapses of consciousness. Absence seizures often have no visible symptoms,

although some patients may have purposeless movements during a seizure, such as rapidly

blinking eyes. Absence seizures often have a brief duration, and a person may resume









the previous activity immediately after the seizure [73]. These brief seizures can happen

several times during a d,,-, but in some patients, the frequency of absence seizures can

be as high as hundred of times a di-, which interferes with the daily activities of a child

such as school. In some cases of childhood absence epilepsy, the seizures stop when a child

reaches puberty. Absence seizures exhibit a characteristic spike-and-wave EEG pattern at

a 3 Hz frequency [73].

Figure 4-3 di-pl'i-1 a multichannel EEG recording that includes an absence seizure.

The duration of the seizure is approximately 4 seconds. The figure vividly illustrates a

characteristic spike-and-wave activity during the seizure.

4.2.3 Numerical Study of Synchrony in Multichannel EEG Recordings from
Patients with Absence Epilepsy

The proposed approach to studying synchronization among multiple channels was

applied to analysis of EEG data recorded from the patience with absence epilepsy.

First, the multiple time series of the instantaneous phases were extracted from the

raw EEG data using the Hilbert transform approach as described in Section 3.2. In

particular, we took advantage of the functions hilbert and angle readily available in the

MATLAB R 2006a environment.

The VAR modeling and testing were implemented using the R 2.6.1 statistical

software. In our analysis of the instantaneous phases, we incorporated ar, adf. test,

po.test, caj olst and other functions found in packages series and urea.

Next, we illustrate our approach on the example of the EEG data file that includes

three seizure intervals. The file contains a 16-channel recording of scalp EEG sampled

at the 200 Hz frequency as well as two auxiliary channels, which were discarded. The

instantaneous phase values were estimated from the EEG time series by means of Hilbert

transform, and the resulting phase series were tested using the ADF test introduced in

Section 2.4.1. Specifically, we applied the Augmented Dickey-Fuller procedure to test the









presence of a unit root in the individual univariate components of the multiple time series

of estimated phases.

The results of our experiments for seizures 1, 2, and 3 are presented in Tables 4-1,

4-2, and 4-3, respectively. The channels, for which the ADF unit root test has detected a

presence of a unit root at the significance level a = 0.01, are listed as integrated. Whereas

the channels, for which the null hypothesis of a unit root has been rejected by the ADF

at the 1 percent level, are denoted by stat.:. ; : The channels for which the p-values of

the ADF test exceed 2.5' are marked with *. Notice that all three seizure segments are

considered stable, when the ADF is applied at a 0.025 significance level.

Next, we fit vector autoregression to the multiple time series of phase estimates, for

each of three different segments (before, during, and after a seizure) in order to determine

appropriate lag length parameter p. To find appropriate lags p, the Akaike Information

Criteria (AIC) was used. This led us to choose several lag length for each segment and

each seizure. Finally, Johansen cointegration rank procedure was applied to determine the

values of cointegration rank r for each case. The results are summarized in Tables 4-4, 4-5,

and 4-6.

Notice that during the seizure the system becomes stable, especially when modeled

using a short estimate of the lag parameter. Since the durations of the seizure 1 and

seizure 2 are rather short, and only include 440-500 sample points, the models estimated

under a long lag parameter may not adequately represent the underlying processes in

seizure 1 and 2. On the other hand, seizure 3 is estimated based on almost 1200 sample

values, and therefore the long lag model of a longer seizure 3 may be more realistic, than

the long lag models for shorter seizures 1 and 2. Overall, the models based on a short

lag p for all three seizures provide an evidence of absolute synchronization among the

channels. Whereas, the the pre-seizure and post-seizure models are more likely to be less

restricted, and seem to exhibit a cointegration rank between 9 and 16.










Results of the ADF unit root tests for each channel during three segments (2
seconds immediately before seizure, during seizure, and 2 seconds after seizure)
for seizure 1. Note that the significance at 2.5'. level is denoted by *


pre-seizure
3,4,7,9,15
1*,2*,5,6*,8* ,10,11,12*, 13* ,14* ,16*


seizure
1-3,5-14
4,15,16


post-seizure
1,3,5,6,8,10,12,15
2*,4*,7,9,13*, 14,16*


Table 4-1.



Seizure #
stationar'
integrate

Table 4-2.


pre-seizure
3,4,7,9-11,13,16
1*,2*,5*,6,8*,12*,14*,15


seizure
1-16
none


post-seizure
3,7,11-16
1*,2,4*,5,6*,8,9*,10


Results of the ADF unit root tests for each channel during three segments (2
seconds immediately before seizure, during seizure, and 2 seconds after seizure)
for seizure 3. Note that the significance at 2.5'. level is denoted by *


pre-seizure
7,11,13,14
1*,2*,3,4*,5*,6*,8*,9,10*,12,15,16*


seizure
1-16
none


post-seizure
2,4,5,11,13,15
1,3*,6*,7,8,9*,10*,12,14*,16


Results of the Johansen cointegration rank procedure for the multiple series
during three segments (2 seconds immediately before seizure, during seizure,
and 2 seconds after seizure) for seizure 1. Significance level is 1 Full rank is
denoted by t


Segment
long lag
short lag

Table 4-5.




Seizure #
long lag
short lag

Table 4-6.




Seizure #
long lag
short lag


pre-seizure
p = 22,r = 12
p 2,r 13


p 23, r
p


seizure
11, p= 20, r = 13
S2,r 16t


post-seizure
p 20, r = 12
p 2, r = 10


Results of the Johansen cointegration rank procedure for the multiple series
during three segments (2 seconds immediately before seizure, during seizure,
and 2 seconds after seizure) for seizure 2. Significance level is 1 full rank is
denoted by t


pre-seizure
S21, r 14
2, r 16t


seizure
p =26, r = 12, p= 20, r 9
p =3, r = 16t


post-seizure
p 20, r = 10
p 2, r 13


Results of the Johansen cointegration rank procedure for the multiple series
during three segments (2 seconds immediately before seizure, during seizure,
and 2 seconds after seizure) for seizure 3. Significance level is 1 full rank is
denoted by t


pre-seizure
p 24, r 13
p 2, r 9


seizure
p = 26, r = 16t, p = 20,r = 16t
p 2, r = 16t


post-seizure
p 20, r = 13
p 2,r 16t


Results of the ADF unit root tests for each channel during three segments (2
seconds immediately before seizure, during seizure, and 2 seconds after seizure)
for seizure 2. Note that the significance at 2.5'. level is denoted by *


y
d


Seizure #
stationary
integrated


Table 4-3.


Seizure #
stationary
integrated


Table 4-4.





















100


0 0


Figure 4-1. Numbers n of roots of the reverse characteristic polynomial (RCP) for
VAR(p), which lie on and inside the complex unit circle, computed for different
sample sizes T and for different model orders p using the raw data


0 0


Figure 4-2.


Numbers n of roots of the reverse characteristic polynomial (RCP) for
VAR(p), which lie on and inside the complex unit circle, computed for different
sample sizes T and for different model orders p using the 0-30 Hz band filtered
data


























-Jl~


=r



-J1






L~ i'


- -7 I


s
N
m

0



-o

0
0;



0
ccO





0



0
r^
Cs








0b




0
0
0
U


0



o







0
0;
Cs







0




V




0
o
E
5D
r^j

--)


(i~l i I iiii iiiiii


C


C









CHAPTER 5
CONCLUSION

Investigation of spatio-temporal properties of the EEG data by data mining and

optimization approaches posts various challenges. Numerous features and methods have

been proposed for studying the multivariate series that is EEG. The analysis of EEG time

series is often approached from two different points of view, the one that treats EEG data

as produced by a deterministic chaotic dynamical system, and the other more traditional

approach of linear autoregressive modeling.

In this work, we investigated several statistical approaches that are recently

introduced for data mining brain dynamics. In particular, we examined the application of

vector autoregressive modeling and linear Granger causality to raw and filtered EEG data.

Motivated by recent success in application of phase synchronization to analysis of

dynamic processes in epileptic brain, we developed a concept of generalized ;.'. I,.-

nization based on the novel idea of extending the classical synchronization condition

of a bounded linear combination of two phases. This simple bivariate condition in the

multivariate case is analogous to a cointegrating relationship in the multiple time series.

Thus, we can analyze the synchrony among different parts of the common interrelated

system (such as a human brain), by modeling the phases extracted from a finite number

of signals in the systems by means of cointegrated vector autoregression. Moreover, we

showed that the cointegration rank in the cointegrated VAR model of the phase time series

can be viewed as a measure of synchrony among the phases of different components of the

EEG signal.

Not only this new measure of multivariate phase synchrony can be tested on various

biomedical data, such as multichannel EEG recorded from an epileptic brain, but also

the new multiple phase synchronization can be employ, 1 in different areas of applied and

theoretic research (including physics, communication, electronics, laser dynamics, and









control) for studying synchronization among several dynamical systems or a system that

consists of several parts.









REFERENCES


[1] L. Baccala and K. Sameshima, "Partial directed coherence: a new concept in neural
structure determination," Biol. C;,I1,,. i vol. 84, pp. 463-74, 2001.

[2] C. Bernasconi and P. Koenig, "On the directionality of cortical interactions studied
by structural analysis of electrophysiological recordings," Biol. C;,l ,,., vol. 81, pp.
199-210, 1999.

[3] R. D. Bickford, "Electroencephalography," in Adelman G. ed. E,. ;; 1.>pedia of
Neuroscience. Cambridge, MA: Birkhauser, pp. 371-373, 1987.

[4] C. D. Binnie, "Long term EEG recording and its role in clinical practice," IEE
Colloquium on Data Logging of Pi,;,'..:'.....: S.: ,il- pp.5/1-5/2, 1995.

[5] P. S. Bradley, O. L. Mangasarian, and W. N. Street, "Clustering via concave
minimization," in M. C. Mozer, M. I. Jordan, and T. Petsche, eds.,Advances in
Neural Information Processing S;r -/. vol. 9, pp. 368-374, 1997.

[6] P. S. Bradley and O. L. Mangasarian, 1. iI ire selection via concave minimization
and support vector machines," in J. Shavlik, Machine Learning Proceedings of the
Fifteenth International Conference, pp. 82-90, 1998.

[7] P. S. Bradley, U. M. Fayyad, and O. L. Mangasarian, \ II .!, ii ii Programming
for Data Mining: Formulations and C('! 11. es," INFORMS Journal on Co,,l',/,.ul
vol. 11, no. 3, pp. 217-238, 1999.

[8] S. L. Bressler, R. Coppola, and R. Nakamura, "Episodic multiregional cortical
coherence at multiple frequencies during visual task performance," Nature, vol. 366,
pp. 153156, 1993.

[9] R. Brown and L. Kocarev, "A unifying definition of synchronization for dynamical
systems," Chaos, vol. 10, pp. 344-349, 2000.

[10] T. H. Bullock and M. C. McC'lti. "Lateral coherence of the electrocorticogram: a
new measure of brain synchrony," Electr. ,'.. 1,,i].r Clin N '. *'iu,;;-,.: vol. 73, pp.
479-498, 1989.

[11] T. L. Carroll and L. M. Pecora, "Cascading synchronized chaotic systems," PI,,..
D, vol. 67, pp. 126, 1993.

[12] G. C. Carter, "Coherence and time delay estimation," Proc IEEE, vol. 75, pp.
236-255, 1987.

[13] Y.-W. C('!. t- K. S. Lai, "Lag Order and Critical Values of the Augmented
Dickey-Fuller Test," Journal of Business & Economic Statistics, vol. 13, no. 3, pp.
277-280, 1995.









[14] R. Cont, "Statistical Properties of Financial Time Series," in R. Cont and J. Yong
editors, Mathematical Finance: Theory and Practice, Lecture Series in Applied
Mathematics, vol. 1, 1999.

[15] A. R. Damasio, "Synchronous activation in multiple cortical regions: a mechanism
for recall," Semin Neurosci, vol. 2, pp. 287-296, 1990.

[16] D. A. Dickey, and W. A. Fuller, "Distribution of the Estimators for Autoregressive
Time Series With a Unit Root," Journal of the American Statistical Association, vol.
74, pp. 427-431, 1979.

[17] M. Eichler, Graphical modelling of multivariate time series, University of Heidelberg:
Preprint, 2001.

[18] M. Eichler, R. Dahlhaus, and J. Sandkuhler, "Partial correlation analysis for the
identification of synaptic connections," Biol. C.;Il. ,,. vol. 89, pp. 289-302, 2003.

[19] M. Eichler, "A graphical approach for evaluating effective connectivity in neural
systems," Philos. Transact. R. Soc. B vol. 360, pp. 953-67, 2005.

[20] L. F 1,1 i:w, P. Colet, and R. Roy, "Coherence and phase dynamics of spatially
coupled solid-state lasers," Phys. Rev. A, vol. 47, pp. 4287, 1993.

[21] W. J. Freeman, "Spatial properties of an EEG event in the olfactory bulb and
cortex," Electr., '1',hi .,r Clin N,-jir ;.?ih.1 vol. 44, pp. 586-605, 1978.

[22] W. A. Fuller, Introduction to Statistical Time Series, 2nd ed. New York: John Wiley,
1996.

[23] D. Gabor. "Theory of communication," in Proc. IEEE London, vol. 93, pp. 429,
1946.

[24] P. Georgiev, A. Cichocki, and H. Bakardii ,i "Optimization techniques
for Independent Component Analysis with Applications to EEG data," In
P. M. Pardalos, J. C. Sackellares, P. R. Carney, L. D. Iasemidis, eds., Quantita-
tive Neuroscience: Models, Algorithms, Diagnostics, and Theraputic Applications,
Kluwer Academic Publishers, pp. 53-68, 2004.

[25] J. Geweke, \! I-, cement of linear dependence and feedback between multiple time
series," J. American Statis. Association, vol. 77, pp. 304-313, 1982.

[26] C. W. J. Granger, "Economic Processes Involving Feedback," Information and
Control, vol. 6, pp. 28-48, 1963.

[27] C. W. J. Granger, lii., -i ii ig Causal Relations by Econometric Models and
Cross-Spectral Methods," Econometrics, vol. 37, pp. 424-438, 1969.

[28] C. W. J. Granger, "Some Properties of Time Series Data and Their Use in
Econometric Model Specification," J of Econometrics, vol. 16, pp. 121-130, 1981.









[29] M. J. Griffiths, P. Grainger, M. V. Cox, and A. W. Preece, "Recent advances
in EEG monitoring for general anaesthesia, altered states of consciousness and
sports performance science," in The 3rd IEE International Seminar on Medical
Applications of S.:j,,,l Processing, pp. 1-5, 2005.

[30] J. D. Hamilton, Time Series A,..l;;-'-. Princeton, 1994.

[31] M.-J. Hoeve, R. D. Jones, G. J. Carroll, and H. Goelz, "Automated detection of
epileptic seizures in the EEG," in Engineering in Medicine and B.:.. .,i'/; S '.. ./.;/
Proceedings of the :' :'.1 Annual International Conference of the IEEE, vol. 1, p1' "! ;
946, 2001.

[32] R. Horst, P. M. Pardalos, and N. V. Thoai, Introduction to Global Optimization, 2nd
edition. Kluwer Academic Publishers, 2000.

[33] C. Huygens. Horoloquium Oscilatorium.Paris, 1673.

[34] L. D. Iasemidis, "On the dynamics of the human brain in temporal lobe epilepsy,"
Ph.D. Dissertation, University of Michigan, Ann Arbor, 1991.

[35] L. D. Iasemidis, and J. C. Sackellares, "The evolution with time of the spatial
distribution of the largest Lyapunov exponent on the human epileptic cortex," in
D.W. Duke, and W.S. Pritchard, eds., Measuring chaos in the human brain, pp.
49-82, World Scientific, 1991.

[36] L. D. Iasemidis, D.-S. Shiau, J. C. Sackellares, P. M. Pardalos, and A. Prasad,
"Dynamical resetting of the human brain at epileptic seizures: application of
nonlinear dynamics and global optimization techniques IEEE Transactions on
Biomedical Engineering, vol. 51, no. 3, pp. 493-506, 2004.

[37] H. H. Jasper, "The ten-twenty electrode system of the International Federation,"
Electr .. ., ,,]hal.-,p. Clin. N, a,., ;-...1 vol. 10, pp. 370-375, 1958.

[38] S. Johansen and K. Juselius, \!. ::i:;i ii1 Likelihood Estimation and Inference on
Cointegration With Applications to the Demand for Money," Oxford Bulletin of
Economics and Statistics, vol. 52, pp. 169-210, 1990.

[39] E. Juergens, A. Guettler, and R. Eckhorn. "Visual stimulation elicits locked and
induced gamma oscillations in monkey intracortical- and EEG-potentials, but not in
human EEG," Experimental Brain Research, vol. 129, pp. 247-259, 1999.

[40] Y. Kaji, M. Akutagawa, F. Shichijo, H. N ., i-iip,, Y. Kinouchi, and S. N .', iliro,
"Analysis for Brain Activities during Operations Using Measured EEG," in Conf
Proc IEEE Eng Med Biol Soc. vol. 6, pp. 6003-6006, 2005.

[41] T. K i1 ii i and O. Ozdamar, \V i. !. t preprocessing for automated neural network
detection of EEG spikes," Engineering in Medicine and B. :..,'i; l.if', .-.:,.; IEEE,
vol. 14, no. 2, pp.160-166, 1995.









[42] M. Kaminski, M. Ding, W. A. Truccolo, and S. L. Bressler, "Evaluating causal
relations in neural systems: Granger causality, directed transfer function and
statistical assessment of significance," Biol. C;I,., ,i vol. 85, pp. 145-57, 2001.

[43] A. Kammerdiner, P. Xanthopoulos, and P. M. Pardalos, \ii[II 1, 1 I1 limitations
in application of vector autoregressive modeling and Granger causality to in 1,i-; of
EEG time series," In AIP Conf. Proc., vol. 953, pages 232-245, 2007.

[44] A. Kammerdiner, and P. M. Pardalos, "Analysis of multichannel EEG recordings
based on generalized phase synchronization and cointegrated VAR," In Proc. of
Conference on Computational Neuroscience :' "/'' 2008 (submitted).

[45] E. Keogh, and S. Kasetty, "On the need for time series data mining benchmarks: a
survey and empirical demonstration," in Proceedings of the Eighth AC'MI SIGKDD
international Conference on Knowledge Discovery and Data Mining, pp. 102-111,
2002.

[46] M.E. Kirlangic, D. Perez, S. Kudryavtseva, G. Griessbach, G. Henning, and
G. Ivanova, Fli I i! dimension as a feature for adaptive electroencephalogram
segmentation in epilepsy," in Engineering in Medicine and B'.: '..,i,; S... I.; Proceed-
ings of the *',, Annual International Conference of the IEEE, vol. 2, pp.1573-1576,
2001.

[47] W. A. Kittel, C. M. Epstein, and M. H. H v-- "EEG monitoring based on fuzzy
classification," Circuits and S;,-1. ',- Proceedings of the 35th Midwest Symposium,
vol. 1, pp.699-702, 1992.

