The Geometry of the Wilks s Λ Random Field

Transcription

1 The Geometry of the Wilks s Λ Random Field F. Carbonell Institute for Cybernetics, Mathematics and Physics Havana L. Galan Cuban Neuroscience Center Havana August 19, 2005 K. J. Worsley McGill University Montreal Abstract The statistical problem addressed in this paper to approximate the P-value of the maximum of a smooth random field of Wilks s Λ statistics. So far results are only available for the usual univariate statistics (Z, t, χ 2, F ) and a few multivariate statistics (Hotelling s T 2, maximum canonical correlation, Roy s maximum root). We derive results for any differentiable scalar function of two independent Wishart random fields, such as Wilks s Λ and Pillai s trace. We apply our results to a problem in brain shape analysis. Key words and phrases: Multivariate Random Fields, Excursion Sets, Euler Characteristic, Derivatives of Matrix Functions. MSC 2000: 60G60, 60D05,62M40, 1 Introduction The motivation for this work came from practical applications in brain imaging. Changes in brain shape can be inferred from the 3D non-linear deformations required to warp a patient s MRI scan to an atlas standard. Using these vector deformations as a dependent variable, we can use a multivariate multiple regression model to detect regions of the brain whose shape is correlated with external regressors, such as disease state. The statistic of choice is Wilks s Λ, the likelihood ratio statistic (Anderson 1984), evaluated at each voxel in the brain. The challenging statistical problem is how to assess the P-value of local maxima of a smooth 3D random field (RF) of Wilks s Λ statistics. This problem of assessing the P-value of local maxima was studied in the seminal works of Adler & Hasofer (1976) and Hasofer (1978) for the case of Gaussian RFs. The key idea consists on approximating the P-value that the local maximum exceeds a high threshold value via the geometry of the excursion set, the set of points the RF exceeds the threshold value. For high thresholds, the excursion sets consists of isolated regions, each containing a local maxima. The Euler characteristic (EC) of the excursion sets counts the number of such connected components. For high thresholds near the maximum of the RF, the EC takes the value one if the maximum is 1

2 above the threshold, and zero otherwise. Thus, for high thresholds, the expected EC approximates the P-value of the maximum of the RF. During the last decade this approach has became an important tool in different areas of applied probability. This is due to the many applications that have emerged principally in medical imaging and astrophysics (Gott, Park, Juskiewicz, Bies, Bennett, Bouchet & Stebbins 1990, Beaky, Scherrer & Villumsen 1992, Vogeley, Park, Geller, Huchira & Gott 1994, Worsley 1995a, Worsley 1995b, Worsley, Cao, Paus, Petrides & Evans 1998, Cao & Worsley 2001). So far, results are available for some fields of standard univariate statistics such as Z, t, χ 2 and F RFs (Worsley 1994). Some other useful extensions were obtained in Cao & Worsley (1999a) and Cao & Worsley (1999b) for Hotelling s T 2 and the cross-correlation RFs, respectively. However, the extension of these results to other RFs of classical multivariate statistics has not so far been fully studied (Worsley, Taylor, Tomaiuolo & Lerch 2004). RFs of multivariate statistics appear when dealing with multivariate RF data and multiple contrasts, the classical example is the Wilks s Λ statistic mentioned above. The purpose of this work is to extend Adler s results to a wide class of RFs of multivariate statistics to which the Wilks s Λ RF belongs. We build these random fields from Gaussian random fields in the same way as we build the statistics themselves. First let Z = Z(t), t R N, be a smooth isotropic Gaussian random field with zero mean, unit variance, and with Var(Ż) = I N, (1) dot represents derivative with respect to t, and I N is the N N identity matrix. Note that if we have another isotropic Gaussian ransom field Z with Var(Ż ) = v 2 I N and v 1, then we can always re-scale the parameter t by dividing by v to define Z(t) = Z (t/v) so that (1) is satisfied. Now suppose that Z(t) is an ν m matrix independent Gaussian random fields each with the same distribution as Z(t). Then the Wishart random field with ν degrees of freedom is defined as W (t) = Z(t) Z(t) Wishart m (I m, ν) (2) at each fixed t (Cao & Worsley 1999b). Let U(t) and V(t) be independent Wishart random fields with ν and η degrees of freedom, respectively, whose component Gaussian random fields all have the same distribution as Z(t) above. We shall be concerned with the class of RFs Λ(t) defined as Λ (t) = Λ (U (t), V (t)), (3) Λ is any differentiable function of two matrix arguments. Our aim is to provide approximate values for the expected EC of the excursion sets of Λ(t). The plan of the paper is the following. In Section 2 we shall state the notation as well as some preliminaries facts about RF theory and matrix differential theory. In the Section 3 we obtain expressions for the derivatives up to second order of the random field Λ (t), while, in the Sections 4 and 5 we get the EC densities of Λ (t). Finally, we applied these results to detecting brain damage of a group of 17 non-missile trauma patients compared to a group of 19 age and sex matched controls in the Section 6. 2 Preliminaries In this section we shall settle the notation to be used throughout the paper, as well as some basic facts about RFs and matrix differentiation. 2

