Numerical Functional Analysis and Optimization

ISSN: 0163-0563 (Print) 1532-2467 (Online) Journal homepage: http://www.tandfonline.com/loi/lnfa20

Analysis of Symmetric Matrix Valued Functions

H. Mohebi & A. Salemi

To cite this article: H. Mohebi & A. Salemi (2007) Analysis of Symmetric Matrix Valued
Functions, Numerical Functional Analysis and Optimization, 28:5-6, 691-715, DOI:
10.1080/01630560701380999

To link to this article: https://doi.org/10.1080/01630560701380999

Published online: 17 May 2007.

Submit your article to this journal

Article views: 132

Citing articles: 3 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=lnfa20
Numerical Functional Analysis and Optimization, 28(5–6):691–715, 2007
Copyright © Taylor & Francis Group, LLC
ISSN: 0163-0563 print/1532-2467 online
DOI: 10.1080/01630560701380999

ANALYSIS OF SYMMETRIC MATRIX VALUED FUNCTIONS

H. Mohebi and A. Salemi
Mahani Mathematical Research Center,

University of Kerman, Kerman, Iran

For any symmetric function f :
n −→
n , one can define a corresponding function

on the space of n × n real symmetric matrices by applying f to the eigenvalues of the
spectral decomposition. We show that this matrix valued function inherits from f the
properties of continuity, Lipschitz continuity, strict continuity, directional differentiability, Fréchet
differentiability, and continuous differentiability.

Keywords Continuous differentiability; Directional differentiability; Fréchet

differentiability; Lipschitz continuity; Spectral function; Strict continuity; Symmetric
matrix valued function.

AMS Subject Classification 49R50; 47A56; 15A18.

1. INTRODUCTION
There has been growing interest in the variational analysis of spectral
functions. This growing trend is due to the following reasons. On one
hand, spectral functions have important applications to some fundamental
problems in applied mathematics such as semidefinite programs and
engineering problems (see [7, 13] for many such applications). On the
other hand, efficient nonsmooth analysis tools have only been available in
the past few years (see [21]). Classical interest in spectral functions arose
from their important role in quantum mechanics [9, 22]. Nowadays they
are an inseparable part of optimization [13] and matrix analysis [4, 5].
In modern optimization, the key example is semidefinite programming,
where one encounters problems involving spectral functions. There are
many examples where a property of the spectral function is actually
equivalent to the corresponding property of the underlying symmetric
function. Among them are first-order differentiability [10], convexity

Address correspondence to H. Mohebi, Mahani Mathematical Research Center, University of

Kerman, P.O. Box 76135-133, Kerman, Iran; E-mail: hmohebi@mail.uk.ac.ir

691
692 H. Mohebi and A. Salemi

[12], generalized first-order differentiability [10, 11], analyticity [28],

and various second-order properties [25–27]. Our study is inspired by
the recent progress on spectral functions [10, 11, 19] and progress on
symmetric matrix valued functions [2, 3, 6, 23, 24, 29].
Let  be the space of real symmetric matrices. We denote by 
the group of all real orthogonal matrices. For any A ∈  , its (repeated)
eigenvalues
1 ≥ · · · ≥
n are real, and it admits a spectral decomposition
of the form

A = P diag[
1 ,    ,
n ]P T , (1.1)

for some P ∈ , where diag[
1 ,    ,
n ] is the diagonal matrix with its ith
diagonal entry
i  Note that (1.1) is independent of the choice of P ∈ 
(see [1, 5]).
Let
(·) :  −→
n be the eigenvalue function such that
i (A), i =
1,    , n, yield eigenvalues of A for any A ∈  and are ordered in a
nonincreasing order, that is,
1 (A) ≥ · · · ≥
n (A) A function f :
n −→

is called symmetric if f is invariant under coordinate permutation, that
is, f (x) = f (Px) for any permutation matrix P and any x ∈
n  A spectral
function is a composition of a symmetric function f :
n −→
and the
eigenvalue function
(·) :  −→
n , that is, the spectral function f ◦
:
 −→
is defined by

(f ◦
)(A) := f (
(A)) (A ∈  )

In the sequel, let  denote the set of all permutation matrices in
n×n 
For given
∈
n , 
denotes the stabilizer of
defined by


:= P ∈  : P
=


We say that a function f :
n −→
n satisfies the property () if

∈
n and P ∈ 
, then Pf (
) = f (P
)

It is well-known that various continuity and differentiability properties

of spectral functions have been studied in [14, 19], and [23]. Recently,
Lewis and Sendov (see [14]) proved that a spectral function is twice
differentiable at a matrix if and only if the corresponding symmetric
function is twice differentiable at the vector of eigenvalues. In [3], Chen,
Qi, and Tseng showed that for any function f :
−→
, one can define a
corresponding function f
:  −→  (see [1, 5]) by

f
(A) := P diag[f (
1 ),    , f (
n )]P T , A ∈ ,

 Analysis of Symmetric Matrix Valued Functions 693

such that the properties of continuity, strict continuity, Lipschitz

continuity, directional differentiability, differentiability, and continuous
differentiability are each inherited by f
from f 
Now, corresponding with (1.1), for each symmetric function f :

n −→
n with

f (x) = (f1 (x),    , fn (x))T (x ∈
n ),

which satisfies the property (), we define a function f˜ :  −→  by

f˜ (A) := P diag[(f1 ◦
)(A),    , (fn ◦
)(A)]P T (A ∈  ) (1.2)

Note that f˜ (A) is independent of the choice of p ∈  and that belongs to  .

In this paper, we study some nonsmoothness properties of f˜ . Indeed,
we show that the properties of continuity, strict continuity, Lipschitz
continuity, directional differentiability, differentiability, and continuous
differentiability are each inherited by f˜ from f  In fact, the aim of
this paper is to study various continuity and differentiability properties
of a class of symmetric matrix valued functions, which are natural
extensions of vector valued functions to matrix valued functions. We use
techniques in [3] to obtain new results (for example, see Theorems 3.1,
4.1, Proposition 4.1, and Theorem 5.1) for vector valued mappings. We
shall apply our results to the semidefinite complementarity problem (see
[15, 17, 18]) for vector valued mappings, and leave the application of our
results to the semidefinite complementarity problem for future research.

