See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/282944747

Conditional Granger causality and partitioned

Granger causality: differences and similarities

Article in Biological Cybernetics October 2015

DOI: 10.1007/s00422-015-0665-3

CITATIONS READS

0 33

2 authors:

Sheida Malekpour William A. Sethares

University of WisconsinMadison University of WisconsinMadison
3 PUBLICATIONS 24 CITATIONS 199 PUBLICATIONS 2,604 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Texture classification View project

Musica Facta View project

All content following this page was uploaded by William A. Sethares on 25 May 2017.

The user has requested enhancement of the downloaded file.

Dear Author,

Here are the proofs of your article.

You can submit your corrections online, via e-mail or by fax.

For online submission please insert your corrections in the online correction form. Always
indicate the line number to which the correction refers.

You can also insert your corrections in the proof PDF and email the annotated PDF.

For fax submission, please ensure that your corrections are clearly legible. Use a fine black
pen and write the correction in the margin, not too close to the edge of the page.

Remember to note the journal title, article number, and your name when sending your
response via e-mail or fax.

Check the metadata sheet to make sure that the header information, especially author names
and the corresponding affiliations are correctly shown.

Check the questions that may have arisen during copy editing and insert your answers/
corrections.

Check that the text is complete and that all figures, tables and their legends are included. Also
check the accuracy of special characters, equations, and electronic supplementary material if
applicable. If necessary refer to the Edited manuscript.

The publication of inaccurate data such as dosages and units can have serious consequences.
Please take particular care that all such details are correct.

Please do not make changes that involve only matters of style. We have generally introduced
forms that follow the journals style.
Substantial changes in content, e.g., new results, corrected values, title and authorship are not
allowed without the approval of the responsible editor. In such a case, please contact the
Editorial Office and return his/her consent together with the proof.

If we do not receive your corrections within 48 hours, we will send you a reminder.

Your article will be published Online First approximately one week after receipt of your
corrected proofs. This is the official first publication citable with the DOI. Further changes
are, therefore, not possible.

The printed version will follow in a forthcoming issue.

Please note
After online publication, subscribers (personal/institutional) to this journal will have access to the
complete article via the DOI using the URL: http://dx.doi.org/[DOI].
If you would like to know when your article has been published online, take advantage of our free
alert service. For registration and further information go to: http://www.link.springer.com.
Due to the electronic nature of the procedure, the manuscript and the original figures will only be
returned to you on special request. When you return your corrections, please inform us if you would
like to have these documents returned.
 Metadata of the article that will be visualized in
OnlineFirst

Please note: Images will appear in color online but will be printed in black and white.
ArticleTitle Conditional Granger causality and partitioned Granger causality: differences and similarities
Article Sub-Title
Article CopyRight Springer-Verlag Berlin Heidelberg
(This will be the copyright line in the final PDF)
Journal Name Biological Cybernetics
Corresponding Author Family Name Malekpour
Particle
Given Name Sheida
Suffix
Division Department of Electrical and Computer Engineering
Organization University of Wisconsin-Madison
Address 2556 Engineering Hall, 1415 Engineering Dr., 53706, Madison, WI, USA
Email malekpour2@wisc.edu

Author Family Name Sethares

Particle
Given Name William A.
Suffix
Division Department of Electrical and Computer Engineering
Organization University of Wisconsin-Madison
Address 2556 Engineering Hall, 1415 Engineering Dr., 53706, Madison, WI, USA
Email sethares@wisc.edu

Received 25 April 2015

Schedule Revised
Accepted 5 October 2015
Abstract Neural information modeling and analysis often requires a measurement of the mutual influence among
many signals. A common technique is the conditional Granger causality (cGC) which measures the
influence of one time series on another time series in the presence of a third. Geweke has translated this
condition into the frequency domain and has explored the mathematical relationships between the time and
frequency domain expressions. Chen has observed that in practice, the expressions may return
(meaningless) negative numbers, and has proposed an alternative which is based on a partitioned matrix
scheme, which we call partitioned Granger causality (pGC). There has been some confusion in the
literature about the relationship between cGC and pGC; some authors treat them as essentially identical
measures, while others have noted that some properties (such as the relationship between the time and
frequency domain expressions) do not hold for the pGC. This paper presents a series of matrix equalities
that simplify the calculation of the pGC. In this simplified expression, the essential differences and
similarities between the cGC and the pGC become clear; in essence, the pGC is dependent on only a subset
of the parameters in the model estimation, and the noise residuals (which are uncorrelated in the cGC) need
not be uncorrelated in the pGC. The mathematical results are illustrated with a simulation, and the
measures are applied to an EEG dataset.
Keywords (separated by '-') Conditional Granger causality (cGC) - Partitioned Granger causality (pGC) - Multivariate autoregressive
(MVAR) model
Footnote Information
 Biol Cybern
DOI 10.1007/s00422-015-0665-3

ORIGINAL ARTICLE

Conditional Granger causality and partitioned Granger causality:

of
differences and similarities
Sheida Malekpour1 William A. Sethares1

pro
Author Proof

Received: 25 April 2015 / Accepted: 5 October 2015

Springer-Verlag Berlin Heidelberg 2015

1 Abstract Neural information modeling and analysis often Keywords Conditional Granger causality (cGC) 27

requires a measurement of the mutual influence among many Partitioned Granger causality (pGC) Multivariate
2

8
signals. A common technique is the conditional Granger
causality (cGC) which measures the influence of one time
series on another time series in the presence of a third.
Geweke has translated this condition into the frequency
domain and has explored the mathematical relationships
between the time and frequency domain expressions. Chen
ted autoregressive (MVAR) model

1 Introduction

Many studies in neural information processing (Kamiski

28

29

30

31

9 has observed that in practice, the expressions may return et al. 2001; Bernasconi and Knig 1999), in the mathemat- 32
rrec
10 (meaningless) negative numbers, and has proposed an alter- ical description of neural functioning (Wang et al. 2007; 33

11 native which is based on a partitioned matrix scheme, which Gourvitch et al. 2006) and in the statistical methods used 34

12 we call partitioned Granger causality (pGC). There has to analyze neural data (Bressler et al. 2007; Brovelli 2012), 35

13 been some confusion in the literature about the relationship use the concept of Granger causality to measure the influence 36

14 between cGC and pGC; some authors treat them as essen- of neural signals on each other. Granger (1969) introduced 37

15 tially identical measures, while others have noted that some the concept of causality that bears his name as a statistical 38

16 properties (such as the relationship between the time and fre- hypothesis test which determines whether one time series 39

17 quency domain expressions) do not hold for the pGC. This can help to predict another. The causal influence measured 40

18 paper presents a series of matrix equalities that simplify the by Granger causality has been further studied in the time and 41
o

19 calculation of the pGC. In this simplified expression, the frequency domain in Geweke (1984a,b), which extends to the 42

20 essential differences and similarities between the cGC and mutual influence of many time series. Barnett has shown in 43

21 the pGC become clear; in essence, the pGC is dependent on Barnett et al. (2009) that for Gaussian random variables, con- 44
unc

