Professional Documents
Culture Documents
Article
Generalization of the Lieb–Thirring–Araki Inequality and
Its Applications
Yonggang Li 1 , Jing Wang 2, * and Huafei Sun 3
Abstract: The matrix eigenvalue is very important in matrix analysis, and it has been applied to
matrix trace inequalities, such as the Lieb–Thirring–Araki theorem and Thompson–Golden theorem.
In this manuscript, we obtain a matrix eigenvalue inequality by using the Stein–Hirschman operator
interpolation inequality; then, according to the properties of exterior algebra and the Schur-convex
function, we provide a new proof for the generalization of the Lieb–Thirring–Araki theorem and
Furuta theorem.
On the basis of this definition, we have a formula relating the trace of matrix A and
the eigenvalue of A:
Copyright: © 2021 by the authors. n
Licensee MDPI, Basel, Switzerland. Tr[ A] = ∑ λ i ( A ),
i =1
This article is an open access article
distributed under the terms and where λi ( A) is the eigenvalue of A. It is well known that Tr[ AB] = Tr[ BA] for any
conditions of the Creative Commons A, B ∈ Hn+ . However, Tr[e A+ B ] = Tr[e A e B ] only if AB = BA. Fortunately, in 1965,
Attribution (CC BY) license (https:// Thompson and Golden independently discovered an inequality called the Thompson–
creativecommons.org/licenses/by/ Golden theorem (refer to [5–7]):
4.0/).
Tr[e A+ B ] ≤ Tr[e A e B ].
In general, the following limit holds (called the Lie–Trotter formula [8]):
pA 1
pA p
lim Tr[ e 2 e pB e 2 ] = Tr[e A+ B ].
p →0
h i h α α 1
i
Tr ∧k e A+ B ≤ Tr ∧k ( A 2 Bα A 2 ) α .
2. Preliminary
In this section, we recall some notions and definitions from matrix analysis, and
introduce some important results of the matrix-monotone function, which are used through
the article (refer to [14–17]).
· · · a1n B
a11 B
.. .. ..
A⊗B = . ,
. .
am1 B · · · amn B
where A = aij 1≤i≤m,1≤ j≤n .
The tensor product is different from matrix multiplication, and one of the differences
is commutativity:
( I ⊗ B)( A ⊗ I ) = ( A ⊗ I )( I ⊗ B) = A ⊗ B.
Mathematics 2021, 9, 723 3 of 11
AC ⊗ BD = ( AC ⊗ I )( I ⊗ BD )
= ( A ⊗ I )(C ⊗ I )( I ⊗ B)( I ⊗ D )
= ( A ⊗ I )( I ⊗ B)(C ⊗ I )( I ⊗ D )
= ( A ⊗ B)(C ⊗ D ).
⊗k A = |A ⊗ A ⊗
{z
· · · ⊗ A} .
k
In addition to the tensor product, there is another common product named exterior
algebra ([18]). Exterior algebra, denoted by ∧, is a binary operation for any An×n that is
1
ξ i1 ∧ ξ i2 · · · ∧ ξ i k = √
n!
∑ (−1)π ξ π(i1 ) ⊗ ξ π(i2 ) · · · ⊗ ξ π(ik ) ,
π ∈σn
shows that
k
λ1 (∧k A) = ∏ λ i ( A ). (1)
i =1
If ~x and ~y satisfy
k k
∑ x [i ] ≤ ∑ y [i ] , k ≤ n − 1;
i =1 i =1
n n
∑ x [i ] = ∑ y [i ] , k = n.
i =1 i =1
∂f ∂f
( xi − x j )( − ) ≥ 0 f or all ~x ∈ I n .
∂xi ∂x j
Mathematics 2021, 9, 723 4 of 11
A = P∗ Λ A P,
where P is the unitary matrix and Λ A := diag{λ1 , ..., λn } is a diagonal matrix with eigen-
values as elements. When ~x ∗ ( A − B)~x > 0 for any ~x ∈ Cn , we denote A ≥ B.
