QUASI-FACTORIZATION AND MULTIPLICATIVE COMPARISON OF SUBALGEBRA-RELATIVE ENTROPY

NICHOLAS LARACUENTE

Abstract. Purely multiplicative comparisons of quantum relative entropy are desirable but challenging to prove. We show such comparisons for relative entropies between comparable densities, including subalgebra-relative entropy and its perturbations. These inequalities are asymptotically tight in approaching known, tight inequalities as perturbation size approaches zero. Following, we obtain a quasi-factorization inequality, which compares the relative entropies to each of several subalgebraic restrictions with that to their intersection. We apply quasi- factorization to uncertainty-like relations and to conditional expectations arising from graphs. Quasi-factorization yields decay estimates of optimal asymptotic order on mixing processes de- scribed by finite, connected, undirected graphs.

1. Introduction The strong subadditivity (SSA) of von Neumann entropy states that for a tripartite density ρABC on systems ABC,

H(AC)ρ + H(BC)ρ ≥ H(ABC)ρ + H(C)ρ , (1)

where the von Neumann entropy H(ρ) := − tr(ρ log ρ), and H(A)ρ := H(trBC (ρ)). Lieb and Ruskai proved SSA in 1973 [LR73]. SSA’s impacts range from quantum Shannon theory [Wil13] to holographic spacetime [CKPT14]. A later form by Petz in 1991 generalizes from subsystems to subalgebras. Let M be a von Neumann algebra and N be a subalgebra. Associated with N is a unique conditional expectation EN that projects an operator in M onto N . We will denote by EN the trace-preserving, unital, completely positive conditional expectations from M to N . If S, T ⊆ M as subalgebras, we call them a commuting square if ES ET = ET ES = ES∩T , the conditional expectation onto their

arXiv:1912.00983v5 [quant-ph] 31 Aug 2021 intersection. Petz’s theorem [Pet91] implies:

Theorem 1.1 (Petz’s Conditional Expectation SSA). Let ES , ET be conditional expectations to subalgebras S, T ⊆ M. If S and T form a commuting square, then for all densities ρ,

H(ES (ρ)) + H(ET (ρ)) ≥ H(ES∩T (ρ)) + H(ρ) . The Umegaki relative entropy for matrices is defined by D(ρkσ) = tr(ρ(ln ρ − ln σ)) when supp(ρ) ⊆ supp(σ) and is infinite otherwise. In terms of Umegaki’s relative entropy, strong

NL is supported by IBM as a Postdoctoral Scholar at UChicago & the Chicago Quantum Exchange. NL was previously supported by the Department of Physics at the University of Illinois at Urbana-Champaign. 1 2 NICHOLAS LARACUENTE subadditivity can be rewritten as

D(ρkES (ρ)) + D(ρkET (ρ)) ≥ D(ρkES∩T (ρ)) (2) for S and T forming a commuting square. As we note with Gao and Junge in [GJL17], theorem 1.1 implies SSA and an uncertainty relation for mutually unbiased bases, joining these concepts. In [GJL17], it is shown that the condition [ES , ET ] = 0 is necessary as well as sufficient in finite dimensions. D(ρkES (ρ)) + D(ρkET (ρ)) ≥ D(ρkES∩T (ρ)) − c (3) can only hold for all ρ when [ES , ET ] 6= 0 if c > 0. When conditional expectations don’t commute, there are nonetheless inequalities approximating strong subadditivity. Perturbations of entropy and its inequalities often take additive forms [BCC+10, Win16, GJL18a, GJL18b]. If entropies are sufficiently small, then equation (3) becomes trivial and weaker than the statement that the left hand side is non-negative, as do most entropy inequalities with additive constants. Multiplicative bounds on relative entropy exist [AE05, AE11, Ver19], but these are constrained by the infinite divergence of entropy relative to states with smaller support. The special case of D(ρk1ˆ/d), however, has range [0, ln d] on a system of dimension d. More generally, for a doubly stochastic conditional expectation E, we refer to the function on a density ρ given by D(ρkE(ρ)) as subalgebra-relative entropy. Having bounded range, subalgebra-relative entropy may support much stronger inequalities than general relative entropy. D(ρk1ˆ/d) is a special case of subalgebra-relative entropy, where the algebra is C1, that of the complex scalars. Subalgebra-relative entropy appears as measures of resources such as quantum coherence [WY16] and reference frame asymmetry [VAWJ08, GMS09, MS14]. One may write the conditional mutual information as D(ρkES (ρ)) + D(ρkET (ρ)) − D(ρkES∩T (ρ)) for subsystem-restricted algebras S and T , which then applies to derived entanglement measures [Tuc99, Tuc02, CW04]. Subalgebra- relative entropy is a natural measure of decoherence from processes that are self-adjoint with respect to the Hilbert-Schmidt inner product [BR18, GJL20a]. The maximum subalgebra-relative entropy for a given conditional expectation connects closely to the theory of subalgebra indices [GJL20b]. Hence subalgebra-relative entropy is fundamental to quantum information, motivating inequalities on this form. More broadly, relative entropy should be comparable between densities that are comparable up to constants in the Loewner order, where these constants determine the strength of such comparisons. Again, it is not hard to find bounds with additive corrections of this form. Mul- tiplicative comparisons are more challenging, but multiplicative relative entropy comparison is more powerful in many settings, especially when all of the relative entropies involved could be arbitrarily small. The notion of quasi-factorization for classical entropies was introduced and shown in [Ces01], showing a multiplicative generalization of strong subadditivity for non-commuting conditional expectations. Several works consider quantum generalizations or related properties in a variety of settings [CLPG18, BCL+19, BCPH21]. This form of inequality is also known as approximate SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 3 tensorization [CMT14, BCR20] ∗, including a “strong” form that is fully multiplicative, and a “weak” form that includes an additive correction term. In this paper, we refer primarily to the multiplicative form, which we generalize to any finite number of subalgebras:

Definition 1.2 (Multiplicative Quasi-factorization). Let {Ej : j ∈ 1...J ∈ N} be a set of con- ditional expectations, and E the conditional expectation onto their intersection algebra. We say that this set satisfies a strong quasi-factorization (SQF, or specifically α-SQF) inequality if X α D(ρkEj(ρ)) ≥ D(ρkE(ρ)) j for some α > 0. We say that it satisfies complete, strong quasi-factorization (CSQF, or specifi- C cally α-CSQF) if {1ˆ ⊗ Ej} has α-SQF for any finite-dimensional extension C. More generally, we say the set has {αj}-(C)SQF if X αjD(ρkEj(ρ)) ≥ D(ρkE(ρ)) . j

The original version of this paper [LaR21, v1] showed Theorem 1.3 and a special case of Theorem 1.8 for subalgebras with scalar intersection. In that version, it was left as a conjecture that the dimension could be replaced by a subalgebra index, the result generalized to arbitrary sets of subalgebras, and the inequality made tensor-stable. Later, a new technique developed by Gao and Rouz´e[GR21a] showed general, tensor-stable, multiplicative quasi-factorization with constant determined by a subalgebra index C, rather than the system’s dimension. Their result shows the existence of quasi-factorization for all finite-dimensional quantum systems. Incorporating one of the techniques of that result, we find a quasi-factorization that is asymptotically tight in the following sense: for a pair of conditional expectations E1, E2 with intersection conditional expectation E such that kE1E2 − Ek3 → 0, our bound approaches strong subadditivity. Our results have an asymptotic α ∼ O(ln C) dependence on the index for large C, improving on the asymptotic dependence from preceding versions of [GR21a]. We find a strong, complete quasi-factorization like that in [GR21a], which also has asymptotic tightness like in [BCR20] or [LaR21, v1], and logarithmic index scaling like that in [LaR21, v1]. Furthermore, we explicitly show this for any finite number of conditional expectations. Decay and decoherence are some of the most vexing challenges to quantum technology. As a primary application, α-(C)SQF allows us to derive exponential relative entropy decay rates of continuous-time quantum processes having subprocesses with strong decay. SQF implies merging of a particularly strong form of decay estimate known as a modified logarithmic Sobolev inequality (MLSI). This does not follow from perturbations of strong subadditivity that include additive constants as in equation (3).

∗We thank the authors of [BCR20] for access to an early draft that considered such an inequality in parallel with the writing of this manuscript. Their present version derives a comparable, multiplicative form in some cases. 4 NICHOLAS LARACUENTE

It is shown in [BJL+21] that CMLSI upper bounds capacities of quantum channels, which are famously difficult to calculate due to superadditivity and hardness of numerics for high- dimensional quantum entropies. Furthermore, we note in that work that CMLSI implies tensor- stable decoherence time estimates, an important problem for quantum computing and memory.

1.1. Primary Contributions. A quantum channel is a completely positive, trace-preserving map. By E we may denote a channel or a conditional expectation that is self-adjoint with respect to the trace, in which case it is the unique restriction map to some subalgebra. By Eσ, Eσ∗ we denote a conditional expectation weighted by state σ as in Section 3. By D(·k·) we denote the relative entropy, and by H(·) the von Neumann entropy. By 1ˆ we denote the identity matrix. For systems A, B, C, ... or von Neuman algebras M, N , ... we denote by |A| or |M| the dimension. The A A subsystem entropy is denoted H(A)ρ := H(ρ ), where ρ denotes the restriction to subsystem A of a multipartite state ρ on A ⊗ B ⊗ C ⊗ .... The state 1ˆ/|A| or 1ˆ/|M| is the repsective complete mixture on A or M. Though this notation assumes finite dimension, one may interpret 1ˆ/|M| as a normalized trace on a finite von Neumann algebra. Unless otherwise noted, results of this paper assume the existence of a finite trace. First, we prove a basic bound on relative entropy to complete mixture:

Theorem 1.3. Given two densities ρ, σ in dimension d such that ρ σ (ρ majorizes σ), ˜ D(ρk(1 − ζ)1ˆ/d + ζσ) ≥ βa,b,ζ D(ρk1ˆ/d) ˜ for any a ∈ [0, 1], b ∈ (0, 1), βa,β,ζ = (1 − a), and n 1 − b b o ζ ≤ a min , . d + a(1 − b) + 1 (1 − ab)d + ab + 1 We may optimize a and b in Theorem 1.3 for given values of d and ζ. If we wish to avoid optimization, b = 1/2 is a reasonable value, and one may use the calculated bound of Corollary 2.4. We further see from this Corollary that 1.3 is asymptotically tight: as ζ → 0 for fixed d, ˜ βa,b,ζ → 1 for suitable choices of a, b. Theorem 1.3 relies on a comparison using telescopic relative entropy [Aud14]. As an intermediate result proven in Section 2.1:

Proposition 1.4. Let ρ and σ be two densities such that ρ σ (ρ majorizes σ), and ζ ∈ [0, 1]. Then D(ρk(1 − ζ)1ˆ/d + ζσ) ≥ D(ρk(1 − ζ)1ˆ/d + ζρ) .

Using Lemmas shown in [GR21a], we obtain a generalized form of 1.3 that replaces complete mixture by the output of a channel E and dimension-dependence by comparison of input and output densities.

Theorem 1.5 (Triangle-like Relative Entropy Comparison). Let ρ, σ, ω be densities such that (1 − ζ)σ ≤ ω ≤ (1 + ζ(c − 1))σ for constants ζ ∈ (0, 1) and c ≥ 1. Let η := (ω − (1 − ζ)σ)/ζ, so that ζη + (1 − ζ)σ = ω. Then η is a density, (1 − ζ)2D(ρkσ) ≤ (1 + ζ(c + 1) + 2ζ2(c − 1))D(ρkω) + ζ(2 + ζ)cD(ρkη) , SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 5 and  (c − 1)2  (1 − ζ)2D(ρkσ) ≤ (1 + ζ(c + 1) + 2ζ2(c − 1))D(ρkω) + 3ζ(2 + ζ) 1 + D(ηkσ) . c(ln c − 1) + 1 This Theorem is asymptotically tight in approaching the equality D(ρkσ) = D(ρkω) as ζ → 0. Its connection to quasi-factorization and subalgebra-relative entropy is apparent via a Corollary:

Corollary 1.6 (Multiplicative Perurbation of Subalgebra-relative Entropy). Let ρ be a density and E, Φ be quantum channels such that (1 − ζ)E(ρ) + ζΦ(ρ) ≤ (1 + ζ(c − 1))E(ρ) (when the 1st inequality is satisfied, the 2nd is equivalent to Φ(ρ) ≤ cE(ρ)) for constants ζ ∈ (0, 1) and c ≥ 1. Furthermore, let ΦE = E. Then

D(ρk(1 − ζ)E(ρ) + ζΦ(ρ)) ≥ βc,ζ D(ρkE(ρ)) for 1  (c − 1)2  β := 1 − 3ζ(2 + ζ) − 8ζ − 2ζ2 . c,ζ 1 + ζ(c + 1) + 2ζ2(c − 1) c(ln c − 1) + 1

As with the Theorem, this Corollary is asymptotically tight in that βc,ζ → 1 as ζ → 0. Conversely:

Proposition 1.7. Let E, Φ be quantum channels and ρ be a density such that κE(ρ) ≤ Φ(ρ) ≤ cE(ρ), and ρ ≤ cE(ρ) for some κ, c > 0. Assume that D(Φ(ρ)kE(ρ)) ≤ D(ρkE(ρ)). Then a(c − 1)2 D(ρkΦ(ρ)) ≤ D(ρkE(ρ)) κ(c(ln c − 1) + 1) ˜ ˜ for constant a. In general, a ≤ 4. If Φ = (1 − b)E + b√Φ for some b ∈ [0, 1] and Φ such that Φ˜E = E, then we may improve the constant to a = 1 + 2 b + b.