22 only a subset of the parameters in the model estimation, and ditional Granger causality (cGC) is equal to transfer entropy 45

23 the noise residuals (which are uncorrelated in the cGC) need multiplied by 2. Chen et al. (2006) derive a causality measure 46

24 not be uncorrelated in the pGC. The mathematical results are in the frequency domain (which we call partitioned Granger 47

25 illustrated with a simulation, and the measures are applied to causality, or pGC) based on partitioned matrix techniques to 48

26 an EEG dataset. overcome the negative (and hence meaningless) estimation 49

of cGC. Many studies such as Malekpour et al. (2012), Leung 50

et al. (2010), Bollimunta et al. (2008), Dhamala et al. (2008), 51

B Sheida Malekpour Zhou et al. (2009), Liao et al. (2010) have utilized the pGC 52

malekpour2@wisc.edu to study the connectivity and information flow between dif- 53

William A. Sethares ferent regions of brain. 54

sethares@wisc.edu Barnett et al. suggest that the pGC may not exactly repre- 55

1 sent the cGC as defined by Geweke in the frequency domain 56

Department of Electrical and Computer Engineering,
University of Wisconsin-Madison, 2556 Engineering Hall, (Barnett and Seth 2014). In particular, they observe that the 57

1415 Engineering Dr., Madison, WI 53706, USA pGC does not satisfy an inequality [Eq. (18)] that relates the 58

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

59 time and frequency domains that is known to hold for the !
Z (t) = DZX j X (t j) + DZZ j Z (t j) + (t), 103
60 cGC.
j=1
61 This paper simplifies the equation for pGC given in Chen
62 et al. (2006) and discusses a possible reason that negative cov( (t)) = ! , (1) 104

63 values were encountered in the calculation of cGC as carried

out in Chen et al. (2006). Studying the simplified equation where the noises (t) and (t) are zero mean and uncorre- 105

of
64

65 reveals why the pGC is different from Gewekes cGC. To esti- lated over time, i.e., E{(t)(t )} = 0, and E{ (t) (t 106

66 mate the cGC network connectivity in a three-node network )} = 0 for all = 0, and where the covariance between the 107

67 (i.e., with data at nodes defined by X (t), Y (t), and Z (t)), all noises is 108

68 of the estimated parameters in a multivariate autoregressive "

! !
#
(MVAR) model are needed. In contrast, it is possible to esti- ! noise(1) = . (2)

pro
69 109
! ! 110
70 mate the pGC of the same time series using only a subset of
71 the estimated parameters. The relevant formulas are clarified In (1), X (t) and Z (t) are known, and the parameters 111
by the simplified pGC equation. Furthermore, the simplified
Author Proof

72
DXX j , DXZ j , DZX j , DZZ j , and ! noise(1) are to be estimated. 112
73 equation reveals that although the pGC estimates are always Next, an MVAR model between all three random processes 113
74 nonnegative, they do not always fulfill the timefrequency X (t), Y (t), and Z (t) is assumed 114
75 domain inequality.
76 Section 2 summarizes the definition of cGC in the time
!
77 and frequency domains. Section 3 discusses a way that nega- X (t) = BXX j X (t j) + BXY j Y (t j) 115

tive values may arise in the calculation of the cGC. Section 4 j=1
78

79

80

81

82

83

84
summarizes the assumptions and calculations that lead to the
pGC, the partitioned matrix method introduced in Chen et al.
(2006). Chens expression for cGC is presented in (17), which
is calculated via (21)(24). Section 5 simplifies this to (27)
using the theorem of Sect. 5. The same section shows in an
example the equality of the original pGC and the simplified
ted Y (t) =
+BXZ j Z (t j) + E X (t),

!

!
BYX j X (t j) + BYY j Y (t j)
j=1
+BYZ j Z (t j) + E Y (t),
116

117

118

85 expressions. Section 6 discusses the differences between the Z (t) = BZX j X (t j) + BZY j Y (t j) 119
rrec
86 pGC and the cGC, which are demonstrated by two exam- j=1

87 ples in Sect. 8. The intuition behind the pGC is discussed +BZZ j Z (t j) + E Z (t), (3) 120

88 in Sect. 7, and the paper concludes in Sect. 9. Proofs of the

89 lemmas and theorems appear in the Appendices. Through- where the noises E X (t), E Y (t), and E Z (t) are again assumed 121

90 out the paper, uppercase and boldface uppercase symbols to be zero mean and uncorrelated over time; the covariances 122

91 represent vectors and matrices, respectively. Lowercase sym- of the noise terms are 123

92 bols represent scalars. Superscripts * and 1 denote complex

! XX ! XY ! XZ
93 conjugate transpose and matrix inverse, respectively. I rep-
! noise(2) = ! YX ! YY ! YZ . (4) 124
94 resents the identity matrix.
o

! ZX ! ZY ! ZZ 125

In (3), X (t), Y (t), and Z (t) are known, and the parame- 126
unc

ters BXX j , BXY j , BXZ j , BYX j , BYY j , BYZ j , BZX j , BZY j , BZZ j , 127
95 2 Overview of conditional Granger causality and ! noise(2) are to be estimated. The conditional Granger 128

causality from Y (t) to X (t) given Z (t) is then defined as 129

96 2.1 Conditional Granger causality in the time domain
|! |
FY X |Z = ln . (5) 130
97 Given three vector stochastic (random) processes X (t) |! XX | 131
98 Rm , Y (t) Rn , and Z (t) Rk , the cGC from Y (t) to X (t)
99 given Z (t) is calculated by combining two MVAR estima- If X (t) and Y (t) are independent, |! | = |! XX | and 132

100 tions. First, X (t) and Z (t) are assumed to be related by FY X |Z = 0; otherwise, |! | > |! XX | and FY X |Z > 0. 133

Hence, the cGC is always nonnegative (Schelter et al. 2006). 134

It may seem odd that (t) can contain the influence of 135

! Y (t) on X (t) and yet still be uncorrelated over time. Observe 136

101 X (t) = DXX j X (t j) + DXZ j Z (t j) + (t), that although the same time series X (t), Y (t), and Z (t) are 137

j=1 used in (1) and (3), the estimation of the two MVAR models 138

102 cov((t)) = ! , is done independently. Accordingly, any influence of Y (t) 139

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

140 on X (t) in (1) will be reflected in the different coefficient

141 estimates.