Associated with a function f ( x ) on (0, +∞), the matrix function f ( A) is defined as
f ( A) = P∗ f (Λ A ) P,
where f (Λ A ) = diag{ f (λ1 ), ..., f (λn )}. Then, the function f is said to be matrix-monotonic
if it satisfies
f ( A) ≥ f ( B) for all A ≥ B > 0.
Since the matrix-monotone function is a special type of operator monotone function,
we present the following general conclusion about the operator-monotone function, which
can be found in [20,21].
Lemma 2. The following statements for a real-valued continuous function f on (0, +∞) are equiv-
alent:
( i) f is operator-monotone;
( ii) f admits an integral representation
Z0
f (λ) = α + βλ + (1 + λt)(t − λ)−1 d µ(t), f or any λ > 0, (2)
−∞
where α is a real number, β is non-negative and µ is a finite positive measure on (−∞, 0).
Theorem 1. For any 0 < α ≤ β and A, B ∈ Hn+ , the following inequality holds
α α 1 β β 1
λ1 ( A 2 B α A 2 ) α ≤ λ1 ( A 2 B β A 2 ) β . (3)
If we denote
max {< ~x, A~x >}
x ∈Cn
λ1 ( A ) = = max {< ~x, A~x >}
< ~x, ~x > ~x ∈Cn ,<~x,~x >=1
Mathematics 2021, 9, 723 5 of 11
Furthermore,
h k k
i1 h k +1 k +1
i 1
k k +1
λ1 ( A 2 B k A 2 ) ≤ λ 1 ( A 2 B k +1 A 2 ) .
This implies
h m1 n2 m1 n2 i 1 h m2 n1 m2 n1 i m 1n
m1 n2
λ1 ( A 2 B m1 n2 A 2 ) ≤ λ 1 ( A 2 B m2 n1 A 2 ) 2 1 ,
1 1
for any m1 , m2 , n1 , n2 > 0 and m1 n2 ≤ m2 n1 . Let A = A n1 n2 , B = B n1 n2 ; we obtain
m1 m1 m1
n1 m2 m2 m2
n2
m1 m2
λ1 ( A 2n1
B n1
A 2n1
) ≤ λ1 ( A 2n2
B n2
A 2n2
) .
Although Theorem 1 has been obtained from the Cauchy inequality, the frequency
of retractions improves the inequality. In the following, we obtain Theorem 1 by using
operator interpolation. First, let us introduce the Stein–Hirschman operator interpolation
inequality ([12]).
Lemma 3. Supposing Gz tobe an analytic family of linear operators of admissible growth defined
in the strip 0 ≤ R(z) ≤ 1 and 1 ≤ p1 , p2 , q1 , q2 ≤ ∞, when 1p = 1p−t + pt2 , 1q = 1q−t + qt2 (0 ≤
1 1
t ≤ 1) , k Giy ( f ) kq1 ≤ A0 (y) k f k p1 and k G1+iy ( f ) kq2 ≤ A1 (y) k f k p2 , we can obtain
Z
ln k Gt ( f ) kq ≤ (ω (1 − t, y) ln A0 (y) + ω (t, y) ln A1 (y)) d y + ln k f k p , (4)
R
tan( tπ
2 )
where ln | A0 (y) |, ln | A1 (y) |≤ Aeα|y| (α < π ) and ω (t, y) = tπ yπ yπ .
2[tan ( 2 )+tanh2 ( 2 )]cosh2 ( 2 )
2
Mathematics 2021, 9, 723 6 of 11
From Lemma 3, we can improve the result in Theorem 1 and obtain the following the-
orem.
ω (t, y)
Z
t t 1 1 1 1 1 1
λ1 ( A 2 B t A 2 ) t ≤ ( λ1 (( A 2 B1−iy AB1+iy A 2 )) 2 ) d y ≤ λ1 ( A 2 BA 2 ). (5)
R t
t t
ln k Gt ( f ) k 1 = ln | λ1 ( A 2 Bt A 2 ) | + ln k f k 1
t t
Z
≤ (ω (1 − t, y) ln A0 (y) + ω (t, y) ln A1 (y)) d y + ln k f k 1
t
ZR
1 1 1
≤ (ω (1 − t, y) ln A0 (y) + ω (t, y) ln λ1 (( A B1−iy AB1+iy A 2 )) 2 ) d y + ln k f k 1 2
R t
Z
1 1
≤ (ω (1 − t, y) ln 1 + ω (t, y) ln | λ1 ( A BA ) | d y + ln k f k 1 2 2
R t
1 1
= t ln | λ1 ( A BA ) | + ln k f k 1 .
