Chapter 15

Matrix Differentiation

[26/100%]
Show
LaTeX
Problem 15.1

Using the result of Part (c) of Exercise 6.2, verify that every neighborhood of a point xx in Rm×1\mathbb {R}^{m \times 1} is an open set.

?
Problem 15.2

Let ff represent a function, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables, suppose that the set SS contains at least some interior points, and let cc represent an arbitrary one of those points. Verify that if ff is kk times continuously differentiable at cc, then ff is kk times continuously differentiable at every point in some neighborhood of cc.

?
Problem 15.3

Let X={xij}X = \{ x_{ij}\} represent an m×nm \times n matrix of mnmn variables, and let xx represent an mnmn-dimensional column vector obtained by rearranging the elements of XX (in the form of a column vector). Further, let SS represent a set of XX-values, and let S∗S^{*} represent the corresponding set of xx-values (i.e., the set obtained by rearranging the elements of each m×nm \times n matrix in SS in the form of a column vector). Verify that an mnmn-dimensional column vector is an interior point of S∗S^{*} if and only if it is a rearrangement of an m×nm \times n matrix that is an interior point of SS.

?
Problem 15.4

Let ff represent a function whose domain is a set SS in Rm×1\mathbb {R}^{m \times 1} (that contains at least some interior points). Show that the Hessian matrix HfHf of ff is the gradient matrix of the gradient vector (Df)′(Df)' of ff.

?
Problem 15.5

Let gg represent a function, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables, let S∗={x∈S:g(x)≠0}S^{*} = \{ x \in S : g(x) \neq 0\}, and let cc represent any interior point of S∗S^{*} at which gg is continuously differentiable. Show that (for any positive integer kk) g−kg^{-k} is continuously differentiable at cc and

∂g−k∂xj=−kg−k−1∂g∂xj. \frac{\partial g^{-k}}{\partial x_j} = -kg^{-k-1}\frac{\partial g}{\partial x_j} .
?
Remark.
?

Do so based on the result that

∂(1/g)∂xj=−(1/g)2∂g∂xj,(*) \frac{\partial (1/g)}{\partial x_j} = -(1/g)^2 \frac{\partial g}{\partial x_j} , \tag {*}

and, letting ff (like gg) represent a function (defined on SS) of xx, the results that if ff is continuously differentiable at cc, then the ratio f/gf/g is also continuously differentiable at cc, that, for any positive integer kk, fkf^k is continuously differentiable at any point at which ff is continuously differentiable, and that

∂fk∂xj=kfk−1∂f∂xj.(**) \frac{\partial f^k}{\partial x_j} = kf^{k-1}\frac{\partial f}{\partial x_j} . \tag {**}
Problem 15.6

Let FF represent a p×pp \times p matrix of functions, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables. Let cc represent any interior point of SS at which FF is continuously differentiable. Show that if FF is idempotent at all points in some neighborhood of cc, then (at x=cx = c)

F∂F∂xjF=0. F \frac{\partial F}{\partial x_j} F = 0 .
?
Problem 15.7

Let gg represent a function, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables, and let ff represent a p×1p \times 1 vector of functions (defined on SS) of xx. Let cc represent any interior point (of SS) at which gg and ff are continuously differentiable. Show that gfgf is continuously differentiable at cc and that (at x=cx = c)

∂(gf)∂x′=f∂g∂x′+g∂f∂x′. \frac{\partial (gf)}{\partial x'} = f\frac{\partial g}{\partial x'} + g\frac{\partial f}{\partial x'} .
?
Problem 15.8
?
(a)

Let X={xij}X = \{ x_{ij}\} represent an m×nm \times n matrix of mnmn ``independent'' variables, and suppose that XX is free to range over all of Rm×n\mathbb {R}^{m \times n}.

  1. Show that, for any p×mp \times m and n×pn \times p matrices of constants AA and BB,

    ∂tr⁡(AXB)∂X=A′B′. \frac{\partial \operatorname {tr}(AXB)}{\partial X} = A'B' .
  2. Show that, for any mm- and nn-dimensional column vectors aa and bb,

    ∂(a′Xb)∂X=ab′. \frac{\partial (a'Xb)}{\partial X} = ab' .
(b)

Suppose now that XX is a symmetric (but otherwise unrestricted) matrix (of dimensions m×mm \times m).