[48] L. Kocarev and U. Parlitz, "General approach for chaotic synchronization with
applications to communication," Phys. Rev. Lett., vol. 74, pp. 5028, 1995.

[49] T. Kreuz, Measuring Synchronization in Model S, ii and Electr .. ,... .i'il' g ..J,1 .
Time Series from E'.:1 /i",;/ Patients, Dissertation. NIC Series, vol. 21, 2003.

[50] J. P. Lachaux, E. Rodriguez, J. Martinerie, and F. J. Varela, \!. ,-iiiiig phase
synchrony in brain signals," Hum. Brain Mapp., vol. 8, pp. 194, 1999.

[51] J. P. Lachaux, E. Rodriguez, J. Martinerie, C. Adam, D. Hasboun and F. J. Varela,
"Gamma-band activity in human intracortical recordings tii'..: rI by cognitive
tasks," Eur J Neurosci, vol. 12, pp. 2608-2622, 2000.

[52] J. P. Lachaux, E. Rodriguez, M. Le Van Quyen, A. Lutz, J. Martinerie, and
F. J. Varela, "Studying single-trials of phase-synchronous activity in brain,"
Int. J. Bifurcation Chaos Appl. Sci. Eng., vol. 10, pp. 2429, 2000.

[53] A. D. Legatt, J. Arezzo, and H. G. Vaughan, "Averaged multiple unit activity as
an estimate of phasic changes in local neuronal activity: effects of volume-conducted
potentials," J. of Neuroscience Methods, vol. 2 no. 2,pp. 203-217, 1980.









[54] M. Le Van Quyen, C. Adam, J. P. Lachaux, J. Martinerie, M. Baulac, B. Renault,
et al., "Temporal patterns in human epileptic activity are modulated by perceptual
discriminations" Neuroreport, vol. 8, pp. 1703-1710, 1997.

[55] M. Le Van Quyen, J. Martinerie, C. Adam, and F. J. Varela, "Nonlinear
spatio-temporal interdependencies of interictal intracranial EEG recordings from
patients with temporal lobe epilepsy: Localizing of epileptogenic foci," P,;/,-.. D,
vol. 127, pp. 250-65, 1999.

[56] M. Le Van Quyen, J. Foucher, J.-P. Lachaux, E. Rodriguez, A. Lutz, J. Martinerie,
and F. J. Varela, "Comparison of Hilbert transform and wavelet methods for the
analysis of neuronal synchrony," Journal of Neuroscience Methods, vol. 111, pp.
83-98, 2001.

[57] H. Liang, M. Ding, R. Nakamura, and S. L. Bressler, "Causal influences in primate
cerebral cortex during visual pattern discrimination," Neuroreport, vol. 11, pp.
2875-80, 2000.

[58] Q. Liu, M. Sun, M. L. Scheuer, and R. J. Sclabassi, "Patient tracking for video/EEG
monitoring based on change detection in DCT domain," in Bioengineering Confer-
ence, 2005. Proceedings of the IEEE 31st Annual Northeast, pp.114-115, 2005.

[59] H. Liitkepohl, Intoduction to multiple time series i,,,;l;;--. Berlin, Heidelberg:
Springer-V i1 ,1 1991.

[60] J. G. MacKinnon, "Critical Values for Cointegration Tests," in R. F. Engle and
C. W. J. Granger, eds.,Long-Run Economic Relationships: Readings in Cointegra-
tion,New York: Oxford University Press, pp. 266-276, 1991.

[61] J. G. MacKinnon, "Approximate Asymptotic Distribution Functions for Unit- Root
and Cointegration Tests," Journal of Business & Economic Statistics, vol. 12, pp.
167-176, 1994.

[62] V. Menon, W. J. Freeman, B. A. Cutillo, J. E. Desmond, M. F. Ward, S.L. Bressler,
et al., "Spatio-temporal correlations in human gamma band electrocorticograms,"
Electr ',., 'l,,/,/, Clin N, ; ',,i'i;-l,'.: vol. 98, no. 2, pp. 89-102, 1996.

[63] M. Mezard, G. Parisi, and M. A. Virasoro, Spin glass theory and beyond. World
Scientific, 1987.

[64] S. Mitra and T. Acharya, Data mining: multimidea, soft .i,,;,j'li: and 1'....: .,[.,r-
matics. Hoboken, NJ: John Wiley & Sons, Inc., 2003.

[65] F. Mormann, K. Lehnertz, P. David, and C. E. Elger, .\!, i phase coherence as
a measure for phase synchronization and its application to the EEG of epileptic
patients," Ph;I-;... D, vol. 144, pp. 358-369, 2000.

[66] M. P. Murray "A Drunk and Her Dog: An Illustration of Cointegration and Error
Correction," The American Statistician, vol. 48, no. 1 pp. 37-39, 1994.









[67] E. Niedermeyer, F. H. Lopes da Silva, Electr ,** ',' i.',., l''i,,'i,." Basic principles,
clinical applications and related fields, 3rd edition. Philadelphia: Lippincott,
Williams & Wilkins, 1993.

[68] P. L. Nunez, Neocortical D;,i/,i,. and Human EEG Rl.iln,, New York: Oxford
Univ. Press, 1995.

[69] N. H. Packard, J. P. Crutchfield, J. D. Farmer, and R. S. Shaw, "Geometry from a
Time Series," Phys. Rev. Lett., vol. 45, no. 9, pp.712-716, 1980.

[70] P. Panter. Modulation, noise, and spectral ,.iil,-.:- New York: McGraw-Hill, 1965.

[71] H. S. Park, Y.H. Lee, S. Doo S, et al., "Detection of epileptiform activity using
wavelet and neural network," Engineering in Medicine and B.:,. .,i/;/, S1 ,,. .I;, vol.3,
pp. 1194, 1997.

[72] U. Parlitz, L. Junge, W. Lauterborn, and L. Kocarev, "Experimental observation of
phase synchronization," Phys. Rev. E, vol. 54, pp. 2115, 1996.

[73] J. B. Patten, N, ,. ,',,.: ,l Differential Diagnosis, 2nd ed., Springer-V, i1 .: 1996.

[74] D. W. Peterman, M. Ye, and P. E. Wigen, "High frequency synchronization of
chaos," Phys. Rev. Lett., vol. 74, pp. 1740, 1995.

[75] P. C. B. Phillips, "Time series regression with unit root," Econometrica, vol. 55,
pp. 277-301, 1987.

[76] P. C. B. Phillips and S. Ouliaris, "Asymptotic Properties of Residual Based Tests for
Cointegration," Econometrica, vol. 58, pp. 165-193, 1990.

[77] A. S. Pikovsky, "Phase synchronization of chaotic oscillations by a periodic external
field," Sov J Commun Technol Electron, vol. 30, pp. 85, 1985.

[78] A. S. Pikovsky, M. G. Rosenblum, and J. Kurths. Synchronization. A universal
concept in nonlinear sciences. Cambridge, UK: Cambridge Univ. Press, 2001.

[79] O. A. Prokopyev, V. Boginski, W. C'! ,valitwongse, P. M. Pardalos,
J. C. Sackellares, and P. R. Carney, \, Iwork-based Techniques in EEG Data
Analysis and Epileptic Brain Modeling," in P. M. Pardalos, V. Boginski and
A. Vazacopoulos, eds., Data Mining in Biomedicine, Springer, pp. 559-573, 2007.

[80] K. Pyragas, "Continuous control of chaos by self-controlling feedback," Phys. Lett.
A, vol. 170, pp. 421, 1992.

[81] R. Quian Quiroga, A. Kraskov, T. Kreuz, and P. Grassberger, "Performance
of different synchronization measures in real data: A case study on
electroencephalographic signals," Phys. Rev. E, vol. 65, pp. 041903, 2002.









[82] E. Rodriguez, N. George, J. P. Lachaux, J. Martinerie, and F. J. Varela,
"Perceptions shadow: Long-distance synchronization in the human brain," Na-
ture, vol. 397, pp. 340-343, 1999.

[83] P. R. Roelfsema, A. K. Engel, P. Konig, and W. Singer, "Visuomotor integration is
associated with zero time-lag synchronization among cortical areas," Nature, vol. 385,
pp. 157-161, 1997.

[84] E. Rosa Jr., W. B. Pardo, C. M. Ticos, J. A. Walkenstein, and M. Monti, "Phase
synchronization of chaos in a plasma discharge tube," Int. J. Bifurc. Chaos, vol. 10,
pp. 2551, 2000.

[85] M. Rosenblum, A. Pikovsky, J. Kurths, "Phase synchronization of chaotic
oscillators," Phys Rev Lett, vol. 76, pp. 1804-1807, 1996.

[86] M. G. Rosenblum, A. S. Pikovsky, C. Schafer, P. Tass, and J. Kurths, "Phase
synchronization: From theory to data analysis," in: S. Gielen, F. Moss F, eds.
Handbook of B.: '1.:'..rl PhI,'. ; vol. 4. Elsevier Science, Neuro-informatics, 1999.

[87] R. Roy and K. S. Thornburg, "Experimental synchronization on chaotic lasers,"
Phys. Rev. Lett., vol. 72, pp. 2009, 1994.

[88] N. F. Rulkov, L. S. T-iIvir,.- and H. D. I. Abarbanel, "Tracking unstable orbits in
chaos using dissipative feedback control," Phys. Rev. E, vol. 50, pp. 314, 1994.

[89] J. C. Sackellares, L. D. Iasemidis, R. L. Gilmore, and S. Roper, "Epilepsy when
chaos fails," in K. Lehnertz, J. Arnold, P. Grassberger, and C. E. Elger, eds., Chaos
in the brain?, World Scientific, 2002.

[90] J. C. Sackellares, L. D. Iasemidis, D.-S. Shiau, P. M. Pardalos, and P. R. Carney,
"Spatio-temporal transitions in temporal lobe epilepsy," in P. M. Pardalos,
J. C. Sackellares, P. R. Carney, L. D. Iasemidis, eds., Quantitative Neuroscience:
Models, Algorithms, Diagnostics, and Theraputic Applications, Kluwer Academic
Publishers, pp. 223-238, 2004.

[91] S. E. Said and D. A. Dickey, "Testing for Unit Roots in Autoregressive-Moving
Average Models of Unknown Order," Biometrika, vol. 71, pp. 599-607, 1984.

[92] K. Sameshima and L. Baccala, "Using partial directed coherence to describe
neuronal ensemble interactions," J. Neurosci. Methods, vol. 94, pp. 93-103, 1999.

[93] C. Schafer, M. G. Rosenblum, J. Kurths, and H. H. Abel, "Heartbeat synchronized
with ventilation," Nature, vol. 392, pp. 239-240, 1998.

[94] F. Lopez de Silva, "EEG analisys:Theory and practice; Computer-assisted EEG
diagnosis: Pattern recognition techniques," in E. Niedermeyer, F. H. Lopes da Silva,
eds. Electr '.. '1'1A.. ',i''l,/;, Basic principles, clinical applications and related fields,
pp. 871-919, 1987.









[95] C. A. Sims, "Money, Income and Causality," American Economic Review, vol. 62,
pp. 540-552, 1972.

[96] W. Singer and C. M. Gray, "Visual feature integration and the temporal correlation
hypothesis," Annu Rev Neurosci, vol. 18, pp. 555-586, 1995.

[97] C. Tallon-Baudry, O. Bertrand, C. Delpuech, and J. Pernier, "Oscillatory
gamma-band (30-70 Hz) activity induced by a visual search task in human," J
Neurosci, vol. 17, pp. 722-734, 1997.

[98] D. Y. Tang, R. Dykstra, M. W. Hamilton, and N. R. Heckenberg, "Experimental
evidence of frequency entrainment between coupled chaotic oscillations," Phys. Rev.
E, vol. 57, no. 3, pp. 3649, 1998.

[99] P. Tass, M. G. Rosenblum, J. Weule, J. Kurths, A. Pikovsky, J. Volkmann,
et al., "Detection of n:m phase locking from noisy data: application to
magnetoencephalography," Phys Rev Lett, vol. 81, pp. 3291-3294, 1998.

[100] M. Teplan, "Fundamentals of EEG measurement," Measurement science review, vol.
2, no. 2, pp. 112-23, 2002.

[101] G. Tononi and G.M. Edelman, "Consciousness and complexity," Science, vol. 282,
pp. 1846-1851, 1998.

[102] R. D. Traub, M. A. Whittington, E. Buhl, J. G. Jefferys, and H. J. Faulkner, "On
the mechanism of the gamma to beta frequency shift in neuronal oscillations induced
in rat hippocampal slices by tetanic stimulation," J Neurosci, vol. 19, pp. 1088-1105,
1999.

[103] B. van der Pol and J. van der Mark, "The heartbeat considered as a relaxation
oscillation, and an electrical model of the heart," Phil. Mag., vol. 6, pp. 763, 1928.

[104] F. J. Varela, "Resonant cell assemblies: A new approach to cognitive functions and
neuronal synchrony," Biol Res, vol. 28, pp. 81-95, 1995.

[105] F. J. Varela, J. P. Lachaux, E. Rodriguez, and J. Martinerie, "The brain web: phase
synchroniztion and large-scale integration," Nature Rev Neurosci, vol. 2, pp. 229-39,
2001.

[106] N. Wiener, "The Theory of Prediction," in E. F. Beckenbach, ed. Modern Mathe-
matics for the Engineer, New York: McGraw-Hill, 1956.

[107] A. Wolf, J. B Swift, H. L. Swinney, and J. A. Vastano, "Determining Lyapunov
exponent from a time series," P/,-,K.:. a D, vol. 16, pp. 285-317, 1985.

[108] K. Yang, and C. Shahabi, "On the Stationarity of Multivariate Time Series for
Correlation-Based Data Analysis," in Proceedings of Fifth IEEE International
Conference on Data Mining (ICDM'05), pp. 805-808, 2005.









BIOGRAPHICAL SKETCH

Alla Revenko Kammerdiner was born in Kiev, Ukraine. An older of two children,

she grew up in Kiev, Ukraine, graduating from School #32 in 1994. She earned her B.S.

in Probability Theory and Mathematical Statistics from the National Taras Shevchenko

University of Kyiv in 1998.

In January 2001, Alla joined a graduate program in the Mathematics Department at

the University of Florida. Upon graduating in May 2004 with her M.S. in mathematics,

Alla entered the Ph.D. program in industrial and systems engineering at the University of

Florida.

Alla has been happily married to Jason R. Kammerdiner for the last 3 years. On

February 24, 2008 she completed her first marathon in 3:53:09.





PAGE 1

1

PAGE 2

2

PAGE 3

3

PAGE 4

IwouldliketoshowmyheartfeltappreciationtomyadvisorDr.PanosM.Pardalosforhissupportandmentoring.Workingwithhimhashelpedmegrownotonlyprofessionally,butalsoasaperson.Iamalsoverygratefultoothermemberswhoservedonmysupervisorycommittee,J.ColeSmith,WilliamW.Hager,VladimirL.Boginski,andH.EdwinRomeijn,fortheirvaluablecommentsonmyresearchforthisdissertation.Lastbutnotleast,IthankmyhusbandJason,myparentsOlgaandAleksandr,mybrotherMikhail,andalltherestofmygreatfamilyfortheirunconditionalloveandsupport. 4

PAGE 5

page ACKNOWLEDGMENTS ................................. 4 LISTOFTABLES ..................................... 7 LISTOFFIGURES .................................... 8 LISTOFSYMBOLS .................................... 9 ABSTRACT ........................................ 11 CHAPTER 1INTRODUCTION .................................. 12 1.1StatisticalMethodsforDataMining ..................... 12 1.2ElectroencephalographicRecordings ...................... 16 1.3Featureextraction ............................... 19 1.4ContributionSummary ............................. 23 1.4.1TestingApplicabilityofFrequencyDomainEstimatesofGrangerCausalityforEEGtimeseries ..................... 23 1.4.2GeneralizationofPhaseSynchronizationviaCointegratedVAR ... 25 2AUTOREGRESSIVEMODELINGOFMULTIPLETIMESERIES ....... 28 2.1MultivariateAutoregressiveModelinginEEGDataMining ......... 28 2.2TestsofGrangerCausality ........................... 29 2.3VectorAutoregressiveModels(VAR) ..................... 31 2.3.1MethodsforVARParameterEstimation ............... 31 2.3.2VAROrderSelectionCriteria ..................... 34 2.3.3StabilityConditionandOtherAssumptionsofVAR ......... 35 2.4InegratedandCointegratedVAR ....................... 37 2.4.1AugmentedDickey-FullerTestforTestingtheNullHypothesisofthePresenceofaUnitRoot ...................... 40 2.4.2Phillips-OuliarisCointegrationTest .................. 43 2.4.3EstimationofCointegratedVAR(p)Processes ............ 45 2.4.4TestingfortheRankofCointegration ................. 48 3PHASESYNCHRONYINBRAINDYNAMICS .................. 51 3.1TheRoleofPhaseSynchronizationinNeuralDynamics ........... 51 3.2PhaseEstimationusingHilbertTransform .................. 55 3.3PhaseEstimationviaWaveletTransform ................... 57 3.4ComparisonbetweenTwoApproachestoPhaseExtraction. ......... 58 3.5MeasuresofPhaseSynchrony ......................... 59 5

PAGE 6

....................................... 60 4.1NumericalIssuesinEstimatingParametersofVectorAutoregressionfromEEG ....................................... 60 4.2MultivariateApproachtoPhaseSynchronyviaCointegratedVAR ..... 62 4.2.1CointegrationRankasaMeasureofSynchronizationamongDierentEEGChannels .............................. 64 4.2.2AbsenceSeizures ............................ 67 4.2.3NumericalStudyofSynchronyinMultichannelEEGRecordingsfromPatientswithAbsenceEpilepsy ................. 68 5CONCLUSION .................................... 73 REFERENCES ....................................... 75 BIOGRAPHICALSKETCH ................................ 83 6

PAGE 7

Table page 2-1ADFtest:CriticalvaluesoftheT(bc1)andbstatistics ............. 49 2-2Phillips-Ouliarisdemeaned:CriticalvaluesoftheZstatistic .......... 49 2-3Johansentest:CriticalvaluesoftheLR(r;K)statistic .............. 49 2-4Johansentest:CriticalvaluesoftheLR(r;r+1)statistic ............ 50 4-1Seizure1:ResultsoftheADFunitroottests ................... 70 4-2Seizure2:ResultsoftheADFunitroottests ................... 70 4-3Seizure3:ResultsoftheADFunitroottests ................... 70 4-4Seizure1:ResultsoftheJohansencointegrationranktests ............ 70 4-5Seizure2:ResultsoftheJohansencointegrationranktests ............ 70 4-6Seizure3:ResultsoftheJohansencointegrationranktests ............ 70 7

PAGE 8

Figure page 1-1TheInternational10-20systemforplacementofEEGelectrodes ......... 27 4-1Rawdata:NumbersofunstablerootsfordierentTandp 71 4-2Filtereddata:NumbersofunstablerootsfordierentTandp 71 4-3EEGsegmentwithabsenceseizure ......................... 72 8