3 2.1 Basic Notations We use Normal m n (µ, Σ) to denote the multivariate normal distribution of a matrix X R m n with mean µ and covariance matrix Σ, and write Σ = M(I mn ) when cov(x ij, X kl ) = ε(i, j, k, l) δ ij δ kl with ε(i, j, k, l) symmetric in its arguments and with δ ij = 1 if i = j and 0 otherwise. Let W ishart m (ν, Σ) be represent the Wishart ( distribution of a m m matrix with expectation νσ and degrees of freedom ν, and use Beta 1 m ν, 1η) for the multivariate beta distribution of a m m 2 2 matrix with degrees of freedom ν and η. For any vector we shall use subscripts j and j to denote the jth component and the first j components, respectively. In a similar way, for any matrix, the subscript j shall denote the matrix of the first j rows and columns. For any scalar b, let b + = b if b > 0 and zero otherwise. For any symmetric matrix B, let B = B if B is negative definite and zero otherwise and let detr j (B) be the sum of the determinants of all j j principal minors of B and define detr 0 (B) = 1. We shall denote by B 1/2 the square root of the matrix B, which is defined by B 1/2 (B 1/2 ) = B. For any m m symmetric matrix B and any multi-indices κ = (k 1, k 2,..., k m ), k 1 k 2 k m, we denote with C κ (B) the zonal polynomials corresponding to κ (see details in Muirhead (1982)), which satisfy tr(b) k = C κ (B). (4) κ =k For any real a and natural k, (a) k shall be defined by (a) k = a(a + 1)... (a + k 1), (a) 0 = 1. For the multi-index κ = (k 1, k 2,..., k m ) define (a) κ = m (a 1(i 1)) 2 k i. i=1 Finally, for any real a > 1 2 (m 1), Γ m(a) will denote the multivariate Gamma function, defined by Γ m (a) = π m(m 1)/4 Γ(.) denotes the usual Gamma function. 2.2 Random Field Theory m i=1 Γ(a 1 (i 1)), 2 Let Y = Y (t), t S R N be a real valued isotropic random field and let S be a compact subset of R N with a twice differentiable boundary S. Denote the first derivative of Y by Y. and the N N matrix of second-order partial derivatives of Y by Ÿ. For j > 0 define the j-dimensional Euler Characteristic (EC) density as ρ j (y) = E(Ẏ + j det( Ÿ j 1) Ẏ j 1 = 0, Y = y) θ j 1 (0, y) (5) E ( ) denotes mathematical expectation and θ j 1 (, ) is the joint probability density function of Ẏ j 1 and Y. For j = 0 define ρ 0 (y) = P(Y y). Let a j = 2π j/2 /Γ(j/2) be the surface area of a unit (j 1) sphere in R N. Define µ N (S) to be the Lebesgue measure of S and µ j (S), j = 0,..., N 1 to be proportional to the j dimensional Minkowski functional or intrinsic volume µ j (S) = 1 a N j S 3 detr N 1 j (C) dt

4 with C denoting the inside curvature matrix of S at the point t. The expected EC χ of the excursion set of Y above the threshold y inside S (A y (Y, S)) is given by E(χ(A y (Y, S))) = N µ j (S) ρ j (y). j=0 Then, the P-value of the maximum of Z inside S is well approximated by ( ) P max Y (t) y E(χ(A y (Y, S))) (6) t S (Adler 2000, Taylor, Takemura & Adler 2005). In this paper Y is a function of independent RFs each with the same distribution as the standard Gaussian RF Z, and the EC densities are ρ j (y) are calculated assuming unit variance of the derivative (1). If instead Var(Ż) = v2 I m (7) then t can be re-scaled by dividing t by v to satisfy (1). Thus is equivalent to multiplying µ j (S) by v j, to give ( ) N P max Y (t) y E(χ(A y (Y, S))) = µ j (S)v j ρ j (y). (8) t S For this reason all the subsequent theory for the EC densities will assume v = Matrix Functions Throughout the next section we shall need the following result, which follows easily from the properties of the operators vec ( ) (vectorization) and (Kronecker tensor product). If the p q matrix X is such that X Normal p q (0, I pq ) then for any s p matrix A and q r matrix B, j=0 AXB Normal s r (0, (B B) (AA )). (9) We shall also make intensive use the differential theory of matrix argument functions proposed in Magnus & Neudecker (1988).). It is based on expressing the derivative of a matrix-valued function of a matrix argument F (X) : R p q R m n by the mn pq matrix D X F (X) = d (vec (F (X))) d ( vec (X)). 3 Representations of Derivatives In this section we obtain expressions for the first and second derivatives of the random field Λ (t) with respect to t, denoted by Λ and Λ respectively. These expressions generalize, for instance, the formulas for the derivatives of the F field obtained in Worsley (1994). Theorem 1 The two first derivatives of the random field Λ can be expressed as Λ = A 1 vec(m 0 (W, B)), 4

5 Λ = tr(m 1 (W, B))I N + tr(m 2 (W, B)) 1/2 H + tr(m 3 (W, B))P + tr(m 4 ( W, B))Q +A 1 M 5 (W, B)A 1 + A 1 M 6 (W, B)A 2 + A 2 M 7 (W, B)A 1 + A 2 M 8 (W, B)A 2, equality means equality in law, A 1, A 2 Normal N m 2(0, I Nm 2), H Normal N N (0, M (I N 2)), P W ishart N (ν m, I N ), Q W ishart N (η m, I N ), W W ishart m (ν + η, I m ), B Beta m ( 1 2 ν, 1 2 η), independently, and M i (W, B), i = 0,..., 8 are matrices that depends on B and W. Proof. For each j = 1,..., N we have Λ j = D U Λ (U, V) D tj U + D V Λ (U, V) D tj V. (10) According to Neudecker (1969) one is to able to find two matrix functions G 1 (U, V), G 2 (U, V), such that G 1, G 2 : R m m R m m R m m D U Λ (U, V) = vec(g 1 ), D V Λ (U, V) = vec(g 2 ). From Lemma 3.2 in Cao & Worsley (1999a) we make the substitutions independently. Thus, D tj U = vec(u 1/2 A U j + (U 1/2 A U j ) ), (11) D tj V = vec(v 1/2 A V j + (V 1/2 A V j ) ), (A U 1,..., A U N), (A V 1,..., A V N) Normal m mn (0, I m 2 N), Λ j = vec(g 1 ) vec(u 1/2 A U j + (U 1/2 A U j ) ) + vec(g 2 ) vec(v 1/2 A V j + (V 1/2 A) ) = tr(((g 1 ) + G 1 )U 1/2 A U j ) + tr(((g 2 ) + G 2 )V 1/2 A V j ) = tr(gu 1/2 A U j + FV 1/2 A V j ) = tr(c 1 1 A 1 j), G = (G 1 ) + G 1, F = (G 2 ) + G 2, C 1 = (GUG + FVF) 1/2, A 1 j = C 1 (GU 1/2 A U j + FV 1/2 A V j ), j = 1,..., N. It is easily verified from (9) that, conditional on U, V, A 1 j Normal m m (0, I m 2), which yields Λ = A 1 vec(c 1 1 ), 5