2. PRELIMINARIES
Let  be the space of real symmetric matrices. It is clear that 
is a subspace of
n×n with dimension n(n + 1)/2 For any A ∈  with
eigenvalues
1 ≥ · · · ≥
n , define

P diag[
1 ,    ,
n ]P T := P diag[
(A)]P T , P ∈ 

Definition 2.1. A function f :
n −→
n with

f (x) = (f1 (x),    , fn (x))T (x ∈
n ),

is called symmetric if each fi is a symmetric function (i = 1, 2,    , n)

Definition 2.2. We say that a function f :
n −→
n satisfies the property

() if

∈
n and P ∈ 
, then Pf (
) = f (P
)

694 H. Mohebi and A. Salemi

In this paper, we use the following notations: vectors in
n are viewed as

columns, and capital letters such as A, B, and so forth, denote matrices
in   For each
∈
n , define diag[
] := diag[
1 ,    ,
n ].
Throughout,  ·  denotes the Frobenius norm for matrices and the 2-
norm for vectors. For any x ∈
n , A ∈  , and any scalar r > 0, we denote
the ball with center at x and radius r > 0 and the ball with center at A and
radius r > 0, respectively, by

 (x, r ) := y ∈
n : y − x ≤ r ,

and

(A, r ) := B ∈  : B − A ≤ r 

For any
∈
n and P ∈ , we use the following fact:

diag[P
] = P diag[
]P T 

We recall (see [21]) that a mapping F :
k −→
l is continuous at

x ∈
k if F (y) −→ F (x) as y −→ x, and F is continuous if F is continuous
at every x ∈
k  F is strictly continuous (also called, locally Lipschitz
continuous) at x ∈
k , if there exist scalars K > 0 and  > 0 such that

F (y) − F (z) ≤ K y − z ∀y, z ∈
k with y, z ∈  (x, ),

and F is strictly continuous if F is strictly continuous at every x ∈
k  If

 can be taken to be +∞, then F is Lipschitz continuous with Lipschitz
constant K  Define the function lip F :
k −→ [0, +∞] by

F (y) − F (z)
lip F (x) := lim sup 
y,z−→x y − z
y =z

Then F is strictly continuous at x if and only if lip F (x) is finite.

We recall (see [21]) that F is directionally differentiable at x ∈
k if

F (x + th) − F (x)
F (x, h) := lim exists ∀h ∈
k ,
t −→ 0 t
and F is directionally differentiable if F is directionally differentiable at
every x ∈
k  We say that F is semidifferentiable at x ∈
k if the limit

F (x + t ĥ) − F (x)
lim
t −→ 0 t
ĥ −→ h
 Analysis of Symmetric Matrix Valued Functions 695

exists for every direction h ∈
k  It is easy to see that the limit (if it exists)
equals F (x, h) We say that F is differentiable (in the Fréchet sense) at
x ∈
k if there exists a linear mapping F (x) :
k −→
l such that

F (x + h) − F (x) − F (x)h = o(h),

where we write z = o(), with  ∈
, to mean z||

tends to zero as  −→ 0
z
Also, we write z = O(), with  ∈
, to mean || is uniformly bounded as
 −→ 0
F is called continuously differentiable if F is differentiable at every
x ∈
k and F is continuous.
In the sequel, we review some useful perturbation results for the
spectral decomposition of real symmetric matrices. Let  denote the group
of n × n real orthogonal matrices. For each A ∈  , define the set of all
orthogonal matrices giving the ordered spectral decomposition of A by

A := P ∈  : P T AP = diag[
(A)]

It is clear that A is nonempty for each A ∈  .

The following lemma gives a key perturbation result for eigenvectors of
symmetric matrices (see [2, 24]).

Lemma 2.3. For any A ∈  , there exist scalars  > 0 and  > 0 such that

min P − Q  ≤ A − B ∀B ∈ (A, ), ∀Q ∈ B 

P ∈A

We give the following perturbation result of von Neumann (see [10]).

Lemma 2.4. For any A, B ∈  , we have


(A) −
(B) ≤ A − B

and

|
i (A) −
i (B)| ≤ A − B2 ∀i = 1,    , n,

where  · 2 is the 2-norm.

Finally, we present the following result (see [8, 20]) showing that for
any A ∈  and any H ∈  , the orthonormal eigenvectors of A + tH can be
chosen as a power series in t for all t ∈


Lemma 2.5. For any A ∈  and any H ∈  , there exist P (t ) ∈ , t ∈
,

whose entries are power series in t , convergent in a neighborhood of t = 0, and
P (t )T (A + tH )P (t ) is diagonal.
696 H. Mohebi and A. Salemi

3. CONTINUITY PROPERTIES OF SYMMETRIC MATRIX

VALUED FUNCTIONS
In this section, we show that a symmetric function f :
n −→
n that
satisfies the property () is continuous if and only if the symmetric matrix
valued function f˜ defined by (1.2) is continuous.
The following preliminary result, which shall be used in our proofs, are
due to the symmetry of f (see [14, 19]).

Lemma 3.1. Suppose f :
n −→
is symmetric. Then the following assertions

are true:

(1) f is continuous at
∈
n if and only if f is continuous at P
for any P ∈ 

(2) f is directionally differentiable at
∈
n along h if and only if f is
directionally differentiable at P
along Ph for any P ∈ 
(3) f is semidifferentiable at
∈
n if and only if f is semidifferentiable at P
for
any P ∈ 
(4) f is differentiable at
∈
n if and only if f is differentiable at P
for any
P ∈  In particular, f (P
) = P f (
) Moreover, if P ∈ 
, then f (
) =
P f (
)

Remark 3.2. Note that by part (4) of Lemma 3.1, we conclude that
f satisfies the property () for any symmetric differentiable function
f :
n −→
.

Theorem 3.3. Let f :
n −→
n be a symmetric function with f (x) =

(f1 (x),    , fn (x))T (x ∈
n ) such that f satisfies the property (), and let f˜
be defined by (1.2). Let A ∈  with eigenvalues
1 (A),    ,
n (A) and
(A) :=
(
1 (A),    ,
n (A))T ∈
n  Then the following assertions are equivalent:

(1) f is continuous at
(A)

(2) f˜ is continuous at A.

Proof. (1) ⇒ (2) Suppose that f is continuous at
(A) Then f1 ,    , fn

are continuous at
(A) Because the eigenvalue function
(·) is continuous
on  , it follows that f1 ◦
,    , fn ◦
are continuous at A Because
i (A) (i =
1,    , n) are patterned in nonincreasing order, that is,
1 (A) ≥ · · · ≥
n (A),
it follows from Lemma 2.3 that there exist  > 0 and  > 0 such that

min P − Q  ≤ A − B ∀B ∈ (A, ), ∀Q ∈ B  (3.1)

P ∈A

Now, let B ∈ (A, ) and Q ∈ B be arbitrary. Then by (3.1) there exists

P ∈ A such that

P − Q  ≤ A − B (3.2)

 Analysis of Symmetric Matrix Valued Functions 697

Let
1 (B),    ,
n (B) be eigenvalues of B with
1 (B) ≥ · · · ≥
n (B) and

(B) := (
1 (B),    ,
n (B))T  Therefore we have

Q T BQ = diag[
(B)], and P T AP = diag[
(A)]

For simplicity, let D1 := diag[(f1 ◦
)(B),    , (fn ◦
)(B)] and D2 := diag[(f1 ◦
)
(A),    , (fn ◦
)(A)] Consider

f˜ (B) − f˜ (A) = Q diag[(f1 ◦
)(B),    , (fn ◦
)(B)]Q T

− P diag[(f1 ◦
)(A),    , (fn ◦
)(A)]P T
= Q (D1 − D2 )Q T + (Q − P )D2 Q T + PD2 (Q − P )T  (3.3)

Because by the above, f1 ◦
,    , fn ◦
are continuous at A, it follows that

|(fi ◦
)(A) − (fi ◦
)(B)| −→ 0 as A − B −→ 0 (i = 1,    , n)

This implies that

D1 − D2  −→ 0 as A − B −→ 0 (3.4)

It follows from (3.2), (3.3), and (3.4) that f˜ (A) − f˜ (B) −→ 0 as
A − B −→ 0, and hence f˜ is continuous at A.

