This document discusses the convexity of the set of k-admissible functions on a compact Kähler manifold. It begins by introducing k-admissible functions and some necessary convex analysis concepts. It then proves four main results: 1) the log of elementary symmetric functions of eigenvalues is a convex function, 2) the set of matrices with eigenvalues in a convex set is convex, 3) certain functions of eigenvalues are convex, and 4) the set of k-admissible functions is convex. It uses results on conjugation of spectral functions from convex analysis to prove these results. The proofs rely on properties of convex, lower semicontinuous functions and indicator functions.
2. JBILOU Asma
that LM = )KM, M is then called a K-eigenvector of L. Let K ∈ I6(ℂ) be a fixed positive definite matrix. Let us recall the
following basic result :
Lemma 1.2. Let L ∈ I6(ℂ), then:
1. The spectrum of K*L (i.e. the K-spectrum of L) is entirely real.
2. The greatest eigenvalue of K*L (i.e. the greatest K-eigenvalue of L) equals supRS#
TUR,RV
TWR,RV , where . , . denotes the
standard Hermitian product of ℂ6.
3. K*L is diagonalizable.
Since the spectrum of K*L is the spectrum of the Hermitian matrix K*Y
ZLK*Y
Z, the proof is an easy adaptation of the
standard one for symmetric matrices.
For a given Hermitian matrix L, we denote by )W(C) the eigenvalues of L with respect to K. In this article, we prove the
following four results using the Theorem 2.3 and the Corollary 2.4 of Lewis [2] (see Theorem 2.1 of this article):
Theorem 1.3. For each ; ∈ {1, . . . , m}, the function:
^8
W: I6(ℂ) → ℝ ∪ {+∞}, L → ^8
W(L) = b− ln 8D)W(L)G fg L ∈ )W
*(Γ8)
+∞ otherwise,
where )W
*(Γ8) = {L ∈ I6(ℂ), )W(C) ∈ Γ8}, is convex.
Theorem 1.4. If Γ is a (non empty) symmetric convex closed set of ℝ6, then )W
*(Γ): = {L ∈ I6(ℂ), )W(C) ∈ Γ},
is a convex closed set of I6(ℂ). In particular, )W
*(Γ8 ( ) is a convex closed set of I6(ℂ).
By the Theorem 1.3, and since )W
*(Γ8) is convex (Theorem 1.4), we deduce that:
Corollary 1.5. The function:
−^8
W: )W
*(Γ8) → ℝ, L → ^8
W(L) = ln 8D)W(L)G is concave.
The method used here to prove the Corollary 1.5, gives for K = m a different approach from the proof of [3] and the
elementary proof of [4, p. 51] and [5].
As an immediate consequence of Theorem 1.4, we get the following important result, that allows to notably simplify the
proof of ∂uniqueness of the solution of the equation (H8) in comparison with [5]:
Corollary 1.6. For a compact connected Kähler manifold (, , g, ω), the set of ;-admissible functions:
A8 = 9 ∈ Ln(, ℝ), )oDω + i ∂φG ∈ Γ8? is convex.
2 SOME CONVEX ANALYSIS
The space I6(ℂ) has a structure of Euclidean space thanks to the following scalar product ≪ +, K ≫= tr( A(
r B) =
tr(AB), called the Schur product. Let us denote by t#(ℝ6) the set of functions u: ℝ6 → ℝ ∪ {+∞} that are convex, lower
semicontinuous on ℝ6, and finite in at least one point. Given u ∈ t#(ℝ6) symmetric and K ∈ I6(ℂ) positive definite, we
define:
W: I6(ℂ) → ℝ ∪ {+∞}, vw L → %R
%R
W(L): = u()W,(L), … , )W,6(L))
W
where )W,(L) ≥ )W,n(L) ≥ … ≥ )W,6(L)denote the B-eigenvalues of C repeated with their multiplicity. Such functions %R
are called functions of B-eigenvalues or B-spectral functions. Our first aim is to determine the conjugation for such a
function %R
W using the conjugate function of u. Let us remind that the conjugation or the Legendre–Fenchel transform of u is
the function u∗: ℝ6 → ℝ ∪ {+∞} defined by:
∀s∈ ℝ6, u∗(z) = sup{∈ℝE{‹z, M› − u(M)}
where ‹. , . › denotes the standard scalar product on ℝ6.
Theorem 2.1 (A. S. Lewis [2], Conjugation of spectral functions). Let u ∈ t#(ℝ6) be symmetric, then:
1. The conjugate u∗ (∈ t#(ℝ6)) is also symmetric.
ISSN : 2351-8014 Vol. 4 No. 1, Jul. 2014 43
3. Convexity of the Set of k-Admissible Functions on a Compact Kähler Manifold
~ (defined as above) belong to t#DI6(ℂ)G with %R∗
~ and %R∗
2. The functions of eigenvalues %R
~ = (%R
~)∗, so that in particular the
~ is convex and lower semicontinuous.
function of eigenvalues %R
Proof. See the Theorem 2.3 and the Corollary 2.4 of Lewis [2]. ∎
A similar theorem is proved in the case of symmetric matrices in [6] and [7] (you can see also [8] and [9] for some details).