6 Using the fact that A 1 = (vec(a 1 1),..., vec( A 1 N)) Normal N m 2(0, I m 2 N). W = U + V W ishart m (ν + η, I m ), B = (U + V) 1/2 U(U + V) 1/2 Beta m ( 1 2 ν, 1 2 η), independently (Olkin & Rubin 1962), we obtain the first identity by rewriting the matrix C 1 1 in terms of W and B by means of the substitutions For the second identity we obtain from (10) U = W 1/2 BW 1/2, V = W 1/2 (I m B)W 1/2. (12) Λ kj = (D tj U) D 2 UUΛ (U, V) D tk U + (D 2 VUΛ (U, V))D tk V (13) +(D tj V) D 2 UVΛ ( U, V))D tk U + (D 2 VVΛ ( U, V))D tk V + (D U Λ (U, V)) D 2 t k t j U + (D V Λ (U, V)) D 2 t k t j V, From Lemma 3.2 in Cao & Worsley (1999a) we make the substitutions D 2 t k t j U = vec( P jk + P kj + U 1/2 H U kj + (U 1/2 H U kj) + (A U j ) A U k + (A U k ) A U j 2δ kj U), (14) D 2 t k t j V = vec( Q jk + Q kj + V 1/2 H V kj + (V 1/2 H V kj) + (A V j ) A V k + (A V k ) A V j 2δ kj V ), The expressions (11) and (14) yields and P = ( P jk ) W ishart mn (ν m, I mn ), Q = ( Q jk ) W ishart mn (η m, I mn ), H U kj, H V kj Normal m m (0, M (I m 2)). (D tj U) (D 2 UUΛ)D tk U = vec(u 1/2 A U j ) (I m 2 + K m)(d 2 UUΛ) (I m 2 + K m ) vec(u 1/2 A U k ), (15) (D tj U) (D 2 VUΛ)D tk V = vec(u 1/2 A U j ) (I m 2 + K m)(d 2 VUΛ) (I m 2 + K m ) vec(v 1/2 A V k ), (D tj V) (D 2 UVΛ)D tk U = vec(v 1/2 A V j ) (I m 2 + K m)(d 2 UVΛ) (I m 2 + K m ) vec(u 1/2 A U k ), (D tj V) (D 2 VVΛ)D tk V = vec(v 1/2 A V j ) (I m 2 + K m)(d 2 VVΛ) (I m 2 + K m ) vec(v 1/2 A V k ), D U Λ(D 2 t k t j U) = 2δ kj vec(g 1 ) vec (U) + vec(g) vec( P jk ) + vec(g) vec(u 1/2 H U jk) +vec(g) vec((a U j ) A U k ), D V Λ(D 2 t k t j V) = 2δ kj vec(g 2 ) vec (V) + vec(f) vec( Q jk ) + vec(f) vec(v 1/2 H V jk) +vec(f) vec((a V j ) A V k ). (16) We can now find a linear combination of A U j and A V j, j = 1,..., N that is independent of A 1 j, namely, A 2 j = C 2 (G 1 U 1/2 A U j F 1 V 1/2 A V j ), C 2 = ((GUG) 1 + (FVF) 1 ) 1/2. 6

7 This implies that the matrices A U j, A V j can be written in terms of A 1 j and A 2 j by suitable expressions (unequivocally determined), which when substituted in each of the right members of (15) and (16) give Λ kj = 2δ kj tr(g 1 U + G 2 V) + tr(c 1 1 H kj ) + tr(g P jk ) + tr(f Q jk ) (17) +(vec(a 1 j)) S 1 vec(a 1 k) + (vec(a 1 j)) S 2 vec(a 2 k) +(vec(a 2 j)) S 3 vec(a 1 k) + (vec(a 2 j)) S 4 vec(a 2 k), for certain matrices S 1, S 2, S 3, S 4 that only depends on U, V, and H kj given by H kj = C 1 (GU 1/2 H U jk + FV 1/2 H V kj). It can be easily shown that the matrix (tr(c 1 1 H kj )) kj can be rewritten as with (tr(c 1 1 H kj )) kj = tr(c 2 1 ) 1/2 H, H = tr(c 2 1 ) 1/2 (I N (vec(c 1 1 )) )(vec(h kj )) kj, (18) which, conditional on U, V, is distributed as Normal N N (0, M (I N 2)). On the other hand, noting that we define the matrices (tr(g P kj )) kj = (I N vec(g 1/2 ) )(I m P kj ) kj (I N vec( G 1/2 )), (tr(f Q kj )) kj = (I N vec(f 1/2 ) )(I m Q kj ) kj (I N vec( F 1/2 )), (I m P kj ) kj W ishart m 2 N (ν m, I m 2 N), (I m Q kj ) kj W ishart m 2 N(η m, I m 2 N), P = (tr(g 1 )) 1 (I N vec(g 1/2 ) )(I m P kj ) kj (I N vec(g 1/2 )), (19) Q = (tr(f 1 )) 1 (I N vec(f 1/2 ) )(I m Q kj ) kj (I N vec(f 1/2 )), which distribute independently as W ishart N (ν m, I N ) and W ishart N (η m, I N ), respectively. It is also easily verified that (vec(a 1 j) S 1 vec(a 1 k)) kj = A 1 S 1 A 1, (vec(a 2 j) S 2 vec(a 2 k)) kj = A 2 S 2 A 2, (vec(a 1 j) S 3 vec(a 2 k)) kj = A 1 S 3 A 2, (vec(a 2 j) S 4 vec(a 1 k)) kj = A 2 S 4 A 1, with A 2 = (vec(a 2 1),..., vec(a 2 N )) Normal N m 2 (0, I m 2 N) independently of A 1. Combining these last expressions with (18) and (19) and substituting in (17) we obtain Λ = 2tr(G U U + G V V)I N + tr(c 2 1 ) 1/2 H + tr ( G 1) P + tr ( F 1) Q +A 1 S 1 A 1 + A 1 S 2 A 2 + A 2 S 3 A 1 + A 2 S 4 A 2. The substitutions (12) conclude the proof. 7

8 4 Euler Characteristic Densities In this section, we obtain expressions for the EC densities of the random field Λ. We begin with the following proposition. Proposition 2 The following equalities hold Λ j = tr (M 2 ) 1/2 z Λ j ( Λ j = 0) = tr(m 1 )I j + tr (M 2 ) 1/2 H + tr (M 3 ) P + tr (M 4 ) Q + X 1 M 5 X 1 +X 1 M 6 X 2 + X 2 M 7 X 1 + X 2 M 8 X 2, independently, with Σ given by X 1 Normal j m 2(0, Σ I j ), X 2 Normal j m 2(0, I jm 2), z Normal j (0, I j ), H Normal j j (0, M(I j 2)), P W ishart j (ν m, I j ), Q W ishart j (η m, I j ) W W ishart m (ν + η, I m ), B Beta m ( 1 2 ν, 1 2 η) Proof. Define the vector b and the matrix C by Σ = I m 2 m 2 tr(m 2) 1 vec(m 0 )vec(m 0 ). b = A j 1 vec (M 0 ) and C = (A j 1, b), respectively. It is easily seen from the definition of K 0 and K 2 in the theorem above that which implies vec (M 0 ) vec (M 0 ) = tr(m 2 ), b Normal j (0, tr(m 2 ) I j ). The first equality follows by noting that Λ j = b. Moreover it is very easy to check that C Normal j m 2 +1(0, Σ), ( Ijm Σ = 2 jm 2 vec(k ) 0) I j vec(k 0 ). I j tr(k 2 )I j From Theorem and Theorem in Mardia & Bibby (1979) we obtain independently of b, A j 1 (b = 0) Normal j m 2(0, Σ 1 ), (20) Σ 1 = I jm 2 jm 2 (vec(m 0) I j )(tr(m 2 )I j ) 1 (vec(m 0 ) I j ) = I jm 2 jm 2 tr(m 2) 1 vec(m 0 )vec(m 0 ) I j = (I m 2 m 2 tr(m 2) 1 vec(m 0 )vec(m 0 ) ) I j. 8

