Ergodic Central limit theorem for subgeometric Markov chains

arXiv:2512.18255 · math.PR, stat.CO, stat.ML · Submitted 2025-12-20 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Ergodic Central limit theorem for subgeometric Markov chains".

Jane: Here's the gist of the paper. Given that any error could cost millions, this synthesis aims for absolute precision regarding both the theoretical framework and its practical applications.

Tom: Here is the detailed, combined summary: Comprehensive Research Synthesis: Central Limit Theorems for Ergodic Averages of Markov Chains This research paper addresses a fundamental problem in probability theory and computational statistics: establishing necessary and sufficient conditions for the validity of Central Limit Theorems (CLTs) when applied to ergodic averages of Markov chains defined on general state spaces.

Jane: The core theoretical machinery developed relies on drift conditions derived from a triplet of functions: a Lyapunov function V, a scalar function phi, and another scalar function.

Tom: I.

Title and authors: Tom: So, we're looking at this paper called "Ergodic Central limit theorem for subgeometric Markov chains," which is pretty deep stuff regarding when these averages of Markov chains actually converge to a normal distribution in that way we expect.

Jane: Right, and the authors are Miha Bresar, Aleksandar Mijatović C, and Gareth Roberts ´. The title itself tells us this paper isn't just looking for the conditions where the CLT *should* hold, but rather finding those necessary conditions when it might fail.

Lu: It’s important because we’ve seen a lot of literature on sufficient conditions for CLTs in these systems, like KV86 or DGM08, but this work is focused on providing verifiable necessary conditions using something called drift conditions.

Meng: Necessary conditions are good, but what does that mean practically? Does it just give us a set of rules we have to follow before we can even consider using an algorithm?

Tom: Exactly. The core contribution here is the development and application of these L-drift conditions, denoted as L(V, phi,), which are the framework for characterizing when an ergodic CLT holds or fails.

Jane: So, when we look at the summary of this paper called "Ergodic Central limit theorem for subgeometric Markov chains," it seems they’re setting up this structure using a triplet of functions: a Lyapunov function V, and two scalar functions phi and.

Lu: They define these L-drift conditions, Assumption L(V, phi,), which serve as the necessary framework to characterize when an ergodic CLT holds for Markov chains on general state spaces.

Meng: I see they’re building this by proving that a specific process related to the Lyapunov function and these drift functions satisfies certain submartingale properties, which is crucial for bounding the convergence behavior.

Tom: That technical step involves Lemma two point seven, which verifies Assumption (ii) of the L-drift condition by demonstrating that a specific process is a submartingale, V(x) at most n to infinity Ex

V(X n S T(r)): = E

V(X S* T(r)): , which is essential for bounding the convergence behavior.

Jane: If a person listening across from me doesn't read this, they might ask, "So what are the actual conditions that make the CLT fail?" and I think they’d want to know how specific."

Lu: Theorem three point one one shows explicit conditions under which the CLT fails for a function g: R d to R if g(x) at least x 2s for all x sufficiently large, where s > v/two - d/four <ref:2512.18255#pg1>.

Tom: And there’s another way they show failure, Theorem three point two links the failure of the CLT to the divergence of a specific integral involving time and distance functions: integral zero infinity c'c'' t - k/(two(s+two))dt = infinity <ref:2512.18255#pg1>.

Jane: That integral divergence is pretty concrete, but what does that mean for a person just trying to use an MCMC sampler in their work?

Title and authors: Meng: It means if that integral diverges, then the CLT won't hold for the average of whatever you're sampling from. It’s a direct check on the process dynamics.

Lu: Beyond just existence, the paper also provides sharp bounds on how fast these ergodic averages converge to their stationary distribution pi in various metrics, like f-variation and Wasserstein distance.

Tom: They use those L-drift conditions to derive lower bounds on convergence rates in both f-variation distance and Wasserstein distance. These derived bounds match known upper bounds, which is a strong verification of the speed of mixing for different Markov chain families.

Jane: So, if someone is looking at the paper called "Ergodic Central limit theorem for subgeometric Markov chains," they’re not just getting existence theorems; they are getting concrete metrics on how fast things settle down.

Lu: They also establish lower bounds on crucial probabilistic quantities related to the process, like the tails of return times lambda r and the tails of the invariant measure pi.

Meng: That tells me we can get better estimates for how long it takes for a chain to mix or how heavy its tail is in some contexts.