Corollary 2.2 (Conjugation of K-spectral functions). Let u ∈ t#(ℝ6) be symmetric, then:
1. The conjugate u∗ (∈ t#(ℝ6)) is also symmetric.
2. The functions of K-eigenvalues %R
W€Y = (%R
W (defined as above) belong to t#DI6(ℂ)G with %R∗
W and %R∗
W)∗, so that in
W is convex and lower semicontinuous.
particular the function of K-eigenvalues %R
3 PROOF OF THEOREMS 1.3 AND 1.4.
3.1 PROOF OF THEOREM 1.3.
The proof of Theorem 1.3 is a direct application of the Corollary 2.2 to the function:
u : ℝ6 → ℝ ∪ {+∞}, M = (M,… , M6) → u(x) = b− ln 8(M, … , M6) fg M ∈ t8
+∞ ‚ƒℎ„…†fz„
(3.1)
Our function u is symmetric and belongs to t#(ℝ6), indeed:
(i) It is clearly symmetric. It is finite in a least one point of ℝ6 because t8 is non empty. And it is convex, because the
function (8)
Y
F ∶ t8 → ℝ is concave [3, p.269].
(ii) It is lower semicontinuous. Indeed, let ‡ ∈ ℝ, and consider the set:
{M ∈ ℝ6/ +∞ ≥ u(M) ‡} = {M ∈ t8 / u(M) ‡} ∪ {M ∉ t8/u(M) ‡}
= {M ∈ t8 /− Š‹ 8(M) ‡} ∪ (ℝ6 ∖ t8) (3.2)
By continuity, {M ∈ t8 / − Š‹ 8(M) ‡} is an open set of t8, it is then an open of ℝ6 since t8 is an open of ℝ6.
Furthermore, the cone t8 is also a closed set of ℝ6 (as a connected component), consequently ℝ6 ∖ t8 is an open set of ℝ6.
Therefore, {M ∈ ℝ6/ +∞ ≥ u(M) ‡} is an open set of ℝ6 too. This is valid for all ‡ ∈ ℝ, so that u is lower
semicontinuous.
Therefore, we deduce by the Corollary 2.2 that the K-spectral function %R
W is convex, which proves the theorem.
W=^8
Let us remark that the same technique allows to prove for example that the functions
%(L) := the greatest K-eigenvalue of L and
%(L) := the sum of the z greatest K-eigenvalues of L
with z ∈ {1, . . . , }, (3.3)
are convex on I6(ℂ).
3.2 PROOF OF THEOREM 1.4.
The proof of Theorem 1.4 goes by considering the indicatrix function g# ∶= mŽ of the set t, namely:
g# ∶= mŽ ∶ ℝ6 → ℝ ∪ {+∞}, M = (M,… , M6) → mŽ (M) = 0 fg M ∈ t
+∞ ‚ƒℎ„…†fz„ (3.4)
From the assumptions made on t, g# lies in t#(ℝ6) and is symmetric, indeed:
(i) This function is clearly finite in at least one point since t is non empty.
(ii) The inequality mŽ(ƒM + (1 − ƒ)w) ≤ ƒ mŽ(M) + D1 – ƒG mŽ(w) is valid for all M, w ∈ ℝ6 and all ƒ ∈ [0, 1]. Indeed, if
M, w ∈ t then ƒM + (1 − ƒ)w ∈ t by convexity of tand the two sides of the convexity inequality equal 0 in this
case. Furthermore, if M or w does not belong to t then the right side of the inequality equals +∞ and the inequality
is then satisfied in this case too, which proves that mŽ is convex.
(iii) mŽ is lower semicontinuous. Indeed, let “ ∈ ℝ6 : If “ ≥ 0 then {M ∈ ℝ6 /+∞ ≥ mŽ(M) “} = ℝ6 ∖ t is an open
set since t is closed. Besides, if a 0, {M ∈ ℝ6 /+∞ ≥ mŽ(M) “} = ℝ6 is an open set too.
W lies in t#(ℝ6); in particular it is, convex lower semicontinuous.
So Corollary 2.2 implies that the function of K-eigenvalues %~”
But this function is given by:
%~”
W (L) = b0 fg L ∈ )W
W : I6(ℂ) → ℝ ∪ {+∞}, L → %~”
*(Γ)
+∞ ‚ƒℎ„…†fz„
(3.5)
ISSN : 2351-8014 Vol. 4 No. 1, Jul. 2014 44
5. Convexity of the Set of k-Admissible Functions on a Compact Kähler Manifold
[8] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
[9] J. M. Borwein and A. S. Lewis, Convex Analysis and Nonlinear Optimization: Theory and Examples, Springer, 2000.
[10] E. Hebey, Introduction à l’analyse non linéaire sur les variétés, Diderot, 1997.
ISSN : 2351-8014 Vol. 4 No. 1, Jul. 2014 46