  1. Show that, for any p×mp \times m and m×pm \times p matrices of constants AA and BB,

    ∂tr⁡(AXB)∂X=C+C′−diag⁡(c11,c22,…,cmm), \frac{\partial \operatorname {tr}(AXB)}{\partial X} = C + C' - \operatorname {diag}(c_{11}, c_{22}, \ldots , c_{mm}) ,

    where C={cij}=BAC = \{ c_{ij}\} = BA.

  2. Show that, for any mm-dimensional column vectors a={aj}a = \{ a_j\} and b={bj}b = \{ b_j\},

    ∂(a′Xb)∂X=ab′+ba′−diag⁡(a1b1,a2b2,…,ambm). \frac{\partial (a'Xb)}{\partial X} = ab' + ba' - \operatorname {diag}(a_1b_1, a_2b_2, \ldots , a_mb_m) .
Remark.
?

For Part (a)(1), observe that tr⁡(AXB)=tr⁡(BAX)\operatorname {tr}(AXB) = \operatorname {tr}(BAX). For Part (a)(2), observe that a′Xb=tr⁡(a′Xb)a'Xb = \operatorname {tr}(a'Xb).

Problem 15.9
?
(a)

Let X={xst}X = \{ x_{st}\} represent an m×nm \times n matrix of ``independent'' variables, and suppose that XX is free to range over all of Rm×n\mathbb {R}^{m \times n}. Show that, for any n×mn \times m matrix of constants AA,

∂tr⁡[(AX)2]∂X=2(AXA)′. \frac{\partial \operatorname {tr}[(AX)^2]}{\partial X} = 2(AXA)' .
(b)

Let X={xst}X = \{ x_{st}\} represent an m×mm \times m symmetric (but otherwise unrestricted) matrix of variables. Show that, for any m×mm \times m matrix of constants AA,

∂tr⁡[(AX)2]∂X=2[B+B′−diag⁡(b11,b22,…,bmm)], \frac{\partial \operatorname {tr}[(AX)^2]}{\partial X} = 2[B + B' - \operatorname {diag}(b_{11}, b_{22}, \ldots , b_{mm})] ,

where B={bst}=AXAB = \{ b_{st}\} = AXA.

Problem 15.10

Let X={xij}X = \{ x_{ij}\} represent an m×nm \times n matrix of ``independent'' variables, and suppose that XX is free to range over all of Rm×n\mathbb {R}^{m \times n}. Show that, for k=2,3,…,k = 2, 3, \ldots ,

∂tr⁡(Xk)∂X=k(X′)k−1. \frac{\partial \operatorname {tr}(X^k)}{\partial X} = k(X')^{k-1} .
?
Problem 15.11

Let X={xst}X = \{ x_{st}\} represent an m×nm \times n matrix of ``independent'' variables, and suppose that XX is free to range over all of Rm×n\mathbb {R}^{m \times n}.

?
(a)

Show that, for any m×mm \times m matrix of constants AA,

∂tr⁡(X′AX)∂X=(A+A′)X. \frac{\partial \operatorname {tr}(X'AX)}{\partial X} = (A+A')X .
(b)

Show that, for any n×nn \times n matrix of constants AA,

∂tr⁡(XAX′)∂X=X(A+A′). \frac{\partial \operatorname {tr}(XAX')}{\partial X} = X(A+A') .
(c)

Show (in the special case where n=mn = m) that, for any m×mm \times m matrix of constants AA,

∂tr⁡(XAX)∂X=(AX)′+(XA)′. \frac{\partial \operatorname {tr}(XAX)}{\partial X} = (AX)' + (XA)' .
(d)

Use the results of Parts (a)-(c) to devise simple formulas for ∂tr⁡(X′X)/∂X\partial \operatorname {tr}(X'X)/\partial X, ∂tr⁡(XX′)/∂X\partial \operatorname {tr}(XX')/\partial X, and (in the special case where n=mn = m) ∂tr⁡(X2)/∂X\partial \operatorname {tr}(X^2)/\partial X.