Corollary 1.6 forms the base of a quasi-factorization result:

† Theorem 1.8 (Quasi-factorization). Let {(Nj, Ej): j = 1...J ∈ N} be a set of J ∈ N von Neumann algebras and associated conditional expectations. Let E be the conditional expectation (j) J to the intersection algebra, N = ∩jNj. Let (σ )j=1 be weighting densities such that for a density σ, Ej,σ∗Eσ∗ = Eσ∗Ej,σ∗ = Eσ∗ for each j ∈ 1...J. ⊗m s Let S = ∪m∈N{1...J} be the set of finite sequences of indices. For any s ∈ S, let E denote the composition E (j ) ...E (j ) for s = (j , ..., j ). Let µ : S → [0, 1] be a probability j1,σ 1 ∗ jm,σ m ∗ 1 m measure on S and kj,s upper bound the number of times Ej,σ(j)∗ appears in each sequence s. If for all input densities ρ, X s µ(s)E (ρ) = (1 − )Eσ∗(ρ) + Φ(ρ) (4) s∈S

†After a version of Theorem 1.8 appeared in v3 of this paper, [GR21a] added comparable results (theorem 5.4 & corollary 5.5 in that paper). Nonetheless, the techniques of our Section 3 originally appeared in v1 of this paper. It is these techniques that yield both the logarithmic dependence on c described in Remark 1.10 and the extension from two to many conditional expectations. Furthermore, Theorem 1.8 has the advantage of asymptotic tightness for conditional expectations approaching a commuting square. 6 NICHOLAS LARACUENTE for some channel Φ, then for βc,ζ as in Corollary 1.6, X X µ(s) ks,jD(ρkEj,σ(j)∗(ρ)) ≥ βc,ζ D(ρkEσ∗(ρ)) . s∈S j ˜ If E(ρ) = 1ˆ/d for all ρ in dimension d, then we may replace βc,ζ by βa,b,ζ as defined in and constrained by Theorem 1.3 when σ = σj = 1ˆ/d for all j. Recall the subalgebra indices

C(M : N ) = inf{c > 0|ρ ≤ cEN (ρ)∀ρ ∈ M∗}

Ccb(M : N ) = sup C(M ⊗ Mn : N ⊗ Mn) n∈N as considered in [GJL20b, GR21a] and originally by Pimsner and Popa [PP86] as a finite- dimensional analog of the Jones index [Jon83]. When E is the (doubly stochastic) conditional expectation from M to N , and Φ is a channel that leaves N invariant, ρ, Φ(ρ) ≤ C(M : N )E(ρ), and the bound holds up to arbitrary extensions with Ccb(M : N ) replacing C(M : N ).

Corollary 1.9. Let M denote a von Neumann algebra containing ρ, Nj denote the subalgebra projected to by each doubly stochastic conditional expectation Ej, and N denote the intersection AB algebra projected to by E. We may extend ρ to a bipartite density ρ , where Ej act on A as B B does E. Let E˜j = Ej ⊗ 1ˆ , and E˜ = E ⊗ 1ˆ . Let µ(s), (ks,j), ζ be as in Theorem 1.8. If |B| ≤ |A| (including if B is a trivial system of dimension 1), and X µ(s)Es(ρ) = (1 − ζ)E + ζΦ , s∈S for some channel Φ, then for any  ≤ ζ,

dlogζ e X X µ(s) ks,jD(ρkE˜j(ρ)) ≥ D(ρkE˜(ρ)) βc, s∈S j with c = C(M ⊗ B : N ⊗ B), and B being the algebra of bounded operators on B. If |B| ≥ |A|, then we may set c = Ccb(M : N ) for a complete, strong quasi-factorization. For weighted conditional expectations in finite dimension d, one may scale C(M ⊗ B : N ⊗ B) or Ccb(M : N ) by 1/d divided by the minimal eigenvalue of σ to upper bound on c. Though c can be upper bounded by the index as in Corollary 1.9, sometimes there is a better upper bounded based on specific knowledge of the channels involved. As explained in Section 4.2, Theorems 1.5 and 1.8 are naturally strong in contexts reminiscent of transference, an idea present in [GJL18a, GJL20a, LJL20, BJL+21]. Transference may compare quantum channels through analogous classical channels. When several quantum channels are weighted averages of the same unitary conjugations, we may often derive Loewner order inequalities by studying how operations and compositions affect the weights. These inequalities naturally map to the conditions of Theorems 1.5 and 1.8. Though it might not be obvious that all sets of subalgebras satisfy equation (4) for some finite k and ζ > 0, Proposition 2.11 shows that when the operator norm distance between a unital SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 7 channel Φ and a conditional expectation E acting on half a Bell pair is sufficiently small, and EΦ = E, Φ must be a convex combination of E with another, unital channel. As shown in [JX07], it is always possible to find a convex combination of chains of conditional expectations from a finite set that approaches the conditional expectation to their intersection. Hence:

J Remark 1.10. For any set of finite-dimensional conditional expectations {Ej}j=1, a quasi- factorization inequality in the form of Theorem 1.8 holds. Quasi-factorization is asymptotically tight in the following sense. As

k(1ˆ ⊗ (E1...EJ − E))(|ψihψ|)k∞ → 0 , where |ψi is a Bell state on a specified bipartite system, and E remains fixed, then {Ej} has α-(C)SQF with α → 1, approaching strong subadditivity when J = 2. A concrete estimate for doubly stochastic conditional expectations follows from Proposition 2.11 and Theorem 1.8. In particular, the condition is satisfied if kE1...EJ − Ek3 is sufficiently small, which is in turn satisfied if k[Ej, El]k3 is sufficiently small for all j, l ∈ 1...J, as long as they do not approach a different intersection algebra. When the minimum c in Theorem 1.8 is large, it is possible to obtain a constant scaling as Ω(ln c) when ζ = 1/c2, which one can achieve by iterating from a larger value of ζ. For any j ∈ 1...J, we may replace C(M : N ) or Ccb(M : N ) respectively by C(Nj : N ) or Ccb(Nj : N ) as in Corollary 1.9 at the cost of taking kj → kj + 2. As a primary application, quasi-factorization allows us to combine a particular class of decay estimates known as modified logarithmic-Sobolev inequalities (MLSI). Let Φt be a family of + s t s+t quantum channels in dimension d parameterized by t ∈ R , such that Φ ◦ Φ = Φ for all + s, t ∈ R . This family of channels is thereby a semigroup under composition, and there exists t t −tL a Lindbladian generator given by L = limt→0(1ˆ − Φ ) such that Φ = e . As studied in [KT13, Bar17, GJL20a], we say that L has λ-MLSI if for any input density ρ, D(Φt(ρ)kΦ∞(ρ)) ≤ e−λt D(ρkΦ∞(ρ)) . (5) As defined in [GJL20a], we say that L has λ-CMLSI (complete, modified logarithmic-Sobolev inequality) if for any bipartite density ρAB, D((Φt ⊗ 1ˆB)(ρ)k(Φ∞ ⊗ 1ˆB)(ρ)) ≤ e−λt D(ρk(Φ∞ ⊗ 1ˆB)(ρ)) . (6) We (re)prove:

t Proposition 1.11 (MLSI/CMLSI Merge). Let {Φj : j ∈ 1...J ∈ N} be self-adjoint quantum t ∞ Markov semigroups such that Φj = adexp(−Lj t) with fixed point conditional expectation Φj . Let t P t Φ be the semigroup generated by L = j αjLj + L0, where L0 generates Φ0 with fixed point t algebra containing the fixed point algebra of {Φj}. Let {Ej} have {α}-(C)SQF, and Φj have t λj-(C)MLSI for each j. Then Φ has minj{λj/α}-(C)MLSI. Proposition 1.11 is not surprising, and the historical use of quasi-factorization in proving log- Sobolev inequalities relies on essentially equivalent results. A simple proof of Proposition 1.11 with Φ0 = 1ˆ emerges from the Fisher information formulation of MLSI detailed in [GJL20a], as 8 NICHOLAS LARACUENTE the Fisher information is additive in Lindbladian, and the relative entropy assumed subadditive up to a constant. A full proof that does not use Fisher information appears in Appendix A. While this assumes α-(C)SQF rather than the more fine-grained {αj}-(C)QSF, one can adjust constant prefactors in {Lj} to equalize these coefficients, correspondingly changing the needed, individual (C)MLSI constants. Also notable is that Proposition 1.11 applies when conditional expectations do commute. When multiple subsets of constituent conditional expectations lead to the same intersection algebra, α < 1 is possible. The use of bipartite quasi-factorization to prove MLSI appears in [BCR20] and [CRF20], so similar methods exist in the literature. Remark 1.12. As shown in [BCR20, Section 3.2], one can convert (C)SQF inequalities from the doubly stochastic setting to non-trivially weighted conditional expectations. The results of [JLR19] compare (C)MLSI constants of Lindbladians with non-tracial invariant states to those with tracial invariant states, and the methods therein underlie the comparison in [BCR20, Section 3.2]. One may also use Lemma 1.7 together with Theorem 1.3 or 1.5 similarly. We demonstrate uses of quasi-factorization in two primary examples. First, we show asymp- totically tight, uncertainty-like relations for pairs of measurement bases. In particular, when A is a d-dimensional subsystem of bipartite system AB with respective matrix algebras A and B, and S, T ⊆ A correspond respectively to measurement bases {|iSi : i = 1...d} and {|iT i : i = 1...d} 2 such that ξ = mini,j | hiS|jT i | > 0, quasi-factorization implies that

βd2, H(ES⊗B(ρ)) + H(ET ⊗B(ρ)) ≥ 2H(ρ) + (ln d + H(B)ρ − H(ρ)) dln1−dξ e for any  ≤ 1 − dξ, where βd2, is as in Corollary 1.6. This form of inequality, detailed in Subsection 4.1, strengthens the usual, entropic uncertainty principle for highly mixed states. Second we use quasi-factorization to show new entropy inequalities and decay estimates for mixing channels described by finite groups and graphs. In particular: Theorem 1.13. Assume a family of m-regular, connected, undirected graphs with n vertices and edge sets En each have normalized adjacency matrices with largest subleading eigenvalue magnitude γ(n). Let β(c, ζ) = βc,ζ as in Theorem 1.5, {u(i, j):(i, j) ∈ En} represent G on some Hilbert space, and Ei,j = (1ˆ + ui,j)/2 be the conditional expectation to the invariant subalgebra of u(i, j). Then k X D(ρkE (ρ)) ≥ 1 − exp(−(m − 1)k(ln(5/2)/5 + 7/60)
D(ρkE (ρ)) , β(c, γ(m−1)k/5) i,j G (i,j)∈E where c ≤ (n2 − n)/2 + 1. Hence the Lindbladian given by 1 X L(ρ) = (ρ − u(i, j)ρu(i, j)†) m (i,j)∈E has CMLSI with constant Ω(1/m logγ(n)(1/n)). The technical version of Theorem 1.13 appears as Theorem 4.4 and shows concrete bounds given a value of γ and minimal n. SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 9

Remark 1.14. Let L be a Lindbladian on the classical, n-dimensional, 1-normed vector space n n denoted l1 , in which we may interpret positive, normalized vectors as probabilities. We biject l1 n with the diagonal subspace of n-dimensional, Schatten 1-normed matrices denoted S1 . For a clas- sical Lindlbadian L generating a semigroup of stochastic matrices on l1, that L has CMLSI with constant α is equivalent (up to choice of normalization) to the corresponding quantum semigroup n L having α-CMLSI on diagonal densities in S1 . Let LG be the degree-normalized Laplacian matrix corresponding to a finite, connected, undi- rected graph G with n vertices, and let LG be the Lindbladian as constructed in Theorem 1.13. n n LG is a Lindbladian generator on l1 , and for any corresponding probability vector ~p ∈ l1 and n n n density ρ ∈ S1 such that p = diag(ρ), L(~p) = diag(LG(ρ)), and diag : S1 → l1 denotes restriction n to the diagonal. Furthermore, if ρ˜ ∈ S1 ⊗ B has the partially diagonal form X B ρ˜ = |xihx| ⊗ ρx , x∈1...n so will (L ⊗ 1ˆB)(ρ) for any finite-dimensional extension B, so restriction to the diagonal re- mains bijective. In this way, Theorem 1.13 bounds CMLSI constants for finite graphs in a sense compatible with that of [DSC96, LY98, GJL20a, LJL20].

Theorem 1.13 answers a problem left open by [GJL20a, Remark 7.5], showing that the fastest, regular expander graphs have CMLSI with constant no worse than one over logarithmic in n, in line with expectations based on classical mixing times of expanders and classical, modified logarithmic-Sobolev inequalities [BT06]. More broadly, quasi-factorization yields uncertainty-like inequalities between relative entropies to the invariant subalgebras of finite groups. Section 2.1 proves Theorem 1.3, and 2.2 proves Theorem 1.5. Section 3 proves Theorem 1.8. Section 4 describes applications that use quasi-factorization to tighten entropic uncertainty relations and to derive new inequalities on graphs and groups. Section 5 concludes with some still open problems.