142 2.2 Conditional Granger causality in the frequency

143 domain

of
144 This section summarizes the steps needed to calculate the
145 cGC in the frequency domain following the procedures in
146 Geweke (1984a,b). To eliminate the instantaneous causality
147 [the correlation between the noises in (1) and (3)] and to

pro
148 ensure that the causality is due to interactions of the regres-
149 sion terms in (3), both (1) and (3) are normalized as in
150 Schelter et al. (2006), i.e., both equations in (1) are left- Fig. 1 Block diagram of Eq. (12)
Author Proof

151 multiplied by

" # HXX ( f ) HXY ( f ) HXZ ( f )
Im 0 If H( f ) = HYX ( f ) HYY ( f ) HYZ ( f ) is invert-
152 P1 = (6) 175

153
! ! 1
Ik HZX ( f ) HZY ( f ) HZZ ( f )

AXX ( f ) AXY ( f ) AXZ ( f )
154 and both sides of (3) are left-multiplied by ible with inverse A( f ) = AYX ( f ) AYY ( f ) AYZ ( f ) , 176

155
156

157
P = P2 P1 ,

where

Im 0 0

(7)
ted
then (11) can be written as

EX( f )

AXX ( f )
E Y ( f ) = AYX ( f )
EZ ( f ) AZX ( f )
AZX ( f ) AZY ( f ) AZZ ( f )

AXY ( f )
AYY ( f )
AZY ( f )
AXZ ( f )

AZZ ( f )

X( f )

AYZ ( f ) Y ( f )
Z( f )
177

178

158 P1 = ! YX ! 1

XX In 0 (8) (12)
rrec
! ZX ! 1
179
159 XX 0 Ik
where Aik ( f ) is the weight from signal k( f ) to E i ( f ). 180
160 and
Aik ( f ) is also a function of the effective weights from sig- 181

Im 0 0
nal k( f ) to signal i( f ), i.e., it is a function of Bik j for all 182

161 P2 = 0

)
In
*) *1
0 . time j and all states i {X, Y, Z } and k {X, Y, Z }. 183

0 ! ZY ! ZX ! 1
XX ! XY ! YY ! YX ! 1
XX ! XY Ik Figure 1 shows a block diagram of (12). Note that although 184

162
(9) A( f ) = H1 ( f ), the individual blocks are not equal, that 185
1
is, Aik ( f ) = Hik ( f ). Geweke uses some matrix identi- 186

163 To find the effect of Y (t) on X (t), given Z (t), (1) can be ties (Geweke 1984b), which are formalized in the following 187
o

164 written in the frequency domain as (Chen et al. 2006) lemma. 188

"
X( f )
# "
GXX ( f ) GXZ ( f )
#"
( f )
# m n
unc

165 = , (10) "

m A( f ) B( f )
#
166
Z( f ) GZX ( f ) GZZ ( f ) (f) Lemma 1 If T( f ) = , where A( f ) 189
n C( f ) D( f )
167 where Gik ( f ) is the transfer function constructed from Dik j and B( f ) are square matrices, and T1 ( f ) = 190

168 for all i {X, Y }, k {X, Y }, and j {1, 2, 3, . . .}. m n

m A ( f ) B ( f )
" #
169 Similarly, (3) can be written in the frequency domain as
, where A ( f ) and B ( f ) are square 191

n C ( f ) D ( f )
X( f ) HXX ( f ) HXY ( f ) HXZ ( f ) EX ( f ) matrices, then 192

170 Y ( f ) = HYX ( f ) HYY ( f ) HYZ ( f ) E Y ( f ) ,

Z( f ) HZX ( f ) HZY ( f ) HZZ ( f ) EZ ( f ) 1
B ( f )

A( f ) 0 B( f ) A(f) 0
(11) 0
171 I( f ) 0 = 0 I( f ) 0 . 193

C( f ) 0 D( f ) C ( f ) 0
D(f)
172 where Hik ( f ) describes the effect of E i ( f ) on k( f ) for all
(13)
173 i {X, Y, Z } and k {X, Y, Z }, and are functions of Bik j 194

174 for all time j and all states i and k.

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

195 Proof In Appendix 1. 3 Negative values of cGC? 226

196 Using Lemma 1, Eq. (11) can be substituted into (10) to To calculate the cGC in the time domain, the unknown para- 227

197 give meters in (1) and (3) can be calculated and substituted into 228

(5). Alternatively, these parameters can be substituted into 229

1
(10) and (11) to find (14). In the previous section, it is men-

( f ) GXX ( f ) 0 GXZ ( f ) X( f )

of
230

198 Y( f ) = 0 I 0 Y( f ) tioned that (15) and (17) are, in principle, the same. Since all 231

(f) GZX ( f ) 0 GZZ ( f ) Z( f ) the terms in (16) are positive semidefinite (PSD) matrices, 232

1 (16) itself is PSD and its determinant is greater than or equal 233
GXX ( f ) 0 GXZ ( f )
to the determinant of each of the terms in the right-hand side 234
199 = 0 I 0
of (16) (Zhang 2011). Hence, the numerator of (17) is always

pro
235
GZX ( f ) 0 GZZ ( f )
greater than or equal to the denominator. Therefore, irrespec- 236

HXX ( f ) HXY ( f ) HXZ ( f ) EX( f ) tive of how the parameters have been estimated, this equation 237
HYX ( f ) HYY ( f ) HYZ ( f ) E Y ( f ) is always 0. Hence, the cGC should always be nonnegative.
Author Proof

200 238

HZX ( f ) HZY ( f ) HZZ ( f ) EZ ( f ) However, although the numerator of (15) should be equal 239

QXX ( f ) QXY ( f ) QXZ ( f )

EX( f )
to the numerator of (17), estimation errors may occur. In par- 240

201 = QYX ( f ) QYY ( f ) QYZ ( f ) E Y ( f ) . (14) ticular, if the relationship between the estimated numerators 241

QZX ( f ) QZY ( f ) QZZ ( f ) EZ ( f ) and denominators of (15) and (17) happen to fulfill 242
202

|! | < |QXX ( f )! XX QXX ( f )| 243

203

204

205

206
Geweke (1984b) showed that the cGC from Y (t) to X (t)
given Z (t) can be expressed in the frequency domain as

fY X |Z ( f ) = ln
|! |
|QXX ( f )! XX QXX ( f )|
, (15)
ted < |QXX ( f )! XX QXX ( f ) + QXY ( f )! YY QXY ( f )
+ QXZ ( f )! ZZ QXZ ( f )|,

using the estimation of |! | and |QXX ( f )! XX QXX ( f )| to

calculate cGC will lead to negative values even though (17)
(19)
244

245

246

247

207 where ! is remains nonnegative. The following simulation demonstrates 248

rrec
how this may occur in a simple setting. 249

208 ! = QXX ( f )! XX QXX ( f ) + QXY ( f )! YY QXY ( f )

Example 1 Suppose three signals x(t), y(t), and z(t) are 250
209 +QXZ ( f )! ZZ QXZ ( f ). (16)
related through the following MVAR model, where the noises 251

ex (t), e y (t), and ez (t) are independent Gaussian white noise 252
210 Thus, (15) can be rewritten processes with zero means and variances x2 = 0.3, y2 = 253

0.1, and z2 = 0.2. 254

211 fY X |Z ( f )
|QXX ( f )! XX QXX ( f ) + QXY ( f )! YY QXY ( f ) + QXZ ( f )! ZZ QXZ ( f )|
212 = ln . x(t) = 0.8x(t 1) 0.5x(t 2) + 0.2y(t 2) 255
|QXX ( f )! XX QXX ( f )|
+0.4z(t 1) + ex (t),
o