9 According to Theorem 1 the condition ( Λ j = 0) is equivalent to b = 0. This implies, by (20), Λ j ( Λ j = 0) = tr(m 1 )I N + tr (M 2 ) 1/2 H + tr (M 3 ) P + tr (M 4 ) Q + X 1 M 5 X 1 + X 1 M 6 X 2 +X 2 M 7 X 1 + X 2 M 8 X 2, X 1 Normal j m 2(0, Σ), X 2 Normal j m 2(0, I jm 2) and Σ = I m 2 m 2 tr(m 2) 1 vec(m 0 )vec(m 0 ), (21) which concludes the proof The next theorem is the main result of this section. The EC densities are obtained only for j = 1, 2, 3, which are the most important cases in practical applications. Theorem 3 The EC densities ρ j (λ), j = 1, 2, 3 for the random field Λ are given by ϕ j (λ) are expressions of the type ρ 1 (λ) = (2π) 1/2 θ 0 (λ) ϕ 1 (λ), ρ 2 (λ) = (2π) 1 θ 0 (λ)ϕ 2 (λ) ρ 3 (λ) = (2π) 3/2 θ 0 (λ) ϕ 3 (λ), ϕ j (λ) = E(f j (B, W) Λ(B, W) = λ), for certain scalar values f j (B, W), and θ 0 (λ) denotes the probability density function of Λ. Proof. We shall evaluate the expectations in (5) by first conditioning on Λ = λ, W and B, and then taking expectations over W, B conditional on Λ(B, W) = λ.this is, ρ j (λ) = E(E( Λ + j Λ j 1 = 0,Λ = λ; B.W)E(det( Λ j 1 ) Λ j 1 = 0,Λ = λ; B, W) (22) θ j 1 (0; g, B, W) Λ(B, W) = λ) θ 0 (λ), the corresponding derivatives should be rewritten according to Lemma 2. From Lemma in Adler (1981) and Theorem 2 we have E( Λ + j Λ j 1 = 0,Λ = λ; B, W) = (2π) 1/2 tr(m 2 ) 1/2, (23) and the density of Λ j 1 at zero conditional on Λ = λ and B, W given by θ j 1 (0; λ, B, W) = (2πtr(M 2 )) 1 2 (j 1), (24) According to Proposition 2, Lemma 7 and Lemma 8 we obtain E(det( Λ 0 ) Λ 0 = 0,Λ = λ; B, W) = 1, (25) E(det( Λ 1 ) Λ 1 = 0,Λ = λ; B, W) = ((ν m)tr (M 3 ) + (η m)tr (M 4 ) (26) +tr(m 1 + M 5 Σ + M 8 )), E(det( Λ 2 ) Λ 2 = 0,Λ = λ; B, W) = ( ) ν m 2 tr (M3 ) 2 + ( ) η m 2 tr (M4 ) 2 +tr(m 1 ) 2 (27) +(ν m)(η m)tr (M 3 ) tr(m 4 ) +[(ν m)tr (M 3 ) + (η m)tr(m 4 )]tr(m 1 ) + tr(m 5 Σ + K 8 ) 2 +[(ν m)tr (M 3 ) + (η m)tr(m 4 ) + tr(m 1 )]tr(m 5 Σ + M 8 ) tr(m 5 ΣM 5 Σ + M M 6 M 7 Σ + M 7 M 6 Σ + M 2 ), 9

10 respectively, with Σ given by (21). Using (23), (24) and (25)-(27) we finally obtain ϕ 1 (λ) = E(tr(M 2 ) 1/2 ), ρ 1 (λ) = (2π) 1/2 θ 0 (λ) ϕ 1 (λ), ρ 2 (λ) = (2π) 1 θ 0 (λ)ϕ 2 (λ) ρ 3 (λ) = (2π) 3/2 θ 0 (λ) ϕ 3 (λ), ϕ 2 (λ) = E((ν m)tr (M 3 + (η m)tr (M 4 ) + tr(m 1 ) + tr(m 5 Σ + M 8 )), ϕ 3 (λ) = E(tr(M 2 ) 1/2 ( ( ) ν m 2 tr (M3 ) 2 + ( ) η m 2 tr (M4 ) 2 +(ν m)(η m)tr (M 3 ) tr (M 4 ) +tr(m 1 ) 2 + tr(m 5 Σ + M 8 ) 2 + ((ν m)tr (M 3 ) + (η m)tr(m 4 ))tr(m 1 ) +((ν m)tr (M 3 ) + (η m)tr (M 4 ) + tr(m 1 ))tr(m 5 Σ + M 8 ) tr(m 5 ΣM 5 Σ +M M 6 M 7 Σ + M 7 M 6 Σ + M 2 ))), all expectations above have to be taken by conditioning on Λ(B, W) = λ. 5 Wilks s Λ Random Field In this section we shall apply the previous results for obtaining the EC densities of the Wilks s Λ random field. Definition 4 Let U(t), V(t) be two independent m dimensional Wishart RFs with ν and η degrees of freedom, respectively. The Wilk s Λ RF is then defined as Λ(t) = det(u(t)) det(u(t) + V(t)). Note that although the Wilks s Λ statistic is well defined at a point t provided ν+ η m, this may not be true for all points inside a compact subset of R N. The reason is that both the numerator and denominator if Wilks s Λ could be zero. For the particular case of m = 1 it was shown in Worsley (1994) that ν+η N to avoid this, with probability one. For general m, following the same ideas in Cao & Worsley (1999a), we obtain ν +η m+n 1 as a necessary and sufficient condition for having Λ well defined. When used in hypothesis testing problems, one rejects null hypotheses based on small values of the Wilk s Λ statistics. So, in what follows, we shall work with the RF Y (t) = log(λ(t)). It is well-known that for any symmetric matrix function U(t) the identities det(u) j = det(u)tr(u 1 U j ), det(u) kj = det(u) 1 det(u) k det(u) j det(u)tr(u 1 U k U 1 U j + U 1 Ü kj ) hold, so that Theorem 1 gives Ẏ = 2A 1 vec(m 1/2 0 ) Ÿ = 2(tr(M 0 ) 1/2 H tr(m 0 )P + tr(m 4 )Q + A 1 M 5 A 1 A 1 M 6 A 2 A 2 M 7 A 1), 10