Problem 15.12

Let X={xij}X = \{ x_{ij}\} represent an m×mm \times m matrix. Let ff represent a function of XX defined on a set SS comprising some or all m×mm \times m symmetric matrices. Suppose that, for purposes of differentiation, ff is to be interpreted as a function of the [m(m+1)/2][m(m+1)/2]-dimensional column vector xx whose elements are xijx_{ij} (j≤i=1,…,m)(j \leq i = 1, \ldots , m). Suppose further that there exists a function gg, whose domain is a set TT of not-necessarily-symmetric matrices that contains SS as a proper subset, such that g(X)=f(X)g(X) = f(X) for X∈SX \in S, so that gg is a function of XX and ff is the function obtained by restricting the domain of gg to SS. Define S∗={x:X∈S}S^{*} = \{ x : X \in S\}. Let cc represent an interior point of S∗S^{*}, and let CC represent the corresponding value of XX. Show that if CC is an interior point of TT and if gg is continuously differentiable at CC, then ff is continuously differentiable at cc and that (at x=cx = c)

∂f∂X=∂g∂X+(∂g∂X)′−diag⁡(∂g∂x11,∂g∂x22,…,∂g∂xmm). \frac{\partial f}{\partial X} = \frac{\partial g}{\partial X} + \left(\frac{\partial g}{\partial X}\right)' - \operatorname {diag}\left(\frac{\partial g}{\partial x_{11}}, \frac{\partial g}{\partial x_{22}}, \ldots , \frac{\partial g}{\partial x_{mm}}\right) .
?
Problem 15.13

Let h={hi}h = \{ h_i\} represent an n×1n \times 1 vector of functions, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables. Let gg represent a function, defined on a set TT, of a vector y=(y1,…,yn)′y = (y_1, \ldots , y_n)' of nn variables. Suppose that h(x)∈Th(x) \in T for every xx in SS, and take ff to be the composite function defined (on SS) by f(x)=g[h(x)]f(x) = g[h(x)]. Show that if hh is twice continuously differentiable at an interior point cc of SS and if [assuming that h(c)h(c) is an interior point of TT] gg is twice continuously differentiable at h(c)h(c), then ff is twice continuously differentiable at cc and

Hf(c)=[Dh(c)]′Hg[h(c)]Dh(c)+∑i=1nDig[h(c)]Hhi(c). Hf(c) = [Dh(c)]'Hg[h(c)]Dh(c) + \sum _{i=1}^{n} D_ig[h(c)]Hh_i(c) .
?
Problem 15.14

Let X={xij}X = \{ x_{ij}\} represent an m×mm \times m matrix of m2m^2 ``independent'' variables (where m≥2m \geq 2), and suppose that the range of XX comprises all of Rm×m\mathbb {R}^{m \times m}. Show that (for any positive integer kk) the function ff defined (on Rm×m\mathbb {R}^{m \times m}) by f(X)=∣X∣kf(X) = \left|X\right|^k is continuously differentiable at every XX and that

∂∣X∣k∂X=k∣X∣k−1[adj⁡(X)]′. \frac{\partial \left|X\right|^k}{\partial X} = k\left|X\right|^{k-1}[\operatorname {adj}(X)]' .
?
Problem 15.15

Let F={fis}F = \{ f_{is}\} represent a p×pp \times p matrix of functions, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables. Let cc represent any interior point (of SS) at which FF is continuously differentiable. Use the results of Exercise 13.10 to show that

?
(a)

if rk⁡(F(c))=p−1\operatorname {rk}\left(F(c)\right) = p-1, then (at x=cx = c)

