Spectral sets and functions on Euclidean Jordan algebras

Spectral sets and functions on Euclidean Jordan algebras Juyoung Jeong Department of Mathematics and Statistics University of Maryland, Baltimore County Baltimore, Maryland 21250, USA juyoung1@umbc.edu and M. Seetharama Gowda Department of Mathematics and Statistics University of Maryland, Baltimore County Baltimore, Maryland 21250, USA gowda@umbc.edu February 4, 2016 Abstract Spectral sets (functions) in Euclidean Jordan algebras are generalizations of permutation invariant sets (respectively, functions) in R n. In this article, we study properties of such sets and functions and show how they are related to algebra automorphisms and majorization. We show that spectral sets/functions are indeed invariant under automorphisms, but the converse may not hold unless the algebra is R n or simple. We study Schur-convex spectral functions and provide some applications. We also discuss the transfer principle and a related metaformula. Key Words: Spectral sets, Spectral functions, Euclidean Jordan algebra, Algebra automorphisms, Majorization, Schur-convexity, Transfer principle, Metaformula. AMS Subject Classification: 17C20, 17C30, 52A41, 90C25. 1

1 Introduction This paper deals with some interconnections between spectral sets/functions, algebra automorphisms, and majorization in the setting of Euclidean Jordan algebras. Given a Euclidean Jordan algebra V of rank n, a set E in V is said to be a spectral set [1] if it is of the form E = λ 1 (Q), where Q is a permutation invariant set in R n and λ : V R n is the eigenvalue map (that takes x to λ(x), the vector of eigenvalues of x with entries written in the decreasing order). A function F : V R is said to be a spectral function [1] if it is of the form F = f λ, where f : R n R is a permutation invariant function. How these are related to algebra automorphisms of V (which are invertible linear transformations on V that preserve the Jordan product) and majorization is the main focus of the paper. The above concepts are generalizations of similar concepts that have been extensively studied in the setting of R n (where the concepts reduce to permutation invariant sets and functions) and in S n (H n ), the space of all n n real (respectively, complex) Hermitian matrices, see for example, [3], [4], [5], [10], [11], [12], [13], [14], [20], and the references therein. In the case of S n (H n ), spectral sets/functions are precisely those that are invariant under linear transformations of the form X UXU, where U is an orthogonal (respectively, unitary) matrix. There are a few works that deal with spectral sets and functions on general Euclidean Jordan algebras. Baes [1] discusses some properties of Q that get transferred to E (such as closedness, openness, boundedness/compactness, and convexity) and properties of f that get transferred to F (such as convexity and differentiability). Sun and Sun [22] deal with the transferability of the semismoothness properties of f to F. Ramirez, Seeger, and Sossa [18] and Sossa [21] deal with a commutation principle and a number of applications. In this paper, we present some new results on spectral sets and functions. In addition to giving some elementary characterization results, we show that spectral sets/functions are indeed invariant under automorphisms of V, but the converse may not hold unless the algebra is R n or simple (e.g., S n or H n ). We will also relate the concepts of spectral sets and functions to that of majorization. Given two elements x, y in V, we say that x is majorized by y and write x y if λ(x) is majorized by λ(y) in R n (see Section 2.1 for the definition); we say that x and y are spectrally equivalent and write x y if λ(x) = λ(y) (or equivalently, x y and y x). We show that spectral sets are characterized by the condition x y, y E x E 2

and spectral functions are characterized by x y F (x) = F (y). Under the assumption of convexity, by replacing x y by x y and F (x) = F (y) by F (x) F (y), we derive similar characterizations. In particular, we observe that convex spectral functions are Schur-convex, that is, x y F (x) F (y), and, as a consequence, show that F (Ψ(x)) F (x) (x V) whenever F is a convex spectral function and Ψ is a doubly stochastic transformation on V. We also prove a converse of this latter statement. We conclude the paper with a discussion on the Transfer Principle which asserts that (numerous) properties of Q (of f) get transferred to λ 1 (Q) (respectively, f λ). We also prove a metaformula # = #, where # is a set operation (such as the closure, interior, boundary, convex hull, etc.) and is the set operation defined by Q = λ 1 (Q) for Q R n and E = Σ n (λ(e)) for E V, with Σ n denoting the set of all n n permutation matrices. The organization of the paper is as follows. In Section 2, we cover basic notation, definitions, and preliminary results. Sections 3 and 4 deal with some characterization and automorphism invariance properties of (convex) spectral sets and functions. Section 5 deals with Schur-convex spectral functions. Finally, in Section 6 we deal with the transfer principle and a metaformula. 2 Preliminaries 2.1 Some notation An n n permutation matrix is a matrix obtained by permuting the rows of an n n identity matrix. Every row and column of such a matrix therefore contains a single 1 with 0s elsewhere. The set of all n n permutation matrices is denoted by Σ n. A set Q in R n is said to be permutation invariant if σ(x) Q whenever x Q and σ Σ n. (We assume that the empty set vacuously satisfies this condition, hence permutation invariant.) A function f : R n R is permutation invariant if f(σ(x)) = f(x) for all σ Σ n. The word symmetric is also used in the literature to describe permutation invariance of a set/function. 3

Recall that for two vectors u and v in R n with their decreasing rearrangements u and v, we say that u is majorized by v and write u v if k k u i v i for 1 k n 1 and u i = v i. We have, by setting k = 1 and k = n 1, u v max u i max v i and min u i min v i. (1) 2.2 Euclidean Jordan algebras A Euclidean Jordan algebra is a triple (V,,, ), where (V,, ) is a finite dimensional inner product space over R and the Jordan product (x, y) x y : V V V is a bilinear mapping satisfying the following conditions: (i) x y = y x for all x, y V, (ii) x 2 (x y) = x (x 2 y) for all x, y V where x 2 = x x, (iii) x y, z = x, y z for all x, y, z V. It is known that in such an algebra, there exists a (unique) unit element e such that x e = x for all x V. Henceforth, we let V denote a Euclidean Jordan algebra. In V, we define the (closed convex) cone of squares by V + := {x 2 : x V}. Some examples of Euclidean Jordan algebras are: The Euclidean Jordan algebra of n-dimensional real vectors: V = R n, x, y := x i y i, x y := x y, where x y denotes the component-wise product of vectors x and y. Here, the unit element is e = (1,..., 1) R n. The Euclidean Jordan algebra of n n real symmetric matrices: V = S n, X, Y := tr(xy ), X Y := 1 (XY + Y X). 2 Here, tr denotes the trace of a matrix (which is the sum of its eigenvalues). The identity matrix I S n is the unit element of this algebra. The Jordan spin algebra L n : V = R n (n 2), x, y := x i y i, x y := ( x, y, x 1 y + y 1 x), where x = (x 1, x), x R n 1, etc. In this algebra, the unit element is e = (1, 0,..., 0) R n. 4