2. Multiplicative Perturbations of Relative Entropy Since relative entropy is biconvex, D(ρk(1 − ζ)ω + ζσ) ≤ (1 − ζ)D(ρkω) + ζD(ρkσ) for any densities ρ, ω, σ and ζ ∈ [0, 1]. In the other direction, there are examples of densities ρ, ω, σ for which D(ρk(1 − ζ)ω + ζσ) = 0, while D(ρkω),D(ρkσ) > 0 . Similarly, it is possible that D(ρk(1 − ζ)ω + ζσ) is finite when D(ρkω) or D(ρkσ) is infinite. Hence in full generality, there is no way to multiplicatively restrict the extent of non-concavity of relative entropy. In Section 2.1, we nonetheless show that when ω = 1ˆ/d, and ρ majorizes σ, there is a multi- plicatively adjusted form of concavity-like relation. In Section 2.2, we show an analogous bound 10 NICHOLAS LARACUENTE when ω = E(ρ) and σ = Φ(ρ) for channels Φ and E under certain conditions. These conditions are satisfied when E is a conditional expectation to an invariant subalgebra of Φ. 2.1. Perturbation of Relative Entropy to Complete Mixture. If we restrict to ω = 1ˆ/d, then we see that D(ρk(1−ζ)1ˆ/d+ζσ) and D(ρk1ˆ/d) are necessarily both finite. These conditions are still insufficient to lower bound D(ρk(1 − ζ)1ˆ/d + ζσ) in terms of D(ρk1ˆ/d). We may for instance take σ pure, and ρ = (1 − ζ)1ˆ/d + ζσ. The final condition we add is that ρ majorizes σ (ρ σ), in which case we obtain Theorem 1.3. The results of this subsection precede [GR21a] and use a different approach from that of section 2.2. Many of the results herein are subsumed by results of that section at least up to constants. Nonetheless, we include this subsection as illustrative of a more computationally inspired line of proof that obtains better constants in some cases. Furthermore, this method yields intermediate results of potentially independent interest and gives more intuition for the subsequent generalization. The results of this section are stated for finite dimension. Lemma 2.1 (Flattening). Let ρ, ω be simultaneously diagonal densities of dimension d. Let δ > 0. Let i 6= j ∈ 1...d such that ρi ≥ ρj, and ρiωj ≥ ρjωi. Let ω → ω˜ under the replacement ωi → ω˜i = ωi − δ, ωj → ω˜j = ωj + δ. Then D(ρkω) ≤ D(ρkω˜). + Proof. For any a > b ∈ R ,(a−b)/a ≤ ln a−ln b ≤ (a−b)/b, as one can verify from (d/dx)(ln x) = 1/x. Hence

D(ρkω˜) − D(ρkω) = ρi(ln ωi − ln(ωi − δ)) + ρj(ln ωj − ln(ωj + δ))

≥ (ρi/ωi − ρj/ωj)δ ≥ 0 .

Figure 1. Visualization of the cascading redistribution algorithm, which con- verts the distribution on the top-left to that on the bottom-right. SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 11

Proof of Proposition 1.4. The main idea of this proof is that if ρ σ, then flattening ρ until it becomes σ only increases the value of D(ρk(1 − ζ)1ˆ/d + ζρ). First,

D(ρk(1 − ζ)1ˆ/d + ζσ) ≥ D(ρk(1 − ζ)1ˆ/d + ζEρ0 (σ)) , by data processing under Eρ0 , the conditional expectation onto the subalgebra that commutes with ρ. We hence assume that [ρ, σ] = 0, and they are simultaneously diagonal. Let ρζ = (1 − ζ)1ˆ/d + ζρ, and define σζ accordingly. Let ~ρζ and ~σζ be d-dimensional vectors of the eigenvalues of ρζ and σζ respectively, each in non-increasing order. Let ~ω = ~ρζ . We alter ~ω via a cascading probability redistribution procedure consisting of the following steps, which transform it into a copy of ~σζ : (1) Start with the index i set to 1. ζ (2) Let ∆ = ~ωi − ~σi . If ∆i > 0, then (a) Subtract ∆ from ~ωi. Let j = i + 1. ζ ζ (b) If ~ωj < ~σj , then let δ = min{∆, ~σj − ~ωj}. Add δ to ~ωj and subtract it from ∆. (c) If ∆ = 0 (which must happen at or before j = d for normalized densities), go on to step (3). Otherwise, increment j → j + 1, and return to the previous substep (2b). (3) If i < d, increment i → i + 1 and return to step (2). Otherwise, the procedure is done. See figure 1. Since this procedure only subtracts from larger eigenvalues and adds to smaller ones, we apply Lemma 2.1 at each step that transfers probability mass from one index to another. ζ ζ ζ If ~ρi ≥ ~ρj, then ~ρi/~ρi ≥ ~ρj/~ρj . Furthermore, it is always the case that ~ωi ≤ ~ρi if we are moving ζ probability mass out of ~ωi, and always that ~ωj ≥ ~ρj if we are moving probability into j. Hence ζ ζ ζ ~ρi/~ρj ≥ ~ωi/~ωj. This shows that D(ρkρ ) = D(~ρk~ρ ) ≤ D(~ρk~σ ). Finally, using the simultaneous diagonality of ρ and σ, it is easy to see that D(~ρk~σζ ) ≤ D(ρkσζ ). Lemma 2.2. Let b ≥ 0, and a ∈ [0, 1]. Then for ζ ≤ a/(1 + b), a log(1 + b) ≥ log(1 + ζb) . Proof. We solve (1 + b)a ≥ 1 + ζb yielding (1 + b)a − 1 ζ ≤ . b We then estimate (1 + b)a − 1 (1 + b)1+a − 1 − b 1 + (1 + a)b − 1 − b a = ≥ = b b(1 + b) b(1 + b) 1 + b by Bernoulli’s inequality. n Lemma 2.3. Let ρ be given in its diagonal basis by (ρi)i=1, where n is the dimension of the system. Let a, β ∈ (0, 1), ρi ≥ 1/n ≥ ρj, and n 1 − β β o ζ ≤ a min , . n + a(1 − β) + 1 (1 − aβ)n + aβ + 1 12 NICHOLAS LARACUENTE

Then  ∂ ∂  − tr(ρ(a log(nρ) − log((1 − ζ)1ˆ + ζnρ))) ≥ 0 . ∂ρi ∂ρj Since the proof of Lemma 2.3 is technical and the Lemma a more specific alternative to methods in section 2.2, we defer proof to Appendix A. Proof of Theorem 1.3. Let n be the dimension of ρ, since d may be confused with a derivative. The goal is to show that given some a ∈ [0, 1] and densities ρ, σ such that ρ σ, D(ρk(1 − ζ)1ˆ/n + ζσ) − (1 − a)D(ρk1ˆ/n) ≥ 0 , for an appropriate value of ζ ∈ [0, 1]. We apply Lemma 1.4 to replace σ by ρ. For any states ρ and ω, D(nρknω) = tr((nρ)(log ρ + log n − log ω − log n)) = nD(ρkω) . Hence it is sufficient to prove that D(nρk(1 − ζ)1ˆ + ζnρ) − (1 − a)D(nρk1)ˆ ≥ 0 , which expands as ... = n tr(ρ(a log(nρ) − log((1 − ζ)1ˆ + ζnρ))) ≥ 0 . The main insight behind this proof is Lemma 2.3. If ρ = 1ˆ/n, then both terms are 0, and the proof is trivially complete. If ρ 6= 1ˆ/n, then the total probability mass above 1ˆ/n must equal that below 1ˆ/n to maintain normalization. Hence we may apply the variational argument in Lemma 2.3 to continuously redistribute probability from larger to smaller until ρ → 1ˆ/n. Corollary 2.4. Given a ∈ [0, 1] and two densities ρ, σ in dimension d such that ρ σ (ρ majorizes σ), 32(d + 1)ζ D(ρk(1 − ζ)1ˆ/d + ζσ) − (1 − )D(ρk1ˆ/d) ≥ 0 15 + (7d − 24)ζ is achieved whenever 15 ζ < . 25d + 56 With ζ → 0 as d is held fixed, we see that this expression is asymptotically tight. We may choose a = 1 − 1/d and 15d − 1 15d − 1 ζ ≤ = 32d2 − 7d2 + 7d + 32d + 24d − 24 25d2 + 63d − 24 or a = 1/2 and 15 ζ ≤ . 57d + 88 The proof of this Corollary is contained in Appendix A. The proof is essentially a basic calculation with linear approximations. XB P Remark 2.5. For a bipartite classical-quantum state ρ with classical system X, ρ = x px~ex⊗ P ρx, so one can expand D(ρkE(ρ)) = x∈X pxD(ρxkE(ρx)). Hence Theorem 1.3 may include clas- sical extensions. SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 13

2.2. Perturbation of Relative Entropy to a Subalgebra. The primary result of this section is the proof of Theorem1.5, generalizing and strengthening 1.3 using the methods of [GR21a]. Except for Proposition 2.11, the reults of this section hold wherever the weighted 2-norm and relative entropy are valid with the expected inequalities, not requiring additional assumptions of finite dimensionality. For this proof, we recall 4 useful results of that and preceding works. First, as noted in [GJL20a] or inferred from the form of weighted inner product constructed in [TKR+10, CM17]:

Lemma 2.6 (Inverse-Weighted Norm). The function sZ ∞ † −1 −1 kXkρ−1 := tr(X (ρ + r) X(ρ + r) )dr 0 is a norm for bounded, positive ρ, satisfying the triangle inequality in the sense that

kX + Y kρ−1 ≤ kXkρ−1 + kY kρ−1 .

The triangle inequality, positivity, and faithfulness of kX + Y kρ−1 remain true even if ρ is not invertible but has support containing the support of X and Y .

Though knowing how kXkρ−1 is induced by an inner product is in principle sufficient to deduce geometrically that it must be a norm, we here give an elementary proof:

Proof. Let −α −α −α Γρ,r (X) = (ρ + r) X(ρ + r) as a more parameterized version of Γ−1 as in [GR21a]. Then as in [GR21a], Z ∞ Z 2 −1 † −1 −1 kXkρ−1 = hX, Γr,ρ (X)i dr = tr(X (ρ + r) X(ρ + r) )dr . 0 Via the cyclic property of the trace, Z Z 2 −1/2 −1/2 −1/2 2 kXkρ−1 = hΓr,ρ (X), Γr,ρ (X)i dr = kΓr,ρ (X)k2dr , where k · k2 is the usual Schatten or Hilbert-Schmidt 2-norm. This is already enough to show −1/2 2 2 positivity. By inspecting the form of Γr,ρ (X), we can also see that kaXkρ−1 = |a|kXkρ−1 for 2 ˆ all a ∈ C, and that kXkρ−1 = 0 ⇐⇒ X = 0, the zero matrix. Expanding and using the triangle inequality for the Schatten 2-norm, Z 2 2 2 −1/2 −1/2 kX + Y kρ−1 = kXkρ−1 + kY kρ−1 + 2 kΓr,ρ (X)k2kΓr,ρ (Y )k2dr . Proving the triangle inequality for the weighted norm then reduces to showing that Z s Z Z −1/2 −1/2  −1/2 2  −1/2 2  kΓr,ρ (X)k2kΓr,ρ (Y )k2dr ≤ kΓs,ρ (X)k2ds kΓr,ρ (Y )k2dr . 14 NICHOLAS LARACUENTE

−1/2 2 −1/2 2 Let f(r) := kΓr,ρ (X)k2 and g(r) := kΓr,ρ (Y )k2, noting that f, g ≥ 0 are continuous in r. We seek to show that for general, positive, continuous f, g, s Z  Z  Z  pf(r)g(r)dr ≤ f(s)ds g(r)dr . (7)

We consider the square of both sides, as they are both positive. On the right hand side, we may apply Fubini’s theorem to obtain  Z  Z  ZZ f(s)ds g(r)dr = f(s)g(r)dsdr . (8)

On the left hand side,  Z 2 ZZ pf(r)g(r)dr = pf(s)g(s)f(r)g(r)dsdr . (9)

As all functions involved are continuous and Riemann integrable, we analyze these integrals pointwise. We first note that when r = s, the terms are exactly the same. Hence we reduce the inequality to points on which s 6= r. 0 ≤ (pf(s)g(r) − pf(r)g(s))2 = f(s)g(r) + f(r)g(s) − 2pf(s)g(s)f(r)g(r) . For each integrated point defined by (r, s), we thereby note that the value within the integral of equation (8) is at least as large as that in equation (9). Hence equation (7) holds, completing the (squared) triangle inequality and showing that kXkρ−1 is a norm. If ρ is not a norm but has support on the support of X for any given X, we may restrict the space to the support of ρ on which it is a norm.

Shown explicitly as [GR21a, Lemma 1] and implicit from Lemma 2.6 or from the methods of [JLR19] is a comparison property for inverse-weighted norms:

+ 2 2 Lemma 2.7. For densities ρ and σ and c ∈ R such that ρ ≤ cσ, kXkσ−1 ≤ ckXkρ−1 . Though the following Lemma appears in [GR21a], it follows directly from taking a well-known integral representation of the second derivative of relative entropy D(ρtkσ) with respect to t, where ρt is defined in the Lemma:

Lemma 2.8. Let ρt = (1 − t)σ + tρ for t ∈ [0, 1]. Then Z 1 Z s D(ρkσ) = kρ − σk2 dtds ρ−1 0 0 t As a simple observation,

Remark 2.9. For any pair of bounded, non-negative, scalar functions f(t), g(t) and any s > 0, Z s Z s Z s f(t)g(t)dt ≤ max{f(t), g(t)}2dt ≤ (f 2(t) + g2(t))dt . 0 0 0 Finally, we use the following “key” Lemma of [GR21a]: SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 15

Lemma 2.10 (Lemma 2 from [GR21a]). Let ρ ≤ cσ for c > 0, densities ρ and σ. Then c(ln c − 1) + 1 kρ − σk2 ≤ D(ρkσ) ≤ kρ − σk2 . (c − 1)2 σ−1 σ−1 Using these four results from the literature, we derive the new results below.

2 −1 Proof of Proposition 1.7. By Lemma 2.10, D(ρkΦ(ρ)) ≤ kρ−Φ(ρ)kΦ(ρ)−1 . Since E(ρ) ≤ κ Φ(ρ), 2 −1 2 kρ−Φ(ρ)kΦ(ρ)−1 ≤ κ kρ−Φ(ρ)kE(ρ)−1 using Lemma 2.8. By the triangle inequality and Lemma 2.6,

kρ − Φ(ρ)kE(ρ)−1 ≤ kρ − E(ρ)kE(ρ)−1 + kE(ρ) − Φ(ρ)kE(ρ)−1 . p Again applying Lemma 2.10, kρ − E(ρ)kE(ρ)−1 ≤ g(c)D(ρkE(ρ)), and kE(ρ) − Φ(ρ)kE(ρ)−1 ≤ pg(c)D(Φ(ρ)kE(ρ)), where g(c) is given by Lemma 2.10. Hence, 2 kρ − Φ(ρ)kE(ρ)−1 ≤ g(c) D(ρkE(ρ)) + 2pD(ρkE(ρ))D(Φ(ρ)kE(ρ)) + D(Φ(ρ)kE(ρ))
. To complete the Lemma, we use the assumption that D(Φ(ρ)kE(ρ)) ≤ D(ρkE(ρ)).