∂det⁡(F)∂xj=ky′∂F∂xjz, \frac{\partial \det (F)}{\partial x_j} = ky'\frac{\partial F}{\partial x_j}z ,

where z={zs}z = \{ z_s\} and y={yi}y = \{ y_i\} are any nonnull pp-dimensional vectors such that F(c)z=0F(c)z = 0 and [F(c)]′y=0[F(c)]'y = 0 and where [letting ϕis\phi_{is} represent the cofactor of fis(c)f_{is}(c)] kk is a scalar that is expressible as k=ϕis/(yizs)k = \phi_{is}/(y_iz_s) for any ii and ss such that yi≠0y_i \neq 0 and zs≠0z_s \neq 0; and

(b)

if rk⁡(F(c))≤p−2\operatorname {rk}\left(F(c)\right) \leq p-2, then (at x=cx = c)

∂det⁡(F)∂xj=0. \frac{\partial \det (F)}{\partial x_j} = 0 .
Problem 15.16

Let X={xst}X = \{ x_{st}\} represent an m×nm \times n matrix of ``independent'' variables, let AA represent an m×mm \times m matrix of constants, and suppose that the range of XX is a set SS comprising some or all XX-values for which det⁡(X′AX)>0\det (X'AX) > 0. Show that log⁡det⁡(X′AX)\log \det (X'AX) is continuously differentiable at any interior point CC of SS and that (at X=CX = C)

∂log⁡det⁡(X′AX)∂X=AX(X′AX)−1+[(X′AX)−1X′A]′. \frac{\partial \log \det (X'AX)}{\partial X} = AX(X'AX)^{-1} + [(X'AX)^{-1}X'A]' .
?
Problem 15.17
?
(a)

Let XX represent an m×nm \times n matrix of ``independent'' variables, let AA and BB represent q×mq \times m and n×qn \times q matrices of constants, and suppose that the range of XX is a set SS comprising some or all XX-values for which det⁡(AXB)>0\det (AXB) > 0. Show that log⁡det⁡(AXB)\log \det (AXB) is continuously differentiable at any interior point CC of SS and that (at X=CX = C)

∂log⁡det⁡(AXB)∂X=[B(AXB)−1A]′. \frac{\partial \log \det (AXB)}{\partial X} = [B(AXB)^{-1}A]' .
(b)

Suppose now that XX is an m×mm \times m symmetric matrix; that AA and BB are q×mq \times m and m×qm \times q matrices of constants; that, for purposes of differentiating any function of XX, the function is to be interpreted as a function of the column vector xx whose elements are xijx_{ij} (j≤i=1,…,m)(j \leq i = 1, \ldots , m); and that the range of xx is a set SS comprising some or all xx-values for which det⁡(AXB)>0\det (AXB) > 0. Show that log⁡det⁡(AXB)\log \det (AXB) is continuously differentiable at any interior point cc (of SS) and that (at x=cx = c)

∂log⁡det⁡(AXB)∂X=K+K′−diag⁡(k11,k22,…,kqq), \frac{\partial \log \det (AXB)}{\partial X} = K + K' - \operatorname {diag}(k_{11}, k_{22}, \ldots , k_{qq}) ,

where K={kij}=B(AXB)−1AK = \{ k_{ij}\} = B(AXB)^{-1}A.

Problem 15.18

Let F={fis}F = \{ f_{is}\} represent a p×pp \times p matrix of functions, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables, and let AA and BB represent q×pq \times p and p×qp \times q matrices of constants. Suppose that SS is the set of all xx-values for which F(x)F(x) is nonsingular and det⁡(AF−1(x)B)>0\operatorname {det}\left(AF^{-1}(x)B\right) > 0 or is a subset of that set. Show that if FF is continuously differentiable at an interior point cc of SS, then log⁡det⁡(AF−1B)\log \det (AF^{-1}B) is continuously differentiable at cc and (at x=cx = c)

∂log⁡det⁡(AF−1B)∂xj=−tr⁡[F−1B(AF−1B)−1AF−1∂F∂xj]. \frac{\partial \log \det (AF^{-1}B)}{\partial x_j} = -\operatorname {tr}\left[F^{-1}B(AF^{-1}B)^{-1}AF^{-1}\frac{\partial F}{\partial x_j}\right] .
?
Problem 15.19

Let AA and BB represent q×mq \times m and m×qm \times q matrices of constants.