213 (17) 256

y(t) = 0.9y(t 1) 0.8y(t 2) + e y (t), 257

214 Since (t) is uncorrelated over time, the power spectral z(t) = 0.5y(t 1) + 0.5z(t 1) 0.2z(t 2) + ez (t). 258
unc

215 density ! of ( f ) is constant over all frequencies and so

(20) 259
216 the numerator in (17) is constant. Of course, the denominator
217 is a function of frequency. Geweke (1984b) also showed that
This model was simulated to generate a dataset of 500 260
218 the cGC in the time and frequency domains are bound by the
realizations of 10,000 time points each. Then, the MVAR 261
219 inequality
parameters in (1) and (3) were estimated using a least squares 262

, method with MATLAB code from Omidvarnia (2011). The 263

1
220 fY X |Z ( f ) FY X |Z , (18) estimated parameters were substituted into (10) and (11) and 264
2
221
hence (14). Then, the cGC from y(t) to x(t) given z(t) was 265

- . calculated in two ways: first using (15), i.e., substituting the 266
QXX ( f ) QXY ( f ) QXZ ( f )
222 with equality if and only if det QYX ( f ) QYY ( f ) QYZ ( f )
! estimated with the least squares method, and second 267

QZX ( f ) QZY ( f ) QZZ ( f ) using (17), i.e., using the ! calculated in (16). Figure 2 268

223 = 0 for all frequencies f . presents the two results. Using (15) can lead to negative val- 269

224 The next section posits how negative values of cGC may ues at some frequencies. For instance, at frequency 0 Hz, 270

225 have been encountered in Chen et al. (2006). the numerator and the denominator of (17) are 0.3241 and 271

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

0.2 where 290

Equation (15)
Equation (17) " # " # " #1
E X ( f ) EX( f ) HXX ( f ) HXZ ( f )
= + 291
0.15 E Z ( f ) EZ ( f ) HZX ( f ) HZZ ( f )
" #
HXY ( f )
E Y ( f ). (23) 292
HZY ( f )

of
293
0.1
The goal is to make (22) equivalent to (10) by making 294

E X ( f ) and E Z ( f ) uncorrelated with each other (analogous 295

0.05 to ( f ) and ( f ) after normalization). To make this hap- 296

pen, ( f ) is defined as ( f ) = ' 1 , where '

ZX ' ZX and 297

pro
XX
' XX are the covariance matrices between E X and E Z and 298

0 the covariance matrix of E X , respectively. Then, by substi- 299

tuting # Gsub ( f ) is a normalized version of

( f ) into (21), 300
Author Proof

"
HXX ( f ) HXZ ( f )
in (22). 301
-0.05 HZX ( f ) HZZ ( f )
- 0 After calculating GXX ( f ), GXZ ( f ), GZX ( f ), and GZZ ( f ), 302
Normalized Frequency
and using the notation in (14), Q( f ) can be calculated 303

Fig. 2 For Example 1: calculating Gewekes version of cGC [from y(t)

QXX ( f ) QXY ( f ) QXZ ( f )

to x(t) given z(t)] using (15) and (17). The former may attain negative
Q( f ) = QYX ( f ) QYY ( f ) QY Z ( f )

272

273
values due to estimation errors, while the latter remains nonnegative for
all f

0.3190, respectively, and the numerator of (15) is 0.3050.

The result is a negative value for the cGC at f = 0 Hz.
ted =

QZX ( f ) QZY ( f ) QZZ ( f )
GXX ( f )
0
GZX ( f )
0
I( f )
0
GXZ ( f )
0
GZZ ( f )
1

HXX ( f ) HXY ( f ) HXZ ( f )

304

305

HYX ( f ) HYY ( f ) HY Z ( f ) . (24) 306

rrec
274 This example shows that even though (15) and (17) are HZY ( f ) HZY ( f ) HZZ ( f ) 307
275 equal in principle, numerical errors may cause meaningless
276 negative values in (15). These calculated parameters are then substituted into (17). 308

Since this approach uses partitioned matrices, we call this 309

the pGC. Barnett and Seth (2014) claim that, based on 310

unpublished simulations, the pGC appears to be inaccurate 311

277 4 The partition matrix method (meaning that its value does not align with Gewekes cGC) 312

and have observed in simulations that the inequality (18) may 313
278 Motivated at least in part by examples of negative cGC be violated when using the pGC. To investigate Barnett and 314
o

279 estimates, Chen et al. (2006) proposed a way to ensure that Seths claim, the next section uses a series of matrix equalities 315
280 the estimates would always remain nonnegative. Chen begins to simplify the pGC. The simplified equations help provide 316
281 by estimating GXX ( f ), GXZ ( f ), GZX ( f ), and GZZ ( f ) of (10) a better understanding of the ways that the pGC and the cGC 317
unc

282 from differ. 318

" #" #1
HXX ( f ) HXZ ( f ) Im 0
283 5 Simplification of the partition matrix approach 319
HZX ( f ) HZZ ( f ) #( f ) Ik
" #
GXX ( f ) GXZ ( f ) This section provides a simplified expression for the pGC 320
284 = = Gsub ( f ), (21)
285
GZX ( f ) GZZ ( f ) method described in the previous section. The first two lem- 321

mas are linear algebraic in nature and will be applied to steps 322

(21)(24) of the pGC calculation, leading to a simplification 323

286 where #( f ) is calculated based on the following calculations:
of (17). 324
287 The partitioned matrix method based on (11) gives
Lemma 2 If 325

" # " #" # " # " #" #

X( f ) HXX ( f ) HXZ ( f ) E X ( f ) X( f ) A11 ( f ) A12 ( f ) ( f )
288 = (22) = , (25) 326

289
Z( f ) HZX ( f ) HZZ ( f ) E Z ( f ) Y( f ) A21 ( f ) A22 ( f ) ( f ) 327

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

328 then 10 -16

7

X( f ) A11 ( f ) B12 ( f ) A12 ( f ) ( f )
329 T ( f ) = B21 ( f ) B22 ( f ) B23 ( f ) 0 , 6

Y( f ) A21 ( f ) B32 ( f ) A22 ( f ) ( f )

Abs(Difference Spectrum)
(26) 5

of
330

331 for any B12 ( f ), B22 ( f ), and B32 ( f ).B21 ( f ) and B23 ( f ) are 4
332 such that T ( f ) = B12 ( f )( f ) + B23 ( f )( f ).
3
333 Proof In Appendix 2.

pro
334 Lemma 3 Let A( f ) and C( f ) be two m n matrices such 2
335 that for all n k matrices B( f ), A( f )B( f ) = C( f )B( f ),
336 then A( f ) = C( f ).
Author Proof

337 Proof In Appendix 3.

0
338 Lemma 4 QXX ( f ) = I, for all frequencies f . - 0
Normalized Frequency
339 Proof In Appendix 4.
Fig. 3 For Example 1: the absolute value of the difference between two
Lemma 5 QXZ ( f ) = 0, for all frequencies f .
340