A Euclidean Jordan algebra is said to be simple if it is not a direct sum of two nonzero Euclidean Jordan algebras. It is known, see [7], that any nonzero Euclidean Jordan algebra is, in a unique way, a direct sum/product of simple Euclidean Jordan algebras. Moreover, every simple algebra is isomorphic to one of the following five algebras: (i) the algebra S n of n n real symmetric matrices, (ii) the algebra H n of n n complex Hermitian matrices, (iii) the algebra Q n of n n quaternion Hermitian matrices, (iv) the algebra O 3 of 3 3 octonian Hermitian matrices, (v) the Jordan spin algebra L n for n 3. We say that V is essentially simple if it is either R n or simple. An element c V is an idempotent if c 2 = c; it is a primitive idempotent if it is nonzero and cannot be written as a sum of two nonzero idempotents. We say a finite set {e 1, e 2,..., e n } of primitive idempotents in V is a Jordan frame if e i e j = 0 when i j and e i = e. Note that e i, e j = e i e j, e = 0 whenever i j. It turns out that all Jordan frames in V will have the same number of elements. This common number is the rank of V. Proposition 1 (Spectral decomposition theorem [7]). Suppose V is a Euclidean Jordan algebra of rank n. Then, for every x V, there exist uniquely determined real numbers λ 1 (x),..., λ n (x) (called the eigenvalues of x) and a Jordan frame {e 1,..., e n } such that x = λ 1 (x)e 1 + + λ n (x)e n. Conversely, given any Jordan frame {e 1,..., e n } and real numbers λ 1, λ 2,..., λ n, the sum λ 1 e 1 + + λ n e n defines an element of V whose eigenvalues are λ 1, λ 2,..., λ n. We note that V + = {x V : λ i (x) 0 i = 1, 2,..., n}. We define the trace of an element x V by: tr(x) := λ i (x). 5

From now on, λ(x) denotes the vector of eigenvalues of x with components written in the decreasing order. Note that the mapping x λ(x) is well-defined and continuous, see [1]. Recall by definition, x y in V λ(x) = λ(y) in R n. As this is an equivalence relation on V, we define the equivalence class of an x V by: [x] := {y V : y x}. Proposition 2 (Peirce decomposition theorem [7]). Let V be a Euclidean Jordan algebra of rank n and {e 1,..., e n } be a Jordan frame of V. Then, V is the orthogonal direct sum of subspaces V ij (i j), where for i, j {1,..., n}, V ii := {x V : x e i = x} = Re i, V ij := {x V : x e i = 1 2 x = x e j} when i j. Thus, any element x V will have its Peirce decomposition with respect to {e 1,..., e n }: x = x ij = x i e i + x ij, where x i R and x ij V ij. 1 i j n 1 i<j n Let φ and Ψ be linear transformations on V. We way that φ is an algebra automorphism (or automorphism for short) if it is invertible and φ(x y) = φ(x) φ(y) x, y V. The set of all automorphisms of V is denoted by Aut(V). The automorphisms of R n are precisely permutation matrices. Also, automorphisms of S n (H n ) are of the form X UXU, where U is an orthogonal (respectively, unitary) matrix. See [8] for an explicit description of automorphisms of L n. We note that automorphisms map Jordan frames to Jordan frames. Thus, eigenvalues of an element remain the same under the action of an automorphism. In particular, φ(x) x x V, φ Aut(V). We say that a set Ω in V is invariant under automorphisms (or automorphism invariant) if φ(ω) Ω φ Aut(V); similarly, a function F : V R is invariant under automorphisms (or automorphism invari- 6

ant) if F φ = F φ Aut(V). If φ is an automorphism of V, then φ(v + ) = V +. Linear transformations satisfying the latter property are called cone automorphisms; the set of all such transformations will be denoted by Aut(V + ). Ψ is doubly stochastic [9] if it is (i) positive, i.e., Ψ(V + ) V +, (ii) unital, i.e., Ψ(e) = e, and (iii) trace preserving, i.e., tr(ψ(x)) = tr(x) for all x V. When V = R n, such a Ψ is a doubly stochastic matrix (which is an n n nonnegative matrix with each row/column sum one). We write DS(V) for the set of all doubly stochastic transformations on V. The following result will be used in many of our theorems, particularly, in the converse statements. Here and elsewhere, we implicitly assume that a Jordan frame, in addition to being a set, is also an ordered listing of its objects. Recall that V is essentially simple if it is either R n or simple. Proposition 3. Let V be essentially simple. If {e 1,..., e n } and {e 1,..., e n} are any two Jordan frames in V, then there exists φ Aut(V) such that φ(e i ) = e i for all i = 1,..., n. In particular, if x y in V, then there exists φ Aut(V) such that x = φ(y). Proof. We observe that in R n, up to permutations, there is only one Jordan frame, namely, the standard coordinate system. In this case, we can take φ to be a permutation matrix. If V is simple, the existence of an automorphism mapping one Jordan frame to another comes from Theorem IV.2.5, [7]. The last statement comes from an application of the spectral decomposition theorem (Proposition 1). It is interesting to note that the converse in the above result holds: If every Jordan frame can be mapped onto another by an automorphism, then the algebra is essentially simple. While we do not require this result in the main body of the paper, we provide a proof in the Appendix. We also recall the following result from [9]. 7

Proposition 4. Let V 1 and V 2 be two non-isomorphic simple Euclidean Jordan algebras. If φ Aut(V 1 V 2 ), then φ is of the form (φ 1, φ 2 ) for some φ i Aut(V i ), i = 1, 2, that is, φ(x) = (φ 1 (x 1 ), φ 2 (x 2 )), x = (x 1, x 2 ) V 1 V 2. 2.3 Majorization in R n and Euclidean Jordan algebras We start by recalling two classical results in matrix theory. The first one is due to Hardy, Littlewood, and Pólya and the second one is due to Birkhoff. Proposition 5 ([2], Theorem II.1.10). Let x, y R n. A necessary and sufficient condition for x y is that there exists a doubly stochastic matrix A such that x = A(y). Proposition 6 ([2], Theorem II.2.3). The set of all n n doubly stochastic matrices is a compact convex set whose extreme points are permutation matrices. In particular, every doubly stochastic matrix is a convex combination of permutation matrices. The next proposition, which is essential for Section 4, is somewhat classical and well known. It easily follows from the above two propositions. Proposition 7. (i) If Q is convex and permutation invariant in R n, then u v, v Q u Q. (ii) If f : R n R is convex and permutation invariant, then f is Schur-convex, that is, u v f(u) f(v). The result below describes a connection between majorization, automorphisms, and doubly stochastic transformations in the setting of Euclidean Jordan algebras. Recall, by definition, x y in V λ(x) λ(y) in R n. Proposition 8 (Gowda [9]). For x, y V, consider the following statements: (a) x = Φ(y), where Φ is a convex combination of automorphisms of V. (b) x = Ψ(y), where Ψ is doubly stochastic on V. (c) x y. Then, (a) (b) (c). Furthermore, when V is essentially simple, reverse implications hold. 8

