Quenched Growth of Normal Matrices and Harmonic Measure

summary

The gist

The paper investigates the quenched growth of normal matrices by comparing two growth rules—equilibrium growth and quenched batch sampling—to determine their limiting spatial distributions on a…

In short

The study compares two ways normal matrices grow: equilibrium growth and quenched batch sampling. It finds that under specific scaling conditions, both methods converge to the same limiting spatial distribution on a droplet boundary, which is described by exterior harmonic measure. This means both growth rules lead to the same predictable pattern for where new eigenvalues land.

Key concepts

Exterior Harmonic Measure
This is a mathematical concept describing how points distribute themselves on the outside of a shape. In this context, it dictates the limiting spatial distribution where new eigenvalues will settle on the boundary of an existing spectral droplet after growth.
Equilibrium Growth
This growth rule involves sampling all eigenvalues anew at each step, comparing different matrix ranks. When M/N approaches zero, this process shows that the average density of new points settles into a state governed by harmonic measure on the old boundary.
Quenched Batch Sampling
This rule models growth where new points are added based on projections onto polynomials that vanish at previously frozen eigenvalues. It represents a more realistic scenario where the existing spectrum is fixed while new elements are introduced.
Structure Factor Comparison
This analysis compares how the two growth rules affect spectral correlations. The quenched structure factor shows weaker angular repulsion and larger fluctuations in certain modes compared to the equilibrium band, helping distinguish between the two sampling methods.

Terminology used across episodes

This episode discusses

The paper

Quenched Growth of Normal Matrices and Harmonic Measure · Read on arXiv

Chebyshev Laboratory · HSE University

The eigenvalues of a large normal random matrix form a two-dimensional Coulomb gas. At fixed weight, the spectral droplet grows with the matrix rank, and its semiclassical density increment is harmonic measure. We freeze an N-eigenvalue configuration and sample M further eigenvalues using the projection onto polynomials of degree less than N+M that vanish at the frozen points. Their conditional law is an exact polynomial ensemble in the potential of the frozen charges. For a regular analytic droplet, we prove that the spatial distribution of the new eigenvalues converges to harmonic measure on the old boundary when M/ N to infinity and M/N to0. This convergence holds conditionally on the frozen spectrum, outside a set of frozen configurations whose probability tends to zero. The equilibrium process formed from orthogonal polynomials of degrees N,,N+M-1 has the same limit. An exact finite-rank identity relates polynomial moments of the conditional mean to fluctuations of the frozen spectrum, with a remainder from polynomial modes of degree N+M and above.

Transcript

Introduction to the show: ident: Mathematical Physics. The day in mathematical physics: rigorous models of physical systems.

Kai: We're looking at the paper: "Quenched Growth of Normal Matrices and Harmonic Measure".

Mira: Here's the gist of the paper. The paper investigates the quenched growth of normal matrices by comparing two growth rules—equilibrium growth and quenched batch sampling—to determine their limiting spatial distributions on a droplet boundary The main result is that both rules converge to the same exterior harmonic measure when specific scaling conditions are met.

Kai: The Mathematical Framework The eigenvalues of a normal random matrix form a two-dimensional Coulomb gas after integrating out the matrix angular variables The spectral droplet grows with the matrix rank, and its semiclassical density increment is harmonic measure.

Mira: The joint density of the new points contains the factor Y M a=1 Y N i=1 za − yi squared.

Kai: The conditional law for quenched growth is defined by the projection onto polynomials of degree less than N + M that vanish at the frozen points.

Title and authors: Kai: So this paper is called "Quenched Growth of Normal Matrices and Harmonic Measure". It looks like it’s about how the eigenvalues of a normal random matrix spread out when you keep adding more eigenvalues over time.

Mira: Right. The authors are OLEG ALEKSEEV, and this paper tackles a comparison between two ways this growth happens: equilibrium growth versus quenched batch sampling. It’s trying to figure out where these new points end up on the boundary of the existing set of points.

Lev: From an engineering standpoint, I'm wondering what kind of hardware this is testing. We're talking about eigenvalues forming a two dimensional Coulomb gas, which sounds like a system where charges repel each other based on their positions in space.

