Complete Log-Sobolev inequalities
Li Gao
Technical University of Munich (TUM)
ANR Workshop on NC analysis on groups and quantum groups June 17, 2021
Joint work with Cambyse Rouz´e (TUM). arXiv:2102.04146
Quantum Markov semigroup
(M, τ)finite von Neumann algebra with n.f. trace τ.
Aquantum Markov semigroup (QMS)(Tt)t≥0:M → Mis a continuous family of maps satisfying
Ttis normal unital completely positive Tt◦Ts=Ts+t, T0=id
t7→Tt(x)isw∗-continuous for everyx∈ M.
(symmetric): τ(Tt(x)∗y) =τ(x∗Tt(y)).
Generator: Ax= lim
t→0−1
t(Tt(x)−x), Tt=e−At.
Examples: Classical Markov semigroup: M=L∞(Ω, µ)probability space. E.g. A= ∆Laplacian on a compact Riemannian manifold(M, g). Quantum physics: M= (B(H),tr)dissipative open quantum systems. E.g. LindbladianA(x) =−P
j[aj,[aj, x]], aj∈B(H)
Operator algebras: group von Neumann algebras, quantum group and q-deformed Gaussian.
Quantum Markov semigroup
(M, τ)finite von Neumann algebra with n.f. trace τ.
Aquantum Markov semigroup (QMS)(Tt)t≥0:M → Mis a continuous family of maps satisfying
Ttis normal unital completely positive Tt◦Ts=Ts+t, T0=id
t7→Tt(x)isw∗-continuous for everyx∈ M.
(symmetric): τ(Tt(x)∗y) =τ(x∗Tt(y)).
Generator: Ax= lim
t→0−1
t(Tt(x)−x), Tt=e−At. Examples:
Classical Markov semigroup: M=L∞(Ω, µ)probability space.
E.g. A= ∆ Laplacian on a compact Riemannian manifold(M, g).
Quantum physics: M= (B(H),tr)dissipative open quantum systems.
E.g. LindbladianA(x) =−P
j[aj,[aj, x]], aj∈B(H)
Operator algebras: group von Neumann algebras, quantum group and q-deformed Gaussian.
Convergence and decoherence
ConsiderM= (Md,tr), Tt=e−At:Md→Md.
Fixed point space (subalgebra)N ={x∈Md|Tt(x) =x,∀t}.
State spaceS(Md) ={ρ∈Md|ρ≥0,tr(ρ) = 1}. Then Tt(ρ)−→EN(ρ)ast→ ∞
EN :Md→ N conditional expectation, i.e. ∀x∈ N,tr(xρ) =tr(xEN(ρ)).
How fast does the convergenceTt(ρ)−→EN(ρ)happen?
Mixing time
tmix = inf{t >0| kTt(ρ)−EN(ρ)k1≤1/2,∀ stateρ}
Spectral gap
λgap:= inf
EN(x)=0
tr(x∗Ax) kxk2
minimal positive eigenvalueA onL2(Md). kTt(x)−EN(x)k2≤e−λgaptkx−EN(x)k2⇒tmix .λgap1 O(lnd).
Convergence and decoherence
ConsiderM= (Md,tr), Tt=e−At:Md→Md.
Fixed point space (subalgebra)N ={x∈Md|Tt(x) =x,∀t}.
State spaceS(Md) ={ρ∈Md|ρ≥0,tr(ρ) = 1}. Then Tt(ρ)−→EN(ρ)ast→ ∞
EN :Md→ N conditional expectation, i.e. ∀x∈ N,tr(xρ) =tr(xEN(ρ)).
How fast does the convergenceTt(ρ)−→EN(ρ)happen?
Mixing time
tmix = inf{t >0| kTt(ρ)−EN(ρ)k1≤1/2,∀ stateρ}
Spectral gap
λgap:= inf
EN(x)=0
tr(x∗Ax) kxk2
minimal positive eigenvalueA onL2(Md).
kTt(x)−EN(x)k2≤e−λgaptkx−EN(x)k2⇒tmix .λgap1 O(lnd).
Modified Log-Sobolev Inequality
Relative entropy: for two quantum statesρ, σ, D(ρ||σ) :=tr(ρlogρ−ρlogσ) ThenTt(ρ)→EN(ρ) =⇒D(Tt(ρ)||EN(ρ))→0.