Proof of Theorem 1.5. If ω ≥ (1 − ζ)σ, then η := ω − (1 − ζ)σ is a positive, semidefinite matrix. Furthermore, when ω, σ are densities, tr(η) = (tr(ω) − (1 − ζ) tr(σ))/ζ = (1 − 1 + ζ)/ζ = 1 . Hence η is positive semidefinite with trace 1, the conditions for it to be a density matrix. It also then holds that ω = (1 − ζ)σ + ζη . This convex combination form will be useful in understanding the proof. First, let X = ρ − ω = (1 − ζ)(ρ − σ) + ζ(ρ − η), and we rewrite 1 ζ ρ − σ = X + (η − ρ) . 1 − ζ 1 − ζ By the triangle inequality and Lemmas 2.6 and 2.7, 1 ζ kρ − σk −1 ≤ kXk −1 + kη − ρk −1 (10) ξ 1 − ζ ξ 1 − ζ ξ for any density ξ. Let (ξ, φ)t := (1 − t)φ + tρ for any pair of densities ξ, φ. In this particular situation, Z 1 Z s D(ρkσ) = kρ − σk2 dtds (ρ,σ)−1 0 0 t ZZ 1  2 2 2  −1 −1 (11) ≤ 2 kρ − ωk −1 + 2ζkρ − ωk(ρ,σ) kρ − ηk(ρ,σ) + ζ kρ − ηk −1 dtds (1 − ζ) (ρ,σ)t t t (ρ,σ)t ZZ 1  2 2  ≤ 2 (1 + 2ζ)kρ − ωk −1 + ζ(2 + ζ)kρ − ηk −1 dtds , (1 − ζ) (ρ,σ)t (ρ,σ)t where the last line follows from Remark 2.9. 16 NICHOLAS LARACUENTE

We recall that by the Theorem’s assumptions,

(1 − ζ)(ρ, σ)t ≤ (ρ, ω)t ≤ (1 + ζ(c − 1))(ρ, σ)t for all t ∈ [0, 1]. Hence via Lemma 2.7, 2 2 kρ − ωk −1 ≤ (1 + ζ(c − 1)))kρ − ωk −1 . (12) (ρ,σ)t (ρ,ω)t

Since η ≤ cσ,(ρ, η)t ≤ c(ρ, σ)t, so by Lemma 2.7, 2 2 kρ − ηk −1 ≤ ckρ − ηk −1 . (ρ,σ)t (ρ,η)t Returning to equation (11) and applying Lemma 2.8 completes the proof of the first Equation in this Theorem. For the second Equation, the triangle inequality and Remark 2.9 imply that ZZ ZZ 2  2 2  kρ − ηk −1 dtds ≤ 3 kρ − σk −1 + kη − σk −1 dtds . (ρ,σ)t (ρ,σ)t (ρ,σ)t

The first term integrates to D(ρkσ). For the latter term, we note that σ ≤ (ρ, σ)t/(1 − t). Via Lemmas 2.7 and 2.10, 2 2 −1 2 (c − 1) kη − σk −1 ≤ (1 − t) kη − σkσ ≤ D(ηkσ) . (ρ,σ)t (1 − t)(c(ln c − 1) + 1) We calculate Z 1 Z s dtds Z 1 = − ln(1 − s)ds = 1 . 0 0 1 − t 0 Hence ZZ 2 2  (c − 1)  kρ − ηk −1 dtds ≤ 3 1 + D(ηkσ) . (ρ,σ)t c(ln c − 1) + 1 Returning to Equation (11),  (c − 1)2  (1 − ζ)2D(ρkσ) ≤ (1 + ζ(c + 1) + 2ζ2(c − 1))D(ρkω) + 3ζ(2 + ζ) 1 + D(ρkσ) . c(ln c − 1) + 1

Proof of Corollary 1.6. In Theorem 1.5, let σ = E(ρ), and ω = (1−ζ)E(ρ)+ζΦ(ρ). The Corollary follows from the data processing inequality when ΦE = E.

For any (doubly stochastic) conditional expectation E, there is a basis for which ˆ E(ρ) = ⊕l trBl (PlρPl) ⊗ 1/|Bl| , (13) where the identity is in the subsystem Bl, and Pl is a projection to the lth diagonal block. Proposition 2.11. Let E be a doubly stochastic conditional expectation and Φ a unital quantum channel such that EΦ = ΦE = E, both defined on matrices of dimension d. Let |ψi be a d × d Bell state given by |ψi = √1 Pd |ii ⊗ |ii in the computational basis, which we may assume d i=1 without loss of generality is a block diagonal basis for one copy of E. Based on equation (13), let dl denote the dimension of the lth diagonal block of 1ˆ ⊗ E(ρ), and ml denote the dimension of SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 17 the traced subsystem in that block. Then Φ = (1 − ζ)E + ζΦ˜ for some unital channel Φ such that Φ˜E = EΦ˜ = E whenever 2 1 ≥ ζ ≥ max ml dkΦ(|ψihψ|) − E(|ψihψ|)k∞/dl . l If E(ρ) = 1ˆ/d for all ρ, then we may replace the condition on ζ by

1 ≥ ζ ≥ d sup kΦ(|ηihη|) − E(|ηihη|)k∞ , η where the supremum is over normalized densities.

Proof. For k ∈ 1...d and a d×d matrix X, let λk(X) denote X’s kth eigenvalue in non-increasing order. Let ρ = |ψihψ|. Via equation 13, we estimate λ1(E(ρ)) and λd(E(ρ)) of E(ρ). Let us consider E to act on the outer-indexed subsystem of the Bell pair. We may decompose E = CtrEbl, where Ebl(ρ) = ⊕lPlρPl. Each block in Ebl(ρ) is then itself a maximally entangled state, times dl/d. We then apply a partial trace of an ml-dimensional subsystem from each lth block, which because of the maximal 2 2 entanglement, yields ml distinct eigenvalues of magnitude each ml dl/d. To find λ1(E(ρ)) and λd(E(ρ)), we maximize and minimize over l. We recall Weyl’s inequality for Hermitian matrices [HJ91]. For a pair of Hermitian matrices X,Y , λd(Y ) ≤ λk(X + Y ) − λk(X) ≤ λ1(Y ) . Via Weyl’s inequality, for each value of k,

λk(E(ρ) − (1 − ζ)E(ρ)) − λk(Φ(ρ) − (1 − ζ)E(ρ)) ≤ λ1(E(ρ) − Φ(ρ)) . The right hand side is observed to be always positive and upper-bounded by the ∞-norm distance. Re-arranging and combining terms,

λk(Φ(ρ) − (1 − ζ)E(ρ)) ≥ ζλk(E(ρ)) − kΦ(ρ) − E(ρ)k∞ . Via Choi’s theorem on completely positive maps,Φ − (1 − ζ)E is then a completely positive map. Since both Φ and E are trace-preserving, tr((Φ(ρ) − (1 − ζ)E(ρ))(X)) = ζtr(X) for any matrix X. Hence Φ˜ = (Φ − (1 − ζ)E)/ζ is a quantum channel. By linearity and the assumption that EΦ = ΦE = E, 1 EΦ˜ = Φ˜E = (E − (1 − ζ)E) = E . ζ We may then rewrite Φ = (1 − ζ)E + ζΦ˜ as a convex combination of channels. To see that Φ˜ is unital, if Φ(˜ 1)ˆ 6= 1,ˆ then Φ(1)ˆ cannot be 1,ˆ contradicting the assumption that Φ is unital. When E(η) = 1ˆ/d for any η, we can simplify the above argument by noting that Φ(η) and 1ˆ/d are always simultaneously diagonal. Whenever (1−ζ) ≤ dλk(Φ(η)) for all k, Φ(η)−(1−ζ)1ˆ/d ≥ 0. 18 NICHOLAS LARACUENTE

Via the triangle inequality, we have the condition that 1/d − kΦ(η) − E(η)k∞ − (1 − ζ)/d ≥ 0. Combining terms and re-arranging yields the final part of the Proposition.

Even though Proposition 2.11 depends on the dimension, this is the minimal dimension on which the conditional expectation may act, not including any extra subsystems. Hence the same bound applies for E ⊗ 1ˆB, Φ ⊗ 1ˆB independently of B’s dimension. Proposition 2.11 is similar to Li Gao’s [BJL+21, Lemma VI.8].

3. Combining Conditional Expectations We start this section by recalling basic facts about conditional expectations. A doubly- stochastic conditional expectation EN is a projector to matrices in algebra N that is self-adjoint ∗ under the trace. For this reason, we do not distinguish between E and EN or EN ∗, which respec- tively denote the dual and predual of EN . For a subalgebra N , the doubly-stochastic conditional expectation EN is unique. We also consider weighted conditional expectations with block diago- nal decompositions of the form

Ai Ei EN ,σ∗(ρ) = ⊕i(ρ ⊗ σi ) (14) A for a density σ = ⊕i(1ˆ i /|Ai| ⊗ σi). EN ,σ and EN ,σ∗ are respective adjoints under the trace (when a finite trace exists). EN ,σ,∗ is self-adjoint with respect to the GNS inner product given by hX,Y i = tr(σX†Y ). Letσ ˜ be the unnormalized density σ such that in finite dimension,σ ˜ = dσ. It then holds for a density ρ that 1/2 1/2 EN ,σ∗(ρ) =σ ˜ EN (ρ)˜σ and for an operator X that 1/2 1/2 EN ,σ(X) = EN (˜σ Xσ˜ ) .

Via its block diagonal form, we see that EN ,σ∗ is idempotent, as is its adjoint. Finally, EN = EN ,τ,∗ = EN ,τ , where τ is the normalized identity or trace. For convenience of notation, we will sometimes write Eσ and Eσ∗ for a conditional expectation and its predual if the subalgebra is not explicitly referenced. As with the previous section, Lemmas 3.2 and 3.3 do not require finite dimension. Hence Theorem 1.8 and Corollary 1.9 hold in infinite-dimensional settings, as long as their stated assumptions are met, and the corresponding integral form and properties of relative entropy hold.

Lemma 3.1 (Chain Rule). Let omega be a density and Eσ be a conditional expectation such that Eσ∗(ω) = Φ(ω). Then for any density ρ,

D(ρkω) = D(ρkEσ∗(ρ)) + D(Eσ∗(ρ)kω)

This equality is well-known. Nonetheless, we include a simple proof in the tracial case. SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 19

1/2 1/2 Proof. Let E denote the doubly stochastic conditional expectation such that Eσ(X) = E(σ Xσ ) for an operator X.

D(ρkEσ∗(ρ)) + D(Eσ∗(ρ)kω) = tr(ρ log ρ − ρ log(Eσ∗(ρ)) + Eσ∗(ρ) log(Eσ∗(ρ)) − Eσ∗(ρ) log ω)

= tr(ρ log ρ − ρ log(Eσ∗(ρ)) + ρEσ(log(Eσ∗(ρ))) − ρEσ(log ω)) . (15) Examining the logarithm of the block diagonal form in Equation (14),

Ei Ai log Eσ∗(ρ) = ⊕i(log(ρi) ⊗ 1ˆ + 1ˆ ⊗ log(σi)) , and 1/2 1/2 ˆEi ˆAi ˆEi Eσ(log Eσ∗(ρ)) = ⊕i(log(ρi) ⊗ 1 + 1 ⊗ 1 tr(σi log(σi)σi )) . Let 1/2 1/2 ˆAi ˆEi ηρ := Eσ(log Eσ∗(ρ)) − log Eσ∗(ρ) = ⊕i1 ⊗ 1 tr(σi log(σi)σi )) . Since ηρ has no dependence on ρ, we define ηω = ηρ analogously. Comparing to Equation (15),

D(ρkω) = D(ρkEσ∗(ρ)) + D(Eσ∗(ρ)kω) + tr(ρ(ηω − ηρ)) , and ηω − ηρ = 0.

Lemma 3.2. Let Eσ be a conditional expectation and ρ, ω be densities. Then

D(ρkEσ∗(ρ)) + D(ρkω) ≥ D(ρkEσ∗(ω))

Proof. By data processing on the 2nd term,

D(ρkEσ∗(ρ)) + D(ρkω) ≥ D(ρkEσ∗(ρ)) + D(Eσ∗(ρ)kEσ∗(ω)) . By Lemma 3.1 and the idempotence of conditional expectations, we obtain the Lemma.

Lemma 3.3. Let {(Nj, Ej): j = 1...J ∈ N} be a set of J ∈ N von Neumann algebras and asso- ciated conditional expectations. Let E be the conditional expectation to the intersection algebra, (j) J N = ∩jNj. Let (σ )j=1 be weighting densities such that for a density σ, Ej,σ∗Eσ∗ = Eσ∗Ej,σ∗ = Eσ∗ for each j ∈ 1...J. ⊗m s Let S = ∪m∈N{1...J} be the set of sequences of indices. For any s ∈ S, let E denote the composition of conditional expectations E (j ) ...E (j ) for s = (j , ..., j ). Let µ : S → [0, 1] j1,σ 1 ∗ jm,σ m ∗ 1 m be a probability measure on S and kj,s upper bound the number of times Ej,σ(j)∗ appears in each sequence s. If for all input densities ρ, X s µ(s)E (ρ) = (1 − )Eσ∗(ρ) + Φ(ρ) s∈S for some channel Φ, then J X X µ(s) ks,jD(ρkEj(ρ)) ≥ D(ρk(1 − )Eσ∗(ρ) + Φ(ρ)) . s∈S j=1 20 NICHOLAS LARACUENTE

Proof. For each s, we apply Lemma 3.2 iteratively, finding that for each s ∈ S, J X s ks,jD(ρkEj,σ(j) (ρ)) ≥ D(ρkE (ρ)) . j=1 By convexity, we move the weighted average over s inside the relative entropy, completing the Lemma. Proof of Theorem 1.8. We have by the assumptions of the Theorem and Lemma 3.3 that J X X µ(s) ks,jD(ρkEj(ρ)) ≥ D(ρk(1 − ζ)Eσ∗(ρ) + ζΦ(ρ)) s∈S j=1 for some Φ, k, and ζ. We then note that since

((1 − ζ)Eσ∗ + ζΦ)Eσ∗ = (1 − ζ)Eσ∗ + ζΦEσ = Eσ∗ , it must hold that Eσ∗Φ = Eσ∗, and similarly, ΦEσ∗ = Eσ∗. By data processing, D(ρkEσ∗(ρ)) ≥ D(Φ(ρ)kEσ∗(ρ)). We then use Theorem 1.5 or Theorem 1.3.