?
(a)

Let XX represent an m×mm \times m matrix of m2m^2 ``independent'' variables, and suppose that the range of XX is a set SS comprising some or all XX-values for which XX is nonsingular and det⁡(AX−1B)>0\det (AX^{-1}B) > 0. Use the result of Exercise 15.18 to show that log⁡det⁡(AX−1B)\log \det (AX^{-1}B) is continuously differentiable at any interior point CC of SS and that (at X=CX = C)

∂log⁡det⁡(AX−1B)∂X=−[X−1B(AX−1B)−1AX−1]′. \frac{\partial \log \det (AX^{-1}B)}{\partial X} = -[X^{-1}B(AX^{-1}B)^{-1}AX^{-1}]' .
(b)

Suppose now that XX is an m×mm \times m symmetric matrix; that, for purposes of differentiating any function of XX, the function is to be interpreted as a function of the column vector xx whose elements are xijx_{ij} (j≤i=1,…,m)(j \leq i = 1, \ldots , m); and that the range of xx is a set SS comprising some or all xx-values for which XX is nonsingular and det⁡(AX−1B)>0\det (AX^{-1}B) > 0. Use the result of Exercise 15.18 to show that log⁡det⁡(AX−1B)\log \det (AX^{-1}B) is continuously differentiable at any interior point cc of SS and that (at x=cx = c)

∂log⁡det⁡(AX−1B)∂X=−K−K′+diag⁡(k11,k22,…,kqq), \frac{\partial \log \det (AX^{-1}B)}{\partial X} = -K - K' + \operatorname {diag}(k_{11}, k_{22}, \ldots , k_{qq}) ,

where K={kij}=X−1B(AX−1B)−1AX−1K = \{ k_{ij}\} = X^{-1}B(AX^{-1}B)^{-1}AX^{-1}.

Problem 15.20

Let F={fis}F = \{ f_{is}\} represent a p×pp \times p matrix of functions, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables. Let cc represent any interior point (of SS) at which FF is continuously differentiable. By, for instance, using the result of Part (b) of Exercise 13.10, show that if rk⁡(F(c))≤p−3\operatorname {rk}\left(F(c)\right) \leq p-3, then

∂adj⁡(F)∂xj=0. \frac{\partial \operatorname {adj}(F)}{\partial x_j} = 0 .
?
Problem 15.21
?
(a)

Let XX represent an m×mm \times m matrix of m2m^2 ``independent'' variables, and suppose that the range of XX is a set SS comprising some or all XX-values for which XX is nonsingular. Show that (when the elements of X−1X^{-1} are regarded as functions of XX) X−1X^{-1} is continuously differentiable at any interior point CC of SS and that (at X=CX = C)

∂X−1∂xij=−yizj′, \frac{\partial X^{-1}}{\partial x_{ij}} = -y_iz_j' ,

where yiy_i represents the iith column and zj′z_j' the jjth row of X−1X^{-1}.

(b)

Suppose now that XX is an m×mm \times m symmetric matrix; that, for purposes of differentiating a function of XX, the function is to be interpreted as a function of the column vector xx whose elements are xijx_{ij} (j≤i=1,…,m)(j \leq i = 1, \ldots , m); and that the range of xx is a set SS comprising some or all xx-values for which XX is nonsingular. Show that X−1X^{-1} is continuously differentiable at any interior point cc of SS and that (at x=cx = c)

∂X−1∂xij={−yiyi′,if j=i,−yiyj′−yjyi′,if j<i \frac{\partial X^{-1}}{\partial x_{ij}} = \begin{cases} -y_iy_i', & \text{if } j = i , \\ -y_iy_j' - y_jy_i', & \text{if } j < i \end{cases}

(where yiy_i represents the iith column of X−1X^{-1}).

Problem 15.22

Let XX represent an m×mm \times m matrix of m2m^2 ``independent'' variables. Suppose that the range of XX is a set SS comprising some or all XX-values for which XX is nonsingular, and let CC represent an interior point of SS. Denote the ijijth element of X−1X^{-1} by yijy_{ij}, the jjth column of X−1X^{-1} by yjy_j, and the iith row of X−1X^{-1} by zi′z_i'.