11 M 0 = (B 1 I m )W 1, M 4 = W 1, M 5 = (M 1/2 0 M 1/2 0 ), M 6 = (M 1/2 0 W 1/2 ), M 7 = M 6. Theorem 3 yields ρ 1 (y) = (2π) 1/2 θ 0 (y)e(tr(m 0 ) 1/2 det(b) = e y ), (28) ρ 2 (y) = 2(2π) 1 θ 0 (y)e((ν m)tr(m 0 ) (η m)tr (M 4 ) tr(m 5 Σ) det(b) = e y ), (29) ρ 3 (y) = 4(2π) 3/2 θ 0 (y)e(tr(m 0 ) 1/2 ( ( ) ν m 2 tr (M0 ) 2 + ( ) η m 2 tr (M4 ) 2 + tr(m 5 Σ) 2 (30) (ν m)(η m)tr(m 0 )tr(m 4 ) + ((η m)tr(m 4 ) (ν m)tr(m 0 ))tr(m 5 Σ) tr(m 5 ΣM 5 Σ+2M 6 M 7 Σ) tr(m 0 ))) det(b) = e y ), Σ = I m 2 m 2 tr(m 0) 1 vec(m 1/2 0 )vec(m 1/2 0 ). There is no closed form expression for θ 0 (y), the density of Y = log(λ). A host of approximations are available (Anderson 1984), but it is easier numerically to use Fourier methods as follows. We can write m Y = Y i, i=1 exp( Y i ) Beta 1 ( 1 2 (ν + 1 i), 1 2 η) independently, i = 1,..., m (Anderson 1984). Then the product of the Fourier transforms of the densities of Y i is the Fourier transform of the density of Y. Inverting gives the desired density θ 0 (y) of Y. 5.1 EC densities Providing closed expressions for the expectations that appear in the formulas (28-30) is a very difficult task. In this subsection we propose a practical way to approximate such expectations. Specifically, we shall use expansion series for either negative or fractional powers of tr(m 0 ) (see Shapiro & Gross (1981)), that is, tr(m 0 ) r E(tr(M 0 )) r + re(tr(m 0 )) r 1 (tr(m 0 ) E(tr(M 0 ))) +r(r 1)E(tr(M 0 )) r 2 (tr(m 0 ) E(tr(M 0 ))) However, for simplicity in the exposition, we just describe how to deal with the simplest approximation, namely, tr(u 1 ) r E(tr(U 1 )) r. For the EC density ρ 1 (y) we have ρ 1 (y) = (2π) 1/2 θ 0 (y) E(tr(M 0 ) A) 1/2. 11

12 A is the event det(b) = e y. According to Letac & Massam (2001), E(tr(M 0 ) A) = E(tr(B 1 I m ) A, q q = ν + η m 1. Now, by expressing tr(b 1 ) and det(b) in terms of the eigenvalues of B we obtain E(tr(B 1 ) A) = E(tr(B 1 )) + E(det(B ))e y, for some matrix B Beta m 1 ( 1ν, 1 η), namely, one of the principal minors of order m 1 of B. 2 2 Then, by (34) and (35), Therefore c 1 (y) := E(tr(B 1 ) A) = p m (a, b) = (m 1)(q + 1) (ν m) + p m 1 (1, 0)e y, { Γm ( 1 2 ν+a)γ m( 1 2 (ν+η)+b) Γ m( 1 2 ν+b)γm( 1 (ν+η)+a), 2 m 1 0, m 0. ρ 1 (y) = (2π) 1/2 θ 0 (y)k 1 (y) 1/2, k 1 (y) = c 1(y) m. q In a similar way, we use the approximation to obtain E(tr(M 5 Σ)) E(tr(M 0 )) E(tr(M 0 )) 1 E(tr(M 2 0)) ρ 2 (y) = 2(2π) 1 θ 0 (y)((ν m 1)k 1 (y) (η m)mq 1 + k 1 (y) 1 k 2 (y)), c 2 (y) = p m 1 (2, 0)e 2y + k 2 (y) = c 3(y) 2mc 1 (y) + m 2 q 3 q + c 2(y) 2c 1 (y) + m, q 2 1 (q + 1)(q + 3) (ν m)(ν m + 2) C (q + 1)(q + 2) (2)(I m 1 ) 2(ν m)(ν m + 1) C (1,1)(I m 1 ), c 3 (y) = p m 1 (2, 0)e 2y + (m 1)p m 2 (1, 0)e y + (q + 1)(q + 2) + (ν m)(ν m + 1) C (1,1)(I m 1 ). Note that we have used (Letac & Massam 2001) (q + 1)(q + 3) (ν m)(ν m + 2) C (2)(I m 1 ) and E(tr(M 2 0) A) = E(tr(B 1 I m ) 2 A) q 3 q + E(tr((B 1 I m ) 2 ) A) q 2 1 c 2 (y) := E(tr(B 2 ) A) = E(tr(B 2 )) + E(det(B ) 2 )e 2y, c 3 (y) := E(tr(B 1 ) 2 A) = E(tr(B 1 ) 2 ) + 2E( tr(b 1 ) det(b ))e y + E(det(B ) 2 )e 2y, 12