(2) ⇒ (1) Suppose that f˜ is continuous at A Fix any P ∈ A and let

y = (y1 ,    , yn )T ∈
n and y −
(A) −→ 0 It follows that |yi −
i (A)| −→ 0
for all i = 1,    , n This implies that

B := P diag[y]P T −→ P diag[
(A)]P T = A

In view of Lemma 3.1, we may assume without loss of generality that

y1 ≥ · · · ≥ yn . Therefore, because f˜ is continuous at A, we conclude that

P diag[f1 (
(B)),    , fn (
(B))]P T −→ P diag[f1 (
(A)),    , fn (
(A))]P T ,

where
(B) := (y1 ,    , yn )T = y Thus we have |fi (y) − fi (
(A))| −→ 0 for

all i = 1,    , n Then, f1 ,    , fn are continuous at
(A), and hence f is
continuous at
(A), which completes the proof.

Corollary 3.4. Let f :
n −→
n be a symmetric function with f (x) =

(f1 (x),    , fn (x))T (x ∈
n ) such that f satisfies the property () Then f is
continuous on
n if and only if f˜ is continuous on  
698 H. Mohebi and A. Salemi

4. DIFFERENTIAL PROPERTIES OF SYMMETRIC MATRIX

VALUED FUNCTIONS
We shall show that if a symmetric function f :
n −→
n has the
properties of differentiability and continuous differentiability, then so does
the symmetric matrix valued function f˜ defined by (1.2).
In the sequel, let the differentiable function f :
n −→
n be
given with f (x) = (f1 (x),    , fn (x))T , where f1 ,    , fn :
n −→
are real-
valued functions and x = (x1 ,    , xn )T ∈
n . Define the symmetric matrix
f [1] (x) by:


 fi (x) − fj (x)

 , xi = xj
[1]
xi − xj
f (x)ij = (4.1)

 0, xi = xj , i = j

f (x) ,
ii i = j

where i, j ∈ 1,    , n

Theorem 4.1. Let f :
n −→
n be a symmetric function with f (x) =

(f1 (x),    , fn (x))T (x ∈
n ) such that f satisfies the property () Let A ∈  with
eigenvalues
1 (A),    ,
n (A) and
(A) := (
1 (A),    ,
n (A))T ∈
n  Then the
following assertions are equivalent:

(1) f is differentiable at
(A).

(2) f˜ is differentiable at A.

Moreover,  f˜ (A) is given by

 f˜ (A)H = P (f [1] (
(A)) ◦ (P T HP ))P T (A ∈  ), (4.2)

for any P ∈ A , where the matrix f [1] (
(A)) is defined by (4.1).

Proof. (1) ⇒ (2) Suppose that (1) holds. Then, f1 ,    , fn are

differentiable at
(A) We have
i (A) (i = 1,    , n) are patterned in
nonincreasing order, it follows from Lemma 2.3 that there exist  > 0 and
 > 0 such that

min P − Q  ≤ A − B ∀B ∈ (A, ), ∀Q ∈ B  (4.3)

P ∈A

We show that for any H ∈  with H  ≤ , there exists P ∈ A such that

f˜ (A + H ) − f˜ (A) − P (f [1] (
(A)) ◦ (P T HP ))P T = o(H ) (4.4)

 Analysis of Symmetric Matrix Valued Functions 699

Now, fix H ∈  with H  ≤ . Let B := A + H and Q ∈ B be arbitrary.

Then by (4.3) there exists P ∈ A such that

P − Q  ≤ H  (4.5)

Let
1 (B),    ,
n (B) be eigenvalues of B with
1 (B) ≥ · · · ≥
n (B) and

(B) := (
1 (B),    ,
n (B))T  Therefore, we have

B = Q diag[
(B)]Q T , A = P diag[
(A)]P T ,
(4.6)
f˜ (B) = Q diag[(f1 ◦
)(B),    , (fn ◦
)(B)]Q T ,

and

f˜ (A) = P diag[(f1 ◦
)(A),    , (fn ◦
)(A)]P T 

For simplicity, let

E := f˜ (B) − f˜ (A) − P (f [1] (
(A)) ◦ (P T HP ))P T ,


E := P T EP , H := P T HP , X := diag[(f1 ◦
)(B),    , (fn ◦
)(B)], Y :=
diag[(f1 ◦
)(A),    , (fn ◦
)(A)], Z := Q T P , C := f [1] (
(A)),
:=
(A),
:=
(B),
i :=
i (A), and i :=
i (B), i = 1,    , n Then we have

 
E = Z T XZ − Y − C ◦ H (4.7)

It follows from (4.6) that

diag[
1 ,    ,
n ] = P T AP = Z T diag[ 1 ,    , n ]Z
,
−H

and so


n

i , if i = j
zki zkj k − h̃ij = (4.8)
k=1
0, if i = j

where i, j ∈ 1,    , n

Because Z = Q T P , we conclude that Z = (Q − P )T P + I  In view of
(4.5), it follows that

zij = O(H ) ∀i = j , i, j = 1,    , n (4.9)

Because P , Q ∈ , then Z ∈ , and hence Z T Z = I  It follows that


n
1= zii2 + zki2 = zii2 + O(H 2 ) (i = 1,    , n), (4.10)
k =i,k=1
700 H. Mohebi and A. Salemi

and

n
0 = zii zij + zji zjj + zki zkj
k =i,j ,k=1

= zii zij + zji zjj + O(H 2 ) (i = j , i, j = 1,    , n) (4.11)