2 2
t
This implies
ω (t, y)
Z
t t 1 1 1 1 1 1
λ1 ( A 2 B t A 2 ) t ≤ ( λ1 (( A 2 B1−iy AB1+iy A 2 )) 2 ) d y ≤ λ1 ( A 2 BA 2 )
R t
or
ω (t, y)
Z
t t 1 1 t 1 1
λ1 ( A 2 B t A 2 ) ≤ ( λ1 (( A 2 B1−iy AB1+iy A 2 )) 2 ) d y ≤ λ1 ( A 2 BA 2 )t ,
R t
for any 0 < t < 1, and the first ”≤“ is obtained by the Jensen inequality ([22]). This
completes the proof of Theorem 2.
Mathematics 2021, 9, 723 7 of 11
Theorem 2 is very useful. On one hand, when α < β, letting t = αβ , we can obtain
Theorem 1. On the other hand, using the matrix exterior algebra, we obtain
k 1
k 1
∏ λi ( A 2 B α A 2 ) α ≤ ∏ λi ( A 2 B β A 2 ) β .
α α β β
i =1 i =1
whether it is true or not for any k ≤ n, and this inequality can be regarded as a generaliza-
tion of the Lieb–Thirring–Araki theorem.
Furthermore, we have
β 1 β 1
α α 1 α α 1
β β
ln λ1 ( A 2 Bα A 2 ) α , · · · , ln λn ( A 2 Bα A 2 ) α ≺ ln λ1 ( A 2 B β A 2 ) β , · · · , ln λn ( A 2 B β A 2 ) β .
n
Let f ( x1 , x2 · · · , xn ) = ∑ e xi ; direct calculations show that
i =1
∂f ∂f
( xi − x j )( − ) = ( xi − x j )(e xi − e x j ) ≥ 0,
∂xi ∂x j
n
which implies f ( x1 , x2 · · · , xn ) = ∑ e xi is a Schur-convex function. Hence, we have the
i =1
Thompson–Golden theorem
1
n 1
n 1 1
∑ λi ( A 2 B α A 2 ) α ∑ λi ( A 2 B β A 2 ) β
α α α α β β β β
Tr[( A 2 Bα A 2 ) α ] = ≤ = Tr[( A 2 B β A 2 ) β ], (6)
i =1 i =1
1 1
when α ≤ β. Specially, when β = 1 and A = A α , B = B α , we have the Lieb–Thirring–Araki
theorem
1 1 1 1 1 1
Tr[( A 2 BA 2 ) α ] ≤ Tr[( A 2α B α A 2α )]. (7)
We know that the Lieb–Thirring–Araki theorem can be obtained from the Schur-convex
n
function f ( x1 , x2 · · · , xn ) = ∑ e xi . Generally, we can prove the following conclusion.
i =1
∑
γxi
g ( x1 , x2 · · · , x n ) = eγxi1 eγxi2 · · · e k
Proof. Since
∑
γxi
g ( x1 , x2 · · · , x n ) = eγxi1 eγxi2 · · · e k ,
1≤i1 <i2 <···<ik ≤n
Mathematics 2021, 9, 723 8 of 11
we have !
