🔢 mathematics

Thinning Operation via the Poisson-Föllmer Process

This paper presents an alternative proof of Yu's Thinning Lemma and the Law of Thin Numbers using a stochastic variational formula for relative entropy, which further yields new convergence rates that extend existing results.

Original authors: Ioannis Kavvadias

Published 2026-08-05

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Ioannis Kavvadias

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Digital Shrink: How Math Counts the Unseen

Imagine you are trying to understand a massive, chaotic crowd of people. In the world of probability and statistics, this crowd is often modeled by something called a Poisson distribution. Think of this as the "gold standard" for counting random events that happen independently, like raindrops hitting a roof, stars twinkling in a patch of sky, or customers walking into a store. It's the mathematical way nature keeps score when things happen at a steady, random average rate.

But what happens when you can't see the whole crowd? What if you only get to see a random sample of them? This is where a concept called thinning comes in. Imagine you have a bucket of marbles, and you decide to keep only a certain percentage of them—say, you flip a coin for every marble and keep it only if it lands heads. You've just "thinned" your collection. In the math world, this operation is a powerful tool. It turns out that if you start with a Poisson distribution and thin it, you still get a Poisson distribution, just with fewer marbles on average. This is a very stable, predictable behavior.

However, most real-world data isn't perfectly Poisson. It's messy. The big question mathematicians have been asking is: If you take a messy, random collection of data and start thinning it (keeping fewer and fewer items), does it eventually smooth out and look like a perfect Poisson distribution? And if so, how fast does that happen? This isn't just about counting marbles; it's about understanding how information flows and how randomness settles down. The paper you are about to read dives deep into this, using a clever new "lens" to measure exactly how quickly messy data becomes orderly, and proving that the speed of this transformation depends on the specific shape of the messiness to begin with.

The Paper's Story: A New Lens on Randomness

This paper, written by Ioannis Kavvadias, is a detective story about how random numbers behave when you shrink them down. The author isn't just re-telling an old story; he's using a brand-new set of tools to prove some old rules and discover some faster ways to measure change.

The Main Characters: Thinning and the "Poisson-Föllmer" Process
The star of the show is the thinning operation. As mentioned, this is like taking a random variable (a number that pops out of a machine) and randomly deleting some of its value. If you have a number representing a crowd size, thinning it is like asking everyone to leave with a 50% chance.

To study this, the author uses a very fancy, invisible machine called the Poisson-Föllmer process. Think of this process as a magical, time-traveling camera. Instead of just looking at the final result of the thinning, this camera records the entire history of how the numbers change as they are slowly thinned out over time. It connects the starting messy number to the final, clean Poisson number through a continuous journey. The author uses this "movie" of the data to calculate something called relative entropy. In plain English, relative entropy is a score that tells you how "different" or "surprising" one distribution is compared to another. A high score means the data is very messy and far from the perfect Poisson ideal; a score of zero means it's perfect.

The Big Findings: Proving the Rules and Finding the Speed
The paper does two main things. First, it gives a fresh, alternative proof of a famous rule called Yu's Thinning Lemma. This lemma basically says that when you thin a random variable, the "messiness" (relative entropy) drops by at least the same fraction as the thinning itself. If you keep 50% of the data, the messiness drops by at least 50%. The author proves this using the Poisson-Föllmer process, showing that the "movie" of the thinning process naturally leads to this result.

But the paper goes further. It asks: Can we do better? Is the drop in messiness exactly 50%, or is it actually more than 50% if the data has a special shape? The author finds that if the starting data has a specific, smooth shape called ultra log-concave (think of a bell curve that is very nicely rounded and doesn't have weird spikes), then the messiness drops even faster than the basic rule predicts. The paper provides a new, sharper formula that quantifies exactly how much faster this happens, depending on the specific details of the starting data.

The Speed of the "Law of Thin Numbers"
The paper also tackles the Law of Thin Numbers. This is a big idea that says if you take many independent copies of a random variable, thin them out just enough, and add them up, the result will eventually look exactly like a Poisson distribution. The paper asks: How fast does this happen?

Using the new tools, the author derives new, precise rates for this convergence.

For general messy data: The paper shows that the messiness drops at a rate proportional to $1/n$ , where $n$ is the number of copies you are adding up.
For the special "ultra log-concave" data: The paper proves that the messiness drops even faster, at a rate proportional to $1/n^2$ . This is a significant improvement. It means that for this specific type of well-behaved data, the path to becoming a perfect Poisson distribution is much smoother and quicker than previously thought.

The author also provides a new, asymptotic estimate (a prediction for what happens when $n$ gets huge) that matches previous results but is derived without needing the strict "ultra bounded" assumptions that earlier papers required. This makes the result more robust and applicable to a wider range of real-world scenarios.

What the Paper Rules Out and What It Confirms
The paper is very careful about what it claims. It confirms that the "Law of Thin Numbers" is true and that the convergence rates are indeed tied to the Fisher information (a measure of how much information the data carries about its own shape). It explicitly rules out the idea that the convergence is always slow; for the special class of ultra log-concave distributions, it proves the convergence is significantly faster.

The paper does not claim to have solved every problem in probability. It does not suggest that all random variables will behave this way, only those that fit the specific mathematical definitions provided. The results are presented as rigorous mathematical proofs, not just simulations or guesses. The author uses the Poisson-Föllmer process as a proven method to derive these inequalities, showing that the "movie" of the thinning process holds the key to unlocking these rates.