Tom: And this whole theoretical machinery is then applied directly to sampling algorithms, focusing on families called (SIM-ALG), which are often used when dealing with heavy-tailed distributions.

Jane: This leads us into section three of the paper, where they characterize the performance and CLT validity of specific samplers by looking at the decay rate of a function related to the proposal distribution.

Lu: They compare several algorithms here: Random Walk Metropolis samplers, Metropolis-adjusted and unadjusted Langevin algorithms, stereographic projection sampler, and even an independence sampler.

Meng: I’m curious about what they found when comparing vanilla samplers to single-jump samplers. The paper suggests vanilla samplers have "stickiness in the tails," which is quantified by a specific decay rate of a function.

Tom: That stickiness is quantified, and it shows that single-jump samplers offer faster tail exploration but are highly sensitive to proposal misspecification.

Jane: So, if someone only listens to this, they need to understand the trade-off: faster exploration versus robustness against bad proposals.

Lu: The paper also gives definitive guidance for algorithm selection based on the target distribution's moment properties. If the target pi admits d-moments, the stereographic sampler is optimal because of its uniform ergodicity and preservation of CLTs.

Meng: But if you’re in high-dimensional problems where pi only has fewer than d moments, that stereographic sampler gets outperformed by most others in (SIM-ALG), except for the independence sampler.

Tom: And they show that unadjusted Langevin algorithms with infinite variance proposals have an invariant distribution tail decay that matches the target distribution for sufficiently small step sizes.

Title and authors: Jane: So, if you’re using an MCMC sampler, this paper tells you exactly which tool to pick based on what your data looks like in terms of moments.

Lu: The improvements suggested by the authors focus heavily on how to make this theory actionable for algorithm design. They suggest a way to use these L-drift conditions to find optimal proposal kernels for things like RWM or MALA.

Meng: That sounds useful because I’m always trying to tune my proposal distribution for infinite variance proposals, and they suggest picking one with an increment law q alpha that maximizes the exploration of heavy-tailed tails.

Tom: So, this is about adaptive proposal tuning for algorithms like RWM or MALA. It’s about dynamically adjusting your kernel based on what the theory says should be good for exploring those tails faster while keeping things well-behaved.

Jane: And they claim this leads to faster convergence rates and preserves the validity of the CLT for a broader class of ergodic averages, which is a big deal for reliable estimation.

Lu: They also provide asymptotically sharp lower bounds on convergence rates in f-variation and L p-Wasserstein distance. For instance, they establish bounds such as c n - v/two - epsilon at most P n(x, times) - pi f by applying Theorem three point two or Theorem three point four directly to the specific Markov chain dynamics and target distribution > <ref:2512.18255#pg1>

Tom: Those lower bounds are tight, and they use the L-drift conditions to establish them for these specific systems, which is a huge piece of verification for anyone trying to prove convergence speed analytically.

Jane: So, if someone is listening just tuning in on this, they should know that this paper gives you tools to check if your sampler setup has the right theoretical foundation before you even run the expensive simulations.

Lu: Ultimately, this paper provides verifiable necessary conditions for CLT validity and sharp metrics on convergence rates for ergodic averages of Markov chains, which is what we need to understand the speed of mixing in complex systems >

Meng: I’m glad they focused on the practical side, because when you’re shipping models at scale, you don't just care about theoretical bounds; you care about what breaks in production.

Tom: It’s a lot of rigorous machinery applied to heavy-tailed targets that gives us a much clearer picture of why some algorithms stick and others fail when dealing with complex distributions >

Jane: We’ve covered the necessary conditions, the failure criteria, how they compare different MCMC algorithms like RWM versus MALA, and what they suggest for tuning proposals to get better results >

Lu: So "Ergodic Central limit theorem for subgeometric Markov chains" provides a rigorous toolkit to move beyond just hoping an ergodic average estimator has a CLT and gives us the math to verify its convergence rate in those challenging settings >

The paper's summary: Tom: So we’ve seen how they build these L-drift conditions, and now we’re going to talk about what those actually do for the average of a Markov chain.

Jane: Okay, so essentially, this paper provides the necessary mathematical framework to check if you can actually use a Central Limit Theorem on averages from these kinds of Markov chains.

Tom: Right. It’s not just saying "maybe the CLT holds"; it’s giving us a concrete set of rules—the L-drift conditions—that tell us exactly when that convergence is possible or when it fails entirely.

Jane: I think the core idea is that they use a triplet of functions, like a Lyapunov function V, and two scalar functions, phi and, to define these drift conditions.