PAGE 9

Mostofthenotationisunambiguouslydenedinthetextwhereitisintroduced.Toprovidesomegeneralguidelines,weincludethefollowinglistofcommonlyusedsymbols. ARIMA(p;d;q)autoregressiveintegratedmovingaverageprocessAR(p)autoregressiveprocessoforderpCPVCauchyprincipalvalue 9

PAGE 11

Thisstudydiscussesstatisticalapproachesusedfordataminingofmultichannelelectroencephalogramrecordings.Suchrecordingsrepresentmassivedatasetsthatcontainhiddenpatternsofcomplexdynamicalprocessesinthebrain.Formally,multichannelEEGcanbeviewedasamultipletimeseries,andtherefore,anaturalideaforsummarizingsuchdataistoutilizeautoregressivemodelingofmultivariatestochasticprocesses. Inparticular,wethoroughlydiscussvariousconceptsandapproachesrelatedvectorautoregressiveprocesses,includingstablestationaryVARmodelsoforderpandnonstationarysystemswithintegratedandcointegratedvariables,aswellasproceduresforestimatingparametersofthesystems(e.g.order,lag,orcointegrationrank). TheworkhighlightssomestabilityissuesthatmayariseintheapplicationofvectorautoregressiontominingEEGdata,andquestionstheapplicabilityofGrangercausalityinthefrequencydomaintomultichannelEEG. SynchronizationhasbeenfoundtobeanimportantcharacteristicoftheabnormalbraindynamicsmanifestedbyepilepsyandParkinson'sdisease.Wereviewtwoapproachesforextractingtheinstantaneousphasefromtimeseries.Inthisstudy,wegeneralizetheconceptofthephasesynchronization,andproposeanovelapproachbasedonmultivariateanalysisviamodelingcointegratedVAR(p)processes. 11

PAGE 12

Dataminingtogetherwithdatapreprocessingconstitutethecentralpartofamoregeneralprocessofknowledgediscoveryindatabases(KDD).TheKKDprocesscanbedescribedasasequenceofactions,whichselectstherawdataindatawarehousesandtransformstheselecteddatainordertodiscovervalid,understandable,novelandpotentiallyusefulknowledgefromthedata.Datapreprocessingthatisappliedtorawdatatoimprovethequalityofthedataofteninuencestheselectionandfacilitatestheapplicationofdataminingtechniques.Properpreprocessingofrawdataleadstoadecreaseinthetimeneededtominethedata,andbooststheoverallminingeciency. Thetechniquesusedfordatapreprocessingcanberoughlysubdividedintodatacleaning,dataintegrationanddatareduction[ 64 ].Datacleaningtechniqueshandletheproblemofincomplete,inconsistentanderroneousdata,removethenoiseinherentlypresentintherawdata,minimizeredundancyinthedata,etc.Dataintegrationisconcernedwithcombiningheterogeneousdatacollectedfromdierentsourcestoforma 12

PAGE 13

Thetwofundamentaltasksassignedtodataminingareadescriptivetaskofdiscoveringhiddenpatternsandrelationshipsingivendata,andapredictivetaskofforecastingorclassifyingthemodel'sbehaviorfromavailabledata.Dataminingincludesregression,classication,clustering,imagerestoration,learningassociationrulesandextractingfunctionaldependencies,datasummarization,etc.Dataminingiscloselyconnectedtootherresearchareassuchasstatistics,machinelearningandarticialintelligence,optimization,visualizationanddatabases.Dataminingutilizesmanyimportantresultsfromtherelatedelds,whilekeepingthemainfocusonthealgorithmsandarchitectures,scalabilityofthenumberoffeaturesandinstances,andautomatedmanagingofmassivequantitiesofdiversedata. Manyareasofdataminingemployvariousapproachesdevelopedintheeldofoptimization.Inparticular,itisshownin[ 7 ]thatmanybasicproblemsindatamining,includingclassicationandclustering,canbeformulatedasmathematicalprogrammingproblemsandsolvedusingoptimizationtechniques. Infact,Bradleyatel.[ 7 ]demonstratedthataproblemofminimizingthenumberofmisclassiedpointsintwo-classclassicationcanbeviewedasalinearprogramwithequilibriumconstraints(LPEC).LPECisalinearprogram(LP)withasinglecomplementarityconstraint.Suchconstraintimposesaconditionoforthogonalitybetweentwolinearfunctions.LPECformulationarisesintheinstanceswhentheconstraintsoftheproblemincludeanotherLPproblem. Inaddition,theproblemoffeatureselectionintwo-classclassicationbyndingaseparatingplanethatutilizesminimumnumberoffeaturescanbegivenamathematicalprogrammingformulationasaparametricproblem.Furthermore,theclassicationviasupportvectormachines(SVMsthatndtheseparatingplanemaximizingthe 13

PAGE 14

6 ].Moreover,asindicatedin[ 7 ],theabovemathematicalprogrammingformulationshavebeenextendedtobeeectivelyemployedbyotherdataminingapproaches,includingneuralnetworkstraining,calculationofnonlineardiscriminants,andbuildingdecisiontrees.Theclusteringproblemhasacomplexformulationasaminimizationproblemwiththeobjectivegivenbyasumoftheminimumsofasetofconvexfunctions[ 5 ].Ingeneral,thisobjectivefunctionisneitherconvexnorconcave.SeereviewbyBradleyatel.[ 7 ]foradditionalinformationaboutmathematicalprogrammingformulationsforvariousproblemsindatamining,theapplicationofoptimizationtechniques,aswellasthechallengesthattheeldofdataminingoerstooptimization. Manydataminingapproaches,suchasclassication,clustering,indexingandsegmentation,havebeenappliedtotimeseriesanalysis[ 45 ].Manytraditionalstatisticalapproachesarealsoappliedtominingtimeseries.Forinstance,regressionisoneofthemostcommonlyusedtechniquesformodelingandforecastingtimeseries.Amongthestatisticalmodelsappliedtoregressionintimeseriesarelinearautoregression(AR),autoregressivemovingaverageprocess(ARMA),autoregressiveintegratedmovingaverageprocess(ARIMA),aswellastheirmultivariateanalogs(i.e.vectorautoregression,etc.) Timeseriesariseinvariousappliedareas,includingeconomicsandnance,meteorology,biomedicine,etc.Forinstance,thestudyofseismicactivityrelatedtoearthquakesproducestwo-dimensionaltimeseries,whereeachmeasurementconsistsofthetimeandthemagnitudeofaregisteredseismicevent.Manybiomedicalsignals,suchaselectrocardiogram(ECG),electroencephalogram(EEG)andelectrooculogram(EOG)representtimeseriesthatcanbeinterpretedviaapplicationofregression,segmentation,neuralnetworks,andotherdataminingmethodologies.Soundsignalsareanotherexampleoftimeseriesthatareeectivelyanalyzedusingdierentdataminingtechniques. 14

PAGE 15

14 ],althoughthenormaltimescaleisaverynaturalchoiceofparameterforthetimeseriesdescribingphysicalprocesses,theregulartimeloosesitsnaturalmeaningwhendealingwithmanynancialtimeseries.Becausemostofnancialtimeseriesareirregularlyspacedinphysicaltime,theconceptof\businesstime"or\intrinsictime"isintroducedtorepresentanewtimeparameterwithrespecttowhichtimeseriesareregularlyscaled.Thisprocedureoftimedeformationallowstherelabeledtimeseriestobeviewedasstationaryonanewtimescale.Financialtimeseriesalsooftenexhibitclearseasonaltrends,whichobviouslycannotbefoundwhenexaminingtimeseriesproducedbyspeech. StatisticaltestingofseveralmultivariatetimeseriesdeterminedthatthetimeseriesfromAUSLANandBCIdatasetscanbeconsideredstationary,whereasBCIMPIandEEGcontainednon-stationarytimeseries[ 108 ].Asaresultofinherentdierencesintimeseriesdatafromdiversesources,somedataminingmethodsthataresuccessfullyappliedtotimeseriesinoneresearchareamaynotnecessarilybeapplicabletoanalysisoftimeseriesthatstemfromanotherappliedeld. Timeseriesobtainedfromelectroencephalogram(EEG)recordingshaveseveralinterestingpropertiesthatdistinguishthemfromothertimeseries.Althoughsomestudiesapplyone-dimensionalmodelingbyconsideringonechannelinEEGrecordingatatime,ingeneral,EEGdatashouldbetreatedasmultivariatetimeseries.Themultivariateapproachbecomesespeciallyimportantinviewofitsabilitytoinvestigatespatio-temporaldependenciesintheEEGdataincontrasttobeinglimitedtoonlytemporalrelationsinaone-dimensionalcase.ThereisadisagreementamongresearchersstudyingEEGdataonwhethertheseriesshouldbemodeledbyanon-linearstochasticprocessortheycanbebetterdescribedbyadeterministicchaoticdynamicalsystem. 15

PAGE 16

100 ]. Generallyspeaking,EEGrepresentsadigitaloragraphicrecordoftheelectricalactivityinthebrain,andcanbemeasuredbyeithernon-invasiveorinvasivemethods.EEG(obtainedduringanon-invasiveprocedure)isdenedasarecordofelectricalactivityofanalternatingtypemeasuredfromthescalpsurfaceafterbeingpickedupbymetalelectrodesandconductivemedia[ 67 ].TherearetwotypesofEEGproducedbyinvasiveprocedures,theelectrocorticogram,whichmeasuresthebrain'selectricalactivitydirectlyfromthecorticalsurface,andtheelectrogram,whichisanEEGobtainedusingdeepprobes. EEGestimatesandrecordstherelativechangeinelectricpotentialsproducedbyalargenumberofelectricdipolesduringaperiodofneuralexcitations.Theactivationofneurons(braincells)generateslocalcurrentowsinthebrain.EEGrecordsmostlytheelectricalcurrentsthatowduringsynapticexcitationsofthedendritesofnumerouspyramidalneuronsinthecerebralcortex.EEGrecordedfromthescalpsurfacecanonlydetecttheelectricalactivityproducedbymassivepopulationsofactiveneurons.Ontheotherhand,EEGrecordedusingdeepprobeelectrodesimplantedintothebraincanpickupasignalfromasmallgroupofneurons,whichcanbefurtherlteredouttoobtainthe 16

PAGE 17

68 ].Thisgivesanaveragedensityofabout104neuralcellspercubicmillimeterofthebrain.Neuralcellsareinterconnectedthroughsynapticconnectionsinthebrainintoneuralnets.Thebrainofanaverageadultcontainsapproximately500trillionsynapses.Thetotalnumberofneuronsdecreaseswithage.Asaresultthetotalnumberofsynapticconnectionsdeclineswithaging,eventhoughthenumberofsynapsesperoneneuronincreaseswithage. ToensuretheconsistencyinreferencinglocationsofelectrodesinEEGexperiments,theInternational10-20systemforEEGelectrodeplacementwasdeveloped[ 37 ].The10-20EEGsystemisusedtodescribetherespectivelocationsofscalpelectrodesduringEEGrecordinginrelationtotheunderlyingareaofcerebralcortex. Accordingtothe10-20system,anatomicallandmarksofaskull,suchnasion,inionandpreauricularpoints,areidentiedforconsecutiveplacementoftheelectrodesatxeddistancesfromthesepointsinstepsofeither10or20percent.Thisapproachisdevisedtotakeintoaccountpossiblevariationsofheadsize.Inaddition,themethodiseasilyapplicableinpracticaluse.Asaresult,the10-20EEGsystembecameverywidelyusedforpositioningelectrodes. Inthe10-20system,thepointsaredenotedwithoneortwoletters,andcanbealsofollowedbyanumber(asshownonFigure 1-1 ).Thelettersroughlyrepresentthelobelocation(withexceptionoflettersCandZ),whereasthenumbersserveforidentifyingthecorrespondinghemisphere.Morespecically,thepointslocatedonthelefthemisphereofthebrainarerepresentedbyoddnumbers(1,3,5,and7),andthesitesontherighthemispherearemarkedwithevennumbers(2,4,6,and8).Thesiteslocatedonthefrontal,temporal,parietalandoccipitallobesaredenotedbythecorrespondinginitials 17

PAGE 18

EEGsignalresemblesacollectionofsinusoidsofvariousamplitudeandfrequency.PowerspectrumisextractedfromtherawEEGdatausingFouriertransformtoobtaintheinformationaboutthecontributionofsinusoidalwavesofdierentfrequency.ThepowerspectrumofEEGiscontinuous,rangingfrom0Hzuptoahalfofthesamplingfrequency.Dependingonthestateofthebrain,certainfrequenciesappeartobemoreprevalent.Therearefourmajorfrequencybands,alpha,beta,deltaandtheta,whichpresenceinEEGduringvariousstatesofconsciousnesshasbeenextensivelystudied.Thesebandsrepresentsinewavesofrelativelylowfrequency,withdeltarangingfrom0.5to4Hz,theta4{8Hz,alpha8{13Hz,andbetaover13Hz.AlphawavesdiscoveredbyAdrianandMatthewsin1934arethebest-knownandthemoststudiedamongthefourfrequencybands[ 100 ].Theyareinducedbyclosingeyesandbyrelaxation,andterminatedwitheyesopeningorduetothinking,calculating,andotheranalyticalactivities.Inparticular,inmostpeople,eyeclosingproducesrapidchangesinbrainactivitymanifestingthemselvesinEEGasanadjustmentofthedominantfrequencybandfrombetatoalpha.EEGiscapableofdiscriminatingbetweendierentstates,suchasresting,alertness,stressstate,varioussleepstages,hypnosis,etc.Presenceofbetabandisdominantduringthestateofalertnesswitheyesopen.Drowsinessortherestingareusuallycharacterizedbytheriseinalphaactivity.Duringthesleep,presenceoflowerfrequencywavesbecomesmoreapparent.AhigherproportionofdeltabandfrequenciesisobservedduringstagesIIIandIVofthenon-rapideyemovementsleep(NREM).EEGrecordedfromdistinctregionsinthebrainexhibitsdierentspectrumofwavefrequencies.Inaddition,thebrainpatternsareuniqueforeveryindividual. PracticalapplicationsofEEGincludeepilepsyresearchandlocalizationofthefocusofepilepticseizures,testingofepilepsydrugeects;determiningareasofdamagedue 18

PAGE 19

3 ]. 94 ].Inordertoproperlyreectthespatio-temporalpropertiesofbraindynamics,theanalysisofEEGdatamustinvolveasimultaneousinvestigationofthedependenciesacrosschannelswithrespecttotime. DierentfeatureshavebeenproposedforanalysisofEEGtimeseries,includingFouriertransform,wavelets,cross-correlation,coherence,Grangercausalityandpartialdirectedcoherence,mutualinformationandtransferentropy,globalandphasesynchronization,Lyapunovexponentsandcorrelationdimension,etc. SinceEEGcanbeviewedasacollectionofsinewaves,EEGseriesareoftenanalyzedinafrequencydomain.Inaddition,somefrequencybandshaveshowntoplayspecicrolesinvariousstatesofconsciousness,andsothefrequencyinformationinEEGcanbeparticularlyimportant.Subsequently,theFouriertransformwitharunningtimewindow,alsoknownasshorttimeFouriertransform(STFT),becameoneofthemostwidelyusedmethodsforextractingfeaturesfromEEG.STFTisobtainedfromFouriertransformbyapplyingatimewindowfunctiongwithatimeshift.Mathematically,STFTisgivenbythefollowingformula: whereS(;f)denotestheSTFTwithtimewindowglocatedattime,correspondingtofrequencyf;andx(t)isasignalattimet.Inotherwords,STFTS(;f)representsthepowerspectrumofthesignalestimatedaroundtime.ThedrawbackofSTFTisthatthereisatradeobetweentimeaccuracyandfrequencyprecision.Bymakingthewindow 19

PAGE 20

AnalternativetoFouriertransformiswavelettransform(WT),whichisatransformationofthesignalbasedonaspecialfunction,calledmotherwavelet(MW).Themotherwaveletisshiftedintimebyalocationparameter,andthenadjustedbyascaleparametera.Moreprecisely,thewavelettransformisdenedbythefollowingformula: a)dt;(1{2) whereisamotherwavelet,aisscaleparameter,isatimelocationparameter,andxisasignal. ThescaleparameterainWTisanalogoustothefrequencyparameterfinSTFT.Inparticular,thelargevaluesofparametera(a1)stretchthewavelet,andsotheyrepresentlowfrequencies,whereasthesmallvaluesofa(a<1)shrinkthewaveletfunction,whichcorrespondstohigherfrequencies.Anadvantageofusingwaveletsisthatthehighfrequencycomponentscanbeanalyzedwithahighertimeaccuracythanthelowerfrequencycomponentsofthesignal. Asfollowsfrom( 1{2 ),W(;a)canbeinterpretedastheprojectionofthesignalontotheappropriatelyshiftedandscaledwavelet,i.e.W(;a)isacontributionofthewavelettothesignalx(t). WhileFourierandwavelettransformsareusuallyappliedtostudyeachchannelofEEGsignalindividually,thecross-correlation,coherenceandGrangercausalitymeasuretheinterdependencybetweendierentchannels.Thecross-correlationfunctionquantiesthelinearcorrelationbetweentwoprocesses.Giventwonormalizedsignalsx(t)andy(t)withzeromeansandunitvariances,thecross-correlationbetweenthesesignalsisestimatedas: 20

PAGE 21

Thecoherencefunctionisafrequencydomainanalogofthecross-correlationmeasure.Coherenceisobtainedfromcross-correlationbyapplyingFouriertransformto( 1{3 ).Theestimateofthecoherencespectrumoftwosignalsiscalledperiodogram.Theperiodogramiscalculatedbysubdividingthesignalsintoanumberofepochsofthesamelength,andthenapplyingthefollowingformula: where Whilethecross-correlationandcoherencearefeatures,whichreectthelineardependencybetweentwochannelsinthedata,theconceptofGrangercausalityiscapableofnotonlyestablishingthelineardependency,butalsospecifyingthedirectionofsuchdependency.Inotherwords,byapplyingGrangercausality,itbecomespossibletoidentifycausalrelationshipamongthechannelsofEEG.Grangercausalityisbasedonthemultivariateautoregressivemodelingoftimeseries.Ithasalsoreceivedanalternativereformulationinthefrequencydomainviaspectraldecompositionforstochasticprocesses. 21