13 the expressions for c 2 (y) and c 3 (y) are obtained from (4), (34) and (35). Finally, from Letac & Massam (2001) and Graczyk, Letac & Massam (2003) we have E(tr(M 0 ) 2 A) = E(tr(B 1 I m ) 2 A) q E(tr((B 1 I m ) 2 ) A), q 3 q k 5 (y) := E(tr (M 4 ) tr (M 0 ) A) = mq + 1 q 3 q E(tr(B 1 I m ) A), k 6 (y) := E(tr (M 4 ) tr(m 2 0) A) = (mq + 2)E(tr(B 1 I m ) 2 A) + (mq 2 + 2q 2m)E(tr((B 1 I m ) 2 ) A), q(q 2 1)(q 2 4) k 7 (y) := E(tr(M 2 0M 4 ) A) = 2(q + m)e(tr(b 1 I m ) 2 A) q(q 2 1)(q 2 4) + (q + m)e(tr((b 1 I m ) 2 ) A), (q 2 1)(q 2 4) which implies ρ 3 (y) = 4(2π) 3/2 θ 0 (y)k 1 (y) 1/2 { ( ) ν m k3 (y) + ( ) η m k4 (ν m)(η m)k (y) +(ν m)[k 3 (y) k 2 (y)] (η m)[k 5 (y) k 1 (y) 1 k 6 (y)] + [k 3 (y) k 2 (y)] 2[k 5 (y) k 1 (y) 1 k 7 (y)] k 1 (y)}, k 3 (y) = c 3(y) 2mc 1 (y) + m 2 q 2 1 m 2 k 4 = q m q 3 q, k 5 (y) = (mq + 1)(c 1(y) m), q 3 q + c 2(y) 2c 1 (y) + m, q 3 q k 6 (y) = (mq + 2)(c 3(y) 2mc 1 (y) + m 2 ) + (mq 2 + 2q 2m)(c 2 (y) 2c 1 (y) + m), q(q 2 1)(q 2 4) k 7 (y) = 2(q + m)(c 3(y) 2mc 1 (y) + m 2 ) q(q 2 1)(q 2 4) + (q + m)(c 2(y) 2c 1 (y) + m). (q 2 1)(q 2 4) 5.2 Simulation Results In this subsection we shall show some simulation results for validating the previous formulae. We generated 200 Y = log(λ) RFs on a 32 3 voxel rectilinear lattice as follows. We first generated lattices of independent standard Gaussian random variables then smoothed them with a Gaussian shaped filter of standard deviation σ = 3.2 voxels, which gives v = 1/( 2σ) = 0.22 as the standard deviation of the spatial derivative (Worsley, Marrett, Neelin, Vandal, Friston & Evans 1996).). The Wilks s Λ RF was generated by ( Y = log det(z 1Z 1 ) det(z 1Z 1 + Z 2Z 2 ) ), (31) Z 1 and Z 2 are ν m and η m matrices of independent smoothed Gaussian random fields as above. The variance of the derivative of the component random fields is The EC χ(a y ) of the 13

14 12 Simulated Theoretical 10 8 Expected Euler Characteristic Threshold Figure 1: Euler Characteristic of the excursion set of Y = log(λ) for ν = 7, η = 4 and m = 2 sampled on a 32 3 lattice smoothed by Gaussian filter with standard deviation 6.37 voxels. excursion sets A y was calculated for each of the 200 Y RFs at y = 0, 0.1, 0.2,..., 6.0. Then, the average of the measurements χ(a y ) was used as an estimator of E(χ(A y )). This value and the theoretical approximation obtained in the subsection above are plotted in Figure 1 for ν = 7, η = 4, and m = 2. Note that for thresholds greater that 2 both the theoretical approximation and the simulated values are very close. The P-value approximation (6) based on the expected EC, the (uncorrected) P-value at a single voxel, and the Bonferroni P-value are plotted in Figure 2. The parameter values were chosen to match the example in Section 6: m = 3, ν = 31 and η = 1, 2, 3, 4. The parameter set S was a 3D ball of radius with radius 68mm sampled on a 2mm voxel lattice. The data was smoothed by a Gaussian filter with standard deviation 5.65mm to give v = In the first plot with m = 1, Wilks s Λ is equivalent to Hotelling s T 2 = ν(exp(y ) 1). Exact results for the expected EC of Hotelling s T 2 from Cao & Worsley (1999a) are added to the plot. We can see that these are in reasonable agreement with our Wilks s Λ approximation, which appears to be too liberal by a factor of 3 to 4. 14

15 P(max Y > y) 10 1 m = 3, ν = 31, η = EC Bon 1 vox T 2 P(max Y > y) 10 1 m = 3, ν = 31, η = y = log Wilks s Λ 10 1 m = 3, ν = 31, η = y = log Wilks s Λ 10 1 m = 3, ν = 31, η = 4 P(max Y > y) 10 2 P(max Y > y) y = log Wilks s Λ y = log Wilks s Λ Figure 2: P-values of the maximum of Y = log(λ) over the brain, approximated by a 3D ball with radius 68mm sampled on a 2mm voxel lattice. The data was smoothed by a Gaussian filter with standard deviation 5.65mm. EC = Euler Characteristic approximation (6), Bon = Bonferroni upper bound, 1 vox = uncorrected P-value at a single voxel (lower bound), T 2 = exact expected EC for equivalent Hotelling s T 2 (η = 1 only). 15

16 6 Application 6.1 Multivariate Linear Models We apply our results to a multivariate linear model. Suppose we have n independent RFs of multivariate observations y i (t) R m, i = 1,..., n, t S R N, and a multivariate linear model (Cao & Worsley 1999a): y i (t) = x iβ(t) + ɛ i (t) Σ(t) 1/2, (32) x i is a p-vector of known regressors, and β(t) is an unknown p m coefficient matrix. The error ɛ i (t) is a m-vector of independent zero mean, unit variance isotropic Gaussian components with the same spatial correlation structure, and Var(y i (t)) = Σ(t) is an unknown m m matrix. We can now detect how the regressors are related to the multivariate data at point t by testing contrasts in the rows of β(t). Classical multivariate test statistics evaluated at each point t then form a random field. Let β(t) be the usual least-squares estimate of β(t). The Wilks s Λ random field is defined in terms of two independent Wishart random fields. The first is the error sum of squares matrix U(t) = n (y i (t) x β(t)) i (y i (t) x β(t)) i Wishart m (Σ(t), ν) i=1 ν = n p. Let X = (x 1,..., x n) be the design matrix. The second is the regression sum of squares matrix for a η p matrix of contrasts C V(t) = (C β(t)) (C(X X) 1 C ) 1 (C β(t)) Wishart m (Σ(t), η) under the null hypothesis of no effect, Cβ = 0. In Wilks s Λ (31), Σ(t) will cancel so under the null hypothesis, (31) will be a Wilks s Λ random field. 6.2 Brain Shape Analysis As an illustration of the methods, we apply the P-value approximations for Wilks s Λ to a data set on non-missile trauma (Tomaiuolo, Worsley, Lerch, Di Paulo, Carlesimo, Bonanni, Caltagirone & Paus 2005) that was analysed in a similar way in Worsley et al. (2004). The subjects were 17 patients with non-missile brain trauma who were in a coma for 3-14 days. MRI images were taken after the trauma, and the multivariate data were the N = 3 component vector deformations needed to warp the MRI images to an atlas standard sampled on a 2mm voxel lattice. The same data were also collected on a group of 19 age and sex matched controls. Damage is expected in white mater areas, so the search region S was defined as the voxels smoothed average control subject white matter density exceeded 5%. For calculating the intrinsic volumes, this was approximated by a sphere with the same volume, 1.31cc, which is slightly liberal for a non-spherical search region. The effective v from (7), averaged over the search region, was The first analysis was to look for brain damage by comparing the deformations of the 17 trauma patients with the 19 controls, so the sample size is n = 36. We are looking at a single contrast, the difference between trauma and controls, so η = 1 and the residual degrees of freedom is ν = 34. In this case Wilks s Λ is Hotelling s T 2. The P = 0.05 threshold, found using the EC densities in Cao & Worsley (1999a) and by equating (8) to 0.05 and solving for y, was y = The thresholded data, 16