When V is a simple Euclidean Jordan algebra, it is known [17] that λ(x + y) λ(x) + λ(y) for all x, y V. Here is a generalization of this result. Proposition 9. Let V be a Euclidean Jordan algebra of rank n. Then for x, y V, there exist doubly stochastic matrices A and B such that λ(x + y) = Aλ(x) + Bλ(y). (2) When V is simple, we can take A = B. Proof. As the result for a simple algebra is known, we assume that V is non-simple, that is, V is a product of simple algebras. For simplicity, we assume V = V 1 V 2, where V 1, V 2 are simple algebras of rank n 1, n 2, respectively. Now, let x = (x 1, x 2 ), y = (y 1, y 2 ), where x i, y i V i for i = 1, 2. Define u i = λ(x i ) R n i for i = 1, 2 and u = (u 1, u 2 ) R n 1+n 2 = R n. As eigenvalues of x come from the eigenvalues of x 1 and x 2, we have λ(x) = σ 1 (u) for some σ 1 Σ n. Similarly, there exist σ 2, σ 3 Σ n such that λ(y) = σ 2 (v) λ(x + y) = σ 3 (w) where v = (v 1, v 2 ), v i = λ(y i ) w = (w 1, w 2 ), w i = λ(x i + y i ) for i = 1, 2. Since V 1, V 2 are simple, w i = λ(x i + y i ) λ(x i ) + λ(y i ) = u i + v i for i = 1, 2. Thus, there are doubly stochastic matrices C i on R n i such that w i = C i (u i + v i ) for i = 1, 2. So, ( ) ( ) ( ) w 1 C 1 0 u 1 + v 1 =. w 2 0 C 2 u 2 + v 2 Letting C denote the the block diagonal matrix that appears above, one can easily verify that C is doubly stochastic on R n. Moreover, we have w = C(u + v) = C(u) + C(v). Now, define matrices A and B by A = σ 3 Cσ 1 1 and B = σ 3 Cσ 1 2. As products of doubly stochastic matrices are doubly stochastic, A and B are doubly stochastic. Finally, we have Aλ(x) + Bλ(y) = λ(x + y). 3 Spectral sets Throughout this paper, V is assumed to be of rank n. Definition 1. A set E in V is a spectral set if there exists a permutation invariant set Q in R n such that E = λ 1 (Q). 9

It is easy to see that arbitrary union and/or intersection of spectral sets is spectral. So is the complement of a spectral set. Clearly, the symmetric cone V + is a spectral set in V as it corresponds to R n +. In this section, we describe some properties of spectral sets. We characterize spectral sets via spectral equivalence. First, we introduce a diamond operation on subsets of R n and V and state a simple proposition that will facilitate our study. This operation is motivated by the question of recovering Q from E. We note that in the case of V = S n, if E is a spectral set, then the corresponding Q is given by Q = {u R n : Diag(u) E}, where Diag(u) is a diagonal matrix with u as the diagonal [11]. Definition 2. For a set Q in R n, we define the set Q in V by For a set E in V, we let E in R n be E := Σ n (λ(e)) = For simplicity, we write Q in place of (Q ), etc. Proposition 10. The following statements hold: Q := λ 1 (Q) = {x V : λ(x) Q}. (3) { } u R n : u = λ(x) for some x E. (4) (i) For any Q that is permutation invariant in R n, Q is a spectral set in V and Q = Q. (ii) For any set E in V, E is permutation invariant in R n. (iii) E = {x V : x y for some y E} = y E [y], where [x] is the equivalence class of x induced by. Proof. For any set Q in R n, we define the core of Q by Q := {u : u Q}. We immediately note the following when Q is permutation invariant: Q Q, Q = Σ n (Q ), and λ(λ 1 (Q)) = Q. (i): Let Q be permutation invariant. By definition, E := Q is a spectral set. We have (Q ) = E = Σ n (λ(e)) = Σ n (λ(λ 1 (Q)) = Σ n (Q ) = Q. (ii): Since Σ n is a group, Σ n (E ) = Σ n (Σ n (λ(e)) = Σ n (λ(e)) = E. Thus, E is permutation invariant. (iii): This follows from (E ) = {x V : λ(x) E } = {x V : λ(x) = λ(y) for some y E}. 10

Example 1. This example shows that in Item (i) above, one needs the permutation invariance property of Q to get the equality Q = Q. Let V = S 2 and Q 1 = {u R 2 + : u 1 u 2 }. Then, it is easy to see that Q 1 = S+ 2 and so Q 1 = R 2 +. Thus, we have Q 1 Q 1. On the other hand, consider Q 2 = {u R 2 + : u 1 u 2 }. Then, Q 2 = {ai : a 0}; therefore Q 2 = {u R 2 + : u 1 = u 2 }. This means Q 2 Q 2. We now characterize spectral sets. Theorem 1. The following are equivalent for any set E in V. (a) E is spectral. (b) x y, y E x E. (c) E = E. Proof. (a) (b): Let E = λ 1 (Q), where Q is permutation invariant. If x y with y E, then λ(x) = λ(y) Q. Hence, x λ 1 (Q) = E. (b) (c): This comes from Item (iii) in Proposition 10. (c) (a): When E = E, we let Q := E. By Item (ii) in Proposition 10, Q is permutation invariant; we see that E = Q = λ 1 (Q) is spectral. Remarks. The above two results show that Q Q sets up is a one-to-one correspondence between permutation invariant sets in R n and spectral sets in V. We now describe spectral sets via automorphisms. This result explains why in S n or H n, spectral sets are completely characterized by automorphism invariance. Theorem 2. Every spectral set in V is invariant under automorphisms. Converse holds when V is essentially simple. Proof. Suppose E is a spectral set, x E, and φ Aut(V). As eigenvalues remain the same under the action of automorphisms, we see that φ(x) x. By Item (b) in Theorem 1, φ(x) E. This proves that E is invariant under automorphisms. To see the converse, assume that E is invariant under automorphisms and V is essentially simple. We verify Item (b) in Theorem 1 to show that E is spectral. To this end, let x y, y E. Then, λ(x) = λ(y). By Proposition 3, there exists φ Aut(V) such that x = φ(y). As E is invariant under automorphisms, we must have x E. This concludes the proof. Example 2. In the theorem above, the converse may not hold for general algebras. To see this, let V = R S 2 and E = R S+. 2 From Proposition 4, every automorphism φ on V is of the form 11