Kai: Exactly. The core idea is that the spectral droplet grows with the matrix rank, and this growth process has an increment that behaves like harmonic measure when certain scaling conditions are met. It suggests that both ways of adding points eventually settle down to the same limit on the boundary if things scale correctly, specifically when M over N goes to zero and M over log N goes to infinity >zero <ref:2610.03912#pg1>.

Mira: That convergence to harmonic measure is the big takeaway here for me. It means that for a regular analytic droplet, the spatial distribution of new points converges weakly to exterior harmonic measure on the old boundary when these scaling conditions are met, according to page two of THIS PAPER <ref:2610.03912#pg2,new points converges weakly to exterior harmonic measure on the old boundary>.

Lev: So if we're building an error correction code or simulating this on real hardware, we need to know that this convergence happens conditionally on the frozen spectrum, which means it holds outside a set of configurations whose probability tends to zero. That’s a crucial caveat for any practical application.

Kai: And then they do a comparison between the two growth rules—equilibrium growth and quenched batch sampling—and they find that when log N is much less than M much less than N, both rules converge to the same limiting spatial distribution on the boundary.

Mira: That's interesting because usually, these two processes behave differently at finite N. The paper notes that at finite N, the two laws differ in angular correlations; quenched batches show weaker angular repulsion and larger fluctuations in arc counts and low angular Fourier modes than the equilibrium band.

Lev: We need to be careful with those fluctuations. If we're trying to model real eigenvalue distributions, these larger fluctuations in low angular Fourier modes might translate into unwanted noise or artifacts on actual hardware measurements.

Kai: The paper then separates the structure factor into the contribution of the conditional mean and the fluctuations of batches sampled at a fixed old spectrum, and it shows that for a three-fold droplet, the quenched structure factor is two point one nine CUE units, while the equilibrium band gives one point zero two <ref:2610.03912#pg2,the structure factor into the contribution of the conditional mean and the>.

Title and authors: Mira: That difference in those structure factors is what tells us how much deviation there is between the two processes even when they both converge to harmonic measure in the limit. It quantifies that difference for a specific geometry, which I think is useful for understanding finite-size effects on real systems.

Lev: So, if we were to run this on a real system and find our structure factor is closer to one point zero two than two point one nine, it would suggest our sampling method is closer to the equilibrium process than the quenched one at that size N <ref:2610.03912#pg2>. That helps us set expectations for our simulation fidelity.

Kai: The paper also points toward future work, suggesting that controlling both the conditional mean and the fluctuations of batches sampled at a fixed spectrum is still open research. They mention extending a one-batch theorem to an evolving configuration, which requires estimates uniform over successive growth steps.

Mira: And they suggest connecting this to particle attachment, which would require quantitative control of the distance between new eigenvalues and the boundary. They even hint that bounds on the quantity R p of Y could turn a finite rank moment identity into quantitative asymptotics for conditional mean moments.

Lev: From an error correction perspective, having those bounds on R p would be essential for running this on real hardware because it would give us a way to control the overhead when we add new eigenvalues sequentially. Right now, we just see the result, but we don't have the tight control needed for robust error correction.

Kai: So to wrap up, OLEG ALEKSEEV’s paper "Quenched Growth of Normal Matrices and Harmonic Measure" shows that both equilibrium growth and quenched batch sampling converge to the same exterior harmonic measure on a regular analytic droplet under specific scaling conditions, like when M over N goes to zero and M over log N goes to infinity.

Mira: This convergence is conditional on the frozen spectrum, meaning it holds outside of configurations with a probability tending toward zero, which is important context for its applicability. It establishes the harmonic measure law for quenched growth through Theorem three point one, controlling fluctuations of Y <ref:2610.03912#pg3>.

Lev: And when comparing the two rules at finite N, we see that quenched batches show weaker angular repulsion and larger fluctuations than the equilibrium band, and they share a common limit identified as the derivative of the equilibrium measure at filling t equals one.

Kai: So for those who are interested in how this applies to real systems, it means if you are adding points sequentially, whether you sample them randomly or based on an equilibrium distribution, in the right scaling regime, you get the same final spatial pattern on that boundary.

Title and authors: Mira: The implications are that we can use harmonic measure as a predictable tool for describing eigenvalue distributions in growing systems, even when the growth is quenched or finite rank. It gives us a way to predict how new eigenvalues will distribute themselves on an existing spectrum when log N is much less than M much less than N.