Proof of Corollary 1.9. First, E projects to a fixed point subspace of {Ej} for all j, so for any Q sequence s of k-many j values, E also projects to a fixed point subspace of j=1...k Ej. For all densities ρ, it must then hold that EΦ(ρ) = ΦE(ρ) = E(ρ), so Φ also has N as a fixed point subalgebra. For convenience of notation, we denote C := C(M : N ), and Ccb := Ccb(M : N ). It follows from the definitions of C and Ccb that (1)Φ( ρ) ≤ CE(Φ(ρ)) = CE(ρ) for all ρ. B B B B AB (2)Φ ⊗ 1ˆ (ρ) ≤ CcbE ⊗ 1ˆ (Φ ⊗ 1ˆ (ρ)) = CcbE ⊗ 1ˆ (ρ) for ρ . For any κ ∈ N, X κ κ κ kjD(ρkEj(ρ)) ≥ D(ρk(1 − ζ )E(ρ) + ζ Φκ(ρ)) j for some Φκ for which ΦκE = EΦκ = E. Via data processing and since EEj = EjE = E for any j ∈ 1...J, taking a non-integer value of κ won’t hurt but may result in an effective exponent of bκc. To achieve a desired , we may take κ = dlogζ e, completing the Corollary for doubly stochastic conditional expectations using Lemma 3.3. For weighted conditional expectations, the block diagonal form in Equation (14) still yields an upper bound on c such that ρ ≤ cEσ∗(ρ) for all ρ. Whenever ΦEσ∗ = Eσ∗Φ = Eσ∗, the same value of c suffices for Φ(ρ). Proof of Remark 1.10. Proposition 2.11 and [JX07] suffice to show that a quasi-factorization inequality with some constant holds. Corollary 1.6 shows a correction term scaling as < O(c2ζ) for small ζ. Hence as ζ → 0, we expect that this constant approaches 1 and achieves this with kdlogζ e = 1. For the last part of the Remark, recall that

D(ρ|E(ρ)) = D(ρkEj(ρ)) + D(Ej(ρ)kE(ρ)) . SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 21

Using a single term of D(ρkEj(ρ)), we cancel the first term in the expansion. Then we observe that if we can use a convex combination of sequences of conditional expectations to create D(ρk(1−ζ)E(ρ)+ζΦ(ρ)), we may consume another D(ρkEj(ρ)) term to ensure that each sequence starts with Ej, and then apply data processing, obtaining a term of the form D(Ej(ρ)k(1−ζ)E(ρ)+ ζEjΦEj(ρ)), which satisfies the conditions of the Theorem with the same value of ζ.

The caveat to Corollary 1.10 is that as ζ → 0, E does not change. For example, if we take two incompatible measurement bases at arbitrarily small angle to each other, they will converge to the same basis, though for any finite angle E will be a trace of the subsystem.

4. Example Applications Well-known uses of quasi-factorization, powered by Proposition 1.11 and similar, include strengthening quantum uncertainty principles and estimating decay rates to thermal equilib- rium in many-body systems [BCR20, CRF20]. MLSI and similar estimates also bound deco- herence times and quantum capacities [BJL+21, GJL20b]. Taking this idea a step further, quasi-factorization gives strong bounds on certain combinations of quantum resources, such as the relative entropy of coherence in incompatible bases [WY16] or the asymmetry with respect to overlapping symmetry groups [VAWJ08, GMS09, MS14]. Similarly, combining MLSI is a power- ful way to estimate decoherence or resource decay for systems undergoing several noise processes simultaneously. In Subsection 4.1 we show how the improved quasi-factorization yields strong, asymptotically tight, uncertainty-like bounds for mixtures.

4.1. Uncertainty Relations. Let S and T correspond to bases in dimension d, not necessarily mutually unbiased. Let {|iSi : i = 1...d} and {|iT i : i = 1...d} index these bases. We compare α-(C)SQF with a conventional Maassen-Uffink uncertainty relation [MU88] of the form 2 D(ρkES (ρ)) + D(ρkET (ρ)) ≥ D(ρk1ˆ/d) − log(d max | hiS|jT i | ) , (16) i,j which may be extended by an auxiliary system [BCC+10]. When ρ ≈ 1ˆ/d and/or when there is high overlap between bases, the right hand side of the conventional uncertainty relation becomes negative, and the bound becomes trivial. In contrast, α-(C)SQF still gives a positive, non-trivial bound on the sum of basis entropies. We also recall the result of [BCR20], 2 D(ρkES (ρ)) + D(ρkET (ρ)) ≥ (1 − 2d max | | hiS|jT i | − 1/d|)D(ρk1ˆ/d) , (17) i,j which gives a stronger bound for bases that are close to mutually unbiased and is asymptotically tight, but it becomes trivial for highly overlapping bases. The quasi-factorization from [GR21b] also shows an uncertainty-like bound with extendibility by an auxiliary system. 2 Let ξ = mini,j | hiS|jT i | . Note that ξ ∈ [0, 1/d]. As a pinch to either basis is completely dephasing in that basis, a subsequent pinch in the other leaves any input state in a classical mixture with probability no less than dξ. Hence ES ET (ρ) = (dξ)1ˆ/d + (1 − dξ)Φ(ρ) for some 22 NICHOLAS LARACUENTE unital Φ. For bipartite ρAB, we may write d X 1ˆ X E E (ρAB) = θ |jihj| ⊗ ρB, E(ρ) = ⊗ θ ρB S T j j d j j j=1 j d AB AB in the S basis with coefficients {0 ≤ θj ≤ 1}j=1. We then see that (dξ)E(ρ ) ≤ ES ET (ρ ) ≤ dE(ρAB), though the latter inequality is better replaced by Φ(ρ) ≤ d2E(ρ) via Corollary 1.9. Letting ζ = 1 − dξ, Theorem 1.8 implies that

βd2, B D(ρkE˜S (ρ)) + D(ρkE˜T (ρ)) ≥ D(ρk1ˆ/d ⊗ ρ ) (18) dlogζ e AB B for any bipartite density ρ with A dimension |A| = d, where E˜ = E ⊗ 1ˆ , and ES , ET are defined analogously. Expanding the relative entropies, this is equivalent to

βd2, H(S ⊗ B)ρ + H(T ⊗ B)ρ ≥ 2H(ρ) + (ln d + H(B)ρ − H(ρ)) , dlogζ e where H(S⊗B)ρ,H(T ⊗B)ρ, and H(B)ρ are defined respectively as the entropies of the outputs of ES , ET , and trA on ρ. When |B| = 1, we may replace βd2, by βd,. Alternatively, using Corollary 2.4 and Theorem 1.8, we may replace βd2, by 32(d + 1)ζ β˜ = 1 − 15 + (7d − 24)ζ as long as ζ ≤ 15/(25d + 46), and |B| = 1. When ζ is not this small, we would instead have D(ρk1ˆ/d) D(ρkES (ρ)) + D(ρkET (ρ)) ≥ . (19) 2dlogζ (15/(57d + 88))e As the two bases approach mutual unbias, ζ approaches 0, and both forms of the inequality approach the entropic uncertainty relation implied by strong subadditivity (see Petz’s version as theorem 1.1). When the two bases approach the same basis, ξ approaches 0. In this regime, 1− dξ is close to

1, so to achieve a sufficiently small  for the inequality to be non-trivial, log1−dξ  must be large. Unlike equation (17), Theorem 1.8 does not become completely trivial until the bases become the same basis. At that point, the intersection conditional expectation suddenly ceases to be that to the B subalgebra, instead becoming the conditional expectation to the basis algebra. As noted in Corollary 1.10, the bases approach a different intersection algebra. SSA for two of the same bases reduces to the trivial statement that 2D(ρkES (ρ)) ≥ D(ρkES (ρ)). Meanwhile, when ξ << 1/d but is still finite, quasi-factorization still compares to the B subalgebra. Hence we should not expect any sort of continuous approach.

4.2. Finite Groups and Transference. A common form of conditional expectation is Z EG = adudg , (20) u∈π(G) SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 23 where dg is taken over the Haar or invariant measure on group G, π(G) is a unitary representation † in some Hilbert space, and adu(X) := uXu for any unitary u and matrix X. Groups may induce collective channels as considered in [BJL+21]. We build on a simplified version of some ideas from that paper and [GJL20a]. When a quantum channel is a convex combination of unitaries, we may rewrite the channel as taking a hidden input that weights these unitaries. Let Φ be a quantum channel of the form n X Φ = piadu(i) i=1 for some probability vector ~p, where u(i) is the ith unitary in a family of n unitaries. Let Ψ(~p) be given by n X † Ψ(~p)(ρ) = piu(i)ρu(i) (21) u=1 for any probability vector ~p. Here Ψ takes as input a probability vector and outputs a quantum channel. The key insight for quasi-factorization of finite groups (and graphs as considered in Subsection 0 4.3) is that for any pair of probability distributions ~p,~q ∈ l1(G), Ψ(~p) ◦ Ψ(~q) = Ψ(Ψ (~p)(~q)). Here G is assumed to be a finite group, and l1(G) is the 1-normed vector space on group elements. 0 Ψ (~p): l1(G) → l1(G) is a channel on probability distributions, acting as 0 X Ψ (~q)(~p) = qgg(~p) , (22) g∈G where g(·) denotes the action of g on itself, promoted to its action on probability distributions in l1(G), and qg denotes the probability of group element g as given by ~q. When Ψ0(~p) = (1 − ζ)~1/|G| + ζΨ˜ 0 for some channel Ψ˜ 0 and ζ ∈ (0, 1), it also holds that

Ψ(~p) = (1 − ζ)EG + ζΨ˜ for some quantum channel Ψ.˜ In the classical probability setting, it is often much easier to prove estimates analogous to those in Proposition 2.11. Also, the convex combination form of equation (21) allows us to upper bound the constant c in Theorem 1.5 by |G|. J A nice case of group quasi-factorization involves subgroups G1, ..., GJ ⊆ G. If ∪j=1Gj contains generators of the entire group, then we may conclude that at least some chain of conditional J expectations from the set {Gj}j=1 of length |G| or shorter will include EG in convex combination. (C)SQF follows with good constants depending on the specific structure of the group. The highlighted example appears in the next section, where we consider transference analogs on conditional expectations and semigroups derived from finite graphs.

Example 4.1 (Symmetric Group of Degree 3). The group of permutations on 3 indices, known both as the symmetric group of degree 3 (S3) and as the dihedral group of degree 3, is the smallest non-abelian group. It contains 6 elements, which we may naturally associate to permutations of ⊗6 6 qudit subsystems on the tensor product Hilbert space Hd for any d ∈ N. We may describe 24 NICHOLAS LARACUENTE convex combinations of unitaries from this group by 6 X Φ~s = sjSj , j=1 where ~s = (s1, ..., s6) is a probability vector, and Sj is a unitary conjugation representing the corresponding element of S3. When s1 = ... = s6 = 1/6, the channel becomes a conditional expectation to the invariant subalgebra of the permutation group. Any channel Φ~s with non- k zero amplitude on two non-redundant generators of S3 will have that Φ → ES3 as k → ∞. By Corollary 1.6, when Φ~s contains non-zero weights on at least two non-redundant generators, k D(ρkΦ~s (ρ)) ≥ βc,minj {s˜j }D(ρkES3 (ρ)) , k wheres ˜j corresponds to the weighting vector of Φ~s , and we know that c ≤ 6 via subalgebra 0 6 6 index. As described in this section, we may construct the channel Ψ (~s)(·): l1(S3) → l1(S3), 6 0 where l1(S3) is the 6-dimensional, 1-normed vector space on elements of S3, and Ψ is as in Equation 22. In the spirit of transference, we have k Φ~s = Φ(Ψ0(~s))k(~1) , where ~1 denotes the probability vector corresponding to the group’s identity element. Hence we may determine ζ and c precisely for any s by finding the distribution induced by k repeated applications of Ψ0(~s) as discrete classical Markov chain to a probability vector. This process may k allow us to bypass the potentially harder problem of calculating Φ~s for arbitrary quantum states with arbitrary extensions. This example is related to Example 4.7, in which using a finite graph model, we obtain asymptotic bounds for symmetric groups in terms of transpositions.

Convex combinations of unitaries are far from the only way to represent groups in Hilbert space, and other representations may admit analogous results. See [BJL+21] for a more compre- hensive treatment and formal notion of group transference.