341

342

343

344
Proof In Appendix 5.
Lemma 6 QXY ( f ) = AXY ( f )A1
f.
Proof In Appendix 6.
YY ( f ), for all frequencies
ted ways of calculating the pGC [the partitioned method and the simplified
equation (27)]. Observe that all values are on the order of 1016 which
is effectively zero given the accuracy of the numerical representation

6 The difference between cGC and pGC 366

Lemma 4 shows that using the partitioned matrix method, 367

rrec
345 Gathering the results of the above lemmas, (21)(24) can QXX ( f ) = I. Accordingly, the denominator of (27) and the 368

346 be simplified, and an equivalent expression for (17) is given denominator of (17) are 369

347 in the following theorem.

|QXX ( f )! XX QXX ( f )| = |! XX |, (28) 370
348 Theorem 1 The pGC from Y (t) to X (t) given Z (t), which 371

349 is the cGC from Y (t) to X (t) given Z (t) calculated using the
which is a constant for all f . If the numerators of (27) and 372
350 partitioned matrix method, is
(17) were also constant for all frequencies [as shown in (16)], 373

351 pfY X |Z ( f ) the pGC would not vary with frequency. This apparent con- 374

) * ) * tradiction encourages a closer look at the calculations. 375

|! XX + AXY ( f )A1 1
o

YY ( f ) ! YY AXY ( f )AYY ( f ) | Given three signals X (t), Y (t), and Z (t), the partitioned 376

352 = ln , matrix method first estimates the MVAR model between all 377
|! XX |
353 (27) three signals, i.e., the parameters of (11). Then, the estimated 378
unc

parameters of (11) are used to estimate the parameters of 379

354 where Aik ( f ) is defined in (12). (10). In other words, (10) is estimated conditioned on (11). 380

In contrast, Granger and Geweke estimate the two models 381

355 Proof In Appendix 7. (10) and (11) independently. Based on (21), 382

356 Observe that by using (27), there is no need to carry out " #
GXX ( f ) GXZ ( f )
357 the calculations in (21)(24) to find (17). Instead, a subset of 383
GZX ( f ) GZZ ( f )
358 the estimated MVAR model parameters in (12) can be used
" #" #1
359 directly in (27). HXX ( f ) HXZ ( f ) Im 0
= 384
360 Figure 3 presents the absolute difference between the pGC HZX ( f ) HZZ ( f ) #( f ) Ik
361 from y(t) to x(t) given z(t), for example 1, calculated using "
HXX ( f ) HXZ ( f )
#"
Im 0
#
362 the partitioned method (21)(24) and the simplified equation = 385
HZX ( f ) HZZ ( f ) #( f ) Ik
363 (27), for f (, ). Clearly, the two plots are the same. In " #
Section 6, the simplified equation (27) will be used to study GXX ( f ) HXZ ( f )
364
= . (29) 386

365 how the pGC differs from the cGC. GZX ( f ) HZZ ( f ) 387

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

(a) (b) (c)

0.44 0.4
Denominator (Prtition Matrix Mehod) pGC
0.34 Denominator (Geweke) cGC
0.42 0.35

0.4 Numerator (Prtition Matrix Mehod) 0.32 0.3

Numerator (Geweke)
MVAR Estimated Power
0.38 0.25

of
0.3
0.36 0.2
0.28
0.34 0.15

0.32 0.26 0.1

pro
0.3 0.05
0.24
0.28 0
- 0 - 0 - 0
Author Proof

Normalized Frequency Normalized Frequency Normalized Frequency

Fig. 4 For example 1: a numerator of (15) calculated using multiple methods, b denominator of (15) calculated using multiple methods, c pGC
and cGC from y(t) to x(t) given z(t)

pGC
388 So, GXZ ( f ) and GZZ ( f ) in (10) are exactly equal to noise term ! in (1) (call this ! ) must be greater than 420

HXZ ( f ) and HZZ ( f ) in (11), showing explicitly how the par- pGC
or equal, i.e, |! cGC
389

390

391

392

393
tition matrix estimations are dependent. In turn, the noise
terms of the partitioned matrix method are
"
( f )
(f)
#
=
"
Im
#( f )
0
Ik
#"
E X ( f )
E Z ( f )
#
.
ted / cGC /
/! /

|! XX |

/
| |! |. Hence, in (5),

/ pGC /
/! /
|! XX |
/

As we have shown, the pGC and cGC are not interchange-

(30)
421

422
423

424

This shows that ( f ) is required to be a specific function able. Accordingly, the integral over frequency of the pGC 425
rrec
394

395 of E X ( f ), E Y ( f ), and E Z ( f ). Thus, these terms are struc- does not need to be equal to the integral over frequency 426

396 turally dependent in the pGC but independent in the cGC. To of the cGC, explaining Barnetts observation that (18) can 427

397 see this in a concrete example, Fig. 4a, b present the numer- be violated for the pGC (Barnett and Seth 2014). Similarly, 428

398 ator and denominator of the pGC (27) for Example 1. The for Gaussian random variables, the cGC is equal to transfer 429

399 numerator, which is the power spectral density of ( f ), is entropy multiplied by 2 (Barnett et al. 2009); again, this need 430

400 not constant over all frequencies. In the cGC, this same term not hold for pGC. Moreover, the pGC is not marginalized 431

401 is almost constant at all frequencies, as in the example 1. over Z (t), since it appears in the approximations of Aik ( f ) 432

402 Since ( f ) estimated by the partition matrix method is not for all i, k {X, Y }. 433

uncorrelated over time, the numerator of (17) will not be con-

o

403

404 stant. Thus, the pGC is not the same as the cGC. Rather, it
405 provides a different measure of the interrelatedness of the 7 What does the pGC mean? 434

signals.
unc

406

407 The pGC, like the cGC, is a ratio of the covariance of the To gain some insight into (27), rewrite (12) as 435

408 noises represented by the MVAR models in (1) and (3). The
409 difference is that the estimated MVAR models for the pGC AXX ( f )X ( f ) = AXY ( f )Y ( f ) AXZ ( f )Z ( f ) + E X ( f ) 436

410 are dependent, and the noises are not necessarily uncorrelated AYY ( f )Y ( f ) = AYX ( f )X ( f ) AY Z ( f )Z ( f ) + E Y ( f ) 437
411 over time. Like the cGC, the pGC is always 0. Like the
AZZ ( f )Z ( f ) = AZX ( f )X ( f ) AZY ( f )Y ( f ) + E Z ( f ). 438
412 cGC, the pGC is zero when Y (t) has no influence on X (t)
and Z (t). In this case, HXY"( f ) and H f ) will be#0 for all (31)
#ZY ( "
439
413

E X ( f ) EX( f )
414 frequencies, hence in (23), = , and so
E Z ( f ) EZ ( f ) Equation (31) can be represented as the block diagram 440