φ = (φ 1, φ 2 ), where φ 1 Aut(R), φ 2 Aut(S 2 ). It is easy to check that E is invariant under automorphisms. Now consider two elements, ( [ ]) ( [ ]) 1 0 1 0 x = 0, / E, and y = 1, E. 0 1 0 0 As λ(x) = λ(y) = (1, 0, 1) T, we have x y. Thus, E violates condition (b) in Theorem 1. Hence, E is not a spectral set. It is easy to verify that x E. Thus, E E even though E is invariant under automorphisms. We now characterize spectral sets that are convex. Theorem 3. For a set E in V, consider the following: (a) E = λ 1 (Q), where Q is permutation invariant and convex in R n. (b) E is convex, and x y, y E x E. (c) E is convex, and x y, y E x E. (d) E is convex and invariant under automorphisms. Then (a) (b) (c) and (c) (d). essentially simple. Furthermore, all statements are equivalent when V is Proof. (a) (b): Let E = λ 1 (Q), where Q is permutation invariant and convex in R n. The convexity of E has already been proved in Theorem 27, [1]. For completeness, we provide a (slightly different) proof. Let x, y E and t [0, 1]. By Proposition 9, λ(tx + (1 t)y)) = tu + (1 t)v, where u = A(λ(x)), v = B(λ(y)) for some doubly stochastic matrices A and B. By Proposition 7, we see that u, v Q. As Q is convex, λ(tx + (1 t)y) Q. Thus, tx + (1 t)y E. This proves the convexity of E. Now, suppose x y and y E. Then, λ(x) λ(y) and λ(y) Q. By Proposition 7, λ(x) Q and hence x E. (b) (c): This is obvious as x y implies x y. (c) (a): When (c) holds, by Theorem 1, E is spectral; let E = λ 1 (Q), where Q is permutation invariant. To show that Q is convex, let u, v Q and t [0, 1]. Then there exist x, y E such that x := n 1 u ie i and y := n 1 v ie i for some Jordan frames {e 1, e 2,..., e n } and {e 1, e 2,..., e n}. Now, define x = n 1 u ie i. Then λ(x) = λ(x), hence x E by (c). As x, y E and E is convex, we get (tu i + (1 t)v i )e i = t u i e i + (1 t) v i e i = tx + (1 t)y E. This proves that tu + (1 t)v Q. Thus, Q is convex. 12

(c) (d): Assume (c), let φ Aut(V) and x E. Then λ(x) = λ(φ(x)) and hence φ(x) x. From (c), we must have φ(x) E, proving (d). Assuming that V is essentially simple, we prove (d) (c). Suppose x, y V such that x y and y E. As V is essentially simple, by Proposition 3, there exists φ Aut(V) such that x = φ(y). By (d), x E. Example 3. We can use Example 2 to show that in a general V, (d) may not imply (a), (b), or (c) in the above theorem. 4 Spectral functions Definition 3. A function F : V R is a spectral function if there exists a permutation invariant function f : R n R such that F = f λ. By way of an example, we observe that F (x) = λ max (x) (the maximum of eigenvalues of x) is a spectral function on V as it corresponds to f(u) = max u i on R n. Proposition 11. A set E in V is a spectral set if and only if its characteristic function χ E (which takes the value one on E and zero outside of E) is a spectral function. Proof. Suppose E is a spectral set, say E = λ 1 (Q), where Q is a permutation invariant set in R n. Then χ E = χ Q λ. As χ Q is a permutation invariant, χ E is a spectral function. Conversely, if χ E = f λ is a spectral function, then let Q be the set where f takes the value one. It is obvious that Q is a permutation invariant set. We now show that E = λ 1 (Q). Suppose x E. Then 1 = χ E (x) = f(λ(x)), which implies λ(x) Q by our construction. Thus, x λ 1 (Q). For the reverse implication, suppose x λ 1 (Q). As λ(x) Q, we get χ E (x) = f(λ(x)) = 1, and so x E. In this section, we describe some properties of spectral functions. We characterize spectral sets via spectral equivalence. As with sets, motivated by the question of recovering f from F, we introduce a diamond operation on functions on R n and V. We note that in the case of V = S n, if F is a spectral function on S n, then the corresponding f is given by f(u) = F (Diag(u)). Definition 4. Given a function f : R n R, we define f : V R by f (x) := f(λ(x)) for x V. (5) Fix a Jordan frame {e 1,..., e n } in V. Then, given a function F : V R, we define F : R n R by ( ) F (u) := F u i e i where u = (u 1,..., u n ) T R n. (6) 13

For simplicity, we write f in place of (f ), etc. In the result below, F and (f ) are defined as in (6) with respect to a (fixed) Jordan frame {e 1,..., e n }. Proposition 12. The following statements hold: (i) For any permutation invariant function f : R n R, f is a spectral function and f = f. (ii) For any function F : V R, F is permutation invariant. Proof. (i): When f is permutation invariant, by definition, f is a spectral function. Now, define (f ) as in (6). Then, for any u R n, we have ( ) ( ( )) f (u) = f u i e i = f λ u i e i = f(u ) = f(u). Since u is arbitrary, f = f. (ii): Take u R n and σ Σ n. As u = σ(u), we get ( ) ( ) F (σ(u)) = F σ(u) i e i = F u i e i = F (u). Hence, F is permutation invariant. Example 4. For the equality f = f in Item (i) above, it is essential to have f invariant under permutations. To see this, take V = S 2 and define f : R 2 R by f(u 1, u 2 ) = u 1. Then, for any x S 2, f (x) = f(λ(x)) = λ 1 (x). Hence, for any u R 2, ( 2 ) f (u) = f u i e i = f(u ) = max{u 1, u 2 }. This proves that f f. We now characterize spectral functions. Theorem 4. The following are equivalent for a function F : V R: (a) F is a spectral function. (b) For each α R, the level set F 1 ({α}) is a spectral set in V. (c) F is constant on each equivalence class of. (d) F = F. Proof. (a) (b): Suppose F is a spectral function, say F = f λ, where f is a permutation invariant function. For any α R n, F 1 ({α}) = λ 1 (f 1 ({α}). As f is a permutation invariant 14

