Rank-One HCIZ Integrals and Multiplicative Convolutions

summary

The gist

We study how spectral averaging interacts with the logarithm in large-dimensional rank-one HarishChandra–Itzykson–Zuber (HCIZ) integrals to determine whether averaging occurs before or after…

In short

The study investigates how spectral averaging interacts with logarithms in large-dimensional rank-one integrals to distinguish between free and classical multiplicative convolution. It finds that averaging before the logarithm leads to free convolution, while taking the logarithm after yields classical convolution. This distinction is crucial for understanding the asymptotic behavior of these complex integrals.

Key concepts

Free Multiplicative Convolution
This type of convolution arises when spectral averaging is performed before taking the logarithm in rank-one integrals. It describes a specific statistical structure where the resulting limiting distribution follows free probability rules, contrasting with classical convolution.
Classical Multiplicative Convolution
This occurs when the logarithm is taken after spectral averaging. It corresponds to classical multiplicative convolution, which is a more standard type of probabilistic operation found in many physical and mathematical models.
S-transform Multiplicativity
The paper proves that the S-transform exhibits multiplicativity for independent matrices, meaning SµBA(z) = SµA(z)SµB(z). This property is a key tool used to establish the relationship between different types of convolutions in the analysis.
Annealed vs. Quenched Energy
The paper compares annealed and quenched free energies. The annealed average yields classical convolution for small parameters, while the quenched average yields free convolution, highlighting how averaging order affects the resulting limiting statistics.

Terminology used across episodes

This episode discusses

The paper

Rank-One HCIZ Integrals and Multiplicative Convolutions · Read on arXiv

Benoĉt Collins, Nicolas Delporte, Manasa Nagatsu, Reiko Toriumi

Department of Mathematics, Graduate School of Science, Kyoto University · Gravity, Quantum Geometry and Field Theory Unit, Okinawa Institute of Science and Technology Graduate University

Transcript

Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.

Ada: We're looking at the paper: "Rank-One HCIZ Integrals and Multiplicative Convolutions".