415 ( f ) = E X ( f ) and ( f ) = E Z ( f ). Thus, both pGC and in Fig. 5. X ( f ), Y ( f ), and Z ( f ), representing the network 441 1
416 cGC are 0. nodes, are shown by the orange circles. Other terms in (31) 442

417 The cGC in (5) is estimated so that the noise term ! which are functions of the main states are also represented by 443

418 [introduced in (1) and calculated in (16), call this ! cGC ] is as orange circles. E X ( f ), E Y ( f ), and E Z ( f ) represent inputs 444

419 small as possible. Accordingly, with pGC, the corresponding to the states and are shown by green boxes. The weights of 445

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

0.5
pCG
0.45 cGC

0.4

0.35

of
0.3

0.25

0.2

pro
0.15

Fig. 5 Block diagram of Eq. (31)

Author Proof

0.1

0.05

0
0 10 20 30 40 50 60
Frequency (Hz)

Fig. 6 Eliminating all nodes and edges related to Z ( f ) from Fig. 5

ted
Fig. 7 For example 2: pGC and cGC from P4 to P3 given C Z

that AYY ( f )Y ( f ) = AYX ( f )X ( f ) + E Y ( f ) is satisfied.

As AYX depends on the physics of the network, and there is
no mathematical constraint on it, it can be any n m matrix.
472

473

474

Therefore, X ( f ) and E Y ( f ) are dependent, and as a result, 475

the edges which are presented in yellow boxes indicate how
rrec
446
AYX ( f )X ( f ) and E Y ( f ) are dependent. This explains why 476
447 the states are related to each other. The pGC (27) is not a pfY X |Z ( f ) = , when AYY ( f ) is not invertible. 477
448 function of AXZ ( f ), AY Z ( f ), AZZ ( f ), AZX ( f ), AZY ( f ), or
449 ! ZZ . Hence, to calculate the pGC from signal Y (t) to X (t)
450 given Z (t), all the edges and nodes related to those weights
451 and parameters shown in Fig. 5 are not needed. So, Fig. 5 sim- 8 Examples 478

452 plifies to Fig. 6. In Fig. 6, it can be seen that AXX ( f )X ( f )

453 is effected by two inputs, i.e., E X ( f ) and E Y ( f ), through This section compares the cGC and pGC for two examples: 479

454 two distinct paths: the former comes directly from E X ( f ), Example 1 and an example containing EEG dataset. 480

455 and the latter comes through AXY ( f )A1 YY ( f )E Y ( f ). Based

For both examples, the cGC is estimated in the frequency 481
o

456 on (27), conditional Granger causality is the ratio of the domain by estimating the MVAR parameters and variances of 482

457 power of the received signals by AXX ( f )X ( f ) from the two the noises in (1) [the MATLAB code used for these calcula- 483

inputs E X ( f ) and E Y ( f ). The numerator is the power of the tions can be found in Omidvarnia (2011)] and then separately 484
unc

458

459 received signals in the presence of E Y ( f ), and the denomi- estimating parameters and noises in (3). The cGC in the time 485

460 nator is the power of the received signals in the absence of domain, i.e., (5), can be calculated using these estimated vari- 486

461 E Y ( f ); therefore, the pGC is nonnegative, as the numerator ance matrices. Then, using the estimated MVAR parameters 487

462 is always greater than or equal to the denominator. in the two models, the cGC is calculated using (10)(17). 488

463 Based on Fig. 6, For the pGC, the estimated MVAR parameters of (3), which 489

were calculated in the previous steps, were used to calculate 490

464
465
AYY ( f )Y ( f ) = AYX ( f )X ( f ) + E Y ( f ). (32) (11) and (12), and then substituted into the simplified pGC 491

(27). 492

466 Assuming that Y ( f ) Rn and AYY ( f ) Rnn , if For example 1, the cGC and pGC from y(t) to x(t) given 493

467 AYY ( f ) is invertible, then for any AYX ( f )X ( f ) and E Y ( f ), z(t) in the frequency domain are shown in Fig. 4c. Clearly, 494

468 there is a unique Y ( f ). However, if AYY ( f ) is not invert- the cGC and the pGC are not the same. In this example, if 495

469 ible, there is a constraint on the dimension of the space of the pGC is substituted into (18) and compared with cGC that 496

470 AYX ( f )X ( f ) + E Y ( f ), hence AYX ( f )X ( f ) and E Y ( f ) was calculated in time domain in the first step, it can be seen 497

471 cannot be any vectors, so they depend on each other such that the inequality (18) is not satisfied. 498

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

499 Example 2 This example applies the pGC and cGC measures Appendix 4: Proof of Lemma 4 533

500 to 20 s of EEG data which is sampled at 128 Hz (Omidvarnia

501 et al. 2011). The cGC and pGC from P4 to P3 given C Z in Proof Since (AB)1 = B1 A1 , (21) can be rewritten as 534

502 the frequency domain are shown in Fig. 7. They are clearly
#1
unequal.
" #"
503
Im 0 HXX ( f ) HXZ ( f )
G1
sub ( f ) = . 535

of
#( f ) Ik HZX ( f ) HZZ ( f )
(34) 536
504 9 Conclusion
" #1
HXX ( f ) HXZ ( f )
505 Chen et al. (2006) proposed the partition matrix technique Express the inverse in form as 537
HZX ( f ) HZZ ( f )

pro
506 pGC as a surrogate for the cGC to overcome possible negative " #
L( f ) M( f )
507 values observed in their studies. This paper shows how the , then 538
N( f ) K( f )
508 pGC differs from the cGC. The primary difference is that the
Author Proof

509 pGC is a function of the limited number of parameters in the " #" #
MVAR model. The pGC thus provides an alternative way to Im 0 L( f ) M( f )
510
G1
sub ( f ) = 539
511 measure a conditional form of causality and our simplified #( f ) Ik N( f ) K( f )
" #
512 expression lays bare exactly where the two measures differ. L( f ) M( f )
= 540
#( f )L( f ) + N( f ) #( f )M( f ) + K( f )
" #1
GXX ( f ) GXZ ( f )
= . (35)
513

514

515
Appendix 1: Proof of Lemma 1

Proof Equation (13) holds because

A( f )
0
0
I( f )
B( f )

0 0
A(f) 0
I( f )
B ( f )
0

ted G( f ) =

GZX ( f ) GZZ ( f )

Moreover, according to (13) in Lemma 1

GXX ( f ) 0 GXZ ( f )
0 I 0
1
541
542

543

544

C( f ) 0 D( f ) C(f) 0 D(f) GZX ( f ) 0 GZZ ( f )
rrec

A(f) 0 B(f) A( f ) 0 B( f ) L( f ) 0 M( f )
516 = 0 I( f ) 0 0 I( f ) 0 = 0 I 0 . 545

C(f) 0 D(f) C( f ) 0 D( f ) #(f) L(f) + N(f) 0 #(f) M(f) + K(f)
517 = I. (33) (36) 546
518