∂g
∑ γxi1 γxi2 γxi
= γeγxi e e ···e k ,
∂xi 1≤i1 <···<<ik ≤n,x̂i
or
h i α α γ
h i
1 1 1 α−iαy α α+iαy 2
R α α
Tr ∧k ( A 2 BA 2 )αγ ≤ Tr ∧k eα R ω ( α ,y) ln( A 2 B A B A )) 2 dy
, ≤ Tr ∧k ( A 2 Bα A 2 )γ (9)
p+2r
where p, r ≥ 0, q ≥ 1 and q ≥ 1+2r .
p+2r
Hence, we suppose p ≥ 1. Meanwhile, we know (1+2r )q
≤ 1, so we only consider
p+2r 1
1+2r = q. Firstly, when 0 ≤ r ≤ 2, we have
1
− 12 −r r p r −r − 12
λ1 B A ( A B A ) A B q
p p p 1 p
− 12 2 − 2 −2r − 2 1− q 2 − 12
= λ1 B B ( B A B ) B B
p+2r p−1 p p p −1 p+2r p−1
− 2 −2r − 2 p+2r
= λ1 ( B 2 ) p + 2r (B A B ) (B 2 ) p + 2r
h p+2r p −1
p p p+2r i p+2r
≤ λ1 B 2 ( B− 2 A−2r B− 2 ) B 2
h i p −1
p+2r
= λ1 ( A−r B2r A−r )
≤ 1.
This implies
1
A−r ( Ar B p Ar ) q A−r ≤ B ≤ A.
That is,
1
( Ar B p Ar ) q ≤ A1+2r .
1
Secondly, let A1 = A1+2r , B1 = ( Ar B p Ar ) q ; then
1 1 1
q
( A12 B1 A12 ) q1 ≤ A21 .
1 1 1
p+4r +1
That is, ( A2r+ 2 B p A2r+ 2 ) q ≤ A2(1+2r) , where q1 = 1+4r +1 .
This implies
1
( As B p As ) q ≤ A1+2s ,
where s = 2r + 12 ∈ [ 12 , 32 ].
Repeating this process, we have finished the proof.
A A
where 0 < α ≤ β and ln λi (e 2 e B e 2 ) ≥ 0 (i = 1, 2 · · · , n). Let
n
d ( x1 , · · · , x n ) = ∏ xi , ( x i ≥ 0).
i =1
∂d( x1 , · · · , xn ) ∂d( x1 , · · · , xn )
( xi − x j )( − ) ≤ 0.
∂xi ∂x j
A A
Corollary 5. Supposing A, B ∈ Hn and λn (e 2 e B e 2 ) ≥ 1,
αA αA 1 βA βA 1
det( A + B) ≥ det(ln(e 2 eαB e 2 ) α ) ≥ det(ln(e 2 e βB e 2 ) β ), (14)
the first ”≥“ obtained by using the Lie–Trotter formula, and 0 < α ≤ β.
A A
Corollary 6. Supposing A, B ∈ Hn and λn (e 2 e B e 2 ) ≥ 1,
αA αA 1 βA βA 1
Tr[∧k ln(e 2 eαB e 2 ) α ] ≥ Tr[∧k ln(e 2 e βB e 2 ) β ], (15)
Proof. Since
β 1 β 1
α α 1 α α 1
β β
ln λ1 ( A 2 Bα A 2 ) α , · · · , ln λn ( A 2 Bα A 2 ) α ≺ ln λ1 ( A 2 B β A 2 ) β , · · · , ln λn ( A 2 B β A 2 ) β ,
a ( x1 , x2 · · · , x n ) = ∑ x i1 x i2 · · · x i k
1≤i1 <i2 <···<ik ≤n
∂a( x1 , · · · , xn ) ∂a( x1 , · · · , xn )
( xii − xi j )( − )
∂xii ∂xi j
= ( xik − xi j )( ∑ x i1 · · · c
x ii · · · x i k − ∑ xi1 · · · xci j · · · xik )
1≤i1 <···ibi <···<ik ≤n 1≤i1 <···ibj <···<ik ≤n
= −( xik − xi j )2 ( ∑ x i1 · · · c
xii xci j · · · xik )
1≤i1 <···ibi ,ibj <···<ik ≤n
≤ 0.
Hence, we know
αA αA 1 βA βA 1
Tr[∧k ln(e 2 eαB e 2 ) α ] ≥ Tr[∧k ln(e 2 e βB e 2 ) β ].