function, Q := f 1 ({α}) is a permutation invariant set in R n. Thus, F 1 ({α}) = λ 1 (Q) is a spectral set. (b) (c): Suppose F satisfies condition (b) and x y. Let α := F (x). Then, F 1 ({α}) is a spectral set and x F 1 ({α}); hence, by Theorem 1, y F 1 ({α}). This implies that F (y) = α; hence, F (x) = F (y) whenever x y. (c) (d): Let x V with its spectral decomposition x = n 1 λ i(x)e i. Define y = n 1 λ i(x)ē i. Then, x y and by (c), F (x) = F (y). This implies, ( ) F (x) = F (λ(x)) = F λ i (x)e i = F (y) = F (x). Hence, F = F. (d) (a): Let F = F. Then, f := F is permutation invariant, by Item (ii) in Proposition 12. Hence, F = f is spectral. Theorem 5. Every spectral function is invariant under automorphisms. Converse holds when V is essentially simple. Proof. Suppose F = f λ for some permutation invariant function f. Note that λ(x) = λ(φ(x)) for every x V and φ Aut(V). Thus, F (φ(x)) = f(λ(φ(x))) = f(λ(x)) = F (x). For the converse, let F be invariant under automorphism and V be essentially simple. If x y, then by Proposition 3, there exists φ Aut(V) such that x = φ(y). This implies that F (x) = F (y). We now use Item (c) in Theorem 4 to see that F is spectral. Example 5. The converse in Theorem 5 may not hold in general. To see this, take V = R S 2 and let F : V R be defined by F ((a, A)) = tr(a), for all a R, A S 2. Since we know the (explicit) description of automorphisms of V (via Proposition 4), we easily see that F is automorphism invariant. However, for x and y of Example 2, λ(x) = λ(y) and so x y. One can easily check that F (x) = 0 1 = F (y). As F violates condition (c) in Theorem 4, F is not a spectral function. A celebrated result of Davis [6] says that a unitarily invariant function on H n is convex if and only if its restriction to diagonal matrices is convex. This result has numerous applications in various fields. A generalization of this result for Euclidean Jordan algebras has already been observed by Baes [1]. In what follows, we consider equivalent formulations of this generalization and give some applications. 15

Theorem 6. Let E be a convex spectral set in V so that E = λ 1 (Q) with Q convex and permutation invariant in R n. Let F : E R. Consider the following statements: (a) F = f λ, where f : Q R is convex and permutation invariant. (b) F is convex, and x y implies F (x) F (y). (c) F is convex, and x y implies F (x) = F (y). (d) F is convex and F φ = F for all φ Aut(V). Then (a) (b) (c) and (c) (d). Moreover, all these statements are equivalent when V is essentially simple. Proof. (a) (b): The convexity of F has already been proved in Theorem 41, [1]. Capturing its essence, we present our proof based on Proposition 9. For t [0, 1] and x, y E, we write λ(tx + (1 t)y) = t Aλ(x) + (1 t) Bλ(y), where A and B are doubly stochastic matrices. As these matrices are convex combinations of permutation matrices, by permutation invariance and convexity of f on Q, we see that f(aλ(x)) f(λ(x)) and f(bλ(y)) f(λ(y)). Thus, F (tx + (1 t)y) = f(taλ(x) + (1 t)bλ(y)) tf(aλ(x)) + (1 t)f(bλ(y)) tf(λ(x)) + (1 t)f(λ(y)) = tf (x) + (1 t)f (y). Now for the second part of (b), suppose x, y E with x y. Writing u = λ(x) and v = λ(y), we get u v in Q and hence (from Propositions 5, 6) u = j α jσ j (v) for some α j > 0 with j α j = 1 and permutation matrices σ j Σ n. As f is convex and permutation invariant, ( ) f(u) = f αj σ j (v) α j f(σ j (v)) = α j f(v) = f(v). Since F = f λ, we get F (x) = f(λ(x)) = f(u) f(v) = f(λ(y)) = F (y). (b) (c): This is obvious as x y implies x y. (c) (a): Given F satisfying (c), we define f : Q R as follows. exists x E such that λ(x) = u. We let f(u) := F (x). For any u Q, there This is well defined: if there is a y E with λ(y) = u, then x y and so by (c), F (x) = F (y). Now, for any permutation σ, u = σ(u) ; hence, f(σ(u)) = F (x) = f(u), proving the permutation invariance of f. We also see F (x) = f(u) = f(u ) = f(λ(x)). Note that these (relations) hold if we start with an x E and let u = λ(x). Thus we have F = f λ. We now show that f is convex. Let u, v Q and t [0, 1]. As w := tu + (1 t)v Q, there exists z E such that λ(z) = w. Let n 1 w i e i be the spectral decomposition of z. Define x = n 1 u i e i and y = n 1 v i e i. Note that λ(x) = u Q and λ(y) = v Q; hence, x, y E with 16

z = tx + (1 t)y. Then, as f is permutation invariant and F is convex, we get Hence, f is convex. f(tu + (1 t)v) = f(w) = f(w ) = F (z) = F (tx + (1 t)y) tf (x)) + (1 t)f (y) = tf(u) + (1 t)f(v). (c) (d): Let φ be an automorphism. For x V, λ(φ(x)) = λ(x) and so φ(x) x. Then, by (c), we must have F (φ(x)) = F (x). Now, we prove (d) (c) when V is essentially simple. Suppose x, y E such that x y. By Proposition 3, there exists an automorphism φ such that x = φ(y). Then, from (d), F (x) = F (φ(y)) = F (y). Thus, condition (c) holds. This completes the proof. Remarks. In the case of a general V, (d) may not imply other statements. This can be seen by taking V and F as in Example 5. Clearly, this F is automorphism invariant and convex (indeed, linear). However, as seen in Example 5, F is not a spectral function. We now provide two applications of the above theorem. Example 6. For p [1, ], let F (x) = x sp, p := λ(x) p, where u p denotes the pth-norm of a vector u in R n. Then, F is a convex spectral function which is also positive homogeneous. Since F (x) = 0 = x = 0, we see that sp, p is a norm on V. This fact has already been observed in [23], where it is proved via Thompson s triangle inequality, majorization techniques, and case-by-case analysis. Example 7. The Golden-Thompson inequality. This inequality says that for A, B H n, tr(exp(a + B)) tr(exp(a)) tr(exp(b)). There is a simple proof of this by Rivin [19] based on the result of Davis [6]. A modification of this proof, given below, leads to the following generalization: For any two elements x, y in a Euclidean Jordan algebra V, tr(exp(x + y)) tr(exp(x)) tr(exp(y)), (7) where exp(x) is defined by exp(x) := n 1 exp(λ i(x)) e i when x has the spectral decomposition x = n 1 λ i(x)e i. We prove (7) as follows. For u R n with components u 1,..., u n, let f(u) = ln( n 1 exp(u i)), 17