Tom: Exactly. And there’s this technical piece where they prove that a specific process related to V and those drift functions acts like a submartingale.

Jane: That submartingale property is super important because it lets them put firm bounds on how fast the averages are converging, or if they even converge at all.

Tom: If that process is a submartingale, they can use that to establish lower bounds on convergence rates in things like f-variation and Wasserstein distance.

Jane: Which means we get sharp numbers telling us how quickly the system mixes—how fast it gets close to its steady state pi.

Tom: They aren't just giving us existence; they’re giving us a mathematical yardstick for speed. They establish those lower bounds, like c n - v/two - epsilon at most P n(x, times) - pi f.

Jane: That’s what changes things for someone using an algorithm. If you can use these bounds to prove your sampler converges at a certain speed, you know your simulation time scales correctly with the desired accuracy.

Tom: And they also have explicit conditions for failure. They show precisely when the CLT breaks down—for example, if a function g grows too fast relative to some parameter s.

Jane: It’s like they’re telling us exactly what kind of distribution we can expect from our samples and what mathematical structure we need to avoid if we want reliable statistics.

Tom: So, the big picture here is taking this abstract L-drift machinery and applying it directly to how AI models sample from complex spaces, especially those with heavy tails.

Jane: And that application is where the real value is for MCMC sampling algorithms. They compare things like Random Walk Metropolis versus Langevin algorithms based on these convergence rates and failure criteria.

Tom: It moves us beyond just picking an algorithm because it’s popular; it lets us pick one because the underlying math guarantees its performance metrics.

Jane: So, this paper gives us the "why" behind the performance numbers we see in simulations, tying them back to fundamental probability theory on Markov chains.

The paper's improvements: Tom: So we’ve covered the paper’s core theory and how it applies to MCMC sampling, and now we look at what they suggest to make this framework even better for real production systems.

Jane: What’s next is how they turn these necessary conditions into actionable steps for designing new AI algorithms.

Tom: They focus on adaptive proposal tuning for things like RWM or MALA when dealing with infinite variance proposals, which is something Meng deals with all the time at scale.

Jane: They propose that we can dynamically adjust the proposal kernel itself—the increment law q alpha—to maximize the exploration of those heavy tails.

Tom: So if you’re running a sampler, instead of picking one fixed proposal distribution, you can use this theory to select one that actually explores the tail better while keeping things well-behaved mathematically.

Jane: That means faster convergence rates and they claim this helps preserve the validity of the Central Limit Theorem for a wider range of average estimators.

Tom: It’s about making sure that when we scale up these models, our theoretical foundation doesn't crumble because the proposal isn't tuned to match the target distribution’s tail shape.

Jane: They also provide asymptotically sharp lower bounds on convergence rates in things like f-variation and L p-Wasserstein distance.

Tom: This is crucial for engineers because it gives us a hard limit on how fast we can expect a simulation to converge before we even start running the code.

Jane: So, they’re giving us concrete numbers on convergence speed, not just vague hopes that the process will eventually settle down.

Tom: This work really solidifies the link between abstract L-drift theory and practical engineering problems in Bayesian statistics.

Jane: It shows that we can use these conditions to rigorously evaluate the trade-offs when choosing between different sampling methods for complex targets.

Tom: The implication is that instead of blindly running algorithms, we can use this theory to select the right tool based on the target distribution’s moment properties and tail decay rate.

Conclusion: Tom: So we’ve walked through how this paper establishes the L-drift conditions for ergodic averages of Markov chains, focusing on necessary conditions for Central Limit Theorems to hold in heavy-tailed settings.

Jane: Exactly. The main point is that they provide a concrete mathematical structure—the L-drift condition—to verify if an average estimator has the CLT, giving us verifiable rules instead of just assumptions.

Tom: It’s about moving from "maybe this works" to "this specific set of functions guarantees convergence or failure."

Lu: I think what this really changes is how we approach model training when the underlying distribution isn't nice, like it has heavy tails.

Meng: For us, it means when we deploy a sampler, we can use these conditions to predict exactly where our simulation might fail at scale because we’re not picking an algorithm randomly anymore.

Jane: And they gave us explicit criteria for failure—like that integral divergence—which is super useful for debugging simulation results when things go wrong.

Tom: So, the practical application is clear: you check if your target distribution fits these drift conditions before you commit to a massive simulation run.