Because E  = E , it is enough to show that 

E = o(H ), and then we
have (4.4). To do this, consider three cases:
Case i. Suppose that i = j and i ∈ 1,    , n Then by the local
boundedness of f1 ,    , fn and that (4.7)–(4.11) hold, it follows that

n 
n
ẽii = zki2 fk ( ) − fi (
) − (Dl fi )(
)h̃ll
k=1 l =1


n 
n  
n
= zki2 fk ( ) − fi (
) − (Dl fi )(
) −
l + zkl2 k
k=1 l =1 k=1


n
= zii2 fi ( ) − fi (
) − (Dl fi )(
) −
l + zll2 l + O(H 2 )
l =1

= (1 + O(H  ))fi ( ) − fi (
)

2


n
− (Dl fi )(
)(−
l + (1 + O(H 2 )) l ) + O(H 2 )
l =1


n
= fi ( ) − fi (
) − (Dl fi )(
)( l −
l ) + O(H 2 )
l =1

= fi ( ) − fi (
) − ((D1 fi )(
),    , (Dn fi )(
))T ( −
) + O(H 2 ) (4.12)

Because, by Lemma 2.4, we have  −
 ≤ H  and by hypothesis each

fi is differentiable at
, hence in view of (4.12), we get ẽii = o(H ), (i =
1,    , n)
Case ii. Assume that i = j ,
i =
j , and i, j ∈ 1,    , n Then, in view
of (4.1), we have cij = 0 Due to the local boundedness of f1 ,    , fn and that
(4.7), (4.9), and (4.11) hold, we obtain

n
ẽij = zki zkj fk ( )
k=1

= zii zij fi ( ) + zji zjj fj ( ) + O(H 2 )

= (zii zij + zji zjj )fi ( ) + zji zjj (fj ( ) − fi ( )) + O(H 2 )
= zji zjj (fj ( ) − fi ( )) + O(H 2 ) (4.13)
 Analysis of Symmetric Matrix Valued Functions 701

Because
i =
j and f satisfies the property (), we conclude that fi (
) =

fj (
) In view of (4.13), we get

ẽij = zji zjj ((fj ( ) − fj (
)) − (fi ( ) − fi (
))) + O(H 2 ) (4.14)

By Lemma 2.4, we have  −
 ≤ H  Because fi and fj are continuous at

, then fj ( ) − fj (
) = O( −
), and hence fj ( ) − fj (
) = O(H ) and
fi ( ) − fi (
) = O(H ) Therefore, it follows from (4.9) and (4.14) that
ẽij = o(H ).
Case iii. Suppose that
i =
j and i, j ∈ 1,    , n Then, in view of
(4.1), we have

fj (
) − fi (
)
cij =  (4.15)

j −
i

Because of the local boundedness of f1 ,    , fn and that (4.7), (4.9), (4.11),

and (4.15) hold, we get


n
ẽij = zki zkj fk ( ) − cij h̃ij
k=1


n 
n
= zki zkj fk ( ) − cij zki zkj k
k=1 k=1

= zii zij fi ( ) + zji zjj fj ( ) − cij (zii zij i + zji zjj j ) + O(H 2 )
= (zii zij + zji zjj )fi ( ) + zji zjj (fj ( ) − fi ( ))
− cij ((zii zij + zji zjj ) i + zji zjj ( j − i )) + O(H 2 )
= zji zjj (fj ( ) − fi ( ) − cij ( j − i )) + O(H 2 )

fj (
) − fi (
)
= zji zjj fj ( ) − fi ( ) − ( j − i ) + O(H 2 )

j −
i
= zji zjj ((fj ( ) − fi ( )) − (fj (
) − fi (
))(1 + (( j −
j ) − ( i −
i ))(
j −
i )−1 )
+ O(H 2 ) (4.16)

By Lemma 2.4, we have  −
 ≤ H , and hence | j −
j | ≤ H  and

| i −
i | ≤ H  Because fi , fj are continuous at
and that (4.9) holds, in
view of (4.16) and that fj ( ) − fj (
) = O(H ) and fi ( ) − fi (
) = O(H ),
we obtain ẽij = o(H ), which completes the proof of the implication
(1) ⇒ (2)
702 H. Mohebi and A. Salemi

(2) ⇒ (1) Suppose that f˜ is differentiable at A First, note that by

Lemma 2.3 there exist  > 0 and  > 0 such that

min P − Q  ≤ A − B ∀B ∈ (A, ), ∀Q ∈ B  (4.17)

P ∈A

Let H ∈  with H  ≤  and Q ∈ A+H be arbitrary. Then due to (4.17)

there exists P ∈ A such that

P − Q  ≤ H  (4.18)

Now, define the function D :  −→  by

D(S ) := P T ( f˜ (A)S )P (S ∈  ) (4.19)

Because by hypothesis f˜ is differentiable at A, it follows that D is well

defined and linear. We also define the function Ci :  −→
by

Ci (S ) := (D(S ))ii (S ∈  ; i = 1,    , n) (4.20)

In view of (4.19), it is not difficult to show that Ci (i = 1,    , n) is linear.

Because by hypothesis f˜ is differentiable at A, then we have

f˜ (A + H ) − f˜ (A) −  f˜ (A)H = o(H ),

and hence by using (4.19) we get

R := Q diag[(f1 ◦
)(A + H ),    , (fn ◦
)(A + H )]Q T

− P diag[(f1 ◦
)(A),    , (fn ◦
)(A)]P T − PD(H )P T = o(H )
(4.21)

For simplicity, let E := diag[(f1 ◦
)(A + H ),    , (fn ◦
)(A + H )], F :=

 := P T RP  Due to (4.21), we obtain
diag[(f1 ◦
)(A),    , (fn ◦
)(A)] and R

 = Z T EZ − F − D(H ) = o(H ),

R (4.22)

where Z := Q T P  Because P , Q ∈ , then Z ∈ , and hence Z T Z = I 

Because Z = Q T P , we conclude that Z = (Q − P )T P + I  In view of (4.18),
it follows that

zij = O(H ) ∀i = j , i, j = 1,    , n (4.23)

 Analysis of Symmetric Matrix Valued Functions 703

Because Z T Z = I , by (4.23) we get


n
1 = zii2 + zki2 = zii2 + O(H 2 ) (i = 1,    , n) (4.24)
k =i,k=1

Now, for any i ∈ 1,    , n, it follows from (4.22) that


n
zki2 (fk ◦
)(A + H ) − (fi ◦
)(A) − (D(H ))ii = o(H )
k=1

Hence we have

n
zii2 (fi ◦
)(A + H ) + zki2 (fk ◦
)(A + H ) − (fi ◦
)(A) − (D(H ))ii
k =i,k=1

= o(H ) (4.25)

Because f˜ is differentiable at A, thus f˜ is continuous at A By Theorem 3.3,

we conclude that f1 ◦
,    , fn ◦
are continuous at A, and hence f1 ◦

,    , fn ◦
are locally bounded around A This, together with (4.20),
(4.24), and (4.25) imply that

(fi ◦
)(A + H ) − (fi ◦
)(A) − Ci (H ) = o(H ) + O(H 2 )

= o(H ) (i = 1,    , n) (4.26)

It follows from (4.26) that f1 ◦
,    , fn ◦
are differentiable at A

and (fi ◦
) (A) = Ci , i = 1,    , n Then, by ([10], Theorem 1.1 and
Corollary 2.5), f1 ,    , fn are differentiable at
(A), which completes the
proof.