Emmy: Here's the gist of the paper. We study how spectral averaging interacts with the logarithm in large-dimensional rank-one HarishChandra–Itzykson–Zuber (HCIZ) integrals to determine whether averaging occurs before or after taking the logarithm, leading to free multiplicative convolution versus classical multiplicative convolution Limiting Descriptions of Averaged Integrals The paper establishes that for independent bi-unitarily invariant matrices with uniformly bounded deterministic singular values, averaging the logarithm of the rank-one spherical integral yields free multiplicative convolution at speed N squared, whereas taking the logarithm after averaging yields classical multiplicative convolution at speed N The paper also proves an exact finite-dimensional equality between this annealed integral and the one for (A∗A) ⊗ (B∗B), while showing that for every fixed number of factors k ≥ 3, counterexamples exist where the analogous equality can fail even asymptotically Comparison of Averaging Orders The authors compare the disorder average of the quenched free energy, 1/N EHN log IN (t; HN) and the annealed free energy, 1/N log EHN IN (t; HN), finding that for small real parameters, the former gives free multiplicative convolution while the latter gives classical multiplicative convolution The annealed limit follows from Theorems 4.2 and 5.1, which provide an exact finite-dimensional annealed identity between global rank-one spherical integrals Multiplicativity of the S-transform The paper derives the known multiplicativity of the S-transform, showing that for independent bi-unitarily invariant matrices A and B, SµBA(z) = SµA(z)SµB(z) holds for all sufficiently small complex z This is established through a radial derivation of Voiculescu’s S-multiplicativity formula using the distributional identity (13) and concentration estimates Tensor Products and Multiplicative Convolutions The paper investigates tensor products, showing that global Haar vectors at speed N k yield classical multiplicative convolution, while product-state vectors at speed N yield free multiplicative convolution under corresponding hypotheses For a fixed number of legs k ≥ 3, counterexamples are constructed showing that the analogous equality can fail asymptotically Global Haar and Product-State Vectors The paper compares global Haar and product-state vectors for tensor products, noting that the empirical spectral measure of XN is exactly the classical multiplicative convolution of the empirical spectral measures of X∗i,N Xi,N The conditional normalized log-spherical integral for a global Haar vector at speed Nk converges in probability to R t 0 Rµ1⊛···⊛µk (s) ds Finite-Dimensional Estimates and Counterexamples The paper provides a uniform finite-dimensional estimate for the S-transform, showing that for independent bi-unitarily invariant matrices AN and BN with bounded singular values, the multiplicative rule holds with uniform error at most any prescribed δ > 0 Furthermore, it constructs counterexamples for fixed k ≥ 3 where the analogous equality fails asymptotically Saddle-Point Derivations The derivation of the S-transform identity involves a saddle-point calculation where the stationarity condition yields y(t) = Φ′B(u) = RB(u), leading to RBA(t) = RA(s)RB(u), which ultimately proves (29) This process relies on the local form of the Laplace principle Convergence in Probability for Product-State Vectors The convergence of the normalized log-spherical integral for a product-state vector x1,N ⊗ · · · ⊗ xk,N is governed at speed N by µ1 ⊠ · · · ⊠ µk The limit is given by Z t 0 Rµ1⊛···⊛µk (s) ds under the independent one-leg hypotheses of Theorem 5.7 Symmetric Tensor Model Failure The paper shows that for fixed k ≥ 3, there are independent bi-unitarily invariant matrices with deterministic singular values for which the normalized annealed log-spherical integral for (X1,N · · · Xk,N) ⊗ 1 ⊗(k−1)N at speed Nk has a strictly larger lower limit than the limit for X1,N ⊗ · · · ⊗ Xk,N This demonstrates that the exact identity of Theorem 4.2 might suggest a similar relation for more factors fails in general for every sufficiently small fixed positive parameter Final Limit Identity The final result shows that under the hypotheses of Corollary 3.9, the normalized annealed log-spherical integrals of AN, BN, and BN AN satisfy ΦA(t) = Z t 0 RµA (s) ds, ΦB(t) = Z t 0 RµB (s) ds, and ΦBA(t) = Z t 0 RµA⊠µB (s) ds This is established by applying the concentration hypotheses of Proposition 3.3 to the empirical spectral measures Second Cumulant Check The second cumulant check reveals that Var(µ1 ⊛ µ2) − Var(µ1 ⊠ µ2) = Var(µ1) Var(µ2), indicating that the classical and free multiplicative convolutions already differ at second order unless at least one factor has zero variance This difference is quantified by Z t 0 RµA⊛µB (s) − RµA⊠µB (s) ds = t 2/2 Var(µA) Var(µB) + O(t 3) Product-State Vector Limit The limit of the exponential moment for the product-state vector is given by Z t 0 Rη(s) ds, where Sη(z) = Sµ1 (z)· · · Sµk (z) This is proven inductively by relating the problem to independent factors and applying Proposition 5.6 The final result is that for real t sufficiently close to 0, lim N→∞ 1/N log Eexp(N tZN) = Z t 0 Rη(s) ds Deterministic Singular Values Conclusion For matrices with deterministic singular values as in Corollary 5.10, the one-leg formula (52) holds, and consequently, after possibly shrinking the neighborhood of 0 required in the iterative saddle-point argument, the conclusion of Theorem 5.7 follows The concentration hypothesis of Proposition 3.3 holds for these deterministic singular values Uniform Estimate for S-transform Theorem 3.11 provides a uniform finite-dimensional estimate for the S-transform, showing that the difference between the empirical S-transforms is bounded by e−γN squared This estimate relies on the uniform support bound and compactness of probability measures Final Comparison Estimate The final comparison estimate (37) proves that P dLµ(BN AN)∗(BN AN), µA⊠µB > ε ≤ 2Ke−cN squared, for all sufficiently large N This result is derived by combining the concentration estimates with the triangle inequality Tensor Product Identity The identity in Theorem 4.2 states that EX1,X2 Z U(N k) expn N k 2t TrX∗N XN UN ΠN U∗N o dUN = EX1,X2 Z U(N k) expn N k 2t TrP∗N PN UN ΠN U∗o dUN This is proved by the distributional identity of squared norms Scaling and Dilation The point-mass factors act as the rescaling t 7→ ait, and this argument extends to any number of such factors If ai > 0, this is consistent with Sδai(z) = a-1 i and with dilation under ⊠ If ai = 0, the logarithmic Laplace limit is identically zero near 0 while the S-transform is not defined Example Application The example application using Marchenko–Pastur limiting Gram measures shows that the resulting limit is expressed through the R-transform of a Fuss–Catalan law, with moments given by the Fuss–Catalan numbers This demonstrates how the theory applies to specific matrix models Failure for Three or More Factors The paper proves that for every fixed k ≥ 3, there are independent bi-unitarily invariant matrices with deterministic singular values for which, for every sufficiently small fixed t > 0, lim inf N→∞ F prod N (t) > lim N→∞ F⊗N (t) This is established by showing that the lower limit of the product integral is strictly larger than the tensor product limit Final Conclusion on Annealed Limit The conclusion of Corollary 3.9 remains valid for random singular values, provided that the uniform norm bound is retained and the empirical spectral measures of A∗N AN and B∗N BN satisfy (19) Weak convergence in probability alone does not control rare spectral events with probability e−o(N), which may still contribute to the annealed quantity N-1 log Ee N(·) Second Cumulant Inequality Equation (46) shows that the classical and free multiplicative convolutions already differ at second order unless at least one factor has zero variance This difference is quantified by Z t 0 RµA⊛µB (s) − RµA⊠µB (s) ds = t 2/2 Var(µA) Var(µB) + O(t 3) This is a key observation regarding the difference between the two convolution types Final Estimate on S-transform Difference Theorem 3.11 proves that there is a radius ρ > 0 such that for all sufficiently large N, P sup z≤ρ Sµ(BN AN)∗(BN AN) (z) − SµA∗N AN (z)SµB∗N BN (z) > δ!