PAGE 22

35 89 ].Inchaoticsystems,trajectoriesoriginatingfromverycloseinitialconditionsdivergeexponentially.Thesystemdynamicsarecharacterizedbytherateofthedivergenceofthetrajectories,whichismeasuredbyLyapunovexponentsanddynamicalphase. ShorttermlargestLyapunovexponent(denotedSTLmax),whichisanestimateofthemaximumLyapunovexponentfornon-stationarydata,isadynamicalmeasureofthechaoticityinthebrain.Next,themethodforestimatingSTLmaxissummarized. First,usingthemethodofdelays[ 69 ],theembeddingphasespaceisconstructedfromadatasegmentx(t)witht2[0;T],sothatthevectorXiofthephasespaceisgivenby whereti2[1;T(p1)],pisachosendimensionoftheembeddingphasespace,anddenotesthetimedelaybetweenthecomponentsofeachphasespacevector. Next,theestimateLoftheshorttermlargestLyapunovexponentSTLmaxiscomputedasfollows: whereNaisthetotalnumberoflocalmaximumLyapunovexponentsthatareestimatedduringthetimeinterval[0;T];4tistheevolutiontimeforthedisplacementvectorX(ti)X(tj);X(ti)representsthepointoftheducialtrajectorysuchthatt=ti,X(t0)=(x(t0);x(t0+);:::;x(t0+(p1))),andX(tj)isanappropriatelyselectedvectorthatisadjacenttoX(ti)intheembeddingphasespace.In[ 34 ],Iasemedisatel.suggestedamethodforestimatingSTLmaxintheEEGdatabasedontheWolf'salgorithmfortimeseries[ 107 ]. TheshorttermlargestLyapunovexponentSTLmaxisprovedtobeanespeciallyusefulEEGfeatureforstudyingthedynamicsoftheepilepticbrain[ 35 89 ].Inparticular, 22

PAGE 23

90 ]. 1.4.1TestingApplicabilityofFrequencyDomainEstimatesofGrangerCausalityforEEGtimeseries 25 ]foranalysisofeconometricseries.Becausefrequencydomaincontainsvaluableinformationaboutthebrainprocesses,Geweke'sdenitionofcausalityseemstobeparticularlyusefulforanalysisofEEGdata. BothdenitionsofGrangercausality,theoriginalonegivenbyGrangerandthefrequencydomaindenition,areintroducedviavectorautoregressivemodelingofmultipletimeseries.Precisely,thelinearvectorautoregressionisusedtotthedata.Basedonthismodel,twocompetinghypothesisaboutthedataareconsideredandtestedstatisticallytodetermine,whichofthesetwoassumptionsissupportedbythedata.Inotherwords,totestthecausality,thehypothesisofdatabeingmodeledaslinearautoregression(thatdoesnotincludeanotherseries)iscomparedtothealternativeofthedatabestdescribedbyincludinginformationfromtheotherseries. SinceGrangercausalityisdenedbasedonlinearvectorautoregressivemodel(VAR),theapplicabilityofGrangercausalitydependsupontheunderlyingassumptionsofthemodel.ThefundamentalassumptionsoftheVARarestability,stationarity,andgaussian 23

PAGE 24

AlthoughBernasconiandKonig[ 2 ]testedthestationarityofdierentEEGdata,andconcludedthatatimeintervalof1secondistheintervalonwhichtheEEGtimeseriescanbeconsideredstationary,theunderlyingassumptionsoftheVARmodelingsuchasstationarityandstabilityarerarelystatisticallytestedinapplicationstoEEGdata. Inparticular,thestabilityassumptionmeansthatthereversecharacteristicpolynomialofthemodeldoesnothaverootsinsideaunitcircle.InordertohighlighttheimportanceofthestabilityconditionforVAR,itisnecessarytopointoutthatwhenstabilityisviolated,themodelmaysimplyfollowarandomwalk,oritmayevenexhibitexplosivebehavior. Tothebestofourknowledge,thestabilityconditionoftheVARestimatedfromtheEEGtimeserieswasnotinvestigatedbeforeourstudy[ 43 ].InmanystudiesthatutilizethevectorautoregressiontoexamineGrangercausalityamongseries,thevericationoftheconditionsassumedbytheVARmodelisoftenomitted. TheresultsofournumericalexperimentsindicatethatthestabilityconditionofvectorautoregressivemodelisoftenviolatedinapplicationtotheEEGdata.Morespecically,wefoundthatthestabilityassumptionimposedonthelinearVARmodelsmaybeviolatedeveninthecasewhenthesamplesizeparameterTismuchlargerthanthelagparameterpoftheestimatedmodel.Inaddition,weshowedthatdespitethefactthatitiscommoninpracticetoltertheEEGdatawithinacertainfrequencyband,lteringtheEEGtimeserieswithinsomerestrictedfrequencybandoftenresultsinsignicantreductionofthe(T;p)domain,wheretheestimatedVAR(p)modelsremainstable. 24

PAGE 25

43 ]. ComprehensivestatisticaltestingisnecessaryinordertomakeconclusionsonwhatmultivariatemodelsarethemostappropriateforextractingthedirectionaldependenciesbetweenchannelsinafrequencydomainfrommultichannelEEGdata. 15 101 104 105 ].Actually,itwasfoundthatoscillationofvariousneuronalgroupsingivenfrequencybandsleadstotemporaryphase-lockingbetweensuchgroupsofneurons.Thisobservationhasstimulatedthedevelopmentofrobustapproachesthatallowonetomeasurethephase-synchronyinagivenfrequencybandfromexperimentallyrecordedbiomedicalsignalssuchasEEG. Inparticular,theimportanceofsynchronizationofneuronaldischargeshasbeenshownbyavarietyofanimalstudiesusingmicroelectroderecordingsofbrainactivity[ 83 96 ],andevenatcoarserlevelsofresolutionbyotherstudiesinanimalsandhumans[ 21 ]. ThephasesynchronizationinthebrainextractedfromEEGdatausingHilbertorwavelettransformshasrecentlybeenshowntobeanespeciallypromisingtoolinanalysisofEEGdatarecordedfrompatientswithvarioustypesofepilepsy[ 86 ]. Inourrecentstudy[ 44 ],weintroduceanovelconceptofgeneralizedphasesynchronization,whichisbasedonvectorautoregressivemodeling.Thisnewnotionofphasesynchronizationisconstructedasanextensionoftheclassicaldenitionofphasesynchronizationbetweentwosystems.Infact,thephasesynchronizationisusuallydenedastheconditionthatsomeintegercombinationoftheinstantaneousphasesoftwosignalsisconstant.Oftenthisconditionisrelaxedbyallowingforaboundedlinearcombinationoftwophases,inorder 25

PAGE 26

Toconstructamoregeneralmultivariateconceptofphasesynchronization,weextendedtheclassicaldenitionbyconsideringalinearcombinationofphasesforanitenumberofsignalsthatrepresentsastationaryprocess.Alltheindividualsignalstogetherformacommonsystemdescribedbysomemultivariateprocess.Wenotethatavectorprocess,suchthatalinearcombinationofitsindividualcomponentsisastationaryprocess,canbemodeledasacointegratedvectorautoregressivetimeseries. Furthermore,itiseasytosee(asshowninSection 4.2.1 )thatthecointegratedrankoftheregressiondetermineshowrestrictedthebehaviorofsuchsystemis.Thismeansthattherankrofcointegratedautoregressivemodel,estimatedfromthemultipletimeseriesoftheinstantaneousphases,measureshowlargethevectorsubspace,whichgeneratesthechangesinthephasevalues,is. ThisnewmeasureofcointegrationwasappliedtoabsenceepilepsyEEGdatain[ 44 ].Thedatasetscollectedfromthepatientswithothertypesofepilepsyarecurrentlybeinginvestigated. 26

PAGE 27

Bproleview TheInternational10-20systemforplacementofEEGelectrodes 27

PAGE 28

26 27 ]formeasuringlineardependenceandfeedbackineconomictimeseries.Later,thisideawasfurtherextendedbyJohnGeweke[ 25 ],whoproposedanequivalentmeasurebasedonthespectralrepresentationoftimeseries.BothGranger'sandGeweke'sapproachesemploythevectorautoregressivemodelingtoderiveestimatesofunderlyingcausalrelationsinthedata.However,thelatterapproachisfoundparticularlyusefulforanalysisofEEGtimeseries,sinceitinvestigatesthecausalrelationinthefrequencydomaininsteadofthetimedomainasintheformerapproach.Inparticular,thespectralmeasureofGrangercausalityproposedbyJohnGewekewasemployedonintracorticallocaleldpotentialsrecordedfrom8electrodesduringgo/no-gotrialsofcat'svisualresponses[ 2 ].Anotherstudy[ 57 ]utilizedasimilarmethodofdirectedtransferfunction(whichisequivalenttothespectralmeasureofGrangercausality)toexaminecausalinuencesintheprimatevisualcortexduringthetaskofvisualpatternrecognition.ThedirecttransferfunctionapproachtoGrangercausalitywasalsoappliedtoanalyzingbrainconnectivitypatternsonhumanEEGdatarecordedduringstage2sleep[ 42 ]. MichaelEichlerproposedagraphicalapproachformodelingGranger-causalrelationshipsinmultivariatetimeseries[ 17 ]andlaterappliedthismethodtostudyingconnectivityinneuralsystems[ 18 19 ].LuizBaccalaandKoichiSameshimaintroducedaconceptofthepartialdirectedcoherenceforinferenceofGrangercausalityinthefrequencydomainbasedonthelinearvectorautoregressivemodeling,andappliedittoinvestigatingthefunctionalinteractionsamongdierentbrainstructures[ 1 92 ]. 28

PAGE 29

30 ]. LetX(t)andY(t),t2Zdenotetwotimeseries(ordiscretetimestochasticprocesses)withthecorrespondingrealizationsxtandyt,t2Z.SupposethatX;tandY;tdenotealltheinformationabouttherealizationsofprocessesXandY,respectively,uptotimet.Then,therelationshipofGrangercausalitybetweensuchseriescanbeformallydenedasfollows: AtimeseriesX(t)issaidtoGranger-causeY(t)ifthereexistsp=1;2;:::suchthatthemeansquarederror(MSE)ofthep-stepforecastofY(t)basedontheinformationX;tandY;tissmallerthantheMSEofthep-stepforecastofY(t)basedonY;talone,i.e. whereY(pj)istheMSEofthep-stepforecastofY(t)basedoninformation. Usingtheabovedenition,wenowpresentthetestforGrangercausalitybasedonthebivariateautoregressivemodel.Supposethatforsomeintegerlagparameterp>0,therealizationsoftimeseriesY(t)aregivenbythemodel where"tisastandardwhitenoise(orinnovationprocess,i.e."thaszeromeanandzeroautocorrelation). 29

PAGE 30

againstthealternative Notethatifthenullhypothesisisaccepted,thenatimeseriesY(t)isbelievedtobeNOTGranger-causedbyX(t).Meanwhile,rejectingthenullhypothesis(i.e.acceptingthealternative)meansthatX(t)isbelievedtocauseY(t)inGranger'ssense. LetTbethesamplesizeparameter.ThemodelparametersforthenullhypothesisH0andtheparametersforthealternativeH1areestimatedfromthesampledatausingtheordinaryleastsquaresmethod(orothermethods)toobtaintheestimatesoftheforecasterrors^"0tand^"1t,respectively,t=1;2;:::;T.ThenthesumofsquaredresidualsRSS0undertheassumptionofnullhypothesisH0is andthesumofsquaredresidualsRSS1underthealternativeH1is ByconductingtheF-testofthenullhypothesis,onecanndtheteststatistic RSS1=(T2p1)Fp;T2p1:(2{7) IftheteststatisticS1exceedsthespeciedcriticalvalue,thenthenullhypothesisthatX(t)doesnotGranger-causesY(t)isrejected.Otherwise,H0isaccepted. AnasymptoticallyequivalenttestofGrangercausalityisgivenbythefollowingstatistic 30

PAGE 31

ThebivariateapproachtotestingGrangercausalitycanbenaturallyextendedtothemultivariatecasebypartitioningthevectorautoregressiveprocessZ(t)intotwocomponentsX(t)andY(t),sothatZ(t)=(X(t);Y(t)),andthentestingthesuitablezeroconstraintsonthecoecientsofvectorautoregression.ForthederivationoftheWaldstatisticandtheF-statisticfortestingGrangercausalityinthemultivariatecase,seethebookonmultipletimeseriesbyLutkepohl[ 59 ]. 2.3.1MethodsforVARParameterEstimation Letpdenoteapositiveinteger,andletytdenotetheK-variatetimeseries(i.e.realizationsofK-dimensionalprocessY(t)).Avectorautoregressivemodeloforderp,denotedVAR(p),isformallydenedasfollows: whereyt=(y1t;:::;yKt)0isa(K1)randomvector,=(1;:::;K)0isaxed(K1)vectorrepresentinganon-zeromeanEY(t),theAi,i=1;:::;parexed(KK)-dimensionalcoecientmatrices,and"t=("1t;:::;"Kt)0isaK-dimensionalwhitenoiseprocess(i.e.E["t]=0,E["s"0t]=0,fors6=t,andE["s"0t]=").Itisassumedthatthecovariancematrix"isnonsingular.Inaddition,threeimportantconditionsareusuallyimposedonthetimeseriesintheVARmodel.TherstconditionisstabilityoftheprocessY(t),thesecondisstationarityofY(t),whilethethirdonesupposesthattheunderlyingwhitenoiseprocess"tisGaussian. 31

PAGE 32

59 ].UndertheassumptionsofstabilityandGaussiandistribution,theseapproachesleadtoestimatorswiththesameasymptoticproperties.However,theasymptoticresultsshouldbeusedcautiouslyininferencefromsmallsamples.Asaresult,dierentapproachesmaysometimesleadtodierentresultswhenestimatingthemodelparametersusingsmallsamples. Letusnowbrieypresentthemultivariateleastsquaresestimation,whichisahigherdimensionalextensionofthewell-knownmethodofordinaryleastsquares.Formoredetaileddiscussion,referto[ 59 ]. Supposethattheavailabledatainclude(T+p)successiverealizationsofestimatedmultipletimeseriesrepresentedbyK-dimensionalvectorsyp+1;:::;y0;y1;:::;yT

PAGE 33

2{9 ),thevectorautoregressivemodeloforderpcanberepresentedinthecompactform: andthecoecientsBofthemodelaregivenbytheleastsquaresestimator: ^B=YZ0(ZZ0)1:(2{12) Thecovariancematrixcanbeestimatedinvariousways.Since"=E["t"0t],theestimator ~"=1 isconsistent.However,thisestimatorofthecovariancematrix"isnotunbiased.Therefore,itisoftenreplacedbythefollowingunbiasedestimator ^"=1 Obviously,bothestimatorsareconsistentestimatorsofthecovariancematrix,andtheyareasymptoticallyequivalent. Whenestimatingthecoecientsofthevectorautoregressivemodelfromdata,weassumedtheorderpoftheVAR(p)tobeknown.Inpractice,however,itisunknown,andtherefore,needstobederivedfromthedata.Sincezerocoecientmatricesareallowed,onecouldsimplysetptosomeupperboundontheVARorder.Ontheotherhand,selectinganunnecessarylargepwouldaecttheforecastprecisionoftheestimatedmodel.Therefore,itisadvantageoustoapplysomesuitablecriteriaforoptimalselectionofthelaglengthparameterp. 33

PAGE 34

59 ]. Let~u(m)denotethemaximumlikelihoodestimatorofucomputedbyttingtheVARmodeloforderm.TheFPEcriterionproposedbyAkaikein1969isbasedontheideathatminimizingthemeansquareerrorimprovestheforecastofthemodel.ForaVAR(p)timeseries,theFPEcriterionisdenedas UsingtheFPEcriterion,theestimate^pFPEofthemodelorderpisselectedsothat whereMdenotessomeupperboundaryonthemodelorder.Inotherwords,rst,foreachm=1;:::;M,thevectorautoregressivemodelofordermisestimatedfromthedata,andtherespectivevaluesoftheFPE(m)arecalculatedusing( 2{15 );thentheorderproducingthesmallestvalueofFPE(m)ischosenamongthepossibleordersm=1;:::;M. AICisanotherpopularorderselectioncriteriathatwasalsointroducedbyAkaike.GivenaVAR(m)model,theAkaikeinformationcriteriaisdenedasfollows: SimilarlytotheFPEcriterion,theVAR(m)modelsareestimatedfordierentm=1;:::;MtoobtainthecorrespondingAIC(m)valuesforeachorder.Thentheestimate^pAICofthemodelorderpwiththesmallestAIC(m)isselected. 34

PAGE 35

T+ln~u(m)(2{18) Asbefore,amongthemodelparametersm=1;:::;M,theparametermhavingthesmallestvalueofHQ(m)ischosenastheestimator^pHQofthetruemodelorderp. Last,butnotleast,wepresentSchwarzcriterion,whichwasderivedusingBayesianarguments.TheSCisformulatedas: T+ln~u(m);(2{19) andtheorderminimizingSC(m)ischosenamongm=1;:::;Mastheestimator^pSCofthemodelorderp. Someinterestingstatisticalpropertiesoftheabovecriteriaareprovedin[ 59 ].Inparticular,itisshownthatAICandFPEcriteriaforVARorderselectionareasymptoticallyequivalent,althoughtheseestimatorsofthemodelorderarenotconsistent.Ontheotherhand,theothertwocriteriaprovideconsistentestimatorsoftheorderparameterp.Moreprecisely,intheunivariatecase(K=1),theHannan-Quinncriterionisconsistent(i.e.limT!+1Prf^p=pg=1).Inaddition,theHQcriterionisstronglyconsistentforK2(i.e.PrflimT!+1^p=pg=1).TheSCisshowntobestronglyconsistentforanydimensionK. ItisimportanttokeepinmindthateventhoughFPEandAICdonotprovideconsistentestimators,theyarenotnecessarilyinferiortoHQandSC.Actually,insmallsamples,andeveninlargersamples,FPEandAICmayproducebetterforecast,althoughtheymaynotestimatethemodelordercorrectly. 2{9 )are 35

PAGE 36

det(IKA1z:::Apzp)6=0forcomplexz:jzj1:(2{20) Inotherwords,theVAR(p)process( 2{9 )satisesthestabilityconditionwhenitsreversecharacteristicpolynomial(givenbydet(IKA1z:::Apzp))hasnorootsonandinsidethecomplexunitcircle. Thestabilityconditionguaranteesthatthereexistsamovingaverage(MA)representationfortheVAR(p)process.Alsostabilityensuresthattheprocessisawell-denedstochasticprocesswiththedistributionsofitsunivariatecomponentsandjointdistributionoftheprocessytuniquelydeterminedbytheinnovationprocess"t.ForastableVAR(p)process,boththeprocessmeanandtheautocovariancearetime-invariant(which,accordingtothedenitionbelow,impliesstationarity). Whenthestabilityconditionisviolated,theprocessvarianceisincreasingwithtimeandunbounded.Specically,ifthereversecharacteristicpolynomialofthetimeserieshasasingleunitroot,andalltheotherrootsareoutsidethecomplexunitcircle,thenthetimeseriesbehaviorissimilartoarandomwalk.Inthisspecialcase,thevarianceincreaseslinearlywithtime,thecorrelationbetweenytandyt+happroaches1,andtheprocessmeanE[Y(t)]exhibitsalineartrendfor6=0.Inaddition,ifoneoftherootsofthereversecharacteristicpolynomialliesstrictlyinsidethecomplexunitcircle,thensuchprocessisexplosive,i.e.theprocessvariancegrowsexponentially.Variousapproachesaredevelopedinthetimeseriesliteraturetoaddressthetimeserieswiththeunitroots.Forexample,theunitrootscanberemovedbytakingdierences.However,theexplosivetimeseriesarenotaswell-studied,becauseitisbelievedthatanexponentialincreaseinthevarianceoftheeconomictimeseriesisnotwellfounded.AsonecanseethestabilityassumptionplaysanimportantroleinVAR(p). Awide-sensestationarityforstochasticprocessesisimposedontheVARtimeseriesasfollows.AstochasticprocessY(t)isconsideredstationaryif 36