4.3. Finite, Connected, Undirected Graphs. In this section, we will use the symbol G = (V,E) to denote a graph with vertex set V and edge set E ⊆ V × V . We reserve G to denote a group. The group and graph scenarios relate closely. An undirected, n-vertex graph G will have a naturally corresponding group G with action on 1...n in which each edge (v1, v2) ∈ V × V corresponds to the swap operation v1 ↔ v2. This association is not unique - we may for instance identify a cyclic graph with a one-generator cyclic group, or with a multi-generator group of self-inverse swaps. Conversely, a finite group G has a corresponding Cayley graph. Most results herein will not be sensitive to all details of a graphs’s representation. For example, n given a basis {|ji}j=1, we may take an n-vertex graph G to represent a confusion graph with swaps |v1i ↔ |v2i. Alternatively, the graph may correspond to swaps between qudit subsystems. In [LJL20], it was shown that connected graphs generally imply CMLSI with constant of O(1/n2). This result arises by comparing all graphs to the broken cycle, the slowest-decaying SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 25 of connected graphs. Similarly, [GJL21, GR21a] showed general CMLSI, resolving at least the existence of CMLSI for graphs. In [GJL20a], complete graphs including two-vertex, single-edge graphs were shown to have CMLSI with constant of O(1), as the channels given by these graphs are already convex combinations that include EG. What’s missing so far is an explicit calculation of the CMLSI constant for expanders and similar graphs, which is expected to be better than Ω(1/n2) but worse than O(1). For an n-vertex, undirected graph G, let u(i, j) denote an operation corresponding to edge (i, j), which can be formalized as a group representation (see Section 4.2). If i = j, then let u(i, j) = 1.ˆ Let 1 E = (1ˆ + ad ) i,j 2 u(i,j) denote the corresponding conditional expectation given by unitary conjugation, and 2  1 X  E = 1ˆ + ad G n2 − n + 2 2 u(i,j) i6=j∈1...n denote the conditional expectation to the invariant subspace of the graph. We say that an undirected graph G is m-regular when each vertex has m incoming/outgoing edges.

1 1 1 2 13 2 8 2 12 3

11 4 7 3 10 5

9 6 4 3 6 4 78

5

Figure 2. Visualizations of a complete graph, Paley graph [Epp21] (a member of a family that is Ramanujan for sufficiently large vertex number), and cyclic graph.

We denote by A the normalized adjacency matrix of a graph G, which for an m-regular graph is the adjacency matrix divided by m. We recall an important result on expander graphs: Theorem 4.2 (theorem 3.3 from [HLW06]). Let G be a connected, undirected, m-regular graph with n vertices. Let the eigenvalues of G’s normalized adjacency matrix A be denoted 1 = λ1 ≥ n ... ≥ λn, and γ = max{|λn|, |λ2|}. Then for any normalized probability vector ~x ∈ l1 and t ∈ N, t t t kA ~x − ~1/nk2 ≤ k~x − ~1/nk2γ ≤ γ . n n Remark 4.3. Let Φ: l1 → l1 be a classical channel on the probabilities over graph vertices. If 1 ≥ ζ ≥ nkΦ(~x) − ~1/nk∞ for every vector ~x, we may rewrite Φ(~x) = (1 − ζ)~1/n + ζΦ(˜ ~x) for which Φ(˜ x) is given by Φ(˜ ~x) ≥ Φ(~x) − (1 − ζ)1ˆ/n 26 NICHOLAS LARACUENTE and is a (classical) channel. Is it easy to check by taking At~i for each basis vector ~i ∈ 1...n that all entries of At must have value at least (1 − ζ). This follows the same line of argument as Proposition 2.11. As noted in Section 4.2, we may transfer this convex combination form to a quantum channel Φ. Hence we find: Theorem 4.4. Assume an m-regular graph with n vertices has γ as defined in theorem 4.2. Then for any k, t ∈ N such that dlogγ(1/n)e ≤ t ≤ (m − 1)k/2, k X
D(ρkEi,j(ρ)) ≥ 1 − exp(−(m − 1)nkD(t/((m − 1)nk)k1/2n)) D(ρkEG(ρ)) , βc,ζ (i,j)∈E where D(pkq) is the relative entropy of a p-coin to a q-coin, and c ≤ (n2 − n)/2 + 1. Hence the P † Lindbladian given by L = (i,j)∈E(ρ − u(i, j)ρu(i, j) ) has CLSI with the same constant. Proof. Consider the channel 1 X 1 1 X Φ (ρ) = E = 1ˆ + ad (23) G |E| i,j 2 2|E| u(i,j) i,j i,j on m-regular graphs with m ≥ 2 and |E| denoting the number of edges. ΦG is a convex com- bination, applying with 1/2 probability the identity, and with 1/2 probability a random u(i, j). n For a single basis vector ~ei ∈ l1 ,ΦG has a probability 1/2n to apply a unitary that changes the value of i. Hence we may think of it as applying one step in a random walk with probability 1/2n. ˜ Let k be the number of applications of ΦG, and t the actual number of steps in a particular random walk. Via Lemma 3.2 and the convexity of relative entropy, 1 X D(ρkE (ρ)) + D(ρkΨ(ρ)) ≥ D(ρkΦ Ψ(ρ)) |E| i,j G (i,j)∈E for any channel Ψ from the input space to itself. Via iteration starting with Ψ = 1ˆ and for any k ∈ N, 1 X |E|k k D(ρkE (ρ)) ≥ D(ρkΦ (ρ)) , |E| i,j G (i,j)∈E so we may set k = |E|k˜. By the Chernoff bound, p(t ≤ νk˜) ≤ exp(−kD˜ (νk1/2n)) . This bound is non-trivial when ν < 1/2n. Via convexity of relative entropy, and using the fact that (m − 1)n = |E|, X |E|kD(t/((m−1)nk)k1/2n))
t k D(ρkEi,j(ρ)) ≥ 1 − e D(ρkΦG(ρ)) . (24) (i,j)∈E for any k ∈ N and t < k. Equation (24) holds for all finite graphs. We will now assume that G’s spectrum has γ = max{|λn|, |λ2|}. Using 4.2 and the fact that the 2-norm is an upper bound for SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 27 the ∞-norm, t sup k~x − ~1/nk∞ ≤ γ , ~x which implies via Remark 4.3 that t t t ˜ D(ρkΦG(ρ)) = D(ρk(1 − nγ )EG(ρ) + nγ Φ(ρ)) for any input density ρ and some Φ˜ such that EΦ˜ = Φ˜E = E. Instead of Theorem 1.8, we directly apply Theorem 1.5. Doing so completes the quasi-factorization part of the proof. Finally, we apply Proposition 1.11 knowing that Ei,j has CMLSI with constant 1.

Theorem 4.4 is the technical underpinning of Theorem 1.13 in the introduction. For its sim- plified version, we choose k = κt/(m − 1) for some constant κ > 1. We estimate D(1/κnk1/2n) =(1/κn)(ln(1/κn) − ln(1/2n)) + (1 − 1/κn)(ln(1 − 1/κn) − ln(1 − 1/2n)) ≥ − ln(2/κ)/κn + (1 − 1/κn)(1/2n − 1/(κn − 1)) . Since n ≥ 3 for the Theorem to be non-trivial, we may choose κ ≥ 5 and find κn − 1 ≥ 4n, so ln(5/2) 7 D(1/5nk1/2n) ≥ + . 5n 60n Hence |E|kD(t/((m − 1)nk)k1/2n)) ≥ (m − 1)k(ln(5/2)/5 + 7/60), which is always less than 1 and scales linearly with k. Hence we have k X D(ρkE (ρ)) β(c, γ(m−1)k/5) i,j (i,j)∈E (25)
≥ 1 − exp(−(m − 1)k(ln(5/2)/5 + 7/60) D(ρkEG(ρ)) .

Finally, we chose k ∼ 2 logγ(1/n). As n → ∞, we see that β → 1, and the subtracted, exponential correction on the right hand side approaches 0 inverse polynomially in n. Theorem 4.4 is especially powerful when γ = O(1) in n, in which case it obtains the expected logarithmic mixing time for Ramanujan and similar graphs. Graphs with γ = O(1) are commonly known as fast expanders. These graphs have the fastest mixing times possible with fixed degree. In contrast to gaps with constant γ is the cycle, which is 2-regular and has the slowest mixing time up to constants of any connected graph. In [LJL20], it was shown that Lindbladians corresponding to cyclic graphs having CMLSI with constant O(1/n2). Based on classical mixing times, we expect this dependence to be optimal. It is well-known that the eigenvalues of a cycle’s normalized adjacency matrix take values cos(2πl/n) for l ∈ 0...n − 1, so γ = cos(2π/n) ≈ 2 2 2 2 4 1 − 4π /n . Hence logγ  ≈ n ln /4π. If we take  < c ≤ O(n ), we find a quasi-factorization α of O(n2 ln n) and CMLSI constant of O(1/n2 ln n). This decay rate is suboptimal, containing an additional ln n factor. This apparent problem leads to the following example:

Example 4.5 (Cyclic Graph). The trick to obtain an asymptotically optimal CMLSI constant through quasi-factorization is a more fine-grained analysis of the channels, which will allow a smaller c than is given by the index. We may compare the mixing on a cyclic graph to a random 28 NICHOLAS LARACUENTE walk. Given a walker starting at an arbitrary vertex, its probability of remaining at the same 2 2 vertex after an steps on an infinite cycle for any a > 0 st. an ∈ N would be given by a binomial distribution with approximation r  2  2 r 2 2 1 an 1an 2 2 1 e−1/3an < < e1/12an (26) aπ n an2/2 2 aπ n via Robbins’s precise form of Stirling’s approximation [Rob55]. Again using the refinements to Stirling’s approximation, for s ≤ an2, r  2  2 2 2 2 1 an 1an e−(4s +1/2)/an < . (27) eπ n an2/2 + s 2 Equation (27) lower bounds the probability for a random walker to land 2s steps to the left or right of its starting point. For a cycle, this probability would be minimal near s = n/2, assuming a fully concentrated walker must reach its farthest point. Ignoring that the walker could loop around to the farthest point from two directions, r 2 2 1 p ≥ e−(g+1/2n )/a , min πa n where g = 1. For a 1D line without periodicity, we would have g = 4. In either case, the value is Ω(1/n). We also must upper bound the maximum probability, which for these graphs (and in general) is given by the probability of remaining at one’s starting position. The binomial distribution assumes open boundaries, not considering the probability that the walker loops around and returns to an earlier position. We must count the contributions of an2/2 + rn for many values of r ∈ Z. For this, we use Hoeffding’s inequality [Hoe63], 2 − 2 2 p(|s| ≥ |r|n) ≤ 2 exp(−2an (+r/an) ) = 2 exp(−2r /a) , (28) where p(|s| ≥ |r|n) is the probability of taking a net s steps in either direction. Since probability declines with net number of steps away from the origin, the probability of |r| = 1 is no more than 1/n. Since the probability to get past this point is in total less than 2 exp(−2/a), the probability + of r = −2 is no more than 2 exp(−2/a)/n, and so on. r ∞ 1  2 2 X 2  p ≤ e1/12an + 2 e−2r /a . max n aπ r=0

We note that pmin(a, n) ≥ Ω(1/n), while pmax(a, n) = O(1/n). We will use these to compare powers of ΦG as defined in Equation (23) with E. We compute that for a = 8, pmax ≤ 5/n, and pmin ≥ 1/8n, though precise numerics could probably find tighter bounds. Since the number of edges in an n-cycle is equal to n, we find that for any b ∈ N, 2 X bn3 bn D(ρkEi,j(ρ)) ≥ D(ρkΦG (ρ)) (i,j)∈E

bn3 with ΦG given by Equation (23). Via the Chernoff bound or Hoeffding’s inequality, ΦG applies at least ban2c steps in a random walk for any a < b/2 with probability approaching 1 as n → ∞. SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 29

Hence for any b > 16 and sufficiently large n, 2 X bn D(ρkEi,j(ρ)) ≥ D(ρk(1 − ζ)EG(ρ) + ζΦ(ρ)) (i,j)∈E with Φ(ρ) ≤ cE(ρ) for ζ = 7/8 as determined by pmin, and c ≤ 6 as determined by pmin and pmax. Because Φ is a convex combination of unitaries and cannot concentrate probability, we k deduce using Remark 4.3 that Φ (ρ) ≤ cE(ρ) for any k ∈ N. We may then increase b to reduce ζ further. Since c = O(1), b need not depend on n, and the quasi-factorization constant is O(n2) and CMLSI constant O(1/n2) in line with expectations.

Remark 4.6. In Example 4.5, we use a combination of pmin and pmax to simultaneously bound the constants ζ and c appearing in Corollary 1.6. The O(n2 ln n) bound follows from bounding pmin but using the trivial bound that pmax = O(1) in n, which obtains a good value for ζ but over- estimates c. A saturating distribution would have a single peak of O(1) magnitude, redistributing the rest of the probability equally. As the probability distributions of graph random walks are well-studied, the literature tells us 2 that pmax is in truth much more strongly bounded to be O(1/n) after O(n ) steps. The Example shows that while we can obtain a bound by controlling pmin alone in the finite index setting, we sometimes obtain a stronger result by explicitly controlling both directions of channel comparison. Another graph that is very unlike the expander or cycle is the complete graph, which is n-regular. CMLSI for complete graphs was studied in [GJL20a]. For a complete graph, 1 X 1 1  1 1  E = 1 − 1ˆ + 1 + E . |E| i,j 2 |E| 2 |E| G (i,j)∈E Hence X  1  D(ρkE (ρ)) ≥ |E|D ρ ((|E| − 1)1ˆ + (|E| + 1)E )(ρ) i,j 2|E| G i,j ≥ O(n)D(ρk((1 − O(2−n))E + O(2−n)1)(ˆ ρ)) . We find a quasi-factorization constant of O(1/n). That the dependence is less than o(1) in n arises because we have not normalized by the degree - were we to construct such a jump process, the probability of making any jump in an infinitesimal time interval would grow proportionally to n. If we normalize with 1/m = 1/n, this leads to an O(1) CMLSI constant. This is entirely unsurprising, as the Lindbladian constructed from a complete graph (again up to normalizing factors) already generates a convex combination including EG. Hence earlier notions and methods [Bar17, BR18, GJL20a] already suffice to show CMLSI for complete graphs. Finally, we note again that the self-inverse u(i, j) are not the only way to construct a graph Lindbladian. For example, [LJL20] uses a single unitary u such that un = 1ˆ to generate the edge transitions of a cyclic group, corresponding to a cyclic graph. In many of these cases, we can still use theorem 4.2 and the same methods of analysis as in Theorem 4.4 to similar effect, and via the classical Laplacian construction of Remark 1.14, we see that in many cases the results are comparable. 30 NICHOLAS LARACUENTE

Example 4.7. We recall the random transposition (RT) graph of [LY98]. Formally, this graph applies to a space of size n! if n ∈ N is the number of distinguishable particles on n distinct sites. n ⊗n We may however map any configuration to an equivalent configuration on (l1 ) , the n-fold tensor product of n-dimensional probability spaces. The latter space is actually larger, and the image of the former space is restricted to convex sums of configurations for which all particles are indeed distinct, which is no longer enforced by the structure of the space. Nevertheless, this mapping allows us to analyze this model by setting u(i, j) to be the swap unitary for positions i, j ∈ 1...n. With n positions, we may then apply the n-vertex graph results shown in this section. If we restrict to nearest neighbor swaps, then we find an O(1/n2) CMLSI constant, whereas with a complete or Ramanujan graph, we obtain the respective O(1) or O(1/ ln n)-CMLSI compatible with the mixing times of classical graphs attributed to Diaconis and Saloff-Coste [DSC96]. This example is related to Example 4.1.