Corollary 4.2. Let f :
n −→
n be a symmetric function with f (x) =

(f1 (x),    , fn (x))T (x ∈
n ) such that f satisfies the property () Then f is
differentiable on
n if and only if f˜ is differentiable on  

Proposition 4.3. Let f :
n −→
n be a symmetric function with f (x) =

(f1 (x),    , fn (x))T (x ∈
n ) such that f satisfies the property () Then f is
continuously differentiable on
n if and only if f˜ is continuously differentiable
on  

Proof. By Lemma 3.1 and Corollary 4.2, we have f is differentiable on

n if and only if f˜ is differentiable on   Now, let A ∈  with eigenvalues

1 (A),    ,
n (A) and
(A) := (
1 (A),    ,
n (A))T ∈
n  Assume that f is
704 H. Mohebi and A. Salemi

continuous at
(A) Then, in view of (4.1) and (4.2), we conclude that f [1]
is continuous at
(A). By Lemma 2.3, there exist  > 0 and  > 0 such that

min P − Q  ≤ A − B ∀B ∈ (A, ), ∀Q ∈ B  (4.27)

P ∈A

Now, let B ∈ (A, ) and Q ∈ B be arbitrary. It follows from (4.27) that

there exists P ∈ A such that

P − Q  ≤ A − B (4.28)

Let
1 (B),    ,
n (B) be the eigenvalues of B with
1 (B) ≥ · · · ≥
n (B) and

(B) := (
1 (B),    ,
n (B))T  Therefore, we have Q T BQ = diag[
(B)] and
P T AP = diag[
(A)]
Let H ∈  with H  ≤ 1, and let D1 := f [1] (
(B)) ◦ Q T HQ and
D2 := f [1] (
(A)) ◦ P T AP . Consider

 f˜ (B)(H ) −  f˜ (A)(H ) = QD1 Q T − PD2 P T

= Q (D1 − D2 )Q T + (Q − P )D2 Q T + PD2 (Q − P )T ,

where

D1 − D2 = f [1] (
(B)) ◦ [Q T HQ − P T HP ] + [f [1] (
(B)) − f [1] (
(A))] ◦ P T HP

= f [1] (
(B)) ◦ Q T H (Q − P ) + f [1] (
(B)) ◦ (Q − P )T HP
+ [f [1] (
(B)) − f [1] (
(A))] ◦ P T HP 

Because P  = Q  = 1, it follows that

( f˜ (B) −  f˜ (A))(H )

≤ 2f [1] (
(B))||Q − P  + f [1] (
(B)) − f [1] (
(A)) + 2Q − P 

Because f [1] is continuous at
(A) and
(·) is continuous at A, in view of

(4.28) we conclude that

 f˜ (B) −  f˜ (A) = sup ( f˜ (B) −  f˜ (A))(H ) −→ 0 as A − B −→ 0

H ≤1

Conversely, suppose that  f˜ is continuous at A Let y =

(y1 ,    , yn )T ∈
n be such that y −
(A) −→ 0 It follows from
Lemma 2.4 that |yi −
i (A)| −→ 0 for all i = 1,    , n This implies
that B := P diag[y]P T −→ P diag[
(A)]P T = A By Lemma 3.1, we
may assume without loss of generality that y1 ≥ · · · ≥ yn  Because
 f˜ (B)(H ) −→  f˜ (A)(H ) for all H ∈  , we conclude (for H = I ) that
 Analysis of Symmetric Matrix Valued Functions 705

fi
xi
(
(B)) −→ xfii (
(A)) for all i = 1,    , n But, we have fi is a symmetric
function (i = 1,    , n) Hence, xfij (
(B)) −→ xfij (
(A)) for all i, j =
1,    , n, and so f is continuous at
(A), which completes the proof.

5. DIRECTIONAL DIFFERENTIABILITY OF SYMMETRIC MATRIX

VALUED FUNCTIONS
In this section, we shall show that if a function f :
n −→
n has
the properties directional differentiability, strict continuity, and Lipschitz
continuity, then so does the symmetric matrix valued function f˜ defined by
(1.2).
We need the following assertion (see [21], Theorem 9.67; and [3],
Lemma 4.5).

Lemma 5.1. Suppose f :
n −→
is strictly continuous. Then there exists

a sequence of continuously differentiable functions gm :
n −→
, m = 1, 2,    ,
converging uniformly to f on any compact set C in
n and satisfying

|gm (x)| ≤ sup lip f (x) ∀x ∈ C , ∀m

x∈C

Theorem 5.2. Let f :
n −→
n be a symmetric function with f (x) =

(f1 (x),    , fn (x))T (x ∈
n ) such that f satisfies the property () Let A ∈  with
eigenvalues
1 (A),    ,
n (A) and
(A) := (
1 (A),    ,
n (A))T ∈
n  Then the
following assertions are true:

(1) If f is semidifferentiable at
(A) Then f˜ is directionally differentiable at A

(2) If f˜ is directionally differentiable at A, then f is directionally differentiable at

(A)
(3) If f is directionally differentiable and strictly continuous at
(A), then f˜ is
directionally differentiable at A

Proof. (1) Suppose that f is semidifferentiable at
(A) Let H ∈  be

arbitrary. Define A(t ) = A + tH , t ∈
 By Lemma 2.5 there exists P (t ) ∈
A(t ) , whose entries are power series in t , convergent in a neighborhood 
of 0, and satisfies

A(t ) = P (t )diag[
1 (t ),    ,
n (t )]P (t )T ,

where

i (t ) = [P (t )T A(t )P (t )]ii , i = 1, 2,    , n,

are power series in t , convergent for any t ∈  

706 H. Mohebi and A. Salemi

Let
(t ) := (
1 (t ),    ,
n (t ))T , t ∈
 Then we have

(t ) =
(0) + t
(0) + o(t ), t ∈
,

where
(0) =
(A) Now, we show that fi ◦
(i = 1, 2,    , n) is directionally

differentiable at A To do this, Consider
(fi ◦
)(A + tH ) − (fi ◦
)(A)
lim
t −→0 t
fi (
(A + tH )) − fi (
(A))
= lim
t −→0 t
fi (
(t )) − fi (
(0))
= lim
t −→0 t
fi (
(0) + t
(0) + o(t )) − fi (
(0))
= lim
t −→0 t
fi
(0) + t
(0) + o(tt ) − fi (
(0))
= lim
t −→0 t
= f (
(0);
(0)),

because fi is semidifferentiable at
(0) =
(A) Therefore, fi ◦
is

directionally differentiable at A and we have

(fi ◦
) (A; H ) = fi (
(0);
(0)), i = 1, 2,    , n

Hence

(fi ◦
)(A + tH ) = (fi ◦
)(A) + tfi (
(0);
(0)) + o(t ), i = 1, 2,    , n

(5.1)