PAGE 37

2. Inotherwords,thestationarityconditionsupposesthattherstandthesecondmomentsaretimeinvariant.Alsonotethattheprocessmeanandtheautocovariancematrixy(h)arenite.Itisshown(seeProposition2.1in[ 59 ])that AstableVAR(p)timeseriesyt,t=0;1;2;:::isstationary. Sincestabilityofatimeseriesimpliesthattheseriesisstationary,thestabilitycondition( 2{20 )issometimescitedintheliteratureasthestationaritycondition.However,itisimportanttorememberthatthesetwoconditionsarenotequivalent.Infact,althoughastablevectorautoregressiveseriesisalwaysstationary,theconverseisnottrue,i.e.anunstabletimeseriesisnotnecessarynon-stationary. TheGaussiandistributionassumptionisintroducedintotheVAR(p)modelthrough"t.Specically,givenrepresentation( 2{9 )oftheVAR(p),theinnovationprocess"tisassumedtobeGaussianwhitenoise.ThisconditionimpliesthatytisaGaussianprocess,i.e.anysubcollectionyt;:::;yt+hfollowsamultivariatenormaldistributionforallpossiblevaluesoftandh. RecallthattheVAR(p)process( 2{9 )satisesthestabilityconditionwhenitsreversecharacteristicpolynomialdet(IKA1z:::Apzp)hasnorootsonandinsideacomplexunitcircle.Ifanunstableprocesshasasingleunitrootandalltheotherrootsoutsideofthecomplexunitcircle,thensuchprocessexhibitsabehaviorsimilartothatofarandomwalk.Inotherwords,thevarianceofsuchprocessincreaseslinearlytoinnity, 37

PAGE 38

Thisrendersthefollowingdenitionofanintegratedprocess. Aone-dimensionalprocesswithdrootsontheunitcircleissaidtobeintegratedoforderd(denotedasI(d)). Itcanbeshown[ 59 ]thattheintegratedI(d)processY(t)oforderdwithallrootsofitsreversecharacteristicpolynomialbeingequalto1canbemadestablebydierencingtheoriginalprocessdtimes.Forexample,theintegratedI(1)processY(t)becomesstableaftertakingtherstdierences(1L)Y(t)=Y(t)Y(t1),whereLrepresentsthelagoperator.Moregenerally,fortheI(d)processY(t),itstransformation(1L)dY(t)isstable. AnexampleofanintegratedI(d)processintheunivariatecaseisanautoregressiveintegratedmovingaverageprocessARIMA(p,d,q),whichissometimescalledfractionallydierencedautoregressivemovingaverageprocessford2(0:5;0:5).Theone-dimensionalprocessY(t)issaidtobeARIMA(p,d,q),ifZ(t):=(1L)dY(t)isastationaryautoregressivemovingaverageARMA(p,q)process,i.e. where"tj'sareindependentnormallydistributedrandomvariableswithmean0andvariance2,andListhedierencingoperatorintroducedabove. Itisnoteworthytopointoutthattakingdierencesmaydistorttherelationshipamongthevariables(i.e.one-dimensionalcomponents)insomeVAR(p)models.Inparticular,thisisthecaseforsystemswithcointegratedvariables.ItturnsoutthatttingVAR(p)modelafterdierencingtheoriginalcointegratedprocessproducesinadequateresults.Next,wediscusssuchprocesses. 38

PAGE 39

SupposethatsampledvaluesyitofKdierentvariablesofinterestYi(t)arecombinedintotheK-dimensionalvectorsyt=(y1t;:::;yKt)0.Inaddition,supposethatthevariablesareinalong-runequilibriumrelation wherec=(c1;:::;cK)0isaK-dimensionalrealvector.Duringanygiventimeinterval,therelation( 2{22 )maynotnecessarilybesatisedpreciselybythesampleyt,insteadwemayhave: where"tisastochasticprocessthatdenotesthedeviationfromtheequilibriumrelationattimet.Ifourassumptionaboutthelong-runequilibriumamongindividualvariablesYi(t),i=1;:::;KisvalidthenitisreasonabletoexpectthatthevariablesYi(t)movetogether,i.e.thestochasticprocess"tisstable.Ontheotherhand,thisdoesnotcontradictthepossibilitythatthevariablesdeviatesubstantiallyasagroup.Therefore,itispossiblethatalthougheachindividualcomponentYi(t)isintegrated,thereisalinearcombinationofYi(t),i=1;:::;K,whichisstationary.Integratedprocesseswithsuchpropertyarecalledcointegrated. Withoutlossofgenerality,weassumethatallindividualone-dimensionalcomponentsYi(t)(i=1;:::;K)areeitherI(1)orI(0)processes.ThenthecombinedK-dimensionalVAR(p)process 39

PAGE 40

=IKA1:::Ap(2{25) hasrankr. Sincesomeone-dimensionalcomponentsofthecointegratedVAR(p)processareintegratedprocesses,onemaybeinterestedintestingthepresenceofaunitrootintheunivariateseries.Inthefollowingsection,wepresentacommonlyusedunitroottest,whichwasderivedbyDickeyandFuller[ 16 ]. 16 ],anditcanbeshownthatthisdistributionisthesamefork>1andfork=1.FullertabulatedtheapproximatecriticalvaluesfortheADFtestwithk1andpk1forspecicsamplesizes. Finite-samplecriticalvaluesfortheADFtestforanysamplesizewereobtainedbymeansofresponsesurfaceanalysisbyMacKinnon[ 60 ],whoalsoshowedthatanapproximateasymptoticdistributionfunctionforthetestcanbederivedviaresponsesurfaceestimationofquantiles[ 61 ]. AlthoughtheasymptoticdistributionoftheADFteststatisticdoesnotdependonthelagorder,itisnotedbyCheungetal.[ 13 ]thatempiricalapplicationsmustdealwithnitesamples,inwhichcasethedistributionoftheADFteststatisticcanbesensitiveto 40

PAGE 41

Aswenotedabove,thelimitingdistributionoftheADFteststatisticisthesamefork>1andk=1.Hence,forsimplicity,weconsiderthecaseofk=1.Infact,letYdenotetheautoregressiveAR(1)model whereY(0)=0,cisarealnumber,and"tN(0;2)(i.e."tisnormallydistributedwithzeromeanandvariance2forallt=1;2;:::). Notethatwhenjcj<1,theprocessY(t)convergestoastationaryprocessast!1;whereas,inthecaseofjcj=1,theprocessY(t)isnotstationarywithvariancet2.Furthermore,whenjcj>1,notonlytheprocessisnotstationary,butthevarianceofY(t)growsexponentiallywithtimet. FromtheAR(1)model( 2{26 ),onecanseethatinthecasewhenc=1,inordertomaketheprocessstationary,theseriescanbeappropriatelytransformedbydierencing.Furthermore,noticethattheconditionc=1in( 2{26 )isclearlyequivalenttotherequirementthatthereversecharacteristicpolynomialdet(1cz)=1zofAR(1)hasaunitroot.Inotherwords,todeterminewhetheranautoregressivetimeseriesAR(1)hasaunitroot,wemusttestthenullhypothesisH0:c=1. Lety1;y2;:::;yTdenoteasampleofTconsecutiveobservationsoftheAR(1)processY(t),thenthemaximumlikelihoodestimatorofcistheleastsquaresestimator Notethatbcisaconsistentestimatoroftheregressioncoecientc. Sinceeachyt,t=1;:::;TisarealizationofanAR(1)process,itfollowsfrom( 2{26 )thatyt=cyt1+"tholds,andsobypluggingthislastconditionintoEquation( 2{27 ),the 41

PAGE 42

=cPTt=1y2t1+PTt=1yt1"t 2{28 andmultiplyingeachsidebyTleadtotheADFstatistic DickeyandFuller[ 16 ]derivedthefollowingrepresentationofthelimitingdistributionforstatisticT(bcc): 21(W21);asT!1(2{30) where =1Xi=1d2iX2i; andrandomvariablesXi,i=1;2;:::,areindependentandidenticallydistributedaccordingtothenormaldistributionwithzeromeanandvariance2.Notethatasymbol)denotesconvergenceindistribution. In[ 16 ],DickeyandFullerconsideredthefollowing\Studentized"statisticbasedonthelikelihoodratiotestofthehypothesisH0:c=1. 2;(2{34) 42

PAGE 43

andbciscomputedfrom( 2{27 ) TablesofthecriticalvaluesfortheasymptoticdistributionsoftheADFteststatisticT(bc1)andthestatisticbcanbefoundinFuller[ 22 ].WesummarizesomeoftheinformationinTable 2-1 ,whichliststhep-valuesforasymptoticdistributionsofT(bc1)andbcorrespondingtopercentilesof90,95,and99percent. 75 ].Inparticular,inhisstudyPhillipsrstconsideredtwostatisticsZandZtfortestingthenullofnocointegrationintimeseries. Becausemanyunitroottests,constructedbefore1987,werefoundedontheassumptionthattheerrorsintheregressionareindependentwithcommonvariance(whichisrarelymetinpractice),Phillipswantedtorelaxtheratherstrictconditionthatthetimeseriesaredrivenbyindependentidenticallydistributedinnovations.Inotherwords,hewantedtodevelopthetestingproceduresbasedontheleastsquaresregressionestimationandtheassociatedregressiontstatistic,whichwouldallowforrathergeneralweaklydependentandheterogeneouslydistributedsequenceoferrorterms. ThepropertiesofasymptoticdistributionsofresidualbasedtestsforthepresenceofcointegrationinmultipletimeserieswerethoroughlyinvestigatedbyPhillipsandOuliaris[ 76 ].Thecharacteristicfeatureofthesetestsisthattheyutilizetheresidualscomputedfromregressionsamongtheunivariatecomponentsofmultivariateseries.TheresidualbasedproceduresdevelopedbyPhillipsandOuliarisaredesignedtotestthenullofnocointegrationbymeansoftestingthenullhypothesisoftheunitrootpresenceintheresidualsagainstthealternativeofarootthatliesinsidethecomplexunitcircle.ThehypothesisH0oftheabsenceofcointegrationisrejected,ifthenullofaunitrootinthe 43

PAGE 44

Asnotedin[ 76 ],theresidualbasedunitroottestsareasymptoticallysimilar,andcanberepresentedviathestandardBrownianmotion.Moreover,theADFandZttestsareprovedtobeasymptoticallyequivalent.However,thesetwotestsarenotaspowerfulasthetestbasedonstatisticZ,becauseitwasshownbyPhillipsandOuliaris[ 76 ]thattherateofdivergenceundercointegrationassumptionisslowerfortheADFandZtthanothertests,suchastheZ-statistictest.Thelatertest(i.e.thecointegrationtestbasedonZ)isalsowidelyknownasthePhillips-Ouliariscointegrationtest. ItisnoteworthythatthenullhypothesisforthePhillips-Ouliaristestisthatofnocointegration(insteadofcointegration).Thisformulationischosenbecauseofsomemajorpitfallsfoundinproceduresthataredesignedtotestthenullofcointegrationinmultipletimeseries.Thesedefects(discussedinmoredetailin[ 76 ])aresignicantenoughtobeastrongargumentagainsttheindiscriminateuseofthetestformulationsbasedonthenullofcointegration,andtosupportthecontinuinguseofresidualbasedunitroottests. ConsidertheK-dimensionalvectorautoregressiveprocessY(t).LetuspartitionY(t)=(Ut;V0t)0intotheunivariatecomponentUt=Y1(t)andthe(K1)-dimensionalVt=(Y2(t);:::;YK(t))0. Theresidualsaredeterminedbyttinglinearcointegratingregression ResidualbasedtestsareformulatedtotestthenullhypothesisthatthemultipletimeseriesY(t)arenotcointegratedusingthescalarunitroottests,suchastheADFtest,whichareappliedtotheresidualst,t=1;2;:::in( 2{36 ). In[ 76 ],theADFtestaswellastwoadditionaltestsZandZt,developedearlierbyPhillips[ 75 ],wereappliedtocheckforthepresenceofaunitrootintheresidualst.In 44

PAGE 45

ThenthestatisticZinPhillips-Ouliaristestisdenedasfollows: 2s2Tls2% whereastheZtstatisticisgivenbythefollowingformula: 2(b1) 2s2Tls2% 2;(2{39) where l+1: Notethats2%andsTlareconsistentestimatorsforthevariance2%of%tandthepartialsumvariance2=limT!1E1 2{37 ). ThecriticalvaluesforZandZtstatisticscanbefoundin[ 76 ](TablesIandII).PhillipsandOuliaristabulatedthevaluesforcointegratingregressionswithatmost5explanatoryvariables.SomeestimatesofthecriticalvaluesforthePhillips-Ouliaristest(Z)arelistedinTable 2-2 45

PAGE 46

2{24 ). Givenarealizationyt,t=1;2;:::,ofY(t),oneseekstodeterminethecoecientsofthefollowingmodel: subjecttotheconstraint rank()=rank(IKA1:::Ap)=r:(2{44) Notethat"tisassumedtobeaGaussianwhitenoisewithanonsingularcovariancematrix".Furthermore,theinitialconditionsyp+1;:::;y0aresupposedtobexed. Inordertoimposethecointegrationconstraint,themodel( 2{43 )isreparameterizedinthefollowingfashion[ 59 ]: yt=D1yt1+:::+Dp1ytp+1+ytp+"t;t=1;2;:::;(2{45) whereyt=ytyt1,andmatrixcanberepresentedasaproduct=HCofmatricesofrankr,i.e.His(Kr)andCis(rK). Consider Y:=[y1;:::;yT];Xt:=266664yt...ytp+2377775; X:=[X0;:::;XT1];D:=[D1;:::;Dp1];Yp:=[y1p;:::;yTp]:

PAGE 47

lnl=KT 2trace(YDX+HCYp)01"(YDX+HCYp): TheproofofthefollowingtheoremonthemaximumlikelihoodestimatorsofacointegratedVARprocesscanbefoundin[ 59 ](Proposition11.1). 59 ]) Dene Denote1:::KtobetheeigenvaluesofGS10S100S01G0, and Thenthelog-likelihoodfunctionin( 2{47 )ismaximizedfor

PAGE 48

1 ,onecaneasilyderivethelikelihoodratiostatisticfortestingacandidatevaluer0ofthecointegrationrankrofaVAR(p)processagainstalargercointegrationrankr1. GivenaVAR(p)processy(t)denedby( 2{24 ),supposewewishtotestahypothesisH0againstanalternativeH1,where Underassumptionthatthenoise"tisaGaussianprocess,themaximumofthelikelihoodfunctionforacointegratedVAR(p)modelwithcointegrationrankriscomputedinTheorem 1 .Fromthatresult,thevalueoftheLRstatisticfortesting( 2{49 )canbedeterminedinthefollowingmanner: (2{50) =T"r1Xi=1ln(1i)+r0Xi=1ln(1i)#=Tr1Xi=r0+1ln(1i); 2{50 )isnonstandard.Specically,theLRstatisticisnotasymptoticallydistributedaccordingto2-distribution.Nevertheless,theasymptoticdistributionofthecointegrationrankteststatisticLRdependsonlyontwofactors: 48

PAGE 49

Asaresult,theselectedpercentagepointsoftheasymptoticdistributionoftheteststatisticLRweretabulatedbyJohansenandJuseliusin[ 38 ].ThepercentagepointsofasymptoticdistributionofLRaregiveninTables 2-3 and 2-4 Table2-1. CriticalvaluesoftheasymptoticdistributionsoftheT(bc1)andbforperformingunitrootcheckbytheADFtest(reproducedfrom[ 22 ]) Statistic90%95%99% Table2-2. CriticalvaluesoftheasymptoticdistributionsoftheZstatisticfortestingthenullofnocointegration(Phillips-Ouliarisdemeaned,reproducedfrom[ 76 ]).Parametern(n=K1)representsthenumberofexplanatoryvariables 1-17.0390-20.4935-28.32182-22.1948-26.0943-34.16863-27.5846-32.0615-41.13484-32.7382-37.1508-47.51185-37.0074-41.9388-52.1723 Table2-3. PercentagepointsoftheasymptoticdistributionsoftheLR(r;K)fortestingthecointegrationrank(reproducedfrom[ 38 ]) 16.698.0811.58215.5817.8421.96328.4431.2637.29 49

PAGE 50

PercentagepointsoftheasymptoticdistributionsoftheLR(r;r+1)fortestingthecointegrationrank(reproducedfrom[ 38 ]) 16.698.0811.58212.7814.6018.78318.9621.2826.15 50

PAGE 51

Inthischapterweintroduceaconceptofphasesynchronization,andconsidertwomethodsforestimatingthephaseofasignal,specicallyusingtheHilberttransformandviathewavelettransform. 78 ].SynchronizationphenomenawerediscoveredintheseventeenthcenturybyC.Huygenswhorstobservedsynchronizationbetweentwopendulumclockshangingfromacommonsupport[ 33 ].Sincethen,thestudyofsynchronizationbetweendynamicalsystemsbecameanactiveeldofresearchinmanyscienticandtechnicaldisciplines,includingsolidstatephysics[ 74 ],plasmaphysics[ 84 ],communication[ 11 48 ],electronics[ 72 77 ],laserdynamics[ 20 87 98 ],andcontrol[ 80 88 ]. Synchronizationphenomenacanalsobefoundinphysiologicalsystems,suchasheartandbrain.SynchronizationprocessesinphysiologicalsystemswerediscoveredbyB.vanderPolinthebeginningofthetwentiethcentury.Inparticular,vanderPolwasthersttoapplyoscillationtheorytothehumanheart[ 103 ]. Oneoftheimportantresearchareasinneuroscienceexplorestheroleofsynchronizationinneuraldynamics.Mucheortisgiventoinvestigationofsynchronizationphenomenaonalldierentlevelsoforganizationofbraintissue,startingwithpairsofindividualneuronstolargerscales,suchaswithinagivenareaofthebrainorbetweendistinctpartsofthebrain. 51