?
(a)

Show that X−1X^{-1} is twice continuously differentiable at CC and that (at X=CX = C)

∂2X−1∂xij∂xst=yjsyizt′+ytiyszj′. \frac{\partial ^2 X^{-1}}{\partial x_{ij}\partial x_{st}} = y_{js}y_iz_t' + y_{ti}y_sz_j' .
(b)

Suppose that det⁡(X)>0\det (X) > 0 for every XX in SS. Show that log⁡det⁡(X)\log \det (X) is twice continuously differentiable at CC and that (at X=CX = C)

∂2log⁡det⁡(X)∂xij∂xst=−ytiyjs. \frac{\partial ^2 \log \det (X)}{\partial x_{ij}\partial x_{st}} = -y_{ti}y_{js} .
Problem 15.23

Let F={fis}F = \{ f_{is}\} represent a p×pp \times p matrix of functions, defined on a set SS, of a vector x=(x1,…,xm)′x = (x_1, \ldots , x_m)' of mm variables. For any nonempty set T={t1,…,ts}T = \{ t_1, \ldots , t_s\}, whose members are integers between 1 and mm, inclusive, define D(T)=∂sF/∂xt1⋯∂xtsD(T) = \partial^s F/\partial x_{t_1} \cdots \partial x_{t_s}. Let kk represent a positive integer and, for i=1,…,ki = 1, \ldots , k, let jij_i represent an arbitrary integer between 1 and mm, inclusive.

?
(a)

Suppose that FF is nonsingular for every xx in SS, and denote by cc any interior point (of SS) at which FF is kk times continuously differentiable. Show that F−1F^{-1} is kk times continuously differentiable at cc and that (at x=cx = c)

∂kF−1∂xj1⋯∂xjk=∑r=1k∑T1,…,Tr(−1)rF−1D(T1)F−1D(T2)⋯F−1D(Tr)F−1,(E.1) \frac{\partial ^k F^{-1}}{\partial x_{j_1} \cdots \partial x_{j_k}} = \sum _{r=1}^{k} \sum _{T_1,\ldots ,T_r} (-1)^r F^{-1}D(T_1)F^{-1}D(T_2) \cdots F^{-1}D(T_r)F^{-1} , \tag {E.1}

where T1,…,TrT_1, \ldots , T_r are rr nonempty mutually exclusive and exhaustive subsets of {j1,…,jk}\{ j_1, \ldots , j_k\} (and where the second summation is over all possible choices for T1,…,TrT_1, \ldots , T_r).

(b)

Suppose that det⁡(F)>0\det (F) > 0 for every xx in SS, and denote by cc any interior point (of SS) at which FF is kk times continuously differentiable. Show that log⁡det⁡(F)\log \det (F) is kk times continuously differentiable at cc and that (at x=cx = c)

∂klog⁡det⁡(F)∂xj1⋯∂xjk=∑r=1k∑T1,…,Tr(−1)r+1tr⁡[F−1D(T1)F−1D(T2)⋯F−1D(Tr)],(E.2) \begin{aligned} & \frac{\partial ^k \log \det (F)}{\partial x_{j_1} \cdots \partial x_{j_k}} \\ & \quad = \sum _{r=1}^{k} \sum _{T_1,\ldots ,T_r} (-1)^{r+1} \operatorname {tr}[F^{-1}D(T_1)F^{-1}D(T_2) \cdots F^{-1}D(T_r)] , \tag {E.2} \end{aligned}

where T1,…,TrT_1, \ldots , T_r are rr nonempty mutually exclusive and exhaustive subsets of {j1,…,jk}\{ j_1, \ldots , j_k\} with jk∈Trj_k \in T_r (and where the second summation is over all possible choices for T1,…,TrT_1, \ldots , T_r).

Problem 15.24

Let X={xij}X = \{ x_{ij}\} represent an m×mm \times m symmetric matrix, and let xx represent the m(m+1)/2m(m+1)/2-dimensional column vector whose elements are xijx_{ij} (j≤i=1,…,m)(j \leq i = 1, \ldots , m). Define SS to be the set of all xx-values for which XX is nonsingular and S∗S^{*} to be the set of all xx-values for which XX is positive definite. Show that SS and S∗S^{*} are both open sets.