17 together with the estimated contrast (mean trauma - control deformations) is shown in Figure 3(a). A large region near the corpus callosum seems to be damaged. The nature of the damage, judged by the direction of the arrows, is away from the center (see Figure 3(b)). This can be interpreted as expansion of the ventricles, or more likely, atrophy of the surrounding white matter, which causes the ventricle/white matter boundary to move outwards. We might also be interested in functional anatomical connectivity: are there any regions of the brain whose shape (as measured by the deformations) is correlated with shape at a reference point? In other words, the explanatory variables are the deformations at a pre-selected reference point, and the test statistic is Wilks s Λ. We chose as the reference the point with maximum Hotelling s T 2 for damage, marked by axis lines in Figure 3. Figure 3(c) shows the resulting -log Wilks s Λ field above the P = 0.05 threshold of 1.54 for the combined trauma and control data sets removing separate means for both groups (ν = 31, η = 3 ). Obviously there is strong correlation with points near the reference, due to the smoothness of the data. The main feature is the strong correlation with contra-lateral points, indicating that brain anatomy tends to be symmetric. A more interesting question is whether the correlations observed in the control subjects are modified by the trauma (Friston, Büchel, Fink, Morris, Rolls & Dolan 1997). In other words, is there any evidence for an interaction between group and reference vector deformations? To do this, we simply add another three covariates to the linear model whose values are the reference vector deformations for the trauma patients, and the negative of the reference vector deformations for the control subjects. The resulting -log Wilks s Λ field for testing for these three extra covariates, thresholded at 1.72 (P = 0.05, η = 3, ν = 28) is shown in Figure 3(d). Apart from changes in the neighbourhood of the reference point, there is some evidence of a change in correlation at a location in the contraleteral side, slightly anterior. Looking at the maximum canonical correlations in the two groups separately, we find that correlation has increased at this location from to 0.927, perhaps indicating that the damage is strongly bilateral. These results are in close agreement with those in Worsley et al. (2004) which used Roy s maximum root rather than Wilks s Λ. A Appendix Lemma 5 (Lemma A.2 in Worsley (1994)) Let H Normal j j (0, M(I j )), let h be a fixed scalar, and let A be a fixed symmetric j j matrix. Then Lemma 6 Let E(det(A + hh)) = j 2 i=0 ( 1) i (2i)! h 2i detr 2 i j 2i (A). i! P W ishart j (ν, I j ), Q W ishart j (η, I j ), X 1 Normal j m (0, Σ I j ), X 2 Normal j m (0, I jm ) independently, a, b, c be fixed scalars, and C 1, C 2, C 3, C 4 be fixed m m matrices. Then = E(detr i (ap + bq + ci j + X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2)) i j! i k ( )( ) ν η k ( ) j l a i k r b r c k l E(detr l (X 1 C 1 X 1 (j i + k)! i k r r k l k=0 r=0 +X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2 )). l=0 17

18 (a) (b) (c) (d) Figure 3: Shape analysis of non-missile brain trauma data. (a) Trauma minus control average deformations (arrows and color bar), sampled every 6mm inside the brain, with Hotelling s T 2 field for significant group differences (threshold y = 54.0, P = 0.05). The reference point of maximum Hotelling s T 2 is marked by the intersection of the three axes. (b) Closeup of (a) showing that the damage is an outward movement of the anatomy, either due to swelling of the ventricles or atrophy of the surrounding white matter. (c) Regions of effective anatomical connectivity with the reference point, assessed by the Wilks s Λ field (threshold y = 1.54, P = 0.05). The reference point is connected with its neighbours (due to smoothness) and with contralateral regions (due to symmetry). (d) Regions the connectivity is different between trauma and control groups, assessed by the Wilks s Λ field (threshold y = 1.72, P = 0.05 ). The small region in the contralateral hemisphere (right) is more correlated in the trauma group than the control group. 18

19 Proof. Holding X 1 and X 2 fixed and using Lemma in Worsley (1994) we obtain = = E(detr i (ap + bq + ci j + X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2)) i E(detr i k (ap + bq))detr k (ci j + X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2) k=0 i k=0 j! (j i + k)! +X 2 C 4 X 2). i k ( r=0 ν i k r )( η r ) a i k r b r detr k (ci j + X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 Taking expectations over X 1, X 2 in the equality above we have = E(detr i (ap + bq + ci j + X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2)) i j! i k ( )( ) ν η k ( ) j l a i k r b r c k l E(detr l (X 1 C 1 X 1 (j i + k)! i k r r k l k=0 r=0 +X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2)), l=0 which ends the proof. Lemma 7 Let P W ishart j (ν, I j ), Q W ishart j (η, I j ), X 1 Normal j m (0, Σ I j ), X 2 Normal j m (0, I jm ), H Normal j j (0, M(I j )) independently, a, b, c, h be fixed scalars, and C 1, C 2, C 3, C 4 be fixed m m matrices. Then = E(det(aP + bq + ci j + X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2 + hh)) ( 1) i j 2i j 2i k (2i)! h 2i j! ( )( ) ν η k a j 2i k r b r 2 i i! (2i + k)! j 2i k r r [ j 2 ] i=0 k=0 E(detr l (X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2 )). r=0 l=0 ( j l k l ) c k l Proof. Holding P, Q, X 1, X 2 fixed and using Lemma 5 we obtain = E(det(aP + bq + ci j + X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2 + hh)) [ j 2 ] i=0 ( 1) i (2i)! h 2i detr 2 i j 2i (ap + bq + ci j + X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X i! 2). Now, taking expectations and using Lemma 6 we have the desired result. Lemma 8 Let X 1 Normal j m (0, Σ I j ), X 2 Normal j m (0, I jm ), and C 1, C 2, C 3, C 4 be fixed symmetric m m matrices.. Then tr(c 1 Σ + C 4 ), j = 1, E(det(X 1 C 1 X 1 + X 1 C 2 X 2 + X 2 C 3 X 1 + X 2 C 4 X 2 )) = tr(c 1 Σ + C 4 ) 2 tr(c 1 ΣC 1 Σ +C C 2 C 3 Σ + C 3 C 2 Σ), j = 2. 19