Lev: And for modeling spectral correlations, the paper allows you to calculate that difference between the quenched and equilibrium mean moments by looking at fluctuations of the old spectrum and a finite rank remainder. That gives a more accurate description of deviations from ideal behavior, which is something we need when dealing with real hardware noise.

Kai: So while the paper confirms this convergence, they flag that controlling angular fluctuations remains an open question for future work, requiring estimates on how batches sampled at fixed spectra behave.

Mira: And the comparison of structure factors in three-fold symmetry gives us a concrete number: two point one nine CUE units for quenched versus one point zero two for the equilibrium band, which is a tangible measure of that difference at finite N <ref:2610.03912#pg2>.

Lev: I think those numbers are useful because they give us a benchmark to test our error correction simulations against, showing exactly how much deviation we expect based on the sampling method.

Kai: So the paper shows that both rules have the same limit, but they differ in how they get there, and these differences are measurable at finite sizes. It’s a good piece of theory for predicting eigenvalue distributions in growing systems.

Mira: Indeed. The final picture is that when you meet those scaling conditions, the spatial distribution settles down to harmonic measure on the boundary for both methods, giving us a consistent prediction across different growth processes.

Lev: And for anyone building real hardware, knowing that this convergence holds outside of vanishing probability sets means we have a solid theoretical foundation to trust when we try to use these models for actual error correction or spectral analysis.

Kai: That’s what it boils down to. We’ve got the results on quenched growth and harmonic measure from OLEG ALEKSEEV, and it gives us some strong predictions about how eigenvalues organize themselves as a system grows.

Mira: It provides a clear path forward by showing that we can use this convergence to predict distributions in growing systems under those precise scaling conditions.

Lev: And for us, it points toward the need for more quantitative control over those fluctuations so we can move from theory to practical hardware implementation with less uncertainty about the overhead.

Kai: We'll keep an eye on how these future directions play out in subsequent work on this topic as we continue our experiments here.

The paper's summary: Kai: So we’re talking about this paper now, "Quenched Growth of Normal Matrices and Harmonic Measure." The main thing they’re showing is that when you look at how a normal matrix grows, whether you grow it in a smooth, equilibrium way or in a random batch sampling way, they both end up with the exact same spatial pattern on the edge of that growing shape if you scale things right.

Mira: That convergence to harmonic measure is the key idea here. It means that for these two different ways of adding new eigenvalues—one being perfectly balanced, the other being randomly sampled in batches—the resulting distribution settles down to a specific geometric measure when it gets big enough and scaled correctly.

Lev: So what does that mean for us on real hardware? It suggests that if we are building something where the spectrum is evolving, whether we add points one by one in a controlled way or in random groups, the final layout of those points on the boundary will look like this harmonic measure.

Kai: Exactly. The paper proves this convergence happens when things scale like M over N goes to zero and M over log N goes to infinity. It sets up these precise conditions where you get that same limit for both growth rules.

Mira: But there’s a catch, Lev, because the convergence isn't automatic. It only holds conditionally on the old spectrum, meaning it only applies outside of configurations that are incredibly rare—configurations with a probability tending toward zero.

Lev: That is important for error correction research. It tells us that while the math predicts this limit holds most of the time, we have to be careful about those very rare configurations where the convergence might fail.

Kai: And when you compare the two growth rules at finite sizes, like when N is not huge yet, they do behave differently in terms of their angular correlations. Quenched batches show weaker repulsion between points and bigger fluctuations in how many points land in certain angular bins compared to the equilibrium band.

Mira: That’s where the structure factor comes into play. They look at this factor to see that for a three-fold droplet, the quenched structure factor is two point one nine CUE units, while the equilibrium band gives a value of one point zero two.

Lev: two point one nine versus one point zero two shows a pretty clear gap in how those two methods are behaving right now at finite sizes. That difference is what we need to model accurately when we try to simulate this on real hardware with finite rank corrections.

Kai: So, the implication for us is that we can use this convergence as a tool to predict eigenvalue distributions in growing systems under those specific scaling rules, even when the growth process itself is quenched or involves adding points in batches.

Mira: It gives us a predictable way to describe how new eigenvalues will distribute themselves on an existing spectrum when log N is much smaller than M and M is much smaller than N.

Lev: We still have to deal with those angular fluctuations, though. The paper admits that controlling the fluctuations of batches sampled at a fixed spectrum is still open research right now. It’s a challenge for us to get tighter control over those overheads when we try to implement this on actual hardware.