We have

A(t ) = P (t )diag[
1 (t ),    ,
n (t )]P (t )T 

Because P (t )T P (t ) = I (t ∈
), we get

P (t )T A(t ) = diag[
1 (t ),    ,
n (t )]P (t )T = (t )P (t )T ,

where (t ) = diag[
1 (t ),    ,
n (t )] It follows that

P (t )T A(t ) + P (t )T A (t ) = (t )P (t )T + (t )P (t )T ,

where (t ) = diag[
1 (t ),    ,
n (t )] This implies that

P (t )T A(t )P (t ) + P (t )T A (t )P (t ) = (t )P (t )T P (t ) + (t )P (t )T P (t )
 Analysis of Symmetric Matrix Valued Functions 707

Because A (t ) = H , we obtain

P (t )T A(t )P (t ) + P (t )T HP (t ) = (t ) + (t )P (t )T P (t )

 (t ) = P (t )T HP (t ) and 
Now, let H P (t ) = P (t )T P (t ) Then we have

 (t ) =
P (t )T A(t )P (t ) + H (t ) + (t )
P (t )

Because A(t ) = P (t ) (t )P (t )T , it follows that

 (t ) =
P (t )T P (t ) (t )P (t )T P (t ) + H (t ) + (t )
P (t )

We conclude that

  (t ) =
P (t ) (t ) + H (t ) + (t )
P (t ),

and hence

 (t ) = 
(t ) − H P (t ) (t ) − (t )
P (t )

This implies that

 (t )ii =
(t )ii ,
H i = 1, 2,    , n, (5.2)

and

 (t )ij = 
H P (t )ij (
j (t ) −
i (t )), i = j , i, j = 1, 2,    , n (5.3)

But, we have P (t )T P (t ) = I  Then

P (t )T P (t ) + P (t )T P (t ) = 0,

and hence 
P T = −
P , where 
P =
P (0) Because P is differentiable at t = 0,
we conclude that

P (t ) = P + tP + o(t ), where P = P (0) and P = P (0) (5.4)

Also, we have P and P are continuous at t = 0, and so

 (t ) = P T HP
lim H and lim 
P (t ) = 
P (5.5)
t −→0 t −→0

It follows from (5.2) and (5.3) that

 (0)ii = (P T HP )ii

i (0) = H
708 H. Mohebi and A. Salemi

and

 (0)ij = 
H Pij (
j (0) −
i (0)), i = j ; i, j = 1, 2,    , n

This implies that

(P T HP )ij

Pij = whenever
i (0) =
j (0), (5.6)

j (0) −
i (0)

and

 (0)ij = (P T HP )ij = 0
H whenever
i (0) =
j (0), i = j  (5.7)

Now, define the symmetric matrix f [1] (
(0); P T HP ) by


 fi (
(0)) − fj (
(0)) T


i (0) =
j (0)

i (0) −
j (0) (P HP )ij ,
[1]
f (
(0); P HP )ij =
T

fi (
(0); ((P T HP )11 ,    , (P T HP )nn )), i = j

0, otherwise,
(5.8)

where i, j ∈ 1,    , n

In view of (5.1), (5.4)–(5.8) and that 
P T = −
P and P T
=
P P T , we get

f˜ (A + tH ) = P (t )diag[f1 (
(A + tH )),    , fn (
(A + tH ))]P (t )T

= (P + tP + o(t ))diag[f1 (
(A + tH )),    , fn (
(A + tH ))]
× (P + tP + o(t ))T
= (P + tP + o(t ))
diag[t f1 (
(0);
(0)) + f1 (
(0)) + o(t ),    , t fn (
(0);
(0))
+ fn (
(0)) + o(t )](P T + t P + o(t ))
T

= P diag[f1 (
(0)),    , fn (
(0))]P T

+ tP diag[f1 (
(0);
(0)),    , fn (
(0);
(0))]P T
+ t P diag[f1 (
(0)),    , fn (
(0))]P T
+ P diag[f1 (
(0)),    , fn (
(0))]P  + o(t )
T

= f˜ (A) + tP diag[f1 (
(0);
(0)),    , fn (
(0);
(0))]P T

+ tP 
P T diag[f1 (
(0)),    , fn (
(0))]
+diag[f1 (
(0)),    , fn (
(0))]
P P T + o(t )
 Analysis of Symmetric Matrix Valued Functions 709

= f˜ (A) + tP diag[f1 (
(0);
(0)),    , fn (
(0);
(0))]P T

+ tP −
P diag[f1 (
(0)),    , fn (
(0))]
+ diag[f1 (
(0)),    , fn (
(0))]
P P T + o(t )
= f˜ (A) + tP diag[f1 (
(A);
(0)),    , fn (
(A);
(0))]P T
+ tP [
Pij (fi (
(A)) − fj (
(A)))]ni,j =1 P T + o(t )
= f˜ (A) + tP diag[f1 (
(A); ((P T HP )11 ,    , (P T HP )nn ),    ,
fn (
(A); ((P T HP )11 ,    , (P T HP )nn )]
+ [
Pij (fi (
(A)) − fj (
(A))]ni,j =1 P T + o(t )
= f˜ (A) + tPf [1] (
(A); P T HP )P T + o(t )

Therefore, we have

f˜ (A + tH ) − f˜ (A)
lim = Pf [1] (
(A); P T HP )P T 
t −→0 t

Hence f˜ is directionally differentiable at A and (f˜ ) (A; H ) = Pf [1] (
(A);

P T HP )P T , for some P ∈ A 

(2) Suppose that f˜ is directionally differentiable at A Fix any P ∈ A

such that A = P diag[
1 (A),    ,
n (A)]P T  Let h = (h1 , h2 ,    , hn ) ∈
n be
arbitrary and

H = P diag[h1 , h2 ,    , hn ]P T 

Then we have (f˜ ) (A; H ) exists and

f˜ (A + tH ) − f˜ (A)
(f˜ ) (A; H ) = lim
t −→0 t
P diag[f1 (
(A + tH )) − f1 (
(A)),    ,
fn (
(A + tH )) − fn (
(A))]P T
= lim 
t −→0 t
Hence,

P T (f˜ ) (A; H )P
f1 (
(A + tH )) − f1 (
(A)) fn (
(A + tH )) − fn (
(A))
= lim diag ,, 
t −→0 t t
(5.9)
710 H. Mohebi and A. Salemi