519 ,
Using (24), where Q( f ) = G( f )H( f ) 547

QXX ( f ) QXY ( f ) QXZ ( f )
o

520 Appendix 2: Proof of Lemma 2 Q( f ) = QYX ( f ) QYY ( f ) QY Z ( f ) 548

QZX ( f ) QZY ( f ) QZZ ( f )

521 Proof X ( f ) and Y ( f ) calculated from (26) are equal to L( f ) 0 M( f )
unc

522 X ( f ) and Y ( f ) calculated from (25).

, = 0 I( f ) 0 549

#(f) L(f) + N(f) 0 #(f) M(f) + K(f)

HXX ( f ) HXY ( f ) HXZ ( f )
HYX ( f ) HYY ( f ) HY Z ( f ) 550
523 Appendix 3: Proof of Lemma 3 HZX ( f ) HZY ( f ) HZZ ( f )

L( f )HXX ( f ) + M( f )HZX ( f ) QXY ( f ) QXZ ( f )
524 Proof If k = n, then an invertible matrix B( f ) can be chosen;
= QYX ( f ) QYY ( f ) QY Z ( f ) . 551
525 hence, A( f )B( f )B1 ( f ) = C( f )B( f )B1 ( f ). Therefore, QZX ( f ) QZY ( f ) QZZ ( f )
526 A( f ) = C( f ) (Kolman and Hill 2008). (37) 552
527 However, if k = n, then we can define Bi, j ( f ) whose
elements are zero except for one element in row i and column So, QXX ( f ) = L( f )HXX "( f ) + M( f )H#
ZX ( f ). On the
528
553
529 j, which is one. In that case, A( f )Bi, j ( f ) = C( f )Bi, j ( f ). L( f ) M( f )
Therefore, the ith column of A( f ) equals the ith column other hand, since the matrix is the inverse 554
530 N( f ) K( f )
531 of C( f ). Since this holds for all i {1, . . . , n}, A( f ) = "
HXX ( f ) HXZ ( f )
#
C( f ). ,
of , then 555
532
HZX ( f ) HZZ ( f )

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

" #" # " #

L( f ) M( f ) HXX ( f ) HXZ ( f ) I 0 E X ( f ) EX( f )
556 = .
N( f ) K( f ) HZX ( f ) HZZ ( f ) 0 I 0 = 0 582

557
(38) E Z ( f ) EZ ( f )
1
HXX ( f ) HXY ( f ) HXZ ( f )
558 Therefore, L( f )HXX ( f ) + M( f )HZX ( f ) = I, and conse- + HYX ( f ) HYY ( f ) HY Z ( f ) 583

of
559 quently QXX ( f ) = I. ,
HZX ( f ) HZY ( f ) HZZ ( f )

HXY ( f )
( f ) E Y ( f ), (43) 584

560 Appendix 5: Proof of Lemma 5 HZY ( f ) 585

pro
561 Proof According to (37), or 586

E X ( f ) EX( f )
Author Proof

562
563
QXZ ( f ) = L( f )HXZ ( f ) + M( f )HZZ ( f ), (39) 0 = 0 587

E Z ( f ) EZ ( f )
564 on the other hand from (38), it can be seen that L( f )HXZ ( f )+
AXZ ( f ) AXY ( f ) AXZ ( f )

565 M( f )HZZ ( f ) = 0; therefore, QXZ ( f ) = 0. ,
+ AYX ( f ) AYY ( f ) AY Z ( f ) 588

AZX ( f ) AZY ( f ) AZZ ( f )

566

567

568
Appendix 6: Proof of Lemma 6

Proof Given (37), QXY ( f ) = L( f )HXY ( f )+M( f )HZY ( f ).

Moreover, by using (38), E X ( f ) in (23), i.e.,
ted
HXY ( f )

HZY ( f )

( f ) E Y ( f ).

Equation (43) holds because, after substituting it into (42),

X ( f ) and Z ( f ) must satisfy (22). ( f ) should be defined
(44) 589

590

591

592

#1 such that 593

rrec
" # " # "
E X ( f ) EX( f ) HXX ( f ) HXZ ( f )
569 = +
E Z ( f ) EZ ( f ) HZX ( f ) HZZ ( f ) AYX ( f )HXY ( f ) + AYY ( f ) ( f ) + AY Z ( f )HZY ( f ) = 0. 594
" #
HXY ( f ) (45)
570 E Y ( f ), (40) 595

571
HZY ( f )
On the other hand as H1 ( f ) = A( f ), so 596

572 can be written as

" # " # " # AYX ( f )HXY ( f ) + AYY ( f )HYY ( f ) + AY Z ( f )HZY ( f ) = I. 597

E X ( f ) EX( f ) L( f ) M( f ) (46)
573 = + 598
o

E Z ( f ) EZ ( f ) N( f ) K( f )
" #
HXY ( f ) So, by adding AYY ( f )HYY ( f ) to both sides of (45) 599
574 E Y ( f ). (41)
HZY ( f )
unc

575

AYX ( f )HXY ( f ) + AYY ( f )HYY ( f ) + AYY ( f ) ( f ) 600

576 So, E X ( f ) = E X ( f ) + QXY ( f )E Y ( f ), where QXY ( f ) + AY Z ( f )HZY ( f ) = AYY ( f )HYY ( f ). (47) 601

577 has been calculated in (37). Based on Lemma 2, (22) can be

578 written as By substituting (46) into (47) 602

X( f ) HXX ( f ) HXY ( f ) HXZ ( f ) E X ( f ) I + AYY ( f ) ( f ) = AYY ( f )HYY ( f ). (48) 603
604
579 T ( f ) = HYX ( f ) HYY ( f ) HY Z ( f ) 0 ,
Z( f ) HZX ( f ) HZY ( f ) HZZ ( f ) E Z ( f )
(42) Therefore, 605
580

AYY ( f ) ( f ) = AYY ( f )HYY ( f ) I, 606

E X ( f )
581 where based on (41), 0 is ( f ) = A1
YY ( f )(AYY ( f )HYY ( f ) I), 607

E Z ( f ) ( f ) = HYY ( f ) A1
YY ( f ). (49) 608

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Biol Cybern

609 Replacing (49) into (44) gives Bollimunta A, Chen Y, Schroeder CE, Ding M (2008) Neuronal mecha- 641

nisms of cortical alpha oscillations in awake-behaving macaques. 642

J Neurosci 28(40):99769988 643
E X ( f ) EX( f ) Bressler SL, Richter CG, Chen Y, Ding M (2007) Cortical functional 644
610 0 = 0 network organization from autoregressive modeling of local field 645

E Z ( f ) EZ ( f ) potential oscillations. Stat Med 26:38753885. doi:10.1002/sim. 646

2935 647

of

AXZ ( f ) AXY ( f ) AXZ ( f ) Brovelli A (2012) Statistical analysis of single-trial granger causality 648
611 + AYX ( f ) AYY ( f ) AY Z ( f ) spectra. Comput Math Methods Med. doi:10.1155/2012/697610 649