which is known to be convex [19]. Since f is also permutation invariant, F := f λ is convex by the above theorem. It follows that ( ) x + y F 2 F (x) + F (y), 2 for all x, y V. As F (x) = ln(tr(exp(x))), this leads to [ ( ( ))] x + y 2 tr exp tr(exp(x)) tr(exp(y)). 2 This, with the observation tr(exp(2z)) [tr(exp(z))] 2 for any z V, yields (7). 5 Schur-convex functions Definition 5. A function F : V R is said to be Schur-convex if x y F (x) F (y). Theorem 6 (together with Proposition 8) immediately yields the following. Theorem 7. Every convex spectral function on V is Schur-convex. In particular, for any doubly stochastic transformation Ψ on V, and for any convex spectral function F on V, we have F (Ψ(x)) F (x) for all x V. We illustrate the above result with a number of examples. Example 8. For p [1, ], as in Example 6, consider F (x) = x sp, p := λ(x) p. Then, F is a convex spectral function and hence, for any doubly stochastic transformation Ψ on V, Ψ(x) sp, p x sp, p. This extends Proposition 2 in [9], where it is shown that Ψ(x) sp, 2 x sp, 2 for all x in a simple Euclidean Jordan algebra. Example 9. Consider an idempotent c ( 0, e) in V and the corresponding orthogonal decomposition [7] V = V(c, 1) + V(c, 1 2 ) + V(c, 0), where V(c, γ) = {x V : x c = γ x}, γ {0, 1 2, 1}. For each x V, we write x = u + v + w where u V(c, 1), v V(c, 1 2 ), w V(c, 0). For any a V, let P a denote the corresponding quadratic representation defined by P a (x) = 2a (a x) a 2 x. Then for ε = 2c e, one verifies that P ε (x) = u v + w and ( Pε+I 2 )(x) = u + w. 18

As ε 2 = e, P ε is an automorphism of V ([7], Prop. II.4.4). Thus, P ε and Pε+I 2 are doubly stochastic on V. Then, for any convex spectral function F on V, F (u v + w) F (x) and F (u + w) F (x). We note that the process of going from x = w + v + w to u + w is a pinching process; in the context of block matrices, this pinching is obtained by setting the off-diagonal blocks to zero. Example 10. Let A be an n n real symmetric positive semidefinite matrix with each diagonal entry one. In V, we fix a Jordan frame {e 1, e 2,..., e n } and consider the corresponding Peirce decomposition of any x V (as in Proposition 2): x = i j x ij = n 1 x ie i + i<j x ij. Then the transformation Ψ(x) := A x = i j a ij x ij is doubly stochastic, see [9]. It follows that when F is a convex spectral function on V, F (A x) F (x). By taking A to be the identity matrix, this inequality reduces to ( ) F x i e i F (x). Example 11. Let g : R R be any function. Then the corresponding Löewner mapping L g : V V [22] is defined as follows: For any x V with spectral decomposition x = λ i (x)e i, L g (x) := g(λ i (x)) e i. Clearly, F (x) := tr(l g (x)) = g(λ i (x)) = f λ, where f(u 1,..., u n ) = g(u i ), is a spectral function. Thus, when g is convex, F is a convex spectral function. Hence, the function x tr(l g (x)) is Schur-convex for any convex function g : R R. The second part of Theorem 7 says that a doubly stochastic transformation decreases the value of a convex spectral function. Below, we present a converse to this statement: Theorem 8. The following statements hold: (i) If x, y V with F (x) F (y) for all convex spectral functions F on V, then x y. (ii) If x, y V with F (x) = F (y) for all convex spectral functions F on V, then x y. (iii) If Ψ is a linear transformation on V such that F (Ψ(x)) F (x) for all x V and for all convex spectral functions F, then Ψ is a doubly stochastic transformation on V. 19

This result is based on two lemmas. The first lemma (noted on page 159 in [16]) is a consequence of a result of Hardy, Littlewood, and Pólya. Lemma 1. Suppose u, v R n with the property that f(u) f(v) for all convex permutation invariant functions f : R n R. Then, u v. Our second lemma is a generalization of a a well-known result: An n n matrix A is doubly stochastic if and only if Ax x for all x R n, see Theorem A.4 in [16]. Lemma 2. A linear transformation Ψ : V V is doubly stochastic if and only if Ψ(x) x for all x V. Proof. First, suppose Ψ is doubly stochastic. Then by Proposition 8, we have Ψ(x) x. To prove the converse, suppose Ψ(x) x for all x V. First, let x V +. Then all the eigenvalues of x are nonnegative. Now, Ψ(x) x implies, from (1) in Section 2.1, λ n (Ψ(x)) λ n (x) 0. As λ n (Ψ(x)) is the smallest of the eigenvalues of Ψ(x), we see that all eigenvalues of Ψ(x) are nonnegative; hence Ψ(x) V +. Thus, Ψ(V + ) V +. Next, we show that Ψ(e) = e. As Ψ(e) e and every eigenvalue of e is one, from (1), the smallest and largest eigenvalues of Ψ(e) coincide with 1. Thus, all eigenvalues of Ψ(e) are equal to one. By the spectral decomposition of e, we see that Ψ(e) = e. We now show Ψ is trace-preserving: For any Jordan frame {e 1,..., e n }, we have Ψ(e i ) e i, so tr(ψ(e i )) = tr(e i ) = 1. Thus for x V with (spectral decomposition) x = i x ie i, ( ( )) tr(ψ(x)) = tr Ψ x i e i = x i tr(ψ(e i )) = x i = tr(x). Thus, Ψ is doubly stochastic on V. We now come to the proof the theorem. Proof. (i) Suppose x, y V with F (x) F (y) for all convex spectral functions F on V. Then, for any f that is convex and permutation invariant on R n, we let F = f λ and get f(λ(x)) f(λ(y)). Now from Lemma 1, λ(x) λ(y). This means that x y. (ii) From the proof of Item (i), λ(x) λ(y) and λ(y) λ(x). From the defining (majorization) inequalities of Section 2.1, λ(x) = λ(y). Thus, x y. (iii) Now suppose that Ψ is linear and F (Ψ(x)) F (x) for all x V and all convex spectral functions F on V. By (i), Ψ(x) x for all x. By Lemma 2, Ψ is doubly stochastic. It has been observed before that automorphisms preserve eigenvalues. The following result shows that in the setting of essentially simple algebras, automorphisms are the only linear transformations that preserve eigenvalues. 20