20 Proof. It follows from the fact that the rows of X 1 and X 2 are independent vectors from Normal m (0, Σ) and Normal m (0, I m ), respectively, and using the well-known identity i, j, k, l = 1,..., r, σ ij = E(x i x j ). E(x i x j x k x l ) = σ ij σ kl + σ ik σ jl + σ il σ jk, Lemma 9 If κ = (k 1,..., k m ), κ = k then for a > k 1 + (m + 1)/2, 0<X<I m det(x) a (m+1)/2 det(i m X) b (m+1)/2 C κ (X 1 ) dx (33) = ( 1)k ( (a + b) + m+1) 2 κ Γ m (a)γ m (b) ( a + m+1) 2 κ Γ m (a + b) C κ(i m ), C κ (I m ) = 2 2k k!( m 2 ) κ and p denotes the nonzero indices in κ. p (2k i 2k j i + j) i<j p (2k i + p i)! Proof. It is proved by using following similar ideas to that of Theorem and Theorem in Muirhead (1982). As a particular case we have that if B Beta m ( 1 2 ν, 1 2 η) then for ν > 2k 1 + m + 1, i=1 E(C κ (B 1 )) = ( 1)k ( ν+η + m+1 2 ( ν + m+1) 2 2 κ 2 ) κ Lemma 10 (Corollary in Muirhead (1982)) If B Beta m ( 1ν, 1 η) then 2 2 References, C κ (I m ). (34) E(det(B) h ) = Γ m( 1ν + h)γ 2 m( 1 (ν + η)) 2 Γ m ( 1ν)Γ 2 m( 1 (35) (ν + η) + h). 2 Adler, R. J. (1981). The Geometry of Random Fields, Wiley, New York. Adler, R. J. (2000). On excursion sets, tube formulae, and maxima of random fields., The Annals of Applied Probability 10(1): Adler, R. J. & Hasofer, A. M. (1976). 4: Level crossings for random fields. annals of probability, Anderson, T. W. (1984). An introduction to multivariate statistical analysis, 2 edn, Wiley, New York. Beaky, M. M., Scherrer, R. J. & Villumsen, J. V. (1992). Topology of large scale structure in seeded hot dark matter models. astrophysical journal, 387:

21 Cao, J. & Worsley, K. J. (1999a). The detection of local shape changes via the geometry of hotelling s t 2 fields, Annals of Statistics 27: Cao, J. & Worsley, K. J. (1999b). The geometry of correlation fields with an application to functional connectivity of the brain, Annals of Applied Probability 9: Cao, J. & Worsley, K. J. (2001). Applications of random fields in human brain mapping, In M. Moore (Ed.) Spatial Statists: Methodological Aspects and Applications, Springer Lecture Notes in Statistics 159: Friston, K., Büchel, C., Fink, G. R., Morris, J., Rolls, E. & Dolan, R. J. (1997). Psychophysiological and modulatory interactions in neuroimaging, NeuroImage 6: Gott, J. R., Park, C., Juskiewicz, R., Bies, W. E., Bennett, D. P., Bouchet, F. R. & Stebbins, A. (1990). Topology of microwave background fluctuations: theory, Astrophysical Journal 352: Graczyk, P., Letac, G. & Massam, H. (2003). The complex wishart distribution and the symmetric group, Annals of Statistics 31(1): Hasofer, A. M. (1978). Upcrossings of random fields, Advances in Applied Probability 10: Letac, G. & Massam, H. (2001). The moments of wishart laws, binary trees, and the triple products, Comptes Rendus de l Académie des Sciences - Series I - Mathematics 333(4): Magnus, J. R. & Neudecker, H. (1988).). Matrix differential calculus with applications in statistics and econometrics, Wiley, New York. Mardia, K. V., K. J. T. & Bibby, J. M. (1979). Multivariate Analysis, Academic Press, New York. Muirhead, R. J. (1982). Aspects of Multivariate Statistical Theory, Wiley, New York. Neudecker, H. (1969). Some theorems on matrix differentiation with special reference to kronecker matrix products, Journal of the American Statistical Association 64 (327): Olkin, I. & Rubin, H. (1962). A characterization of the wishart distribution, Annals of Mathematical Statistics 33(4): Shapiro, S. S. & Gross, A. J. (1981). Statistical modeling techniques, Marcel Dekker, New York. Taylor, J. E., Takemura, A. & Adler, R. J. (2005). Validity of the expected Euler characteristic heuristic, Annals of Probability 33(4): Tomaiuolo, F., Worsley, K. J., Lerch, J., Di Paulo, M., Carlesimo, G. A., Bonanni, R., Caltagirone, C. & Paus, T. (2005). Changes in white matter in long-term survivors of severe non-missile traumatic brain injury: a computational analysis of magnetic resonance images, Journal of Neurotrauma 22: Vogeley, M. S., Park, C., Geller, M. J., Huchira, J. P. & Gott, J. R. (1994). Topological analysis of the CfA redshift survey, Astrophysical Journal 420: Worsley, K. J. (1994). Local maxima and the expected euler characteristic of excursion sets of χ 2, f and t fields, Advances in Applied Probability 26:

22 Worsley, K. J. (1995a). Boundary corrections for the expected euler characteristic of excursion sets of random fields, with an application to astrophysics, Advances in Applied Probability 27: Worsley, K. J. (1995b). Estimating the number of peaks in a random field using the hadwiger characteristic of excursion sets, with applications to medical images, Annals of Statistics 23: Worsley, K. J., Cao, J., Paus, T., Petrides, M. & Evans, A. (1998). Applications of random field theory to functional connectivity, Human Brain Mapping 6: Worsley, K. J., Marrett, S., Neelin, P., Vandal, A. C., Friston, K. J. & Evans, A. C. (1996).). A unified statistical approach for determining significant signals in images of cerebral activation, Human Brain Mapping 4: Worsley, K. J., Taylor, J. E., Tomaiuolo, F. & Lerch, J. (2004). Unified univariate and multivariate random field theory, NeuroImage 23: S