Definition (Bardet ’17)
We sayTt=e−Atsatisfiesλ-modified log-Sobolev inequality(λ-MLSI) for λ >0if
2λ D(ρ||EN(ρ))≤tr(Aρlnρ), ∀ρ∈S(Md).
Fisher information: IA(ρ) :=tr(A(ρ) lnρ) =−dtd|t=0D(Tt(ρ)||EN(ρ)). Thenλ-MLSI means:
d
dtD(Tt(ρ)||EN(ρ))≤ −2λD(Tt(ρ)||EN(ρ))
⇐⇒D(Tt(ρ)||EN(ρ))≤e−2λtD(ρ||EN(ρ)).
λ-MLSI⇐⇒ Exponential decay of relative entropy⇒tmix≤ 1λO(ln(lnd))
Modified Log-Sobolev Inequality
Relative entropy: for two quantum statesρ, σ, D(ρ||σ) :=tr(ρlogρ−ρlogσ) ThenTt(ρ)→EN(ρ) =⇒D(Tt(ρ)||EN(ρ))→0.
Definition (Bardet ’17)
We sayTt=e−Atsatisfiesλ-modified log-Sobolev inequality(λ-MLSI) for λ >0if
2λ D(ρ||EN(ρ))≤tr(Aρlnρ), ∀ρ∈S(Md).
Fisher information: IA(ρ) :=tr(A(ρ) lnρ) =−dtd|t=0D(Tt(ρ)||EN(ρ)).
Thenλ-MLSI means:
d
dtD(Tt(ρ)||EN(ρ))≤ −2λD(Tt(ρ)||EN(ρ))
⇐⇒D(Tt(ρ)||EN(ρ))≤e−2λtD(ρ||EN(ρ)).
λ-MLSI⇐⇒ Exponential decay of relative entropy⇒tmix≤λ1O(ln(lnd))
Connection to other functional inequalities
L2-Log-Sobolev inequality (LSI): for an ergodic semigroupTt=e−At (N =C1),
tr(x2lnx2)−tr(x2) lntr(x2)≤ 1
λtr(x∗Ax)
⇐⇒Hypercontractivity: (Olkiewicz-Zegarlinski ’99) kTt:L2→Lpk≤1forp≤1 +e2λt
(Temme-Kastoryano ’13 & Bardet ’17) For the optimal constant, λLSI≤λMLSI≤λgap
Moreover,
MLSI=⇒Transport cost inequality =⇒concentration of measure.
(Datta-Rouz´e, Carlen-Maas ’18)
Tensorization and complete MLSI
Tensorization
Tt, St:L∞(Ω)→L∞(Ω)hasλ-MLSI=⇒Tt⊗Stλ-MLSI (same for LSI) Unknownfor MLSI in quantum cases. True for LSI onM2(King ’14).
Definition
Tt=e−Atsatisfiesλ-complete log-Sobolev inequality (λ-CLSI)forλ >0 if idMn⊗Ttsatisfiesλ-MLSI for alln≥1.
Tensorization of CLSI
Tt, St:Md→Md both hasλ-CLSI=⇒Tt⊗Sthasλ-CLSI Applications inquantum lattice spin system.
Estimating decay ofentanglement: for a bipartite stateρonMn⊗Md, D(idn⊗Tt(ρ)||idn⊗EN(ρ))→0
Tensorization and complete MLSI
Tensorization
Tt, St:L∞(Ω)→L∞(Ω)hasλ-MLSI=⇒Tt⊗Stλ-MLSI (same for LSI) Unknownfor MLSI in quantum cases. True for LSI onM2(King ’14).
Definition
Tt=e−Atsatisfiesλ-complete log-Sobolev inequality (λ-CLSI)forλ >0 if idMn⊗Ttsatisfiesλ-MLSI for alln≥1.
Tensorization of CLSI
Tt, St:Md→Md both hasλ-CLSI=⇒Tt⊗Sthasλ-CLSI Applications inquantum lattice spin system.
Estimating decay ofentanglement: for a bipartite stateρonMn⊗Md, D(idn⊗Tt(ρ)||idn⊗EN(ρ))→0
Example: Depolarizing semigroup
Qubit depolarizing semigroup:
Tt:M2→M2, Tt(ρ) =e−tρ+ (1−e−t)tr(ρ)1 2 Optimal constant:
1/2≤λCLSI(Tt)≤λMLSI(Tt⊗idM2)<λMLSI(Tt) =λLSI(Tt) = 1. In particular,λCLSI6=λMLSI.