4. Conclusions
1 1 α α
In the paper, we discuss the relationship between λ1 ( A 2 BA 2 )α and λ1 ( A 2 Bα A 2 ) by
using the Stein–Hirschman operator interpolation inequality. Through in-depth study, we
obtain some eigenvalue inequalities such as the generalization Golden–Thompson theorem
and Lieb–Thirring–Araki theorem. Moreover, the Furuta theorem is also shown by using
the eigenvalue inequality. At last, we generalize an important determinant inequality by
using the matrix exterior algebra.
References
1. Treanta, S.; Arana-Jimenez, M.; Antczak, T. A necessary and sufficient condition on the equivalence between local and global
optimal solutions in variational control problems. Nonlinear Anal. 2020, 191, 111640. [CrossRef]
2. Bellman, R. Introduction to Matrix Analysis; McGraw-Hill: New York, NY, USA, 1960.
3. Anderson, W.N., Jr. Shorted operators. SIAM J. Appl. Math. 2006, 20, 520–525. [CrossRef]
4. Carlen, E. Trace inequalities and quantum entropy: An introductory course. Contemp. Math. 2010, 529, 73–140.
5. Golden, S. Lower bounds for Helmholtz function. Phys. Rev. 1965, 137, 1127–1128. [CrossRef]
6. Thompson, C.J. Inequality with applications in statistical mechanics. J. Math. Phys. 1965, 6, 1812–1813. [CrossRef]
7. Lenard, A. Generalization of the Golden-Thompson inequality Tr (e A e B ) ≥ Tre A+ B . Zndiana Univ. Math. 1971, 21, 457–467.
[CrossRef]
8. Ando, T.; Hiai, F. Log majorization and complementary Golden-Thompson type inequalities. Linear Algebra Appl. 1994, 197–198,
113–131. [CrossRef]
9. Araki, H. On an inequality of Lieb and Thirring. Lett. Math. Phys. 1990, 19, 167–170. [CrossRef]
10. Lieb, E.H.; Thirring, W. Inequalities for the moments of the eigenvalues of the Schrodinger Hamiltonian and their relation to
Sobolev inequalities. In Studies in Mathematical Physics; Lieb, E.H., Simon, B., Wightman, A., Eds.; Princeton University Press:
Princeton, NJ, USA, 1976; pp. 269–303.
11. Marshall, A.W.; Olkin, I.; Arnold, B.C. Inequalities: Theory of Majorization and Its Applications; Springer: New York, NY, USA, 2011.
12. Huang, D. Generalizing Lieb’s concavity theorem via operator interpolation. Adv. Math. 2020, 369, 107208. [CrossRef]
1 p+2r
13. Furuta, T. A ≥ B ≥ 0 assures ( Br A p Br ) q ≥ B q for r ≥ 0, p > 0, q ≥ 1 with 1+q2r ≥ p + 2r. Proc. Am. Math. Sot. 1987, 101,
85–88.
14. Bhatia, R. Matrix Analysis; Springer: New York, NY, USA, 1997.
15. Bhatia, R. Positive Definite Matrices; Princeton University Press: Princeton, NJ, USA; Oxford, UK, 2007.
16. Hall, B. Lie Groups, Lie Algebras, and Representations: An Elementary Introduction; Springer: New York, NY, USA, 2003.
17. Zhang, F. Matrix Theory: Basic Results and Techniques; Springer: New York, NY, USA, 1999.
18. Simon, B. Trace Ideals and Their Applications, 2nd ed.; Math. Surveys and Monographs. A.M.S.: Cambridge, UK, 2005.
19. Tropp, J.A. An introduction to matrix concentration inequalities. Found. Trends Mach. Learn. 2015, 8, 1–2. [CrossRef]
20. Davis, C. Notions generalizing convexity for functions defined on spaces of matrices. In Proceedings of the Symposia in Pure
Mathematics; American Mathematical Society: Providence, RI, USA, 1963; Volume 7, pp. 187–201.
21. Donogue, W. Monotone Matrix Functions and Analytic. Continuution; Springer: New York, NY, USA, 1974.
22. Bellman, R. Inequalities; Springer Verlag: Berlin/Gottingen/Heidelberg, Germany, 1961.
23. Hansen, F. Golden-Thompson’s inequality for deformed exponentials. J. Stat. Phys. 2015, 159, 1300–1305. [CrossRef]