Lu: It’s a rigorous tool for representation learning and sampling methods, showing us the deep probabilistic requirements needed for stable inference.

Jane: This paper gives us the math to understand why some MCMC algorithms stick in the tails while others explore them faster, which is really insightful for understanding AI behavior in complex spaces.

Tom: So we've looked at the theory, seen how it applies to MCMC comparison, and understood how they tune proposals for infinite variance scenarios.

Lu: It’s a solid piece of representation learning theory that grounds the practical choices we make when building these large models.

Meng: If you’re trying to deploy something that relies on these averages, knowing the convergence rate bounds gives you a much better estimate of your required computational budget.

Jane: So, this paper provides verifiable necessary conditions and sharp metrics on convergence speed for ergodic averages of Markov chains, which is what we need to understand the speed of mixing in complex systems.

MIHA BRESAR, ALEKSANDAR MIJATOVI Ć C, GARETH ROBERTS

math.PR, stat.CO, stat.ML

Submitted: 2025-12-20

Updated: 2026-10-18

Comments: 55 pages; this version has streamlined proofs for the theoretical results, with applications focusing on ULA-type algorithms (MCMC results of the first version are not included); a short YouTube presentation describes our theory and its applications to ULA-type chains with increments of finite and infinite variance, see https://youtu.be/m2y7U4cEqy4

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: Given that any error could cost millions, this synthesis aims for absolute precision regarding both the theoretical framework and its practical applications.

Key concepts

L-drift Conditions: These are the core theoretical framework used to determine if an ergodic CLT holds. They involve a triplet of functions (V, $\varphi$, $\Psi$) that must satisfy specific inequalities derived from submartingale properties. If these conditions are met, convergence is guaranteed; otherwise, the CLT may fail.
L-drift Theory: This theory provides the mathematical machinery to characterize the necessary structure for ergodic convergence. It involves verifying technical steps, such as proving a related process is a submartingale, which ensures that certain bounds on the expected values of functions involving time and distance are maintained during iteration.
Ergodic CLT Failure Conditions: The paper identifies specific mathematical conditions under which the CLT fails for ergodic averages. These include criteria based on the growth rate of functions (e.g., $g(x) \geq |x|^{2s}$) or the divergence of specific integrals involving time and distance, providing explicit failure modes.

Summary

Given that any error could cost millions, this synthesis aims for absolute precision regarding both the theoretical framework and its practical applications.

Here is the detailed, combined summary:

Comprehensive Research Synthesis: Central Limit Theorems for Ergodic Averages of Markov Chains

This research paper addresses a fundamental problem in probability theory and computational statistics: establishing necessary and sufficient conditions for the validity of Central Limit Theorems (CLTs) when applied to ergodic averages of Markov chains defined on general state spaces. The core theoretical machinery developed relies on drift conditions derived from a triplet of functions: a Lyapunov function V, a scalar function phi, and another scalar function.

I. Theoretical Foundation: Necessary Conditions via L-Drift Theory

The central contribution of the paper is the development and application of L-drift conditions, denoted as L(V, phi,). These conditions serve as the necessary framework to characterize when an ergodic CLT holds for Markov chains.

Characterization via Drift Conditions: The main theoretical result (Theorem 2.1) provides sufficient conditions for the failure of the ergodic CLT in terms of these L-drift conditions, formally described by Assumption L(V, phi,). Conversely, the theory establishes the necessary structure required for convergence.
Submartingale Verification: A critical technical step involves proving that a specific process related to the Lyapunov function and the drift functions satisfies certain properties. For instance, Lemma 2.7 verifies Assumption (ii) of the L-drift condition by demonstrating that a specific process is a submartingale: V(x) at most n to infinity Ex[V(X n S T(r))] at most E[V(X S* T(r)))], which is essential for bounding the convergence behavior.
Failure Conditions: The theory provides explicit conditions under which the CLT fails. For instance, Theorem 3.11 shows that the CLT fails for a function g: R d to R if g(x) at least x 2s for all x sufficiently large, where s > v/2 - d/4. Furthermore, Theorem 3.2 links the failure of the CLT to the divergence of a specific integral involving time and distance functions: integral 0 infinity c'c'' t - k/(2(s+2))dt = infinity.