Corollary 1. Suppose V is essentially simple and φ : V V is linear. If φ(x) x for all x V, then φ Aut(V). Proof. Suppose that φ(x) x for all x V. Then, φ is invertible: If φ(x) = 0 for some x, then 0 x and so x = 0. Now, take any convex spectral function F on V. As φ(x) x, we have F (x) = F (φ(x)). From Theorem 8, we see that φ and φ 1 are doubly stochastic on V. Recall, by definition, these are positive transformations, that is, φ(v + ) V + and φ 1 (V + ) V +. Hence, φ(v + ) = V +, that is, φ belongs to Aut(V + ). Since V is essentially simple, from Theorem 8 in [9] we have Aut(V + ) DS(V) = Aut(V), where DS(V) denotes the set of all doubly stochastic transformations on V. φ Aut(V). This proves that 6 The transfer principle and a metaformula In the context of spectral sets E = λ 1 (Q) and spectral functions F = f λ, the Transfer Principle asserts that (many) properties of Q (of f) get transferred to E (respectively, to F ). Theorems 3 and 6, where convexity gets transferred, illustrate this principle. In addition to this, Baes ([1], Theorem 27) shows that closedness, openness, boundedness and compactness properties of Q are carried over to E. Sun and Sun [22] show that semismoothness property of f gets transferred to F in the setting of a Euclidean Jordan algebras. Numerous specialized results exist in the setting of S n and H n ; see the recent article [5] and the references therein for a discussion on the transferability of C -manifold property of Q and various types differentibility properties of f (e.g., prox-regularity, Clarke-regularity, and smoothness). Related to this, in the context of normal decomposition systems, Lewis [12] has observed that the metaformula #(λ 1 (Q)) = λ 1 (#(Q)) holds for sets Q that are invariant under a group of orthogonal transformations and certain set operations #. In particular, it was shown in Theorem 5.4, [12], that the formula holds when # is the closure/interior/boundary operation. Consequently, because of a result in [15], it is also valid in any (essentially) simple Euclidean Jordan algebra; see [10], [11] for results of this type for spectral cones in S n. Motivated by these, we present the following metaformula in the setting of general Euclidean Jordan algebras. In what follows, for a set S (either in R n or in V), we consider closure, interior, boundary, and convex hull operations, which are respectively denoted by S, S, S, and conv(s). Recall that for a set Q in R n and a set E in V, Q := λ 1 (Q) and E := Σ n (λ(e)). 21

(We remark that when V = R n and E = Q, these two definitions coincide.) Theorem 9. Let # denote one of the operations of closure, interior, boundary, or convex hull. Then, over permutation invariant sets in R n and spectral sets in V, the operations # and commute; symbolically, # = #. In preparation for the proof, we state (and prove, for the record) the following folkloric result: Proposition 13. Let Q be a permutation invariant set in R n and E = λ 1 (Q) in V. Then, (i) E = λ 1 (Q). (ii) E = λ 1 (Q ). (iii) E = λ 1 ( Q). (iv) conv(λ 1 (Q)) = λ 1 (conv(q)). Consequently, the closure/interior/boundary/convex hull of a spectral set is spectral. Proof. (i) By continuity of λ, E = λ 1 (Q) λ 1 (Q). To see the reverse inclusion, let x λ 1 (Q) so that λ(x) Q. Let λ(x) = lim q k, where q k Q. We consider the spectral decomposition x = n 1 λ i(x)e i and define x k := n 1 (q k) i e i, for k = 1, 2,.... Then, λ(x k ) = q k Q (recall that Q is permutation invariant). Thus, x k λ 1 (Q) for all k. As x k x, we see that x E. Hence, E = λ 1 (Q). (ii) As Q c is permutation invariant and λ 1 preserves complements, we see, by (i), that E c = λ 1 (Q c ). Taking complements and using the identity E = (E c ) c, we get E = λ 1 (Q ). (iii) This comes from the previous items using the definition E = E\E and the fact that λ 1 preserves set differences. (iv) We first prove that conv(λ 1 (Q)) λ 1 (conv(q)) in V. It is clear that conv(q) is convex and permutation invariant. Hence, by Theorem 3, λ 1 (conv(q)) is convex. As λ 1 (Q) λ 1 (conv(q)), we see that conv(λ 1 (Q)) λ 1 (conv(q)). To prove the reverse implication, let x λ 1 (conv(q)). Then, λ(x) = N k=1 α k q k, where α k s are positive with sum one and q k Q for all k. Writing the spectral decomposition of x = n λ i(x)e i, we see that ( N ) ) N x = α k (q k ) i e i = (q k ) i e i. k=1 k=1 α k ( 1 Letting x k := n 1 (q k) i e i, we get λ(x k ) = q k Q. Hence, x k λ 1 (Q) and x is now a convex combination of elements of λ 1 (Q). Thus, x conv(λ 1 (Q)) proving the required reverse inclusion. Finally, the last statement follows from Items (i) (iv) together with the observation that the closure/interior/boundary/convex hull of a permutation invariant set in R n is permutation invari- 22

ant. Proof of the Theorem: Suppose Q is any permutation invariant set in R n. Then from Proposition 13 we have λ 1 (#(Q)) = #(λ 1 (Q)), that is, [#(Q)] = #(Q ). This means that on permutation invariant sets in R n, the operations # and commute. Now suppose that E is any spectral set in V. Then, Q := E is permutation invariant and Q = E = E (by Theorem 1). Hence, #(E) = #(Q ) = [#(Q)]. As #(Q) is permutation invariant, from Proposition 10, [#(Q)] = #(Q). Thus, [#(E)] = [#(Q)] = #(Q) = #(E ). This proves that the operations # and commute on spectral sets in V. It will be shown below that the equality E = E holds for any set E V. We now provide examples to show that in the above theorem, permutation invariance of Q and/or spectrality of E is needed to get the remaining results. Example 12. Let V = S n and Q = {u R n + : u 1 < u 2 < < u n }. Clearly, Q is not permutation invariant. As Q = λ 1 (Q) =, we have Q =. On the other hand, we have Q = {u R n + : u 1 u 2 u n } and Q = λ 1 (Q) is the (nonempty) set of all nonnegative multiples of the identity matrix. Thus, Q Q. One consequence of Theorem 3 is that when E is spectral and convex, E is convex. This may fail if E is not spectral: In V = R 2, let e 1 and e 2 denote the standard coordinate vectors. Then, E = {e 1 } is convex, while E = {e 1, e 2 } is not. In particular, [conv(e)] conv(e ). Let V = S 2 and let E be the set of all nonnegative diagonal matrices. Then, one can easily verify that E = R 2 + and so (E ) = R 2 ++, where R 2 ++ denotes the set of all positive vectors in R 2. However, we have E = and (E ) =. Hence, (E ) (E ). The preservation of certain properties is an important feature of the diamond operation. Theorem 10. If a set E is closed/open/bounded/compact in V, then so is E in R n. Proof. We assume, without loss of generality, that E is nonempty. Suppose that E is closed. Let {u k } be a sequence in E such that u k u for some u V. For each u k E, we can find x k E with λ(x k ) = u k. Let x k = i λ i(x k )e (k) i be the spectral decomposition of x k for each k. As the set of all primitive idempotents in V forms a compact set, 23