5. Conclusions and Outlook The underpinnings of this paper’s results combine the entropy-geometry links from [CM17, GJL20a] that have continued to [LJL20, GJL21], functional calculus like in [GR21a], and itera- tive, computation-like techniques as detailed in sections 2.1, 3, and 4.3. The combination of non-linearity of relative entropy and off-diagonal elements of quantum densities makes deriving purely multiplicative relative entropy inequalities extremely challenging. Recently, several works including [BCR20, BCPH21, GR21a] and this paper have begun to crack the problem. Roughly, when the relative entropies involved in an inequality are large, additive constants are not necessarily worse than multiplicative constants. There is generally, however, a small relative entropy regime in which additive inequalities become trivial. The purely multiplicative aspect is especially valuable in proving MLSI and similar bounds, where one must otherwise go to great lengths to remove additive constants. It remains open what the best possible quasi-factorization constants would be, and whether there is a simple, closed form for general sets of finite-dimensional conditional expectations. In [GR21a], a two-sided bound is shown in terms of an L2 → L2 norm difference between a composition of conditional expectations and their intersection, but the bound is not necessarily tight on either side. A still open question is to what extend quasi-factorization should hold for infinite dimensions. Proposition 2.11 breaks down if the minimum dimension of a conditional expectation E is infinite. If we assume a convex combination form or matrix order inequality, then Theorem 1.5 should still hold in some circumstances, though it remains to check that assumed and cited bounds hold with no finite-dimensional or tracial assumptions.

6. Data Availability Statement Data sharing is not applicable to this article as no new data were created or analyzed in this study. SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 31

7. Acknowledgments I thank Marius Junge for early feedback on these results. I also acknowledge interactions with Li Gao as helping to motivate this project and with Angela´ Capel as inspiring some improvements to the final version.

References [AE05] Koenraad M. R. Audenaert and Jens Eisert, Continuity bounds on the quantum relative entropy, Journal of Mathematical Physics 46 (2005), no. 10, 102104, Publisher: American Institute of Physics. [AE11] , Continuity bounds on the quantum relative entropy — II, Journal of Mathematical Physics 52 (2011), no. 11, 112201. [Aud14] Koenraad M. R. Audenaert, Telescopic Relative Entropy, Theory of Quantum Computation, Commu- nication, and Cryptography (Berlin, Heidelberg) (Dave Bacon, Miguel Martin-Delgado, and Martin Roetteler, eds.), Lecture Notes in Computer Science, Springer, 2014, pp. 39–52 (en). [Bar17] Ivan Bardet, Estimating the decoherence time using non-commutative Functional Inequalities, arXiv:1710.01039 [math-ph, physics:quant-ph] (2017), arXiv: 1710.01039. [BCC+10] Mario Berta, Matthias Christandl, Roger Colbeck, Joseph M. Renes, and Renato Renner, The uncer- tainty principle in the presence of quantum memory, Nature Physics 6 (2010), no. 9, 659 (En). [BCL+19] Ivan Bardet, Angela Capel, Angelo Lucia, David P´erez-Garc´ıa,and Cambyse Rouz´e, On the modified logarithmic Sobolev inequality for the heat-bath dynamics for 1D systems (en). [BCPH21] Andreas Bluhm, Angela´ Capel, and Antonio P´erez-Hern´andez, Weak quasi-factorization for the Belavkin-Staszewski relative entropy, arXiv:2101.10312 [math-ph, physics:quant-ph] (2021), arXiv: 2101.10312. [BCR20] Ivan Bardet, Angela Capel, and Cambyse Rouz´e, Approximate tensorization of the relative entropy for noncommuting conditional expectations (en). [BJL+21] I. Bardet, M. Junge, N. LaRacuente, C. Rouz´e,and D. S. Fran¸ca, Group transference techniques for the estimation of the decoherence times and capacities of quantum Markov semigroups., IEEE Transactions on Information Theory (2021), 1–1, Conference Name: IEEE Transactions on Information Theory. [BR18] Ivan Bardet and Cambyse Rouz´e, Hypercontractivity and logarithmic Sobolev Inequality for non- primitive quantum Markov semigroups and estimation of decoherence rates, arXiv:1803.05379 [math-ph, physics:quant-ph] (2018), arXiv: 1803.05379. [BT06] Sergey G. Bobkov and Prasad Tetali, Modified Logarithmic Sobolev Inequalities in Discrete Settings, Journal of Theoretical Probability 19 (2006), no. 2, 289–336 (en). [Ces01] Filippo Cesi, Quasi-factorization of the entropy and logarithmic Sobolev inequalities for Gibbs random fields, Probability Theory and Related Fields 120 (2001), no. 4, 569–584 (en). [CKPT14] Elena Caceres, Arnab Kundu, Juan F. Pedraza, and Walter Tangarife, Strong subadditivity, null energy condition and charged black holes, Journal of High Energy Physics 2014 (2014), no. 1, 84 (en). [CLPG18] Angela´ Capel, Angelo Lucia, and David P´erez-Garc´ıa, Quantum conditional relative entropy and quasi- factorization of the relative entropy, Journal of Physics A: Mathematical and Theoretical 51 (2018), no. 48, 484001 (en), Publisher: IOP Publishing. [CM17] Eric A. Carlen and Jan Maas, Gradient flow and entropy inequalities for quantum Markov semigroups with detailed balance, Journal of Functional Analysis 273 (2017), no. 5, 1810–1869. [CMT14] Pietro Caputo, Georg Menz, and Prasad Tetali, Approximate tensorization of entropy at high temper- ature (en). [CRF20] Angela´ Capel, Cambyse Rouz´e,and Daniel Stilck Fran¸ca, The modified logarithmic Sobolev inequality for quantum spin systems: classical and commuting nearest neighbour interactions, arXiv:2009.11817 [math-ph, physics:quant-ph] (2020), arXiv: 2009.11817. 32 NICHOLAS LARACUENTE

[CW04] Matthias Christandl and Andreas Winter, “Squashed entanglement”: An additive entanglement mea- sure, Journal of Mathematical Physics 45 (2004), no. 3, 829–840. [DSC96] P. Diaconis and L. Saloff-Coste, Logarithmic Sobolev inequalities for finite Markov chains, The Annals of Applied Probability 6 (1996), no. 3, 695–750, Publisher: Institute of Mathematical Statistics. [Epp21] David Eppstein, Paley graph, January 2021, Wikipedia Page Version ID: 1000144544. [GJL17] Li Gao, Marius Junge, and Nicholas LaRacuente, Unifying Entanglement with Uncertainty via Symme- tries of Observable Algebras, arXiv:1710.10038 [quant-ph] (2017), arXiv: 1710.10038. [GJL18a] , Capacity Estimates via Comparison with TRO Channels, Communications in Mathematical Physics 364 (2018), no. 1, 83–121 (en). [GJL18b] , Uncertainty Principle for Quantum Channels, 2018 IEEE International Symposium on Infor- mation Theory (ISIT), June 2018, pp. 996–1000. [GJL20a] , Fisher Information and Logarithmic Sobolev Inequality for Matrix-Valued Functions, Annales Henri Poincar´e 21 (2020), no. 11, 3409–3478 (en). [GJL20b] , Relative entropy for von Neumann subalgebras, International Journal of Mathematics 31 (2020), no. 06, 2050046, Publisher: World Scientific Publishing Co. [GJL21] Li Gao, Marius Junge, and Haojian Li, Geometric Approach Towards Complete Logarithmic Sobolev Inequalities, arXiv:2102.04434 [quant-ph] (2021), arXiv: 2102.04434. [GMS09] Gilad Gour, Iman Marvian, and Robert W. Spekkens, Measuring the quality of a quantum reference frame: The relative entropy of frameness, Physical Review A 80 (2009), no. 1, 012307. [GR21a] Li Gao and Cambyse Rouz´e, Complete entropic inequalities for quantum Markov chains, arXiv:2102.04146 [quant-ph] (2021), arXiv: 2102.04146. [GR21b] , Spectral methods for entropy contraction coefficients, arXiv:2102.04146 [quant-ph] (2021), arXiv: 2102.04146. [HJ91] Roger A Horn and Charles R Johnson, Topics in matrix analysis, Cambridge University Press, Cam- bridge; New York, 1991 (English), OCLC: 14212321. [HLW06] Shlomo Hoory, Nathan Linial, and Avi Wigderson, Expander graphs and their applications, Bulletin of the American Mathematical Society 43 (2006), no. 4, 439–561 (en-US). [Hoe63] Wassily Hoeffding, Probability Inequalities for Sums of Bounded Random Variables, Journal of the American Statistical Association 58 (1963), no. 301, 13–30, Publisher: [American Statistical Associa- tion, Taylor & Francis, Ltd.]. [JLR19] Marius Junge, Nicholas LaRacuente, and Cambyse Rouz´e, Stability of logarithmic Sobolev inequalities under a noncommutative change of measure, arXiv:1911.08533 [math-ph, physics:quant-ph] (2019), arXiv: 1911.08533. [Jon83] V. F. R. Jones, Index for subfactors, Inventiones mathematicae 72 (1983), no. 1, 1–25 (en). [JX07] Marius Junge and Quanhua Xu, Noncommutative Maximal Ergodic Theorems, Journal of the American Mathematical Society 20 (2007), no. 2, 385–439. [KT13] Michael J. Kastoryano and Kristan Temme, Quantum logarithmic Sobolev inequalities and rapid mixing, Journal of Mathematical Physics 54 (2013), no. 5, 052202. [LaR21] Nicholas LaRacuente, Quasi-factorization and Multiplicative Comparison of Subalgebra-Relative En- tropy, arXiv:1912.00983 [quant-ph] (2021), arXiv: 1912.00983. [LJL20] Haojian Li, Marius Junge, and Nicholas LaRacuente, Graph H\”ormander Systems, arXiv:2006.14578 [math-ph] (2020), arXiv: 2006.14578. [LR73] Elliott H. Lieb and Mary Beth Ruskai, Proof of the strong subadditivity of quantum-mechanical entropy, Journal of Mathematical Physics 14 (1973), no. 12, 1938–1941. [LY98] Tzong-Yow Lee and Horng-Tzer Yau, Logarithmic Sobolev inequality for some models of random walks, The Annals of Probability 26 (1998), no. 4, 1855–1873, Publisher: Institute of Mathematical Statistics. [MS14] Iman Marvian and Robert W. Spekkens, Extending Noether’s theorem by quantifying the asymmetry of quantum states, Nature Communications 5 (2014), 3821 (En). SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 33

[MU88] Hans Maassen and J. B. M. Uffink, Generalized entropic uncertainty relations, Physical Review Letters 60 (1988), no. 12, 1103–1106. [Pet91] D´enesPetz, On certain properties of the relative entropy of states of operator algebras, Mathematische Zeitschrift 206 (1991), no. 1, 351–361 (en). [PP86] Mihai Pimsner and Sorin Popa, Entropy and index for subfactors, Annales scientifiques de l’Ecole´ Normale Sup´erieure 19 (1986), no. 1, 57–106 (fr). [Rob55] Herbert Robbins, A Remark on Stirling’s Formula, The American Mathematical Monthly 62 (1955), no. 1, 26–29, Publisher: Mathematical Association of America. [TKR+10] K. Temme, M. J. Kastoryano, M. B. Ruskai, M. M. Wolf, and F. Verstraete, The $\chiˆ2$ - divergence and Mixing times of quantum Markov processes, Journal of Mathematical Physics 51 (2010), no. 12, 122201, arXiv: 1005.2358. [Tuc99] Robert R. Tucci, Quantum Entanglement and Conditional Information Transmission, arXiv:quant- ph/9909041 (1999), arXiv: quant-ph/9909041. [Tuc02] , Entanglement of Distillation and Conditional Mutual Information, arXiv:quant-ph/0202144 (2002), arXiv: quant-ph/0202144. [VAWJ08] Joan Alfina Vaccaro, F Anselmi, Howard Mark Wiseman, and Kurt Jacobs, Tradeoff between extractable mechanical work, accessible entanglement, and ability to act as a reference system, under arbitrary superselection rules, Physical Review A 77 (2008), no. 3, 032114. [Ver19] Anna Vershynina, Upper continuity bound on the quantum quasi-relative entropy, Journal of Mathe- matical Physics 60 (2019), no. 10, 102201, arXiv: 1904.00795. [Wil13] Mark M. Wilde, Quantum Information Theory, 1 edition ed., Cambridge University Press, Cambridge, UK. ; New York, June 2013 (English). [Win16] Andreas Winter, Tight Uniform Continuity Bounds for Quantum Entropies: Conditional Entropy, Relative Entropy Distance and Energy Constraints, Communications in Mathematical Physics 347 (2016), no. 1, 291–313 (en). [WY16] Andreas Winter and Dong Yang, Operational Resource Theory of Coherence, Physical Review Letters 116 (2016), no. 12, 120404, Publisher: American Physical Society.