Why This Matters
Why should a curious teenager care about counting marbles and shrinking numbers? Because this math is the backbone of how we understand information. Whether it's compressing data on your phone, analyzing traffic patterns, or understanding how signals travel through a noisy network, knowing how fast a messy system settles into a predictable pattern is crucial. This paper gives us a better ruler to measure that speed, especially for systems that are already somewhat well-behaved. It tells us that if our data is "nice" (ultra log-concave), we can expect it to become predictable much faster than we thought, which is great news for anyone trying to make sense of the world's randomness.

Technical Summary: Thinning Operation via the Poisson-Föllmer Process

Problem Statement
The paper addresses the quantitative analysis of the "Law of Thin Numbers," a discrete analog of the Central Limit Theorem where suitable thinning and convolution of independent random variables converge to a Poisson measure. Specifically, the author seeks to improve upon existing convergence rates and provide alternative proofs for fundamental inequalities governing this operation, namely Yu's Thinning Lemma and the Law of Thin Numbers. The work focuses on measuring the rate of convergence using relative entropy ( $H$ ) and Fisher information ( $I$ ), particularly in the context of modified log-Sobolev inequalities for Poisson measures.

Methodology
The central methodological innovation is the application of the Poisson-Föllmer process, a stochastic counting process constructed on a probability space driven by a Poisson point process. This process serves as a discrete analog to the Föllmer process used in Gaussian settings.

Key technical components include:

Stochastic Variational Formula: The author utilizes a variational representation for the relative entropy with respect to a Poisson measure, originally developed by Klartag and Lehec. This formula expresses relative entropy as an infimum over the expected integral of a convex function $\phi$ of a predictable intensity process $\lambda_t$ .
Martingale Properties: The proof strategy relies heavily on the martingale properties of the intensity process $\lambda_t$ associated with the Poisson-Föllmer process. By analyzing the evolution of $\lambda_t$ and applying convexity arguments (specifically Jensen's inequality and properties of submartingales), the author derives bounds on entropy and Fisher information deficits.
Ultra Log-Concavity: The analysis distinguishes between general discrete random variables and those with ultra log-concave densities, utilizing specific bounds on the second derivative of the log-density to derive sharper stability estimates.

Key Contributions and Results

Alternative Proof of Yu's Thinning Lemma:
The paper provides a new proof of Yu's Thinning Lemma, which states that for a random variable $X$ and $a \in [0,1]$ , $H(T_a(X)|\pi_{aT}) \leq aH(X|\pi_T)$ . This is achieved by showing that the relative entropy of the thinned variable corresponds to the expected integral of a submartingale over a shortened time interval, thereby establishing the inequality directly.
Improved Stability Estimates (Theorems 1.2 and 1.5):
Under the assumption that the density function $f$ is ultra log-concave, the author derives explicit lower bounds for the "deficits" in both the Thinning Lemma and the Fisher information inequality.
- Entropy Deficit: The paper establishes that the difference $aH(X|\pi_T) - H(T_a(X)|\pi_{aT})$ is bounded below by a term involving the function $\Theta$ , the parameter $T$ , and the ratio $f(0)/f(1)$ .
- Fisher Information Deficit: Similarly, a lower bound is derived for $aI(X|\pi_T) - I(T_a(X)|\pi_{aT})$ involving the function $\Delta$ . These results complement and extend previous stability estimates found in the literature (e.g., by Aryan, Rivera, and Shenfeld).
Convergence Rates for the Law of Thin Numbers:
The paper derives finite- $n$ bounds and asymptotic estimates for the convergence of the sum of thinned variables to the Poisson measure.
- General Case: For general densities, the convergence rate is bounded by $I(X|\pi_T)/n$ .
- Ultra Log-Concave Case: For ultra log-concave densities, the bound is refined to include a term of order $1/n^2$ , specifically involving the constant $C = f(0)/(f(0)+f(1))$ .
- Asymptotic Estimate: The author recovers the asymptotic estimate $\limsup_{n \to \infty} n^2 H(\sum T_{1/n}(X_i) | \pi_T) \leq (\frac{\text{Var}(X)}{T} - 1)^2$ without requiring the restrictive "ultra bounded" assumption previously used by Harremöes, Johnson, and Kontoyiannis.
Characterization of Equality Cases:
The paper characterizes the conditions under which equality holds in Yu's Thinning Lemma and the Fisher information comparison. It proves that equality holds for some $a \in (0,1)$ if and only if the random variable $X$ follows a Poisson distribution.
Connection to Size-Biasing:
The work re-derives a formula relating the derivative of the relative entropy under thinning to the relative entropy between the thinned variable and its size-biased version, confirming a result by Yu under relaxed assumptions (removing the finite support requirement).

Significance and Claims
The paper claims that the stochastic variational approach via the Poisson-Föllmer process offers a unified and powerful framework for analyzing discrete information-theoretic inequalities. By leveraging the martingale structure of the process, the author obtains:

New Convergence Rates: Explicit, improved rates for the Law of Thin Numbers that depend on the convexity properties of the underlying density.
Extension of Existing Results: The removal of the "ultra bounded" assumption for certain asymptotic estimates, broadening the class of distributions for which these rates are known.
Conceptual Clarity: The method provides a transparent mechanism for proving stability estimates and equality cases, drawing a parallel between discrete Poisson inequalities and their continuous Gaussian counterparts (Gross's inequality and Shannon's inequality).

The author explicitly notes that the work is motivated by the stochastic representation of relative entropy in the Gaussian setting (Lehec) and aims to replicate this success in the discrete Poisson domain, thereby complementing and extending the results of Harremöes, Johnson, Kontoyiannis, and others in the field of discrete information theory.

The Great Digital Shrink: How Math Counts the Unseen

The Paper's Story: A New Lens on Randomness

More like this