PAGE 52

15 101 104 105 ].Neuralsynchronyplaysavitalroleinsuchlarge-scaleintegration.Infact,variousneuronalgroupsoscillateinspecicfrequencybandsandbecomephase-lockedoveralimitedperiodoftime.Thisobservationhasstimulatedthedevelopmentofrobustapproachesthatallowonetomeasurethephase-synchronyinagivenfrequencybandfromexperimentallyrecordedbiomedicalsignals,suchasEEG. Inparticular,theimportanceofsynchronizationofneuronaldischargeshasbeenshownbyavarietyofanimalstudiesusingmicroelectroderecordingsofbrainactivity[ 83 96 ].Thendingsinthemicroelectrode-recordingstudiesarealsosupportedatcoarserlevelsofresolutionbyotherstudiesinanimalsandhumans[ 21 ]. Anelectrophysiologicalsignalisrecordedviaalowimpedanceextracellularmicroelectrodebyplacingthemicroelectrodesucientlyfarfromindividuallocalneuronsinordertopreventanyparticularcellfromdominatingthesignal.Next,toobtainthelocaleldpotential(LFP),thesignalislow-passltered,withacutoatapproximately300Hz.Duetothelowimpedanceandpositioningofthemicroelectrode,therecordedsignalispredominantlyinducedbytheactivityofalargenumberofneurons.Theunlteredsignalreectsthesumofactionpotentialsfromcellswithinapproximately50{350micrometersfromthetipoftheelectrode[ 53 ]andslowerioniceventsfromwithin0.5{3millimetersfromthetipoftheelectrode[ 39 ].Thespikecomponentofthesignalisremovedbylow-passlter,whereasthelowerfrequencysignal,theLFP,ispreservedinthesignal.Itisassumedthatthelocaleldpotentialcharacterizesthesynchronizedinputintotheobservedarea,inthecontrasttothespikedata,whichrepresentstheoutputfromthearea. Localeldpotentials(LFPs)ofvariousdegreesofspatialresolutioncanberecordedbyscalpEEGorMEG.Infact,studieshaveshownthatthepresenceofgammaandbeta 52

PAGE 53

97 ]andinsubduralelectrocorticograms[ 50 54 ].Inaddition,somerecentndingssuggestthatlong-rangesynchronizationanalogoustotheonefoundinmicroelectrodestudiesinanimalscanalsobedetectedbetweensurfacerecordings[ 82 ]. Ithasbeenshownthatsynchronizationisasignicantattributeofthesignalrecordedfromthepatientsaectedbyseveralneurologicaldisorders.Inparticular,researchershavefoundthatepilepsy[ 65 ]andParkinsonsdisease[ 99 ]manifestasapathologicalformofthesynchronizationprocess. Itisnotedin[ 56 ]thatalthoughthecross-correlogramsbetweenspikedischargesmaybeadequateformicroelectrodestudies,thequanticationofphasesynchronybetweenmeso-ormacro-electrodes(i.e.EEG/MEG,intracranialrecordings)callsforentirelydierentmethods.Therefore,theyemphasizeanimportanceofclearlydistinguishingbetweensynchronyasanappropriateestimateofphaserelation,andtheclassicalmeasuresofcoherenceorspectralcovariancethathavebeenextensivelyusedinneuroscience[ 8 10 62 ].LeVanQuyenetal.discusstwoimportantlimitationsofcoherence[ 56 ]. Therstlimitationarisesbecausethestandardapproachesformeasuringcoherence[ 12 ]basedonFourieranalysisareknowntobehighlydependentonthestationarityofthemeasuredsignal,whereasthesignalsrecordedfromthebrain,suchasEEG,appeartobeclearlynon-stationary.Applyingthetimefrequencyestimationmethod,whichisnotfoundedontheassumptionofstationarity,couldimprovethislimitationtowardsestimatingastable,instantaneouscoherenceaswellassynchronybetweentwoconcurrentbrainsignals. Thesecondlimitationstemsfromthefactthatclassicalcoherenceisameasureofspectralcovariance.Hence,itisnotabletoseparatetheeectsofamplitudeandphaseintherelationsbetweentwosignals.Becauseweareconcernedwithexaminingthespecichypothesisthatphase-lockingsynchronyisthepertinentbiologicalmechanismoftransientintegrationinthebrain,coherenceservesonlyasanindirectmeasure.Inorder 53

PAGE 54

86 ].Inotherwords,eventhoughourdiscussionofthephasesynchronyisfocusedonEEGdata,itsapplicationscanalsobeextendedtotheeldsotherthanneuroscience. Classicalconceptofthesynchronizationoftwooscillatorsisdescribedasanactiveadjustmentoftheirrhythmicitythatmanifestsinphase-lockingbetweenthesynchronizedoscillators.Specically,giventwosignalsX1(t)andX2(t),andtheircorrespondinginstantaneousphases1(t)and2(t),thebasicdenitionofthephaselockingstatesthat whereintegersnandmspecifythephaselockingratio. Wheninvestigatingphasesynchronyinneurophysiologicalsignals,onemustassumethattheconstantphaselockingratioisvalidwithinalimitedtimeintervalT,whichusuallymeansafewhundredsofmilliseconds.Whenexaminingneuralsignals,onemustkeepinmindthatdiscoveringthepresenceofthephaselockingbetweenEEGrecordingsfromtwodistantpartsofthebrainisnotstraightforward.Thedetectionofphasesynchronyinneuralsignalsisproblematicbecauseofseveralfactorsparticularlywhenworkingnotonthelevelofasingleneuron,butratherwithlargeneuronalpopulations,whoseactivityispickedupbymacroscopicormeso-electrodes.Asnotedin[ 56 ],asaconsequenceofvolumeconductioneectsinbraintissues,theactivityofasingleneuronalpopulationcanberecordedbytwodistantelectrodes,whichresultsinspuriousphase-lockingbetweentheirsignals.Furthermore,innon-invasiveEEG,thetruesynchroniesarehiddeninasignicantbackgroundnoise,andso,inthesynchronous 54

PAGE 55

3{1 )mustbeadjustedtoaccountforthenoiseasfollows: whereCdenotesapositiveconstant. Thefollowingtwokeystepsareinstrumentalininvestigatingthephasesynchrony: 1. estimateinstantaneousphaseofeachsignal; 2. provideastatisticalcriteriatoquantifythedegreeofphase-locking. Twomethodsfordetectingphase-lockingappliedtoneuronalsignalshaverecentlybeenconsideredintheliterature.Tassandcolleagues[ 99 ]extractedtheinstantaneousphasesfromoriginalsignalsbymeansoftheHilberttransform,andthenappliedtomagnetoencephalographic(MEG)motorsignalsinpatientsaectedbyParkinson'sdesease[ 99 ],andalsotothesynchronizationbetweencardiovascularandrespiratoryrhythms[ 93 ].Ontheotherhand,Lachauxetal.[ 50 ]estimatedthephasesfromtheoriginalsignalsbymeansofconvolutionwithacomplexwavelet,andthenappliedittoEEGandintracranialdataduringcognitivetasks[ 51 82 ]. TherststepinquantifyingphasesynchronizationbetweentwotimeseriesXandYisthedeterminationoftheirinstantaneousphasesX(t)andY(t).ThisisachievedeitherviatheHilberttransformorviathewavelettransform.Theseapproachesarepresentedinthenexttwosections. 23 ]andlaterextendedformodelsystemsandexperimentaldata[ 70 86 ]. 55

PAGE 56

^f(t)=CPVZ+1f()g(t)d=CPVZ+1g()h(t)d;(3{3) whereg(t):=1 Noticethat^f(t)canbeviewedasaconvolutiong(t)f(t)oftheoriginalfunctionf(t)withthefunctiong(t).ThismeansthattheHilberttransformcanbeperformedbyapplyinganideallter,whoseamplituderesponseequalsto1,andphaseresponseisaconstant=2lagatallfrequencies. Givenanarbitrarycontinuousreal-valuedtimeseriesX(t),thecorrespondinganalyticsignalisdenedasthefollowingcomplex-valuedfunction: wheretdenotestime,{isaunitonthecomplexaxis,^X(t)denotestheHilberttransformofthetimeseriesX(t),aX(t)isthecorrespondinginstantaneousamplitude,andX(t)representstheinstantaneousphaseofthesignalviaHilbertconvolution. ItfollowsfromEquation 3{4 thattheinstantaneousphaseX(t)ofX(t)canbecomputedas: Akeyadvantageoftheanalyticapproachisthatthephasecanbeeasilycomputedforanarbitrarybroad-bandsignal.Ontheotherhand,instantaneousamplitudeandphasehaveaclearphysicalmeaningonlyifX(t)isanarrow-bandsignal.Therefore,ltrationisrequiredinordertoseparatethefrequencybandofinterestfromthebackgroundbrainactivity. 56

PAGE 57

50 52 ],andissimilartotheHilberttransformmethodpresentedabove.Intheirapproach,Lachauxetal.extracttheinstantaneousphasebyapplyingtheconvolutionoftheoriginalsignalwithacomplexMorletwavelet.TheyconsidertheMorletwavelet(alsoknownasGaborfunction)attimetandfrequency!givenbythefollowingformula: Noticethatt;!()istheproductofasinusoidalwaveatfrequency!,andaGaussianfunctioncenteredattimet,withastandarddeviationproportionaltotheinverseof!.Itdependssolelyon,whichsetsthenumberofcyclesofthewaveletto6!. Accordingto[ 56 ],giventhetimeseriesX(t),thecoecientoftheMorlettransformasafunctionoftimetandfrequency!isdenedasfollows: where ThefollowingslightmodicationoftheMorletwaveletisintroducedin[ 81 ]: whereparameters!0andrepresentthecenterfrequencyandtherateofdecayofthewaveletfunction,respectively.Thisisproportionaltothenumberofcyclesandrelatedtothefrequencyspanbytheuncertaintyprinciple. Similarlytotheabove,acomplextimeseriesofwaveletcoecientsisobtainedviatheconvolutionofX(t)with(t): 57

PAGE 58

AsinthecaseoftheHilberttransform,thephasescanbedeterminedfromEquation( 3{9 )as where<[WX(t)]and=[WX(t)]denotetherealandimaginarypartsofthecomplextransformedtimeseriesWX(t),respectively. 56 ]andalsoexplainedtheoreticallybyQuianQuirogaandcolleagues[ 81 ]. Inanutshell,thephaseeX(t)extractedfromthesignalusingthewavelettransformcorrespondsapproximatelytothephaseX(t)determinedviatheHilbertconvolution,whichwouldbeperformedafterbandpasslteringthetimeseries.Furthermore,ifthephaseestimationbasedonwavelettransformwereperformedbyaconvolutionwithananalyticwavelet,andifthiswaveletwereappliedtodothebandpasslteringintheHilbertapproach,thensuchapproacheswould,infact,beequivalent. Itiseasytoseethatinthemethodbasedonthewaveletconvolution,thecenterfrequency!andtherateofdecayofthewaveletcanserveasparametersthatallowustomodifythefrequencyrangeofinterest.Ontheotherhand,theactualphaseextractionviatheHilberttransformisfreeofparameters,andsothecorrespondentphasepreservesinformationfromtheentirepowerspectrumandnotjustthemainfrequencybandasinthecaseofthewaveletconvolution.Asaresult,itispossibletoachieveacomparisonofnarrowbandandbroadbandsynchronizationsimplybyusingbothmethodsofphaseextractionwithoutperforminganyadditionalltering. 58

PAGE 59

49 56 ].However,mostofthecurrentlyusedmeasuresofphasesynchronizationarebasedonbivariateindexes. InChapter 4 ,weproposeanovelmultivariateapproachtodetectingphasesynchronizationinthephasesextractedfrommultipletimeseries,suchasmultichannelEEG. 59

PAGE 60

2 ]appliedstatisticaltestingtoverifythestationarityofthedata,andestablishedthedurationofthestationaryintervalforEEGtobeapproximately1second.ThestatisticaltestingofunderlyingassumptionsoftheVAR(whichwasthoroughlydiscussedinChapter 4 )isoftenomittedinthelaterstudies.Asshownabove,thestabilityconditionisaveryimportantassumptionofvectorautoregression.Inthisstudy,weexaminehowtheparametersofthemodelorderandsamplesizeinuencethestabilityofthederivedVARmodel. InordertoestimateVAR(p)modelfromdataandinvestigatethepropertiesofthederivedmodel,therodentintracranialEEGdatawereused.Thedatasetconsistedoftheelectroencephalographicrecordingsfrom6electrodes(implantedinleftfrontal,rightfrontal,twolefthippocampalandtworighthippocampalpartsoftherodent'sbrain)sampledat200Hz. ToexaminetheapplicabilityofthevectorautoregressivemodelingtoEEGdata,weestimatedtheVAR(p)modelparametersfordierentvaluesoflagorderpanddierentsamplesizesT.ThesamplesizesT2[1;300]wereused,andthelaglengthparameterpvariedbetween1and30.Inaddition,weltereddatausingaRectangularbandpassHammingwindowwith100coecientsintothefrequencybandsof0{30Hz,30{60Hz,60{90Hz,90{120Hz,and120{150Hz.Therawdataandthevedierentlyltereddatarepresentedseparatedatasets.Foreachdataset,weranthemodelestimationprocedurewithT=1;2;:::;300andp=1;2;:::;30.TheprocedureforestimatingcoecientsofthemodelwasimplementedintheMATLABenvironmentbasedonthemultivariateleastsquaresmethodpresentedabove.ForeveryVARmodelderivedfrom 60

PAGE 61

4-1 and 4-2 ThesurfaceplotinFigure 4-1 supportsthefact[ 59 ]thatthesamplesizeparametershouldsignicantlyexceedthelaglengthp,i.e.Tp.Ontheotherhand,itcanbeseenfromFigure 4-1 thatevenforTp,thestabilityassumptionsmaystillbeviolated.Indeed,forT=132considerablylargerthanp=2,theestimatedVARhastwoRCProotsonorinsidetheunitcircle(n=2). Figure 4-2 showsthatfortheltereddata,the(T;p)region,wherethestabilityconditionoftheestimatedVAR(p)modelisviolated,coversalmostthewholedomain.Whereasthe(T;p)regionthatcorrespondstostableVAR(p)modelsismuchsmallerthanthestableregioninFigure 4-1 ,andcharacterizedbylargeTandverysmallp.VerysimilarresultswereobtainedforthenumberoftheRCProotsinsidetheunitcircle,whenestimatingparametersofVAR(p)withdierentpusingltersinthe30{60Hz,60{90Hz,90{120Hz,and120{150Hzbands. Theexperimentwasrepeatedwithconsistentresultsonvarioussamplesfromthesixdatasets.TheresultsofourexperimentsclearlyshowthatthestabilityconditionimposedonthelinearVARmodelisoftenviolatedevenfortheparametersTp.Furthermore,lteringthedatawithinsomerestrictedfrequencybandoftenleadstoreductionofthe(T;p)domainwheretheestimatedVAR(p)modelsremainstable.Inpractice,theEEGdataareoftenlteredwithinacertainfrequencyband.InmanystudiesthatutilizethevectorautoregressiontoinvestigateGrangercausalityinthebiomedicaltimeseries,thevericationoftheconditionsassumedbytheVARmodelisoftenomitted.Fromourpoint 61

PAGE 62

Variousmodicationsofvectorautoregression(whichrelaxthestabilityconditionofnorootsinsideunitcircle)aredevelopedforanalysisofeconomictimeseries.AlthoughBernasconiandKonig[ 2 ]examinedthestationarityofdierentEEGdata,andconcludedthatatimeintervalofapproximatelyonesecondistheintervalduringwhichtheEEGtimeseriescanbeconsideredstationary,theunderlyingassumptionsoftheVARmodelingsuchasstationarityandstabilityarerarelystatisticallytestedinapplicationstoEEGdata.TheresultsofourexperimentsindicatethatthestabilityassumptionsofvectorautoregressivemodelareoftenviolatedinapplicationtotheEEGdata.Thisobservationsuggeststhatsuitableextensionsofmultivariateautoregressiontounstableprocessesmaytthedatabetter,andasaresult,suchextensionsofVARmaybemoreappropriatefortheEEGtimeseriesanalysisthanthestandardlinearvectorautoregressivemodeling. AdditionalstatisticaltestingisrequiredinordertomakeconclusionsonwhatmultivariatemodelsarethemostsuitableforextractingthedirectionaldependenciesbetweenchannelsinafrequencydomainfrommultichannelEEGdata. GiventhesignalrepresentedformallyasamultipletimeseriesX(t),onecanextracttheinstantaneousphasesXi(t)fromeachone-dimensionalcomponentXi(t)ofthesignalasshowninChapter3(eitherviaaconvolutionwiththeMorletwaveletorbyapplyingtheHilberttransform).ThephaseextractionprocedureproducesanewmultipletimeseriesX(t)ofthecorrespondentphases. Next,wederivenewmeasuresofphasesynchronyofthesignalbasedontheconceptsintroducedinChapter 3 .Letusobservethattheleft-handsideofEquation( 3{1 ) 62

PAGE 63

3{1 ),whichdenesphase-lockingbetweentwosignalsX1(t)andX2(t),needstobemodiedinpracticetoaccountforthenoiseinthesignal.Takingintoaccountpresenceofthestochasticnoiseinthephaseseries,letusintroduceamodiedconceptofthephasesynchronybetweentwosignalsbyrelaxingtheintegralityconditiononthecoecientsinthelinearcombinationasfollows. TwosignalsX1(t)andX2(t)areconsideredtobegenerallyphase-synchronized,ifthecorrespondentinstantaneousphasesX1(t)andX2(t)satisfytheconditionbelow: whereztN(C;0)isastochasticvariablethatrepresentsthedeviationfromtheconstantlevelCasaresultofthenoise.Noticethatinthecontrasttocondition( 3{1 )intheclassicdenitionofphasesynchronization,thecoecientsc1andc2inthedenition( 4{1 )ofgeneralizedphase-synchronydonotneedtobeinteger. Furthermore,itisstraightforwardthatthenewcondition( 4{1 )meansthatatwo-dimensionalprocessX(t)=(X1(t);X2(t))0iscointegrated.Basedonthisobservation,wecanextendourmodiedconceptofphasesynchronizationbetweentwosignalstothemultivariatecaseinthefollowingmanner. ThemultichannelsignalX(t)=(X1(t);:::;XK(t))isconsideredtobephase-synchronizedofrankr,iftheprocessX(t)composedofthecorrespondentinstantaneousphasesXi(t),i=1;:::;Kiscointegratedofrankr. Inthesubsequentsubsections,werstdiscusstheroleofthecointegrationrankintheframeworkofmultivariatephase-synchronization,andthenapplythisapproachtomultichannelEEGdatacollectedfromthepatientswithabsenceepilepsy. 63

PAGE 64

66 ],MichaelMurrayusedanexampleofdrunkardandherdogtoillustratetheconceptofthecointegration.Toexplainourreasoningbehindtherankofcointegrationasameasureofsynchrony,webrieysummarizeandthenfurtherextendhisanalogy. Thenonstationaryprocesses(suchasarandomwalk)areoftenintroducedbyteachersofstatisticsbycomparingthem(it)withthedrunkard'swalk.Thedrunkardwondersaimlessly,sothatthedirectionofeachstepisrandomandcompletelyindependentofherprevioussteps.Inotherwords,themeanderingofthedrunkardisdescribedbyarandomwalk: wherextrepresentsthepositionofthedrunkattimet,and"tisastationarywhite-noise,whichmodelsthedrunk'sstepattimet). AsMurraynoticed[ 66 ],anunleashedpuppyisanothercreature,whosebehaviorremindsarandomwalk.Indeed,eachnewscentthatpuppy'snosecomesupondictatesadirectionforthepup'snextstepsostronglythatthelastscentalongwithitsdirectionisforgottenassoonasthenewscentappears.Havingshownthatthepuppiesfollowtherandomwalkyt,t=1;2;:::,letusrepresentthepuppy'swalkas: wheretisastationarywhite-noise(i.e.puppy'sstepattimet). Thewell-knownfeatureofarandomwalkisthatthebestpredictorofthefuturevalueisthemostrecentlyobservedone.Inotherwords,thelongerithasbeensincewehadseenthedrunk,orthedog,thefurtherawayfromtheinitialplace,onaverage,they 64