Now, let
(t ) =
(A + tH ) We conclude that

i (t ) = [P (t )T (A + tH )P (t )]ii , i = 1, 2,    , n

It is easy to show that

i (t ) =
i (A) + thi , i = 1, 2,    , n

It follows that

(t ) = (
1 (t ),    ,
n (t )) = (
1 (A) + th1 ,    ,
n (A) + thn )

= (
1 (A),    ,
n (A)) + t (h1 ,    , hn ) =
(A) + th

In view of (5.9), we obtain

P T (f˜ ) (A; H )P
f1 (
(A) + th) − f1 (
(A)) fn (
(A) + th) − fn (
(A))
= lim diag ,, 
t −→0 t t
This implies that for each i = 1, 2,    , n

fi (
(A) + th) − fi (
(A))

lim exists
t −→0 t
Then, fi is directionally differentiable at
(A) (i = 1, 2,    , n), and hence f
is directionally differentiable at
(A).
(3) This is an immediate consequence of part (1) because we have
if f is directionally differentiable and strictly continuous at
(A), then f is
semidifferentiable at
(A) (see [21]).

The following example shows that the directional differentiability of f

does not imply the directional differentiability of f˜ 

Example 5.3 (see [10]). Let S be the punctured hyperbola, that is

S = (x, y) ∈
2 : (x, y) = e1 , e2 , and x 2 + y 2 + 3xy − 2x − 2y + 1 = 0

Define a symmetric function f :
2 −→
such that f = 0 on S and f = 1

otherwise. Then clearly f has directional derivative zero at the point e1 =
(1, 0)T , but f˜ is not directionally differentiable at the matrix diag[e1 ]

Theorem 5.4. Let f :
n −→
n be a symmetric function with f (x) =

(f1 (x),    , fn (x))T (x ∈
n ) such that f satisfies the property () Let A ∈  with
eigenvalues
1 (A),    ,
n (A) and
(A) := (
1 (A),    ,
n (A))T ∈
n  Then the
 Analysis of Symmetric Matrix Valued Functions 711

following assertions are true:

(1) f is strictly continuous at
(A) if and only if f˜ is strictly continuous at A

(2) f is strictly continuous on
n if and only if f˜ is strictly continuous on  
(3) f is Lipschitz continuous if and only if f˜ is Lipschitz continuous.

Proof. (1) Suppose that f is strictly continuous at
(A) Then there

exists M > 0 and  > 0 such that

f (y) − f (z) ≤ M y − z ∀y, z ∈  (
(A), )

This implies that

fi (y) − fi (z) ≤ M y − z ∀y, z ∈  (
(A), ); i = 1, 2,    , n (5.10)

Now, for each i ∈ 1, 2,    , n, r ∈
and any unit vector u ∈
n define gi :

n −→
by:

fi (
(A) + ru), 0 ≤ r ≤ 
gi (
(A) + ru) = (5.11)
fi (
(A) + u), r > 

Then gi is Lipschitz continuous, and so there exists m > 0 such that

lip gi (t ) ≤ m for all t ∈
n , (i = 1, 2,    , n) Because  (
(A), ) is compact,
by letting C :=  (
(A), ) in Lemma 5.1, there exists a sequence of
continuously differentiable and symmetric functions gik :
n −→
, k =
1, 2,    , satisfying that gik converges uniformly to gi and
 k 
 g (t ) ≤ m ∀t ∈ C , ∀k ≥ 1; i = 1, 2,    , n (5.12)
i

Let g k = (g1k ,    , gnk ), k = 1, 2,     Then g k :
n −→
n is a continuously

differentiable and symmetric function on C (k = 1, 2,    ) Also, we have
gik k≥1 converges uniformly to fi on C (i = 1, 2,    , n)
Note that
(B) ∈  (
(A), ) for any B ∈ (A, ) Therefore, for any
B ∈ (A, ), any Q ∈ B , and any = ( 1 , 2 ,    , n ) ∈
n such that
B = Q diag[ 1 ,    , n ]Q , we have
T


(g k )(B) − f˜ (B)

= Q diag g1k ( (B)),    , gnk ( (B)) Q T

− Q diag[f1 ( (B)),    , fn ( (B))]Q T 
 
= diag g1k ( (B)) − f1 ( (B)) ,    , gnk ( (B)) − fn ( (B)) , (5.13)

where the second equality follows from the fact that Q T Q = I and
properties of the Frobenius norm  ·  Because gik k≥1 converges
712 H. Mohebi and A. Salemi


uniformly to fi on C (i = 1, 2,    , n), it follows from (5.13) that (g k )
k≥1
converges uniformly to f˜ on (A, ) Moreover, in view of (4.2), (5.12),
and continuous differentiability of g k on C (k ≥ 1) and that
(B) ∈ C for
all B ∈ (A, ), we obtain


 (g 
k )(B) = sup  (g k )(B)(H )
H =1

= sup Q (diag[g k (
(A))diag[Q T HQ ]]

H =1

+ (g k )[1] (
(A)) ◦ (Q T HQ ))Q T 

= sup diag[g k (
(A))diag[Q T HQ ]]
H =1

+ (g k )[1] (
(A)) ◦ (Q T HQ )
≤ sup diag[g k (
(A))diag[Q T HQ ]]
H =1

+ sup (g k )[1] (
(A)) ◦ (Q T HQ )

H =1

≤ m0 sup Q T HQ  = m0 , (5.14)
H =1

for some m0 > 0, which depends only on m

Now, fix D, F ∈ (A, ) with D = F . Because (g  k )
k≥1 converges
˜
uniformly to f on (A, ), it follows that for every  > 0, there exists N > 0
such that for all k ≥ N we have


(g k )(B) − f˜ (B) ≤ D − F  ∀B ∈ (A, ) (5.15)

But, we have g k (k ≥ 1) is continuously differentiable, then by

Proposition 4.3 we conclude that (g  k ) is also continuously differentiable

for all k ≥ 1 Therefore, by (5.14), (5.15), and mean-value theorem for

continuously differentiable functions, we deduce


f˜ (D) − f˜ (F ) = f˜ (D) − (g 
k )(D) + (g k )(D) + (g k )(F ) − (gk )(F ) − f˜ (F )


≤ f˜ (D) − (g k )(D) + (g 
k )(D) − (g k )(F )


+ (g k )(F ) − f˜ (F )
 1 
 
≤ 2D − F  +   
 (g )(F + t (D − F ))(D − F )dt 
k

0

≤ (2 + m0 )D − F  ∀k ≥ N 
 Analysis of Symmetric Matrix Valued Functions 713

This implies that

f˜ (D) − f˜ (F ) ≤ m0 D − F  ∀D, F ∈ (A, )

Hence, f˜ is strictly continuous at A.