?
Problem 15.25

Let XX represent an n×pn \times p matrix of constants, and let WW represent an n×nn \times n symmetric positive definite matrix whose elements are functions, defined on a set SS, of a vector z=(z1,…,zm)′z = (z_1, \ldots , z_m)' of mm variables. Further, let cc represent any interior point (of SS) at which WW is twice continuously differentiable. Show that W−WPX,WW - WP_{X,W} is twice continuously differentiable at cc and that (at z=cz = c)

∂2(W−WPX,W)∂zi∂zj=(I−PX,W)′∂2W∂zi∂zj(I−PX,W)−(I−PX,W)′∂W∂ziX(X′WX)−X′∂W∂zj(I−PX,W)−[(I−PX,W)′∂W∂ziX(X′WX)−X′∂W∂zj(I−PX,W)]′. \begin{aligned} \frac{\partial ^2(W-WP_{X,W})}{\partial z_i \partial z_j} & = (I-P_{X,W})'\frac{\partial ^2 W}{\partial z_i \partial z_j}(I-P_{X,W}) \\ & \quad - (I-P_{X,W})'\frac{\partial W}{\partial z_i}X(X'WX)^{-}X'\frac{\partial W}{\partial z_j}(I-P_{X,W}) \\ & \quad - \left[(I-P_{X,W})'\frac{\partial W}{\partial z_i}X(X'WX)^{-}X'\frac{\partial W}{\partial z_j}(I-P_{X,W})\right]' . \end{aligned}
?
Problem 15.26

Let XX represent an n×pn \times p matrix and WW an n×nn \times n symmetric positive definite matrix, and suppose that the elements of XX and WW are functions, defined on a set SS, of a vector z=(z1,…,zm)′z = (z_1, \ldots , z_m)' of mm variables. And, let cc represent any interior point (of SS) at which WW and XX are continuously differentiable, and suppose that XX has constant rank on some neighborhood of cc. Further, let BB represent any p×np \times n matrix such that X′WXB=X′WX'WXB = X'W. Then, at z=cz = c,

∂(WPX,W)∂zj=∂W∂zj−(I−PX,W)′∂W∂zj(I−PX,W)+W(I−PX,W)∂X∂zjB+[W(I−PX,W)∂X∂zjB]′.(*) \begin{aligned} \frac{\partial (WP_{X,W})}{\partial z_j} & = \frac{\partial W}{\partial z_j} - (I-P_{X,W})'\frac{\partial W}{\partial z_j}(I-P_{X,W}) \\ & \quad + W(I-P_{X,W})\frac{\partial X}{\partial z_j}B + \left[W(I-P_{X,W})\frac{\partial X}{\partial z_j}B\right]' . \tag {*} \end{aligned}
?
Remark.
?

Derive result (*) by using the result

PX,W′WPX,W=WPX,W(**) P_{X,W}'WP_{X,W} = WP_{X,W} \tag {**}

to obtain the representation

∂(WPX,W)∂zj=PX,W′W∂PX,W∂zj+PX,W′∂W∂zjPX,W+(∂PX,W∂zj)′WPX,W, \frac{\partial (WP_{X,W})}{\partial z_j} = P_{X,W}'W\frac{\partial P_{X,W}}{\partial z_j} + P_{X,W}'\frac{\partial W}{\partial z_j}P_{X,W} + \left(\frac{\partial P_{X,W}}{\partial z_j}\right)'WP_{X,W} ,

and by then making use of the result

(∂PX,W∂zj)′WPX,W=(I−PX,W)′∂W∂zjPX,W+W(I−PX,W)∂X∂zjB.(⋆) \left(\frac{\partial P_{X,W}}{\partial z_j}\right)'WP_{X,W} = (I-P_{X,W})'\frac{\partial W}{\partial z_j}P_{X,W} + W(I-P_{X,W})\frac{\partial X}{\partial z_j}B . \tag {$\star $}