PAGE 65

However,ifthepuppybelongstothedrunkard,thentheywillremainrelativelyclosetoeachotheratallthetime,similarlytotheindividualintegratedprocessesthattogetherformacointegratedprocess.Indeed,thedrunkwouldstillwonderaimlesslyinarandomwalkfashion,aswouldherpuppy.However,fromtimetotimeshewouldrememberaboutherdogandcallforit,thepuppywouldrecognizehervoice,andbark.Theywouldheareachotherandmaketheirnextstepineachother'sdirection. Thepathsofthedrunkandherdogarestillnonstationary,buttheyarenolongerindependentfromeachother.Asamatteroffact,ateachtime,thepuppyanditsmasterarelikelytobefoundnotfarfromeachother.Ifthisistrue,thenthedistancebetweentwopathsisstationary,andthewalksofthedrunkxtandherdogytaresaidtobecointegrated,i.e.xtandytareintegratedI(1),andthereisalinearcombinationofxtandyt(withnon-zeroweights)thatisI(0),i.e.stationary. Mathematically,thecointegratingrelationshipbetweenaladyandherpuppycanbewrittenas: attimet=1;2;:::.Notethat"tandt,asbefore,representthestationarywhite-noisestepsofthedrunkandherdog. SinceEquation 4{4 canbeeasilyrewritteninformof( 2{45 )asfollows: 264xtyt375=264"tt375264ccdd375264xt1yt1375; 65

PAGE 66

=264ccdd375; Notethatrank()=0,ifandonlyifc=d=0.Insuchcase,( 4{4 )becomessimplyasystemofequations( 4{2 )and( 4{3 ),whichmodelstwoindependentrandomwalksdrivenbyindependentwhitenoiseprocesses"and.Ontheotherhand,whenatleastoneofthecoecientscanddisnon-zero,thenbymultiplyingsystem( 4{6 )byavector[d;c]0,wehave: whichmeansthatthemodelisdrivenbyasinglecommonstochastictrendd"t+ct. AlthoughtheexampledescribedbyMurrayisclearlyabivariatecointegratedVAR(1),itcanbeextendedtoanillustrationofthemultivariatecointegratedprocess.Consider,forexample,aheardofsheepguardedbytwodogs,wherethesheepwonderaimlesslyintheeld,whilethedogsrunaroundandbringthesheepthathavestrayedtoofarbackintotheock.Say,forexample,afasterdogguardssheepfromtheeast,south,andwest,whereasaslowerdog-fromthenorth,thenthecointegratedprocessappeartohavethecointegrationrankof2.Clearly,twodogsareabletokeepaockofsheepclosertogether,thanasingledogcan.Inotherwords,thehighercointegrationrankthemorerestrictiveitis. Infact,letusconsideraK-dimensionalcointegratedvectorautoregressiveprocess,andletrdenotethecointegrationrankoftheprocess.Similarlytothebivariateexampleabove,wecanseethatwhentherankiszero(r=0),theunivariatecomponentsoftheprocessareindependent,andthemodelisdrivenbyKindependentwhitenoiseprocesses(i.e.thereisnocointegration).Inthecaseofr=1,wecandecomposethemultivariate 66

PAGE 67

Therefore,thesmalleristhecointegrationrankr,thelargeristhenumberKroftheunderlyingindependentstochastictrends,andso(thelarger)isthevectorspaceinwhichourcointegratedmodelcantravel.Andtheotherwayaround,increasingthecointegrationrankofthemodelshrinkstheunderlyingdomainoftheprocess,i.e.makesitboundedtoasmallerhyperplane.Forr=K,theVAR(p)isastableprocess,whichclearlyhasthemostconstraineddomain.Forr=0,theVARprocessisnotcointegratedandunrestricted. Thus,intheframeworkofgeneralizedphase-synchronizationintroducedabove,thecointegrationrankrepresentsafundamentalmeasureofsynchronyinthemulti-channelsignal,suchasEEG.Inparticular,wesaythatthesignaliscompletelyasynchronous,ifthecointegrationrankriszero.Ontheotherhand,whenthemultivariateprocessisstable(i.e.therankcoincideswiththedimensionoftheprocess,r=K),thesignalissaidtobeperfectlysynchronous. 67

PAGE 68

73 ].Thesebriefseizurescanhappenseveraltimesduringaday,butinsomepatients,thefrequencyofabsenceseizurescanbeashighashundredoftimesaday,whichinterfereswiththedailyactivitiesofachildsuchasschool.Insomecasesofchildhoodabsenceepilepsy,theseizuresstopwhenachildreachespuberty.Absenceseizuresexhibitacharacteristicspike-and-waveEEGpatternata3Hzfrequency[ 73 ]. Figure 4-3 displaysamultichannelEEGrecordingthatincludesanabsenceseizure.Thedurationoftheseizureisapproximately4seconds.Thegurevividlyillustratesacharacteristicspike-and-waveactivityduringtheseizure. First,themultipletimeseriesoftheinstantaneousphaseswereextractedfromtherawEEGdatausingtheHilberttransformapproachasdescribedinSection 3.2 .Inparticular,wetookadvantageofthefunctionshilbertandanglereadilyavailableintheMATLABR2006aenvironment. TheVARmodelingandtestingwereimplementedusingtheR2.6.1statisticalsoftware.Inouranalysisoftheinstantaneousphases,weincorporatedar,adf.test,po.test,cajolstandotherfunctionsfoundinpackagestseriesandurca. Next,weillustrateourapproachontheexampleoftheEEGdatalethatincludesthreeseizureintervals.Thelecontainsa16-channelrecordingofscalpEEGsampledatthe200Hzfrequencyaswellastwoauxiliarychannels,whichwerediscarded.TheinstantaneousphasevalueswereestimatedfromtheEEGtimeseriesbymeansofHilberttransform,andtheresultingphaseseriesweretestedusingtheADFtestintroducedinSection 2.4.1 .Specically,weappliedtheAugmentedDickey-Fullerproceduretotestthe 68

PAGE 69

Theresultsofourexperimentsforseizures1,2,and3arepresentedinTables 4-1 4-2 ,and 4-3 ,respectively.Thechannels,forwhichtheADFunitroottesthasdetectedapresenceofaunitrootatthesignicancelevel=0:01,arelistedasintegrated.Whereasthechannels,forwhichthenullhypothesisofaunitroothasbeenrejectedbytheADFatthe1percentlevel,aredenotedbystationary.Thechannelsforwhichthep-valuesoftheADFtestexceed2.5%aremarkedwith*.Noticethatallthreeseizuresegmentsareconsideredstable,whentheADFisappliedata0.025signicancelevel. Next,wetvectorautoregressiontothemultipletimeseriesofphaseestimates,foreachofthreedierentsegments(before,during,andafteraseizure)inordertodetermineappropriatelaglengthparameterp.Tondappropriatelagsp,theAkaikeInformationCriteria(AIC)wasused.Thisledustochooseseverallaglengthforeachsegmentandeachseizure.Finally,Johansencointegrationrankprocedurewasappliedtodeterminethevaluesofcointegrationrankrforeachcase.TheresultsaresummarizedinTables 4-4 4-5 ,and 4-6 Noticethatduringtheseizurethesystembecomesstable,especiallywhenmodeledusingashortestimateofthelagparameter.Sincethedurationsoftheseizure1andseizure2arerathershort,andonlyinclude440-500samplepoints,themodelsestimatedunderalonglagparametermaynotadequatelyrepresenttheunderlyingprocessesinseizure1and2.Ontheotherhand,seizure3isestimatedbasedonalmost1200samplevalues,andthereforethelonglagmodelofalongerseizure3maybemorerealistic,thanthelonglagmodelsforshorterseizures1and2.Overall,themodelsbasedonashortlagpforallthreeseizuresprovideanevidenceofabsolutesynchronizationamongthechannels.Whereas,thethepre-seizureandpost-seizuremodelsaremorelikelytobelessrestricted,andseemtoexhibitacointegrationrankbetween9and16. 69

PAGE 70

ResultsoftheADFunitroottestsforeachchannelduringthreesegments(2secondsimmediatelybeforeseizure,duringseizure,and2secondsafterseizure)forseizure1.Notethatthesignicanceat2.5%levelisdenotedby* Seizure#pre-seizureseizurepost-seizure stationary3,4,7,9,151-3,5-141,3,5,6,8,10,12,15integrated1*,2*,5,6*,8*,10,11,12*,13*,14*,16*4,15,162*,4*,7,9,13*,14,16* Table4-2. ResultsoftheADFunitroottestsforeachchannelduringthreesegments(2secondsimmediatelybeforeseizure,duringseizure,and2secondsafterseizure)forseizure2.Notethatthesignicanceat2.5%levelisdenotedby* Seizure#pre-seizureseizurepost-seizure stationary3,4,7,9-11,13,161-163,7,11-16integrated1*,2*,5*,6,8*,12*,14*,15none1*,2,4*,5,6*,8,9*,10 Table4-3. ResultsoftheADFunitroottestsforeachchannelduringthreesegments(2secondsimmediatelybeforeseizure,duringseizure,and2secondsafterseizure)forseizure3.Notethatthesignicanceat2.5%levelisdenotedby* Seizure#pre-seizureseizurepost-seizure stationary7,11,13,141-162,4,5,11,13,15integrated1*,2*,3,4*,5*,6*,8*,9,10*,12,15,16*none1,3*,6*,7,8,9*,10*,12,14*,16 Table4-4. ResultsoftheJohansencointegrationrankprocedureforthemultipleseriesduringthreesegments(2secondsimmediatelybeforeseizure,duringseizure,and2secondsafterseizure)forseizure1.Signicancelevelis1%.Fullrankisdenotedbyy longlagp=22;r=12p=23;r=11,p=20;r=13p=20;r=12shortlagp=2;r=13p=2;r=16yp=2;r=10 Table4-5. ResultsoftheJohansencointegrationrankprocedureforthemultipleseriesduringthreesegments(2secondsimmediatelybeforeseizure,duringseizure,and2secondsafterseizure)forseizure2.Signicancelevelis1%,fullrankisdenotedbyy longlagp=21;r=14p=26;r=12,p=20;r=9p=20;r=10shortlagp=2;r=16yp=3;r=16yp=2;r=13 Table4-6. ResultsoftheJohansencointegrationrankprocedureforthemultipleseriesduringthreesegments(2secondsimmediatelybeforeseizure,duringseizure,and2secondsafterseizure)forseizure3.Signicancelevelis1%,fullrankisdenotedbyy longlagp=24;r=13p=26;r=16y,p=20;r=16yp=20;r=13shortlagp=2;r=9p=2;r=16yp=2;r=16y

PAGE 71

Numbersnofrootsofthereversecharacteristicpolynomial(RCP)forVAR(p),whichlieonandinsidethecomplexunitcircle,computedfordierentsamplesizesTandfordierentmodelorderspusingtherawdata Numbersnofrootsofthereversecharacteristicpolynomial(RCP)forVAR(p),whichlieonandinsidethecomplexunitcircle,computedfordierentsamplesizesTandfordierentmodelorderspusingthe0{30Hzbandltereddata 71

PAGE 72

SegmentofmultichannelEEGrecordingthatcontainsafoursecondlongabsenceseizure

PAGE 73

Investigationofspatio-temporalpropertiesoftheEEGdatabydataminingandoptimizationapproachespostsvariouschallenges.NumerousfeaturesandmethodshavebeenproposedforstudyingthemultivariateseriesthatisEEG.TheanalysisofEEGtimeseriesisoftenapproachedfromtwodierentpointsofview,theonethattreatsEEGdataasproducedbyadeterministicchaoticdynamicalsystem,andtheothermoretraditionalapproachoflinearautoregressivemodeling. Inthiswork,weinvestigatedseveralstatisticalapproachesthatarerecentlyintroducedfordataminingbraindynamics.Inparticular,weexaminedtheapplicationofvectorautoregressivemodelingandlinearGrangercausalitytorawandlteredEEGdata. Motivatedbyrecentsuccessinapplicationofphasesynchronizationtoanalysisofdynamicprocessesinepilepticbrain,wedevelopedaconceptofgeneralizedsynchro-nizationbasedonthenovelideaofextendingtheclassicalsynchronizationconditionofaboundedlinearcombinationoftwophases.Thissimplebivariateconditioninthemultivariatecaseisanalogoustoacointegratingrelationshipinthemultipletimeseries.Thus,wecananalyzethesynchronyamongdierentpartsofthecommoninterrelatedsystem(suchasahumanbrain),bymodelingthephasesextractedfromanitenumberofsignalsinthesystemsbymeansofcointegratedvectorautoregression.Moreover,weshowedthatthecointegrationrankinthecointegratedVARmodelofthephasetimeseriescanbeviewedasameasureofsynchronyamongthephasesofdierentcomponentsoftheEEGsignal. Notonlythisnewmeasureofmultivariatephasesynchronycanbetestedonvariousbiomedicaldata,suchasmultichannelEEGrecordedfromanepilepticbrain,butalsothenewmultiplephasesynchronizationcanbeemployedindierentareasofappliedandtheoreticresearch(includingphysics,communication,electronics,laserdynamics,and 73

PAGE 74

74

PAGE 75

[1] L.BaccalaandK.Sameshima,\Partialdirectedcoherence:anewconceptinneuralstructuredetermination,"Biol.Cybern.,vol.84,pp.463-74,2001. [2] C.BernasconiandP.Koenig,\Onthedirectionalityofcorticalinteractionsstudiedbystructuralanalysisofelectrophysiologicalrecordings,"Biol.Cybern.vol.81,pp.199{210,1999. [3] R.D.Bickford,\Electroencephalography,"inAdelmanG.ed.EncyclopediaofNeuroscience.Cambridge,MA:Birkhauser,pp.371{373,1987. [4] C.D.Binnie,\LongtermEEGrecordinganditsroleinclinicalpractice,"IEEColloquiumonDataLoggingofPhysiologicalSignals,pp.5/1-5/2,1995. [5] P.S.Bradley,O.L.Mangasarian,andW.N.Street,\Clusteringviaconcaveminimization,"inM.C.Mozer,M.I.Jordan,andT.Petsche,eds.,AdvancesinNeuralInformationProcessingSystems,vol.9,pp.368{374,1997. [6] P.S.BradleyandO.L.Mangasarian,\Featureselectionviaconcaveminimizationandsupportvectormachines,"inJ.Shavlik,MachineLearningProceedingsoftheFifteenthInternationalConference,pp.82{90,1998. [7] P.S.Bradley,U.M.Fayyad,andO.L.Mangasarian,\MathematicalProgrammingforDataMining:FormulationsandChallenges,"INFORMSJournalonComputing,vol.11,no.3,pp.217{238,1999. [8] S.L.Bressler,R.Coppola,andR.Nakamura,\Episodicmultiregionalcorticalcoherenceatmultiplefrequenciesduringvisualtaskperformance,"Nature,vol.366,pp.153156,1993. [9] R.BrownandL.Kocarev,\Aunifyingdenitionofsynchronizationfordynamicalsystems,"Chaos,vol.10,pp.344-349,2000. [10] T.H.BullockandM.C.McClune,\Lateralcoherenceoftheelectrocorticogram:anewmeasureofbrainsynchrony,"ElectroencephalogrClinNeurophysiol,vol.73,pp.479-498,1989. [11] T.L.CarrollandL.M.Pecora,\Cascadingsynchronizedchaoticsystems,"PhysicaD,vol.67,pp.126,1993. [12] G.C.Carter,\Coherenceandtimedelayestimation,"ProcIEEE,vol.75,pp.236-255,1987. [13] Y.-W.Cheung,K.S.Lai,\LagOrderandCriticalValuesoftheAugmentedDickey-FullerTest,"JournalofBusiness&EconomicStatistics,vol.13,no.3,pp.277{280,1995. 75

PAGE 76

R.Cont,\StatisticalPropertiesofFinancialTimeSeries,"inR.ContandJ.Yongeditors,MathematicalFinance:TheoryandPractice,LectureSeriesinAppliedMathematics,vol.1,1999. [15] A.R.Damasio,\Synchronousactivationinmultiplecorticalregions:amechanismforrecall,"SeminNeurosci,vol.2,pp.287-296,1990. [16] D.A.Dickey,andW.A.Fuller,\DistributionoftheEstimatorsforAutoregressiveTimeSeriesWithaUnitRoot,"JournaloftheAmericanStatisticalAssociation,vol.74,pp.427{431,1979. [17] M.Eichler,Graphicalmodellingofmultivariatetimeseries,UniversityofHeidelberg:Preprint,2001. [18] M.Eichler,R.Dahlhaus,andJ.Sandkuhler,\Partialcorrelationanalysisfortheidenticationofsynapticconnections,"Biol.Cybern.,vol.89,pp.289-302,2003. [19] M.Eichler,\Agraphicalapproachforevaluatingeectiveconnectivityinneuralsystems,"Philos.Transact.R.Soc.Bvol.360,pp.953-67,2005. [20] L.Fabiny,P.Colet,andR.Roy,\Coherenceandphasedynamicsofspatiallycoupledsolid-statelasers,"Phys.Rev.A,vol.47,pp.4287,1993. [21] W.J.Freeman,\SpatialpropertiesofanEEGeventintheolfactorybulbandcortex,"ElectroencephalogrClinNeurophysiol,vol.44,pp.586-605,1978. [22] W.A.Fuller,IntroductiontoStatisticalTimeSeries,2nded.NewYork:JohnWiley,1996. [23] D.Gabor.\Theoryofcommunication,"inProc.IEEELondon,vol.93,pp.429,1946. [24] P.Georgiev,A.Cichocki,andH.Bakardjian,\OptimizationtechniquesforIndependentComponentAnalysiswithApplicationstoEEGdata,"InP.M.Pardalos,J.C.Sackellares,P.R.Carney,L.D.Iasemidis,eds.,Quantita-tiveNeuroscience:Models,Algorithms,Diagnostics,andTheraputicApplications,KluwerAcademicPublishers,pp.53{68,2004. [25] J.Geweke,\Measurementoflineardependenceandfeedbackbetweenmultipletimeseries,"J.AmericanStatis.Association,vol.77,pp.304{313,1982. [26] C.W.J.Granger,\EconomicProcessesInvolvingFeedback,"InformationandControl,vol.6,pp.28{48,1963. [27] C.W.J.Granger,\InvestigatingCausalRelationsbyEconometricModelsandCross-SpectralMethods,"Econometrics,vol.37,pp.424{438,1969. [28] C.W.J.Granger,\SomePropertiesofTimeSeriesDataandTheirUseinEconometricModelSpecication,"JofEconometrics,vol.16,pp.121{130,1981. 76