Ada: ≤ e−γN squared This uniform estimate is crucial for establishing the multiplicativity of the S-transform Uniform L1 Convergence The uniform estimate also gives uniform L1 convergence, which implies that the S-transforms are uniformly bounded on the common disk w ≤ ρ This uniformity is established by the compactness of probability measures supported in

d, D: Final Identity Theorem The identity theorem extends the equality to the whole disk because all three S-transforms are analytic on the common disk w < ρ This implies that sup w≤ρ Sµ(BN AN)∗(BN AN) (w) − SµA∗N AN (w)SµB∗N BN (w) ≤ δ This proves the desired comparison Final Result on Annealed Limit The final result in Corollary 3.9 is that for all real t sufficiently close to 0, lim N→∞ 1/N squared EAN,BN

log IN squared (t; HN ⊗ 1N): = Z t 0 RµA⊠µB (s) ds This is the annealed limit, contrasting with the quenched limit which is Z t 0 RµA⊛µB (s) ds The difference between these limits satisfies Z t 0 RµA⊛µB (s) − RµA⊠µB (s) ds = t 2/2 Var(µA) Var(µB) + O(t 3) This is the precise quantitative difference between free and classical convolution Product-State Vector Limit Summary The final result for the product-state vector is that for real t sufficiently close to 0, lim N→∞ 1/N log Eexp(N tZN) = Z t 0 Rη(s) ds, where Sη(z) = Sµ1 (z)· · · Sµk (z) This is proven inductively by relating the problem to independent factors and applying Proposition 5.6 The final result is that for real t sufficiently close to 0, lim N→∞ 1/N log Eexp(N tZN) = Z t 0 Rη(s) ds Final Conclusion on Rank-One Asymptotics The final conclusion is that for every real t with t < 1/(64M 2), the one-leg formula (52) holds, and consequently, after possibly shrinking the neighborhood of 0 required in the iterative saddle-point argument, the conclusion of Theorem 5.7 follows This is established by applying Proposition 3.3 with C = M squared The remaining hypotheses of Theorem 5.7 hold by assumption Final Result on Product-State Vector Limit Summary The final result for the product-state vector is that for real t sufficiently close to 0, lim N→∞ 1/N log Eexp(N tZN) = Z t 0 Rη(s) ds This is proven inductively by relating the problem to independent factors and applying Proposition 5.6 The final result is that for real t sufficiently close to 0, lim N→∞ 1/N log Eexp(N tZN) = Z t 0 Rη(s) ds Final Result on Product-State Vector Limit Summary The final result for the product-state vector is that for real t sufficiently close to 0, lim N→∞ 1/N log Eexp(N tZN) = Z t 0 Rη(s) ds This is proven inductively by relating the problem to independent factors and applying Proposition 5.6 The final result is that for real t sufficiently close to 0, lim N→∞ 1/N log Eexp(N tZN) = Z t 0 Rη(s) ds <ref:2610.