Appendix A. Technical Proofs and Calculations Note: in this Lemma and its proof, and thereafter within this section, we use the letter n rather than d for dimension to avoid possible confusion with the derivative.

(Proof of Lemma 2.3). Define δ ≡ ρi − ρj ≥ 0. For the purposes of the derivative, tr(log(nρ)) is equivalent to tr(log ρ), since they differ only by a constant log n. We directly calculate,

∂ ζnρk (ρk log((1 − ζ) + ζnρk)) = log((1 − ζ) + ζnρk) + . (29) ∂ρk (1 − ζ) + ζnρk for any k ∈ 1...n. By setting ζ = 1,  ∂ ∂  − tr(ρ log ρ) = log(ρi/ρj) = log(1 + δ/ρj) . ∂ρi ∂ρj Because x ≤ log(1 + x) ≤ x 1 + x for all x ≥ 0, we have for any b ∈ [0, 1] that

δ/ρj a log(1 + δ/ρj) ≥ a(1 − b) log(1 + δ/ρj) + ab . 1 + δ/ρj 34 NICHOLAS LARACUENTE

This will allow us to deal with the two terms in equation (29) individually. First, we handle the logarithm term, log((1 − ζ) + ζnρk), by finding ζ such that

(1 − ζ) + ζnρj + ζnδ  a(1 − b) log(1 + δ/ρj) ≥ log . (1 − ζ) + ζnρj We rewrite the right hand side as  ζnδ  log 1 + . (1 − ζ) + ζnρj

On the left hand side, since ρj ≤ 1/n, δ/ρj ≥ nδ. We aim to show that  ζ  a(1 − b) log(1 + nδ) ≥ log 1 + nδ , (1 − ζ) + ζnρj which by Lemma 2.2 is achieved when ζ a(1 − b) ≤ . (1 − ζ) + ζnρj 1 + nδ

We know ζ/((1 − ζ) + ζnρj) ≤ ζ/(1 − ζ) for any nρj ≥ 0. Hence any a(1 − b) ζ ≤ (30) 1 + nδ + a(1 − b) is sufficiently small. Next, we handle the fraction terms by finding ζ such that δ/ρ ζn(ρ + δ) ζnρ ab j ≥ j − j . 1 + δ/ρj (1 − ζ) + ζn(ρj + δ) (1 − ζ) + ζnρj By Taylor expansion, ζn(ρ + δ) ζnρ j − j (1 − ζ) + ζn(ρj + δ) (1 − ζ) + ζnρj ∞ ζn(ρj + δ) X  −ζnδ k ζnρj = − . (1 − ζ) + ζnρj (1 − ζ) + ζnρj (1 − ζ) + ζnρj k=0 Canceling the 0th order term, ∞ ∞ ζnρj X  −ζnδ k ζnδ X  −ζnδ k ... = + (1 − ζ) + ζnρj (1 − ζ) + ζnρj (1 − ζ) + ζnδ (1 − ζ) + ζnρj k=1 k=0 ∞ ∞ ζnρj X  −ζnδ k X  −ζnδ k = − (1 − ζ) + ζnρj (1 − ζ) + ζnρj (1 − ζ) + ζnρj k=1 k=1 ∞  ζnρj  X  −ζnδ k = − 1 . (1 − ζ) + ζnρj (1 − ζ) + ζnρj k=1 SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 35

Now to turn this back into the form of a fraction, ∞ −(1 − ζ) X  −ζnδ k ... = (1 − ζ) + ζnρj (1 − ζ) + ζnρj k=1 ∞ (1 − ζ)ζnδ X  −ζnδ k = 2 ((1 − ζ) + ζnρj) (1 − ζ) + ζnρj k=0 (1 − ζ)ζnδ 1 = 2 ζnδ ((1 − ζ) + ζnρj) 1 + (1−ζ)+ζnρj 1 ζnδ = . (1 − ζ) + ζnρ ζ j 1 + 1−ζ (ρj + δ)n

Since 0 ≤ ρj ≤ 1/n, ζ 1−ζ nδ ... ≤ ζ . 1 + 1−ζ nδ We must find a ζ for which ζ δ/ρ nδ nδ ab log(1 + δ/ρ ) ≥ ab j ≥ ab ≥ 1−ζ . j 1 + δ/ρ 1 + nδ ζ j 1 + 1−ζ nδ One can easily check that this is satisfied by ζ ab ≤ , 1 + ζ (1 − ab)nδ + 1 which follows from ab ζ ≤ . (31) (1 − ab)nδ + ab + 1 Finally, any n a(1 − b) ab o ζ ≤ min , 1 + n + a(1 − β) (1 − ab)n + ab + 1 satisfies both inequalities 30 and 31.

Proof of 2.4. Given that n 1 − b b o ζ ≤ a min , , n + a(1 − β) + 1 (1 − aβ)n + ab + 1 we seek (1) A value of b at which both expressions are equal, such that the minimum is maximized; (2) The maximum possible constraint on ζ; (3) For a given ζ, the corresponding best achievable a. (4) A formula for ζ given a reasonable a. The optimal solution may use numerics. Here we derive an approximation. 36 NICHOLAS LARACUENTE

First, we solve for an equalizing b, (1 − b)(n − abn + ab + 1) = b(n + a(1 − b) + 1) n − anb + 1 − bn + anb2 − b = bn + b anb2 − (an + 2n + 2)b + (n + 1) = 0 . By the quadratic formula, 1 p b = an + 2n + 2+ (an + 2n + 2)2 − 4an(n + 1)
. 2an − Since an + 2n + 2 > 1, and b ∈ [0, 1], only the “−” root is relevant. Examining the expression in the square root, (an + 2n + 2)2 − 4an(n + 1) =(2 + a)2n2 + 4(a + 2)n + 4 − 4an2 − 4an =4n2 + 4an2 + a2n2 + 4an + 8n + 4 − 4an2 − 4an =4n2 + a2n2 + 8n + 4 =4(n + 1)2 + a2n2 . Via the binomial approximation, p  a2n2  4(n + 1)2 + a2n2 ≈ 2(n + 1) 1 + . 8(n + 1)2 We then calculate 1  2(n + 1)a2n2  1 an  1 a b ≈ an − = 1 − ≈ 1 − . 2an 8(n + 1)2 2 4(n + 1) 2 4 Plugging this b approximation back into the original minimization, n 4 + a 4 − a o ζ ≤ a min , , 8n + 4a + a2 + 8 (8 − 4a + a2)n + 4a − a2 + 8 We may find a common numerator by multiplying the left by 4 − a and the right by 4 + a. The left denominator becomes 32n + 16a + 4a2 + 32 − 8na − 4a2 − a3 − 8a = 32n − 8na + 32 + 8a − a3 . The right becomes 32n − 16na + 4na2 + 16a − 4a2 + 32 + 8na − 4na2 + na3 + 4a2 − a3 + 8a =32n − 8na + na3 + 32 + 24a − a3 . To eliminate nonlinear terms in a, we approximate 0 ≤ a3 ≤ a2 ≤ a ≤ 1, and overestimate denominators. We conclude that 15a ζ ≤ ζ (a) := . (32) + 32n − 7na + 32 + 24a We directly calculate the derivative with respect to a, d 15 15a(7n − 24) ζ (a) = + ≥ 0 da + 32n − 7na + 32 + 24a (32n − 7na + 32 + 24a)2 SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 37 to see that ζ+(a) is increasing with a on the interval [0, 1]. Hence by comparing to a = 1, it is sufficient to find 15 ζ ≤ ζ (a) ≤ . + 25n + 56 When ζ satisfies this condition, we may find an attainable a by solving a linear equation, yielding 32(n + 1)ζ a = . 15 + (7n − 24)ζ For a non-trivial inequality, we must take the inequality on ζ to be strict, otherwise we solve for a = 1 and the right hand side becomes 0. Fixing a = 1 − 1/n, we approximate 15n − 1 15n − 1 ζ ≤ = . 32n2 − 7n2 + 7n + 32n + 24n − 24 25n2 + 63n − 24 This approximation is expected to be good for large n. Alternatively, if a = 1/2, 15 ζ ≤ . 57n + 88

As n → ∞, our estimated maximal ζ → 3/5n. Here we show an alternate proof of Proposition 1.11 for the self-adjoint case that does not rely on Fisher information. While the proof is more involved, this technique gives some intuition that might be useful in developing further techniques.

Proof of Proposition 1.11. We start by writing Φt(ρ) = Φτ (ρ0), where τ < t, and ρ0 = Φt−τ (ρ). Let    wsb(, N ⊆ M) = 2D(MkN ) + (1 + 2)h , (33) 1 + 2 as in proposition 3.7 of [GJL20b] and based on Lemma 7 in [Win16], where h is the binary entropy, N ⊂ M is a subalgebra, and D(MkN ) = supρ D(EM(ρ)kEN (ρ)). Rather than depend on dimension as would the Fannes-Alicki bound, this bounds entropy in terms of trace distance and subalgebra index. Let N be the fixed point algebra of Φt with conditional expectation E, t and Nj the fixed point algebra of Φj. Let M be the original algebra containing ρ. By the Suzuki-Trotter expansion, the Taylor expansion of an exponential, or by direct commutation, τ τ τ τ 2 τ 0 Φ (ρ) = Φ1...ΦJ Φ0(ρ) + O((J + 1)τ ). Letρ ˆ = Φ0(ρ ). Hence τ τ τ 2 D(Φ (ˆρ)kE(ˆρ)) ≤ D(Φ1...ΦJ (ˆρ)kE(ˆρ)) + wsb(O(Jτ ), N ⊆ M). (34)

Define γj by τ τ τ τ τ D(Φj+1...ΦJ (ˆρ)kEj(Φj+1...ΦJ (ˆρ))) = γjD(Φ (ˆρ)kE(ˆρ)) . (35) 38 NICHOLAS LARACUENTE

Then τ τ D(Φj ...ΦJ (ˆρ)kE(ˆρ)) τ τ τ τ τ τ = D(Φj ...ΦJ (ˆρ)kEj(Φj ...ΦJ (ˆρ))) + D(Ej(Φj+1...ΦJ (ˆρ))kE(ˆρ)) 2 2 τ τ τ τ ≤ (1 − λjτ + O(λj τ ))D(Φj+1...ΦJ (ˆρ)kEj(Φj+1...ΦJ (ˆρ))) τ τ + D(Ej(Φj+1...ΦJ (ˆρ))kE(ˆρ)) 2 2 τ τ τ τ ≤ (1 − λjγjτ + O(λj τ ))(D(Φj+1...ΦJ (ˆρ)kEj(Φj+1...ΦJ (ˆρ))) τ τ + D(Ej(Φj+1...ΦJ (ˆρ))kE(ˆρ))) 2 2 τ τ = (1 − λjγjτ + O(λj τ ))D(Φj+1...ΦJ (ˆρ)kE(ˆρ)) , τ where the first equality follows from Lemma 3.1 and the first inequality from CLSI of Φi . The second inequality follows from equation (35), where we replace the τ τ τ τ −λjτD(Φj+1...ΦJ (ˆρ)kEj(Φj+1...ΦJ (ˆρ))) τ subtracted term by −λjγjτD(Φ (ˆρ)kE(ˆρ)) and then note that this is at least as large as the entire original expression. Iterating, we obtain  J  τ τ X  2 X 2 D(Φi ...ΦJ (ˆρ)kE(ˆρ)) ≤ 1 − λjγjτ + O τ λj D(MkNj) D(ˆρkE(ρ)) . (36) j=1 j P Now we must show that the subtracted sum j λjγj is bounded below. Here we invoke α-SQF. First, note that for any j, τ τ τ τ D(Φj+1...ΦJ (ˆρ)kEj(Φj+1...ΦJ (ˆρ))) 2 ≥ D(ˆρkEj(ˆρ)) + wsb(O((J − j)τ ), Nj ⊆ M) . Hence X τ τ τ τ X D(Φj+1...ΦJ (ˆρ)kEjΦj+1...ΦJ (ˆρ)) = α γjD(ˆρkE(ˆρ)) j j J X 2 ≥ D(ˆρkE(ˆρ)) + wsb(O((J − j)τ ), Nj ⊆ M) . j=1 P t This implies that j γj ≥ 1/α. D(ˆρkE(ρ)) ≥ D(Φ (ρ)kE(ρ)) by data processing, so either the latter is zero, or the former is at least some  that is independent of τ. If D(Φt(ρ)kE(ρ)) = 0, then the Theorem is trivially complete. Otherwise, returning to equations (36) and (34),

τ   D(Φ (ˆρ)kE(ˆρ)) ≤ 1 − τ min{λj}/α + err D(ˆρkE(ˆρ)). (37) j if we let wsb(O(Jτ 2), N ⊆ M) err =  J 2 (38) X wsb(O((J − j)τ ), Nj ⊆ M)  + + O(λ2τ 2)D(MkN ) .  j j j=1 SUBALGEBRA-RELATIVE ENTROPY INEQUALITIES 39

Furthermore, τ 0 0 0 0 D(Φ0(ρ )kE(ρ )) ≤ D(ρ kE(ρ )), (39) as the left hand side follows from data processing on the right hand side. Since Φt = (Φτ )t/τ , t t/τ D(Φ (ρ)kE(ρ)) ≤ (1 − min{λj}τ/α + err) D(ρkE(ρ)) . (40) j 2 P 2 Since err ≤ O(J log Jτ log τ(D(MkN ) + j λj log λjD(MkNj))),

t/τ − minj {λj }t/α lim(1 − min{λj}τ/α + err) = e . τ→0 j This completes the proof.