see [7], page 78, there exists a sequence k m such that e (km) i e i for all i = 1, 2,..., n. Thus, x km = λ i (x km )e (km) i u i e i. Since E is closed, we have x := lim x km E, which implies that λ(x) = u. Hence, u E proving the closedness of E. For any (primitive) idempotent c in V we observe that c 2 = c, c = c c, e = c, e c e and hence c e. Then, for any Jordan frame {e 1, e 2,..., e n } in V and u R n, we have ( ) u i e i u i e n u e (8) where u denotes the 2-norm of u. 1 1 Now assume that E is open in V. To show that E open, let u E. Then there exists x E such that λ(x) = u. As E is open, there exists ɛ > 0 such that B(x, ɛ) := {y V : y x < ɛ} E. Putting δ := ɛ n e, we show that the ball B(u, δ) := {v R n : u v < δ} is contained in E. To this end, let x = i u i e i be the spectral decomposition of x. Then, there exists a permutation matrix σ Σ n such that x = i u ie σ(i). Now, for any v B(u, δ), define y := i v ie σ(i) V. Then, from (8), x y n e u v < ɛ. This proves that y E. As λ(y) = v, we see that v E. This shows that B(u, δ) E ; hence E is open. Now let E be compact. Then λ(e) is compact, by the continuity of λ. As Σ n is compact, we see that E = Σ n (λ(e)) is also compact. Finally, if E is bounded, then E is compact. Hence, (E) is compact. Clearly, E is bounded as it is a subset of (E). Corollary 2. For any set E in V, E = E. Proof. Without loss of generality, we assume that E is nonempty. Since E E and E is closed by the above Theorem, we have E E. To see the reverse inclusion, let u E. Then there exists x E such that λ(x) = u. As u is some permutation of u, there exists a permutation σ Σ n such that σ 1 (u) = u = λ(x). As x E, there exists a sequence {x k } of E converging to x. Then, by the continuity of λ, we get λ(x) = lim k λ(x k ). This implies ( ) u = σ(λ(x)) = σ lim λ(x k) k = lim k σ(λ(x k)). Letting u k := σ(λ(x k )), we get u k = λ(x k); thus u k E for all k. As u k u, we have u E proving E E. This completes the proof. 24

Our final result deals with the double diamond operation. For a set E in V, we call E, the spectral hull of E in V. Here are some properties of the spectral hull. Proposition 14. For any set E V, we have (i) E E. (ii) If E 1 E 2, then E 1 E 2. (iii) E is the smallest spectral set containing E. (iv) E is spectral if and only if E = E. (v) (E 1 E 2 ) = E 1 E 2, (E 1 E 2 ) = E 1 E 2. (vi) E = (E). In particular, if E is closed, then so is E. (vii) If E is open, then E is open. (viii) If E is compact, then E is compact. Proof. Items (i) (v) follow from Item (iii), Proposition 10. (vi): As E (= Q) is permutation invariant, this follows from Proposition 13 and Corollary 2. (vii): As E is permutation invariant, it is clear from Theorem 10 and Proposition 13 that E is open whenever E is open. (viii): Suppose E is compact. invariant. By Theorem 27 in [1], E = λ 1 (Q) is also compact. Then, by Theorem 10, Q := E is compact and permutation 7 Appendix Theorem 11. V is essentially simple (that is, V is either R n or simple) if and only if any Jordan frame of V can be mapped onto another by an automorphism of V. Proof. The only if part has been observed in Proposition 3. We prove the if part. Suppose, if possible, V is a product of simple algebras with at least one of the factors, say, V 1, has rank more than one. Let V 2 denote the product of the remaining factors. Then, V = V 1 V 2. Let e 1, e 2,..., e m (m 2) and f 1, f 2,,..., f k be Jordan frames in V 1 and V 2 respectively. Let c i = (e i, 0) T and d j = (0, f j ) T for i = 1, 2,..., m and j = 1, 2,..., k. Then these m + k objects, in any specified order, form a Jordan frame. We show that V does not have an automorphism taking c 1 to c 1 and c 2 to d 1. Assuming the contrary, let such an automorphism be given in the block form by [ ] A B L =, C D where A : V 1 V 1 is linear, etc. It is easy to verify that A is Jordan homomorphism, that is, it preserves the Jordan product on V 1. Since V 1 is simple, the kernel of A (which is an ideal of 25

V 1 ) must be either {0} or V 1. However, this cannot happen as L(c 1 ) = c 1 implies A(e 1 ) = e 1 and L(c 2 ) = d 1 implies A(e 2 ) = 0. Thus we have the if part. References [1] M. Baes, Convexity and differentiability of spectral functions on Euclidean Jordan algebra, Linear Alg. Appl., 422 (2007) 664-700. [2] R. Bhatia, Matrix Analysis, Springer Graduate Texts in Mathematics, Springer-Verlag, New York, 1997. [3] V. Chandrasekaran, P.A. Parrilo, A.S. Willsky, Convex graph invariants, SIAM Review, 54 (2012) 513-541. [4] A. Daniilidis, A. Lewis, J. Malick, and H. Sendov, Prox-regularity of spectral functions and spectral sets, J. Convex Anal., 15 (2008) 547-560. [5] A. Daniilidis, D. Drusvyatskiy, and A.S. Lewis, Orthogonal invariance and identifiability, SIAM Matrix. Anal., 35 (2014) 580-598. [6] C. Davis, All convex invariant functions of hermitian matrices, Archiv der Mathematik, 8 (1957) 276-278. [7] J. Faraut and A. Korányi, Analysis on Symmetric Cones, Clarendon Press, Oxford, 1994. [8] M.S. Gowda, R. Sznajder, and J. Tao, Some P-properties of linear transformations on Euclidean Jordan algebras, Linear Alg. Appl., 393 (2004) 203-232. [9] M. S. Gowda, Positive and Doubly Stochastic maps, and Majorization in Euclidean Jordan Algebras, Research Report, Department of Mathematics and Statistics, University of Maryland Baltimore County, Baltimore, Maryland 21250, USA, August 2015. [10] R. Henrion and A. Seeger, Inradius and circumradius of various convex cones arising in applications, Set Valued Analysis, 18 (2010) 483-511. [11] A. Iusem and A. Seeger, Angular analysis of two classes of non-polyhedral convex cones: the point of view of optimization theory, Comp. Applied Math., 26 (2007) 191-214. [12] A. S. Lewis, Group invariance and convex matrix analysis, SIAM J. Matrix Anal., 17 (1996) 927-949. [13] A. S. Lewis, Convex analysis on the Hermitian matrices, SIAM J. Optim., 6 (1996) 164-177. 26