λLSI(Tt⊗idM2) = 0. In general, LSI/Hypercontractivity fails for non ergodicTtwheneverN is noncommutative, hence for idMn⊗Tt.
Question
Does every finite dimensionalTt:Md→Md has λCLSI:= inf
n λMLSI(idMn⊗Tt)>0?
(Junge-Li-LaRacuente, ’20) True for classical casesTt:l∞d →ld∞.
Example: Depolarizing semigroup
Qubit depolarizing semigroup:
Tt:M2→M2, Tt(ρ) =e−tρ+ (1−e−t)tr(ρ)1 2 Optimal constant:
1/2≤λCLSI(Tt)≤λMLSI(Tt⊗idM2)<λMLSI(Tt) =λLSI(Tt) = 1. In particular,λCLSI6=λMLSI.
λLSI(Tt⊗idM2) = 0. In general, LSI/Hypercontractivity fails for non ergodicTtwheneverN is noncommutative, hence for idMn⊗Tt.
Question
Does every finite dimensionalTt:Md→Md has λCLSI:= inf
n λMLSI(idMn⊗Tt)>0?
(Junge-Li-LaRacuente, ’20) True for classical casesTt:l∞d →ld∞.
Finite dimensional results
Theorem. (G.-Rouz´ e, preprint ’21)
LetTt=e−At:Md →Md be a symmetric Quantum Markov semigroup and N be the fixed point subalgebra. Then
λgap
2Ccb(Md:N) ≤λCLSI ≤λgap
C(Md:N)Pimsner-Popa index.
Remain valid forGNS-symmetricsemigroup: for a faithful stateσ, tr(Tt(x)∗yσ) =tr(x∗Tt(y)σ)
Pimsner-Popa index
N ⊂ Mfinite von Neumann algebra. EN :M → N cond. expectation.
Pimsner-Popa index
C(M:N) := inf{C >0|ρ≤CEN(ρ)∀ρ∈ M+}
(Pimsner-Popa, ’86)C(M:N) = [M:N] for II1 subfactor
& Explicit formula for finite dimensionalM,N.
CB-version: Ccb(M:N) = supnC(Mn(M) :Mn(N)). E.g. C(Md:C) =d, Ccb(Md:C) =d2and
Ccb(Md:N)≤Ccb(Md:C) =d2.
Pimsner-Popa index
N ⊂ Mfinite von Neumann algebra. EN :M → N cond. expectation.
Pimsner-Popa index
C(M:N) := inf{C >0|ρ≤CEN(ρ)∀ρ∈ M+}
(Pimsner-Popa, ’86)C(M:N) = [M:N] for II1 subfactor
& Explicit formula for finite dimensionalM,N.
CB-version: Ccb(M:N) = supnC(Mn(M) :Mn(N)).
E.g. C(Md:C) =d, Ccb(Md:C) =d2and
Ccb(Md:N)≤Ccb(Md:C) =d2.
Quantum χ
2-divergence
Forρ∈S(Md), defineρweightedL2-norm kXk2ρ−1=
Z ∞
0
tr X∗ 1
ρ+sX 1 ρ+s
ds .
commutative case,kXk2ρ−1=R |X|2 ρ dµ
for two quantum states,kρ−σk2σ−1:=χ2(ρ, σ)quantumχ2-divergence.
Key Lemma
Recall the relative entropyD(ρ||σ) =tr(ρlogρ−ρlogσ). i) D(ρ||σ)≤kρ−σk2σ−1
ii) Ifρ≤Cσ,CkXk2ρ−1≥kXk2σ−1
Quantum χ
2-divergence
Forρ∈S(Md), defineρweightedL2-norm kXk2ρ−1=
Z ∞
0
tr X∗ 1
ρ+sX 1 ρ+s
ds .
commutative case,kXk2ρ−1=R |X|2 ρ dµ
for two quantum states,kρ−σk2σ−1:=χ2(ρ, σ)quantumχ2-divergence.
Key Lemma
Recall the relative entropyD(ρ||σ) =tr(ρlogρ−ρlogσ).
i) D(ρ||σ)≤kρ−σk2σ−1
ii) Ifρ≤Cσ,CkXk2ρ−1≥kXk2σ−1
Proof of Theorem.