II. Convergence Rates and Metric Bounds

Beyond mere existence, the theory provides sharp bounds on how fast ergodic averages converge to their stationary distribution (pi) in various metrics:

Convergence Rates: The L-drift conditions are used to derive lower bounds on convergence rates in both f-variation distance and Wasserstein distance. These derived bounds match known upper bounds, providing rigorous characterizations of the speed of mixing for different Markov chain families.
Tail Bounds: The paper establishes lower bounds on crucial probabilistic quantities related to the process, such as:

The tails of return times (lambda r).
The tails of the invariant measure (pi).

III. Application to Sampling Algorithms (MCMC)

The theoretical framework is rigorously applied to a broad spectrum of Markov Chain Monte Carlo (MCMC) sampling algorithms, particularly those targeting heavy-tailed distributions, where standard assumptions often break down. The analysis focuses on families denoted as (SIM-ALG).

Algorithm Characterization: The performance and CLT validity of specific samplers are characterized by the decay rate of a function related to the proposal distribution: pi G-1 1(sqrt r) / sqrt r as r to infinity.
Sampler Comparison: The study provides detailed comparisons across several algorithms, including:

Random Walk Metropolis (RWM) samplers (with finite- and infinite-variance proposals).
Metropolis-adjusted and unadjusted Langevin algorithms (MALA/ULA).
The stereographic projection sampler.
The independence sampler.

Key Algorithmic Insights: The analysis reveals crucial performance trade-offs:

Vanilla Samplers: Exhibit stickiness in the tails due to long excursion durations, which is quantified by the decay rate of a specific function.
Single-Jump Samplers: Offer faster tail exploration but are highly sensitive to proposal misspecification.
Metropolis Correction: Is deemed indispensable for eliminating discretization bias when sampling heavy-tailed targets.

Optimal Selection Criteria: The paper provides definitive guidance for algorithm selection based on the target distribution's moment properties:

If the target pi admits d-moments, the stereographic sampler is optimal due to its uniform ergodicity and preservation of CLTs.
If pi admits fewer than d moments (common in high-dimensional problems), the stereographic sampler is outperformed by all other algorithms in (SIM-ALG), except for the independence sampler.
Unadjusted Langevin algorithms with infinite variance proposals are shown to have an invariant distribution tail decay that matches the target distribution for sufficiently small step sizes.

Conclusion

In summary, this paper moves beyond simply stating that a CLT might hold; it provides sharp, verifiable conditions (L-drift conditions) that dictate the validity of the CLT for ergodic averages of Markov chains. This theory is not merely abstract; it furnishes rigorous tools to analyze and compare practical MCMC algorithms used in Bayesian statistics and machine learning when dealing with challenging heavy-tailed targets. The results offer precise metrics for understanding algorithm performance, convergence rates, and the fundamental requirements for probabilistic convergence in these complex systems.

Improvements for AI systems

Sharp CLT Failure Characterization for MCMC Sampling Systems: The improved AI can rigorously determine when an ergodic average estimator fails to satisfy a Central Limit Theorem by checking if a function in L2(π) satisfies the L-drift conditions (L(V, φ, Ψ)). This allows the AI to provide verifiable necessary conditions for CLT validity rather than relying solely on sufficient conditions like [KV86] or classical drift conditions [DGM08].
Optimal Algorithm Selection for Heavy-Tailed Targets: The improved AI can compare diverse MCMC algorithms (RWM, MALA, ULA, SPS, IS) by calculating Jump and convergence rate metrics derived from the L-drift calculus. For a given heavy-tailed target distribution π, the AI can select the algorithm that minimizes the decay rate of the function r 7→ rφ(1/r) as r → ∞ to achieve optimal performance in terms of convergence rates in total variation and Wasserstein metrics.
Adaptive Proposal Tuning for Infinite Variance Sampling: The improved AI can dynamically adjust proposal kernels for algorithms like RWM or MALA by leveraging results from infinite-variance proposals (iv-RWM). This allows the system to select a proposal distribution with an increment law qα that maximizes the exploration of heavy-tailed tails, leading to faster convergence rates and preserve[ing] the validity of the CLT for a broader class of ergodic averages.
Guaranteed Convergence Rate Bounds: The improved AI can provide asymptotically sharp lower bounds on convergence rates in f-variation and L p-Wasserstein distance for ergodic Markov chains. For instance, it can establish bounds such as cn− v/2 −ε ≤ ∥P n(x, ·) − π∥f by applying Theorem 3.2 or Theorem 3.4 directly to the specific Markov chain dynamics and target distribution.

Sources

An adaptive ensemble filter for heavy-tailed distributions: tuning-free inflation and localization