AZX ( f ) AZY ( f ) AZZ ( f ) Chen Y, Bressler SL, Ding M (2006) Frequency decomposition 650

of conditional Granger causality and application to multivari- 651

HXY ( f ) ate neural field potential data. J Neurosci Methods 150(2): 652

612 HYY ( f ) A1YY ( f ) . (50) 228237 653

pro

HZY ( f ) Dhamala M, Rangarajan G, Ding M (2008) Analyzing information flow 654
613
in brain networks with nonparametric Granger causality. Neuroim- 655

age 41(2):354362 656

614 The equation for E X ( f ) from (50) can be simplified as Geweke J (1984a) Inference and causality in economic time series. 657
Author Proof

Handb Econom II:11021143 658

Geweke J (1984b) Measures of conditional linear dependence and feed- 659

615 E X ( f ) = E X ( f ) + (AXZ ( f )HXY ( f ) + AXY ( f )(HYY ( f ) back between time series. J Am Stat Assoc 79(388):907915 660

616 A1
YY ( f )) + AXZ ( f )HZY ( f ))E Y ( f )
Gourvitch B, Le Bouquin-Jeanns R, Faucon G (2006) Linear and 661

nonlinear causality between signals: methods, examples and neu- 662

617 = E X ( f ) + (AXZ ( f )HXY ( f ) + AXY ( f )HYY ( f ) rophysiological applications. Biol Cybern 95:349369. doi:10. 663

618 AXY ( f )A1

YY ( f ) + AXZ ( f )HZY ( f ))E Y ( f )
1007/s00422-006-0098-0 664

619

620

621

622
= E X ( f ) AXY ( f )A1
YY ( f )E Y ( f ).

On the other hand, based on (41), E X ( f ) = E X ( f ) +

QXY ( f )E Y ( f ) for all E Y ( f ), so based on Lemma 3,
QXY ( f ) = AXY ( f )A1YY ( f ) .
,
ted
(51)
Granger CWJ (1969) Investigating causal relations by econometric
models and cross-spectral methods. Econometrica 37(3):424438
Kamiski M, Ding M, Truccolo WA, Bressler SL (2001) Evaluating
causal relations in neural systems: Granger causality, directed
transfer function and statistical assessment of significance. Biol
Cybern 85:145157. doi:10.1007/s004220000235
Kolman B, Hill DR (2008) Elementary linear algebra with application,
9th edn. Pearson Education, Pearson
665

666

667

668

669

670

671

672

Leung B, Cheung P, Riedner BA, Tononi G, Van Veen BD (2010) Esti- 673
rrec
mation of cortical connectivity From EEG using state-space mod- 674
623 Appendix 7: Proof of Theorem 1 els. IEEE Trans Biomed Eng 57(9):21222134. http://ieeexplore. 675

ieee.org/stamp/stamp.jsp?tp=&arnumber=5471144 676

624 Proof Based on Lemmas 46, the equation for conditional Liao W, Mantini D, Zhang Z, Pan Z, Ding J, Gong Q, Yang Y, 677

Chen H (2010) Evaluating the effective connectivity of resting 678

625 Granger causality, i.e., (17), can be simplified as follows
state networks using conditional Granger causality. Biol Cybern 679

|! |
102(1):5769 680
626 fY X |Z ( f ) = ln // Malekpour S, Li Z, Cheung BLP, Castillo EM, Papanicolaou AC, 681
QXX ( f )! XX QXX ( f )/
/
Kramer LA, Fletcher JM, Van Veen BD (2012) Interhemi- 682
/QXX ( f )! XX Q ( f ) + QXY ( f )! YY Q ( f ) + QXZ ( f )! ZZ Q ( f )/
/ /
XX XY XZ spheric effective and functional cortical connectivity signatures of 683
627 = ln
/QXX ( f )! XX Q ( f )/
/ /
XX
spina bifida are consistent with callosal anomaly. Brain Connect 684

2(3):142154
o

/ ) * ) * / 685
/! XX + AXY ( f )AYY ( f ) ! YY AXY ( f )A1
1
YY ( f ) /
/ /
Omidvarnia A (2011) Time-varying EEG connectivity: a time- 686
628 = ln
|! XX | frequency approach. http://www.mathworks.com/matlabcentral/ 687
/ ) * / fileexchange/33721-time-varying-eeg-connectivity-a-time- 688
/! XX + AXY ( f )A1 ( f ) ! YY (AXY ( f )A1 /
YY ( f )) /
/
unc

YY frequency-approach/content/WOSSPA_Mathworks_v2/mvar.m
629 = ln . (52) 689 2
|! XX | Omidvarnia A, Mesbah M, OToole JM, Colditz P, Boashash B 690

(2011) Analysis of the time-varying cortical neural connectivity 691

630 ,
in the newborn EEG: a time-frequency approach. In: 7th interna- 692

tional workshop on systems, signal processing and their applica- 693

tions (WoSSPA 2011), pp 179182. doi:10.1109/WOSSPA.2011. 694

5931445 695
631 References Schelter B, Winterhalder M, Timmer J (2006) Handbook of time series 696

analysis. Wiley-VCH, Weinheim. doi:10.1002/9783527609970 697

632 Barnett L, Seth AK (2014) The MVGC multivariate Granger causality Wang X, Chen Y, Bressler SL, Ding M (2007) Granger causal- 698

633 toolbox: a new approach to Granger-causal inference. J Neurosci ity between multiple interdependent neurobiological time series: 699

634 Methods 223:5068. doi:10.1016/j.jneumeth.2013.10.018 blockwise versus pairwise methods. Int J Neural Syst 17(2):7178. 700

635 Barnett L, Barrett AB, Seth AK (2009) Granger causality and trans- doi:10.1142/S0129065707000944 701

636 fer entropy are equivalent for gaussian variables. Phys Rev Lett Zhang F (2011) Matrix theory basic results and techniques, 2nd edn. 702

637 103(23):238701 Springer, New York 703

638 Bernasconi C, Knig P (1999) On the directionality of cortical interac- Zhou Z, Chen Y, Ding M, Wright P, Lu Z, Liu Y (2009) Analyzing brain 704

639 tions studied by structural analysis of electrophysiological record- networks with PCA and conditional granger causality. Hum Brain 705

640 ings. Biol Cybern 81:199210. doi:10.1007/s004220050556 Mapp 30:21972206. doi:10.1002/hbm.20661 706

123
Journal: 422 MS: 0665 TYPESET DISK LE CP Disp.:2015/10/9 Pages: 11 Layout: Large
 Journal: 422
Article: 665

Author Query Form

of
Please ensure you fill out your response to the queries raised below
and return this form along with your corrections

Dear Author

pro
During the process of typesetting your article, the following queries have arisen. Please check your typeset proof carefully
against the queries listed below and mark the necessary changes either directly on the proof/online grid or in the Authors
Author Proof

response area provided below

Query Details required Authors response

1. As per the information provided by the publisher, Figure
3 will be black and white in print; please interpret the

2.
ted
color information in text near the figure citation.
Please provide publisher name and the location for the
reference Omidvarnia (2011).
o rrec
unc

View publication stats