Kai: So, it’s a good prediction about the final spatial pattern, but we still need better math to control the noise along the way.

The paper's improvements: Kai: We just talked about how both equilibrium growth and quenched batch sampling eventually converge to the same harmonic measure on the boundary under specific scaling conditions. Now, let's look at what’s next—the authors are suggesting ways to make this math more useful in practice.

Mira: They are pointing out that while the convergence is good, it doesn't tell us everything about how these systems behave when they aren't perfectly scaled. The paper suggests improving the modeling by incorporating finite rank correction terms for quenched growth.

Lev: That makes sense from an engineering standpoint because we need to model deviations from ideal behavior on real hardware, and this method lets us calculate that difference between the quenched and equilibrium mean moments by looking at fluctuations of the old spectrum plus a finite rank remainder.

Kai: So, instead of just getting a single limit, this approach lets us quantify exactly how far away our real system is from that ideal limit by measuring those corrections. It’s about calculating the error based on how much the old spectrum fluctuates.

Mira: That’s useful because it moves us beyond just knowing *if* it converges to harmonic measure to understanding *how fast* we get there and what the specific deviations look like for different geometries.

Lev: For me, that means we can start building better predictive models for error correction by using those fluctuation measures instead of just assuming the equilibrium result is perfect. It gives us a more accurate description of how noisy our real system will be.

Kai: So, the implication here is that we move from just observing the final result to understanding the dynamics and errors during the process itself. We get a finer-grained picture of how things actually grow on that boundary.

Mira: The authors also flag some open research areas they want to tackle next, specifically controlling angular fluctuations. They need better estimates for both the conditional mean and the fluctuations of batches sampled at a fixed spectrum.

Lev: That’s where we need more rigorous math before we can reliably predict how these systems will perform in real-world error correction schemes. We need those uniform estimates over successive growth steps, not just one step at a time.

Kai: And they also suggest a connection with particle attachment, which would require us to have quantitative control over the distance between new eigenvalues and the boundary. That’s a big step toward linking this spectral theory to physical attachment models.

Mira: So, while the current paper gives us the limit, these improvements are about building tools that let us control and quantify those messy intermediate steps better. It’s about getting from a theoretical limit to a more controllable simulation setup.

Conclusion: Tom: So to wrap up, we’ve covered how the paper "Quenched Growth of Normal Matrices and Harmonic Measure" shows that both equilibrium growth and quenched batch sampling eventually converge to the same harmonic measure on a regular analytic droplet when scaling conditions are met.

Kai: Basically, for those specific scaling rules, you get one single prediction for how new eigenvalues will spread out on an existing spectrum, whether you sample them smoothly or in batches.

Mira: It's really about establishing a predictable tool using harmonic measure to describe eigenvalue distributions in growing systems under those precise conditions.

Lev: For error correction, that means we have a solid theoretical foundation for what the final layout should look like when the system is large enough, as long as we respect those conditional probability caveats.

Kai: And we saw that at finite sizes, like when N isn't huge yet, quenched batches show weaker angular repulsion and larger fluctuations than the equilibrium band.

Mira: That difference in structure factor numbers—two point one nine versus one point zero two—shows how much deviation there is between the two methods right now.

Lev: Those numbers are useful because they give us a benchmark to test our error correction simulations against, showing exactly how much we expect to deviate based on which sampling method we use.

Kai: So the paper confirms that both growth rules share this limit, but they differ in how they get there at finite sizes, and these differences are measurable right now.

Mira: The main implication is that when you meet those scaling conditions, the spatial distribution settles down to harmonic measure for both methods, giving us a consistent prediction across different growth processes.

Lev: And for us, it points toward the need for more quantitative control over those angular fluctuations so we can move from theory to practical hardware implementation with less uncertainty about the overhead.

Kai: We’ve got this result on quenched growth and harmonic measure, and it gives us some strong predictions about how eigenvalues organize themselves as a system grows.

Mira: It provides a clear path forward by showing that we can use this convergence to predict distributions in growing systems under those precise scaling conditions.

Lev: And for us, it points toward the need for more rigorous math before we can reliably predict how these systems will perform in real-world error correction schemes.

Kai: We’ll keep an eye on how those future directions play out in subsequent work on this topic as we continue our experiments here.

More episodes

← Home