DenoteρN =EN(ρ). Recallλ-MLSI is
2λ D(ρ||ρN)≤IA(ρ) OnMd,Tt=e−At is given byA(x) =−Pk
j=1[aj,[aj, x]]. Then A=δ∗δ , δ(x) =⊕kj=1[aj, x]
is a derivation. Then
IA(ρ) =tr(Aρlogρ) =tr(δ(ρ)δ(logρ))
= Z ∞
0
tr(δ(ρ) 1
ρ+sδ(ρ) 1
ρ+s)ds=kδ(ρ)k2ρ−1
D(ρ||ρN)≤ kρ−ρNk2ρ−1 N
≤ 1
λgap kδ(ρ−ρN)k2ρ−1 N
= 1
λgap kδ(ρ)k2ρ−1 N
= C
λgap kδ(ρ)k2ρ−1 (ρ≤CρN)
Discussion on optimality
SinceC(Md:N)≤d, Ccb(Md:N)≤d2, we have shown λgap
2d ≤ λgap
2C(Md:N)≤λMLSI≤λgap
λgap
2d2 ≤ λgap
2Ccb(Md:N) ≤λCLSI ≤λgap
ergodic case: N =C1 λgap(Tt)
logd+ 2 ≤λMLSI(Tt)≤λgap(Tt) (Temme-Kastoryano ’13)
Question
Can we haveλMLSI≥λgapO(log1d)or evenλCLSI≥λgapO(log1d)?
(Brannan-G.-Junge, ’20) Ture forSchur multiplierand some Random unitary.
Discussion on optimality
SinceC(Md:N)≤d, Ccb(Md:N)≤d2, we have shown λgap
2d ≤ λgap
2C(Md:N)≤λMLSI≤λgap
λgap
2d2 ≤ λgap
2Ccb(Md:N) ≤λCLSI ≤λgap
ergodic case: N =C1 λgap(Tt)
logd+ 2 ≤λMLSI(Tt)≤λgap(Tt) (Temme-Kastoryano ’13)
Question
Can we haveλMLSI≥λgapO(log1d)or evenλCLSI≥λgapO(log1d)?
(Brannan-G.-Junge, ’20) Ture forSchur multiplierand someRandom unitary.
Heat Semigroup
(Li-Junge-LaRacuente,’20) For a compact Riemannian Manifold(M, g), ifRicci(M)≥λ >0, the Heat semigroupHt=e−∆t satisfyλ-CLSI.
(Brannan-G.-Junge, ’20) The Heat semigroupHt=e−∆ton every compact Riemannian Manifold hasλCLSI>0.
Example: Unit Circle
T={z∈C| |z|= 1}andHt(zm) =e−m2tzm. 1
4 ln 3 ≤λCLSI(Ht)≤λMLSI(Ht) =λLSI(Ht) = 1
Question
DoesλCLSI=λMLSI for classical semigroup?
Hypo-elliptic case
(G.-Junge-LaRacuente, ’18) For a sub-Laplacian∆X =−P
jXj2 satisfyingH¨ormander condition, the subordinate semirgoupe−∆θXt satisfy hasλCLSI>0 for0< θ <1.
Example:Possion semigroup
T={z∈C| |z|= 1}andPt(zm) =e−|m|tzm.
λCLSI(Pt) =λMLSI(Pt) =λLSI(Pt) =λgap(Pt) = 1
Question
Does the sub-Laplacian∆X itself satisfy CLSI?
Implies dimension free estimate for all tranference semigroup of∆X
No example known, evenSU(2).
Compact Quantum group
Central semigroup on compact quantum group has Ricci≥0.
kδ(Tt(x))kρ−1≤e−λtkδ(x)kTt(ρ)−1 (Ricci≥λ) Free orthogonal groupON+ & quantum permutation groupSN+:
Heat semigroup haveλCLSI≥O(Nlog1 N).
Group von Neumann algebra: Fourier multiplierTt(λ(g)) =e−φ(g)tλ(g) hasλCLSI>0 under some growth condition onφ.
Theorem (Brannan-G.-Junge & Wirth-Zhang, ’20)
Fd free group andPt(λ(g)) =e−|g|tλ(g).
λRicci(Pt) =λCLSI(Pt) =λgap(Pt) = 1
LSI constantλLSI(Pt)≥1.17(Junge-Palazuelos-Parcet-Perrin-Ricard) (1) Optimal Hypercontractivity forPt?
(2) Positive curvature forO+N?