Conversely, assume that f˜ is strictly continuous at A Then there exists
m > 0 and  > 0 such that

f˜ (D) − f˜ (F ) ≤ mD − F  ∀D, F ∈ (A, ) (5.16)

Choose any P ∈ A such that A = P diag[
1 (A),    ,
n (A)]P T  Let

x = (x1 ,    , xn ), y = (y1 ,    , yn ) ∈  (
(A), ) be arbitrary, and let

D := P diag[x1 ,    , xn ]P T ,

and

F := P diag[y1 ,    , yn ]P T 

Therefore, we have D − A = x −
(A) ≤  and F − A = y −

(A) ≤ , and hence in view of (2.3) and (5.16), we get

f (x) − f (y) = f˜ (D) − f˜ (F )

≤ mD − F  = mx − y

This implies that f is strictly continuous at
(A)

(2) This is an immediate consequence of part (1).
(3) Suppose that f is Lipschitz continuous with constant m
Then lip f (x) ≤ m for all x ∈
n  Now, fix any A ∈  with eigenvalues

1 (A),    ,
n (A) and
(A) = (
1 (A),    ,
n (A)) ∈
n  For any  > 0, let
C :=  (A, ) Then, with a similar proof as in part (1), we conclude that

f˜ (D) − f˜ (F ) ≤ m0 D − F  ∀D, F ∈ (A, ) and some m0 > 0

Because  > 0 was arbitrary and m0 > 0 is independent of the choice of ,

it follows that

f˜ (D) − f˜ (F ) ≤ m0 D − F  ∀D, F ∈  

Hence, f˜ is Lipschitz continuous.

Conversely, assume that f˜ is Lipschitz continuous with constant m0  Let
x = (x1 ,    , xn ), y = (y1 ,    , yn ) ∈
n be arbitrary. For any P ∈ , define

A := P diag[x1 ,    , xn ]P T ,
714 H. Mohebi and A. Salemi

and

B := P diag[y1 ,    , yn ]P T 

Then, we have

f (x) − f (y) = f˜ (A) − f˜ (B) ≤ m0 A − B = m0 x − y

This implies that f is Lipschitz continuous, and so the proof is complete.

ACKNOWLEDGMENT
This research was supported by Mahani Mathematical Research Center.

REFERENCES
1. R. Bhatia (1997). Matrix Analysis. Springer-Verlag, New York.
2. X. Chen and P. Tseng (2003). Non-interior continuation methods for solving semidefinite
complementarity problems. Math. Program. 95:431–474.
3. X. Chen, H.D. Qi, and P. Tseng (2003). Analysis of nonsmooth symmetric-matrix-valued
functions with applications to semidefinite complementarity problems. SIAM J. Optimiz.
13:960–985.
4. R.A. Horn and C.R. Johnson (1985). Matrix Analysis. 2nd ed. Cambridge University Press,
Cambridge.
5. R.A. Horn and C.R. Johnson (1991). Topics in Matrix Analysis. Cambridge University Press,
Cambridge.
6. H. Jiang and D. Ralph (1999). Global and local superlinear convergence analysis of Newton-
type methods for semismooth equations with smooth least squares. In Reformulation: Non-smooth,
Piecewise Smooth, Semismooth and Smoothing Methods, (M. Fukushima and L. Qi, eds.), Appl. Optim.
22. Kluwer Academic Publishers, Dordrecht, pp. 181–209.
7. C. Kanzow and C. Nagel (2002). Semidefinite programs: New search directions, smoothing-type
methods, and numerical results. SIAM J. Optimiz. 13:1–23.
8. T. Kato (1984). Perturbation Theory for Linear Operators. Springer-Verlag, Berlin.
9. E.C. Kemble (1958). The Fundamental Principles of Quantum Mechanics. Dover, New York.
10. A.S. Lewis (1996). Derivatives of spectral functions. Math. Oper. Res. 21:576–588.
11. A.S. Lewis (1999). Nonsmooth analysis of eigenvalues. Math. Program 84:1–24.
12. A.S. Lewis (1996). Convex analysis on the Hermitian matrices. SIAM J. Optimiz. 6:164–177.
13. A.S. Lewis and M.L. Overton (1996). Eigenvalue optimization. In Acta Numerica. Acta Numer. 5.
Cambridge University Press, Cambridge, pp. 149–190.
14. A.S. Lewis and H.S. Sendov (2001). Twice differentiable spectral functions. SIAM J. Matrix Anal.
Appl. 23:368–386.
15. R. Mifflin (1977). Semismooth and semiconvex functions in constrained optimization. SIAM J.
Control Optimiz. 15:959–972.
16. J. von Neumann (1937). Some matrix inequalities and metrization of metric space. Tomsk
University Review 1:286–300.
17. L. Qi (1993). Convergence analysis of some algorithm for solving nonsmooth equations. Math.
Oper. Res. 18:227–244.
18. L. Qi and J. Sun (1993). A nonsmooth version of Newton’s method. Math. Program. 58:353–357.
19. H.D. Qi and X. Yang (2004). Semismoothness of spectral functions. SIAM J. Matrix Anal. Appl.
25:766–783.
20. F. Rellich (1969). Perturbation Theory of Eigenvalue Problems. Gordon and Breach, New York.
 Analysis of Symmetric Matrix Valued Functions 715

21. R.T. Rockafellar and R.J.-B. Wets (1998). Variational Analysis. Springer-Verlag, Berlin.
22. L.I. Schiff (1955). Quantum Mechanics. McGraw-Hill, New York.
23. D. Sun and J. Sun (2002). Semismooth matrix valued functions. Math. Oper. Res. 27:150–169.
24. D. Sun and J. Sun (2003). Strong semismoothness of eigenvalues of symmetric matrices and its
application to inverse eigenvalue problem. SIAM J. Numer. Anal. 40:2352–2367.
25. M. Torki (1999). First- and second-order epi-differentiability in eigenvalue optimization. J. Math.
Anal. Appl. 234:391–416.
26. M. Torki (1999). Valeurs Propres de Matrices Symetriques: Analyses de la Sensibilitc d’Ordre
Superieur et Formulations Variationelles pour Leur Estimation. Ph.D. thesis. Paul Sabatier
University, Toulouse, France.
27. M. Torki (2001). Second-order directional derivatives of all eigenvalues of a symmetric matrix.
Non-linear Anal. 46:1133–1150.
28. N.K. Tsing, M.K.H. Fan, and E.I. Verriest (1994). On analyticity of functions involving
eigenvalues. Linear Algebra Appl. 207:159–180.
29. P. Tseng (1998). Merit functions for semidefinite complemetarity problems. Math. Program
83:159–185.