Paper summary: Ada: So we're looking at this paper today called "Rank-One HCIZ Integrals and Multiplicative Convolutions". It’s about how spectral averaging interacts with the logarithm in large-dimensional rank-one integrals.

Emmy: Basically, the core question they are tackling is whether you average the logarithm before or after you take it. This choice determines whether you get free multiplicative convolution or classical multiplicative convolution.

Kolya: So, what’s the big claim here? They establish that for independent bi-unitarily invariant matrices with uniformly bounded deterministic singular values and weakly convergent empirical spectral measures of their Gram matrices, averaging the logarithm after you average yields free multiplicative convolution at speed N squared.

Ada: That's a key distinction there. And they also show that taking the logarithm before averaging yields classical multiplicative convolution at speed N.

Emmy: Beyond that, they prove an exact finite-dimensional equality between this annealed integral and the one for (A star A) tensor (B star B). But they also found counterexamples for every fixed number of factors k greater than or equal to three where that equality fails even as N gets really large.

Kolya: That's interesting because it shows the nice symmetry they found in this finite case doesn't hold generally when you have more than two components. What about how this relates to the free energy?

Ada: Well, they compare the disorder average of the quenched free energy, which is one over N times that log integral with HN, against the annealed free energy, which is just one over N times that log integral without averaging first.

Emmy: And for small real parameters, those two different quantities give you different results—one gives you free multiplicative convolution while the other gives classical multiplicative convolution. This comparison follows from theorems four point two and five point one, which provide an exact finite-dimensional annealed identity between global rank-one spherical integrals.

Kolya: So it seems like the order of averaging really matters for how these quantities behave in practice, especially when you look at the disorder average versus the annealed average.

Ada: Exactly. Then there’s this point about the S transform being multiplicative. They derive that for independent bi-unitarily invariant matrices A and B, this S transform rule holds: SmuBAz equals SmuA(z) times SmuB(z) for all sufficiently small complex z.

Emmy: That multiplicativity is established through a radial derivation of Voiculescu’s S-multiplicativity formula using a distributional identity. It relies on concentration estimates, which are pretty strong.

Kolya: So they're showing that this multiplicative property of the S transform isn't just an artifact of the setting; it holds under these specific conditions for independent matrices A and B. What does that mean for the overall structure?

Paper summary: Ada: It means the underlying algebraic structure is additive in a way, even though we are dealing with multiplicative convolutions in terms of spectral measures. They also look at tensor products, showing that global Haar vectors at speed N to the power k yield classical multiplicative convolution.

Emmy: But product-state vectors at speed N yield free multiplicative convolution under corresponding hypotheses. And again, for a fixed number of legs k greater than or equal to three, they construct counterexamples where that equality fails as N gets large.

Kolya: So we have this tension between what happens with global Haar vectors versus product-state vectors when you take tensor products, and the fact that the equality breaks down for multiple factors.

Ada: They compare those two vector types specifically, noting that the empirical spectral measure of X N is exactly the classical multiplicative convolution of the empirical spectral measures of X star i N times X i N.

Emmy: And they also look at convergence in probability for a global Haar vector at speed N to the power k. It converges to something called R t zero R mu one cross mu two cross... cross mu k, s ds.

Kolya: So when you look at the product state vector, they find that it converges in probability to Z t zero R mu one times sigma two times... times sigma k, s ds under the independent one-leg hypotheses of theorem five point seven.

Ada: That limit for the product-state vector is governed by sigma one cross sigma two cross... cross sigma k, s ds. And they prove this inductively by relating it to independent factors and applying proposition five point six.

Emmy: Then there's a comparison between global Haar and product-state vectors for tensor products, noting that the empirical spectral measure of X N is exactly the classical multiplicative convolution of the empirical spectral measures of X star i N times X i N.

Kolya: And they also look at the second cumulant check, which shows that Var mu one cross mu two minus Var mu one box mu two equals var mu one times var mu two. This means classical and free multiplicative convolutions already differ at second order unless at least one factor has zero variance.