PAGE 77

M.J.Griths,P.Grainger,M.V.Cox,andA.W.Preece,\RecentadvancesinEEGmonitoringforgeneralanaesthesia,alteredstatesofconsciousnessandsportsperformancescience,"inThe3rdIEEInternationalSeminaronMedicalApplicationsofSignalProcessing,pp.1{5,2005. [30] J.D.Hamilton,TimeSeriesAnalysis.Princeton,1994. [31] M.-J.Hoeve,R.D.Jones,G.J.Carroll,andH.Goelz,\AutomateddetectionofepilepticseizuresintheEEG,"inEngineeringinMedicineandBiologySociety,Proceedingsofthe23rdAnnualInternationalConferenceoftheIEEE,vol.1,pp.943{946,2001. [32] R.Horst,P.M.Pardalos,andN.V.Thoai,IntroductiontoGlobalOptimization,2ndedition.KluwerAcademicPublishers,2000. [33] C.Huygens.HoroloquiumOscilatorium.Paris,1673. [34] L.D.Iasemidis,\Onthedynamicsofthehumanbrainintemporallobeepilepsy,"Ph.D.Dissertation,UniversityofMichigan,AnnArbor,1991. [35] L.D.Iasemidis,andJ.C.Sackellares,\TheevolutionwithtimeofthespatialdistributionofthelargestLyapunovexponentonthehumanepilepticcortex,"inD.W.Duke,andW.S.Pritchard,eds.,Measuringchaosinthehumanbrain,pp.49{82,WorldScientic,1991. [36] L.D.Iasemidis,D.-S.Shiau,J.C.Sackellares,P.M.Pardalos,andA.Prasad,\Dynamicalresettingofthehumanbrainatepilepticseizures:applicationofnonlineardynamicsandglobaloptimizationtecniques,"IEEETransactionsonBiomedicalEngineering,vol.51,no.3,pp.493{506,2004. [37] H.H.Jasper,\Theten-twentyelectrodesystemoftheInternationalFederation,"Electroencephalogr.Clin.Neurophysiol.,vol.10,pp.370{375,1958. [38] S.JohansenandK.Juselius,\MaximumLikelihoodEstimationandInferenceonCointegration-WithApplicationstotheDemandforMoney,"OxfordBulletinofEconomicsandStatistics,vol.52,pp.169{210,1990. [39] E.Juergens,A.Guettler,andR.Eckhorn.\Visualstimulationelicitslockedandinducedgammaoscillationsinmonkeyintracortical-andEEG-potentials,butnotinhumanEEG,"ExperimentalBrainResearch,vol.129,pp.247-259,1999. [40] Y.Kaji,M.Akutagawa,F.Shichijo,H.Nagashino,Y.Kinouchi,andS.Nagahiro,\AnalysisforBrainActivitiesduringOperationsUsingMeasuredEEG,"inConfProcIEEEEngMedBiolSoc.vol.6,pp.6003{6006,2005. [41] T.KalayciandO.Ozdamar,\WaveletpreprocessingforautomatedneuralnetworkdetectionofEEGspikes,"EngineeringinMedicineandBiologyMagazine,IEEE,vol.14,no.2,pp.160{166,1995. 77

PAGE 78

M.Kaminski,M.Ding,W.A.Truccolo,andS.L.Bressler,\Evaluatingcausalrelationsinneuralsystems:Grangercausality,directedtransferfunctionandstatisticalassessmentofsignicance,"Biol.Cybern.,vol.85,pp.145{57,2001. [43] [44] [45] E.Keogh,andS.Kasetty,\Ontheneedfortimeseriesdataminingbenchmarks:asurveyandempiricaldemonstration,"inProceedingsoftheEighthACMSIGKDDinternationalConferenceonKnowledgeDiscoveryandDataMining,pp.102{111,2002. [46] M.E.Kirlangic,D.Perez,S.Kudryavtseva,G.Griessbach,G.Henning,andG.Ivanova,\Fractaldimensionasafeatureforadaptiveelectroencephalogramsegmentationinepilepsy,"inEngineeringinMedicineandBiologySociety.Proceed-ingsofthe23rdAnnualInternationalConferenceoftheIEEE,vol.2,pp.1573{1576,2001. [47] W.A.Kittel,C.M.Epstein,andM.H.Hayes,\EEGmonitoringbasedonfuzzyclassication,"CircuitsandSystems,Proceedingsofthe35thMidwestSymposium,vol.1,pp.699{702,1992. [48] L.KocarevandU.Parlitz,\Generalapproachforchaoticsynchronizationwithapplicationstocommunication,"Phys.Rev.Lett.,vol.74,pp.5028,1995. [49] T.Kreuz,MeasuringSynchronizationinModelSystemsandElectroencephalographicTimeSeriesfromEpilepsyPatients,Dissertation.NICSeries,vol.21,2003. [50] J.P.Lachaux,E.Rodriguez,J.Martinerie,andF.J.Varela,\Measuringphasesynchronyinbrainsignals,"Hum.BrainMapp.,vol.8,pp.194,1999. [51] J.P.Lachaux,E.Rodriguez,J.Martinerie,C.Adam,D.HasbounandF.J.Varela,\Gamma-bandactivityinhumanintracorticalrecordingstriggeredbycognitivetasks,"EurJNeurosci,vol.12,pp.2608-2622,2000. [52] J.P.Lachaux,E.Rodriguez,M.LeVanQuyen,A.Lutz,J.Martinerie,andF.J.Varela,\Studyingsingle-trialsofphase-synchronousactivityinbrain,"Int.J.BifurcationChaosAppl.Sci.Eng.,vol.10,pp.2429,2000. [53] A.D.Legatt,J.Arezzo,andH.G.Vaughan,\Averagedmultipleunitactivityasanestimateofphasicchangesinlocalneuronalactivity:eectsofvolume-conductedpotentials,"J.ofNeuroscienceMethods,vol.2no.2,pp.203{217,1980. 78

PAGE 79

M.LeVanQuyen,C.Adam,J.P.Lachaux,J.Martinerie,M.Baulac,B.Renault,etal.,\Temporalpatternsinhumanepilepticactivityaremodulatedbyperceptualdiscriminations,"Neuroreport,vol.8,pp.1703-1710,1997. [55] M.LeVanQuyen,J.Martinerie,C.Adam,andF.J.Varela,\Nonlinearspatio-temporalinterdependenciesofinterictalintracranialEEGrecordingsfrompatientswithtemporallobeepilepsy:Localizingofepileptogenicfoci,"PhysicaD,vol.127,pp.250-65,1999. [56] M.LeVanQuyen,J.Foucher,J.-P.Lachaux,E.Rodriguez,A.Lutz,J.Martinerie,andF.J.Varela,\ComparisonofHilberttransformandwaveletmethodsfortheanalysisofneuronalsynchrony,"JournalofNeuroscienceMethods,vol.111,pp.83-98,2001. [57] H.Liang,M.Ding,R.Nakamura,andS.L.Bressler,\Causalinuencesinprimatecerebralcortexduringvisualpatterndiscrimination,"Neuroreport,vol.11,pp.2875{80,2000. [58] Q.Liu,M.Sun,M.L.Scheuer,andR.J.Sclabassi,\Patienttrackingforvideo/EEGmonitoringbasedonchangedetectioninDCTdomain,"inBioengineeringConfer-ence,2005.ProceedingsoftheIEEE31stAnnualNortheast,pp.114{115,2005. [59] H.Lutkepohl,Intoductiontomultipletimeseriesanalysis.Berlin,Heidelberg:Springer-Verlag,1991. [60] J.G.MacKinnon,\CriticalValuesforCointegrationTests,"inR.F.EngleandC.W.J.Granger,eds.,Long-RunEconomicRelationships:ReadingsinCointegra-tion,NewYork:OxfordUniversityPress,pp.266{276,1991. [61] J.G.MacKinnon,\ApproximateAsymptoticDistributionFunctionsforUnit-RootandCointegrationTests,"JournalofBusiness&EconomicStatistics,vol.12,pp.167{176,1994. [62] V.Menon,W.J.Freeman,B.A.Cutillo,J.E.Desmond,M.F.Ward,S.L.Bressler,etal.,\Spatio-temporalcorrelationsinhumangammabandelectrocorticograms,"ElectroencephalogrClinNeurophysiol,vol.98,no.2,pp.89-102,1996. [63] M.Mezard,G.Parisi,andM.A.Virasoro,Spinglasstheoryandbeyond.WorldScientic,1987. [64] S.MitraandT.Acharya,Datamining:multimidea,softcomputing,andbioinfor-matics.Hoboken,NJ:JohnWiley&Sons,Inc.,2003. [65] F.Mormann,K.Lehnertz,P.David,andC.E.Elger,\MeanphasecoherenceasameasureforphasesynchronizationanditsapplicationtotheEEGofepilepticpatients,"PhysicaD,vol.144,pp.358-369,2000. [66] M.P.Murray\ADrunkandHerDog:AnIllustrationofCointegrationandErrorCorrection,"TheAmericanStatistician,vol.48,no.1,pp.37{39,1994. 79

PAGE 80

E.Niedermeyer,F.H.LopesdaSilva,Electroencephalography:Basicprinciples,clinicalapplicationsandrelatedelds,3rdedition.Philadelphia:Lippincott,Williams&Wilkins,1993. [68] P.L.Nunez,NeocorticalDynamicsandHumanEEGRhythms.NewYork:OxfordUniv.Press,1995. [69] N.H.Packard,J.P.Crutcheld,J.D.Farmer,andR.S.Shaw,\GeometryfromaTimeSeries,"Phys.Rev.Lett.,vol.45,no.9,pp.712{716,1980. [70] P.Panter.Modulation,noise,andspectralanalysis.NewYork:McGraw-Hill,1965. [71] H.S.Park,Y.H.Lee,S.DooS,etal.,\Detectionofepileptiformactivityusingwaveletandneuralnetwork,"EngineeringinMedicineandBiologySociety,vol.3,pp.1194,1997. [72] U.Parlitz,L.Junge,W.Lauterborn,andL.Kocarev,\Experimentalobservationofphasesynchronization,"Phys.Rev.E,vol.54,pp.2115,1996. [73] J.B.Patten,NeurologicalDierentialDiagnosis,2nded.,Springer-Verlag,1996. [74] D.W.Peterman,M.Ye,andP.E.Wigen,\Highfrequencysynchronizationofchaos,"Phys.Rev.Lett.,vol.74,pp.1740,1995. [75] P.C.B.Phillips,\Timeseriesregressionwithaunitroot,"Econometrica,vol.55,pp.277{301,1987. [76] P.C.B.PhillipsandS.Ouliaris,\AsymptoticPropertiesofResidualBasedTestsforCointegration,"Econometrica,vol.58,pp.165-193,1990. [77] A.S.Pikovsky,\Phasesynchronizationofchaoticoscillationsbyaperiodicexternaleld,"SovJCommunTechnolElectron,vol.30,pp.85,1985. [78] A.S.Pikovsky,M.G.Rosenblum,andJ.Kurths.Synchronization.Auniversalconceptinnonlinearsciences.Cambridge,UK:CambridgeUniv.Press,2001. [79] O.A.Prokopyev,V.Boginski,W.Chaovalitwongse,P.M.Pardalos,J.C.Sackellares,andP.R.Carney,\Network-basedTechniquesinEEGDataAnalysisandEpilepticBrainModeling,"inP.M.Pardalos,V.BoginskiandA.Vazacopoulos,eds.,DataMininginBiomedicine,Springer,pp.559{573,2007. [80] K.Pyragas,\Continuouscontrolofchaosbyself-controllingfeedback,"Phys.Lett.A,vol.170,pp.421,1992. [81] R.QuianQuiroga,A.Kraskov,T.Kreuz,andP.Grassberger,\Performanceofdierentsynchronizationmeasuresinrealdata:Acasestudyonelectroencephalographicsignals,"Phys.Rev.E,vol.65,pp.041903,2002. 80

PAGE 81

E.Rodriguez,N.George,J.P.Lachaux,J.Martinerie,andF.J.Varela,\Perceptionsshadow:Long-distancesynchronizationinthehumanbrain,"Na-ture,vol.397,pp.340-343,1999. [83] P.R.Roelfsema,A.K.Engel,P.Konig,andW.Singer,\Visuomotorintegrationisassociatedwithzerotime-lagsynchronizationamongcorticalareas,"Nature,vol.385,pp.157-161,1997. [84] E.RosaJr.,W.B.Pardo,C.M.Ticos,J.A.Walkenstein,andM.Monti,\Phasesynchronizationofchaosinaplasmadischargetube,"Int.J.Bifurc.Chaos,vol.10,pp.2551,2000. [85] M.Rosenblum,A.Pikovsky,J.Kurths,\Phasesynchronizationofchaoticoscillators,"PhysRevLett,vol.76,pp.1804-1807,1996. [86] M.G.Rosenblum,A.S.Pikovsky,C.Schafer,P.Tass,andJ.Kurths,\Phasesynchronization:Fromtheorytodataanalysis,"in:S.Gielen,F.MossF,eds.HandbookofBiologicalPhysics,vol.4.ElsevierScience,Neuro-informatics,1999. [87] R.RoyandK.S.Thornburg,\Experimentalsynchronizationonchaoticlasers,"Phys.Rev.Lett.,vol.72,pp.2009,1994. [88] N.F.Rulkov,L.S.Tsimring,andH.D.I.Abarbanel,\Trackingunstableorbitsinchaosusingdissipativefeedbackcontrol,"Phys.Rev.E,vol.50,pp.314,1994. [89] J.C.Sackellares,L.D.Iasemidis,R.L.Gilmore,andS..Roper,\Epilepsy-whenchaosfails,"inK.Lehnertz,J.Arnold,P.Grassberger,andC.E.Elger,eds.,Chaosinthebrain?,WorldScientic,2002. [90] J.C.Sackellares,L.D.Iasemidis,D.-S.Shiau,P.M.Pardalos,andP.R.Carney,\Spatio-temporaltransitionsintemporallobeepilepsy,"inP.M.Pardalos,J.C.Sackellares,P.R.Carney,L.D.Iasemidis,eds.,QuantitativeNeuroscience:Models,Algorithms,Diagnostics,andTheraputicApplications,KluwerAcademicPublishers,pp.223{238,2004. [91] S.E.SaidandD.A.Dickey,\TestingforUnitRootsinAutoregressive-MovingAverageModelsofUnknownOrder,"Biometrika,vol.71,pp.599-607,1984. [92] K.SameshimaandL.Baccala,\Usingpartialdirectedcoherencetodescribeneuronalensembleinteractions,"J.Neurosci.Methods,vol.94,pp.93-103,1999. [93] C.Schafer,M.G.Rosenblum,J.Kurths,andH.H.Abel,\Heartbeatsynchronizedwithventilation,"Nature,vol.392,pp.239-240,1998. [94] F.LopezdeSilva,\EEGanalisys:Theoryandpractice;Computer-assistedEEGdiagnosis:Patternrecognitiontechniques,"inE.Niedermeyer,F.H.LopesdaSilva,eds.Electroencephalography:Basicprinciples,clinicalapplicationsandrelatedelds,pp.871{919,1987. 81

PAGE 82

C.A.Sims,\Money,IncomeandCausality,"AmericanEconomicReview,vol.62,pp.540{552,1972. [96] W.SingerandC.M.Gray,\Visualfeatureintegrationandthetemporalcorrelationhypothesis,"AnnuRevNeurosci,vol.18,pp.555-586,1995. [97] C.Tallon-Baudry,O.Bertrand,C.Delpuech,andJ.Pernier,\Oscillatorygamma-band(30-70Hz)activityinducedbyavisualsearchtaskinhuman,"JNeurosci,vol.17,pp.722-734,1997. [98] D.Y.Tang,R.Dykstra,M.W.Hamilton,andN.R.Heckenberg,\Experimentalevidenceoffrequencyentrainmentbetweencoupledchaoticoscillations,"Phys.Rev.E,vol.57,no.3,pp.3649,1998. [99] P.Tass,M.G.Rosenblum,J.Weule,J.Kurths,A.Pikovsky,J.Volkmann,etal.,\Detectionofn:mphaselockingfromnoisydata:applicationtomagnetoencephalography,"PhysRevLett,vol.81,pp.3291-3294,1998. [100] M.Teplan,\FundamentalsofEEGmeasurement,"Measurementsciencereview,vol.2,no.2,pp.112-23,2002. [101] G.TononiandG.M.Edelman,\Consciousnessandcomplexity,"Science,vol.282,pp.1846-1851,1998. [102] R.D.Traub,M.A.Whittington,E.Buhl,J.G.Jeerys,andH.J.Faulkner,\Onthemechanismofthegammatobetafrequencyshiftinneuronaloscillationsinducedinrathippocampalslicesbytetanicstimulation,"JNeurosci,vol.19,pp.1088-1105,1999. [103] B.vanderPolandJ.vanderMark,\Theheartbeatconsideredasarelaxationoscillation,andanelectricalmodeloftheheart,"Phil.Mag.,vol.6,pp.763,1928. [104] F.J.Varela,\Resonantcellassemblies:Anewapproachtocognitivefunctionsandneuronalsynchrony,"BiolRes,vol.28,pp.81-95,1995. [105] F.J.Varela,J.P.Lachaux,E.Rodriguez,andJ.Martinerie,\Thebrainweb:phasesynchroniztionandlarge-scaleintegration,"NatureRevNeurosci,vol.2,pp.229-39,2001. [106] N.Wiener,\TheTheoryofPrediction,"inE.F.Beckenbach,ed.ModernMathe-maticsfortheEngineer,NewYork:McGraw-Hill,1956. [107] A.Wolf,J.BSwift,H.L.Swinney,andJ.A.Vastano,\DeterminingLyapunovexponentfromatimeseries,"PhysicaD,vol.16,pp.285{317,1985. [108] K.Yang,andC.Shahabi,\OntheStationarityofMultivariateTimeSeriesforCorrelation-BasedDataAnalysis,"inProceedingsofFifthIEEEInternationalConferenceonDataMining(ICDM'05),pp.805{808,2005. 82

PAGE 83

AllaRevenkoKammerdinerwasborninKiev,Ukraine.Anolderoftwochildren,shegrewupinKiev,Ukraine,graduatingfromSchool#32in1994.SheearnedherB.S.inProbabilityTheoryandMathematicalStatisticsfromtheNationalTarasShevchenkoUniversityofKyivin1998.InJanuary2001,AllajoinedagraduateprogramintheMathematicsDepartmentattheUniversityofFlorida.UpongraduatinginMay2004withherM.S.inmathematics,AllaenteredthePh.D.programinindustrialandsystemsengineeringattheUniversityofFlorida.AllahasbeenhappilymarriedtoJasonR.Kammerdinerforthelast3years.OnFebruary24,2008shecompletedherrstmarathonin3:53:09. 83