Ada: That difference is quantified by Z t zero R mu one cross mu two s minus R box sigma1 times sigma2 s ds equals t squared over two var mu one var mu two plus some term of order three.

Emmy: It’s a concrete number here showing how the classical and free convolutions separate when you look at the second moment. That separation is directly related to those variance terms.

Kolya: So, what does this mean for someone who only listens to the show? It tells them that if they are dealing with non-zero variances, these two types of convolution aren't identical even at the second order level.

Paper summary: Ada: And then they look at the final result in corollary three point nine. Under those hypotheses, the normalized annealed log-spherical integrals of A star N times B star N and B star N times A star N satisfy specific relations involving R mu A cross mu B s.

Emmy: Specifically, they show that for real t sufficiently close to zero, the annealed limit for AN and BN is Z t zero R mu A box mu B s ds.

Kolya: But that’s different from the quenched limit, which is Z t zero R mu A cross mu B s ds. The difference between these two limits is exactly that term we just talked about: t squared over two var mu A var mu B plus some third order term.

Ada: So the annealed limit for the product of AN and BN is Z t zero R mu A cross sigma B s ds, contrasting with the quenched limit which uses the free convolution R mu A cross mu B s ds.

Emmy: It’s a statement about how averaging interacts with those spectral measures, showing that for small real t, the difference between classical and free convolutions is dictated by these variance terms.

Kolya: And they do give one more concrete result for the product-state vector limit. For real t sufficiently close to zero, the log of the exponential moment converges to Z t zero R eta s ds where S eta z is just sigma one z times sigma two z times... times sigma k z.

Ada: That eta function is simply the product of all those individual S transforms. And they prove this inductively by relating it to independent factors and applying proposition five point six.

Emmy: And finally, they do a lot of work on the deterministic singular values case in corollary five point ten. For matrices with deterministic singular values, the one leg formula holds, which lets them conclude theorem five point seven after shrinking that neighborhood of zero required for the saddle-point argument.

Kolya: So to summarize what we’ve heard about "Rank-One HCIZ Integrals and Multiplicative Convolutions," we have two main things: averaging before or after the logarithm changes the resulting convolution type, and there are specific limits involving cross and box convolutions depending on whether you're looking at annealed or quenched quantities.

Ada: And for someone listening, it means that when you analyze these integrals, if you’re dealing with non-zero variances in your components, those two types of convolution aren't the same even at the second order level.

Emmy: It really boils down to how the spectral averaging interacts with the logarithm. It shows that this interaction is not trivial and requires careful handling when you look at these large-dimensional problems.

Conclusion: Ada: So we've been digging into this paper called "Rank-One HCIZ Integrals and Multiplicative Convolutions," and now we're coming to the finish line, looking at what all this actually means.

Emmy: Right, it’s about how spectral averaging interacts with the logarithm in these large-dimensional rank-one integrals. The main point is that you have two different ways to average—before or after you take the log—and that choice dictates whether you get free or classical multiplicative convolution.

Kolya: So, it's showing a clear separation based on the order of operations, which is interesting because usually in these settings we expect things to be more unified.

Ada: Exactly. The authors prove that for independent matrices with bounded singular values, averaging the logarithm after you average gives free multiplicative convolution at speed N squared. But if you take the log first and then average, you get classical multiplicative convolution at speed N.

Emmy: And they also found a specific finite-dimensional equality relating this annealed integral to something involving tensor products of matrices. That's a nice structural link they established for a moment.

Kolya: But that connection isn't perfect; they showed that for any fixed number of factors, say three or more, there are counterexamples where that equality fails even as N gets huge.

Ada: That’s important because it shows the nice symmetry they found in this finite case doesn't hold when you have more than two components. It points to a real complexity in scaling up these problems.

Emmy: So, what this means for the world is that if you are dealing with complex systems modeled by these random matrices, the way you average things—the disorder or annealing—fundamentally changes the resulting mathematical structure you see.

Kolya: And it suggests that even when you look at very specific limits like those involving product state vectors, there’s a clear distinction between what happens in the quenched case versus the annealed case.

Ada: It really boils down to how spectral averaging interacts with the logarithm, showing that this interaction is not trivial and requires careful handling when you look at these large-dimensional problems. That leads us into how this plays out when we consider more than two factors, which is what we'll be looking at next.

More episodes

← Home