📊 statistics

Asymptotic Analysis and Practical Evaluation of Jump Rate Estimators in Piecewise-Deterministic Markov Processes

This paper proposes a unified framework to rigorously compare non-parametric jump rate estimators for Piecewise-Deterministic Markov Processes, establishing new asymptotic results and demonstrating through simulations and real-world *E. coli* data that no single method uniformly outperforms the others.

Original authors: Romain Azaïs, Solune Denis

Published 2026-08-03

📖 9 min read🧠 Deep dive

CC BY 4.0

Original authors: Romain Azaïs, Solune Denis

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

Imagine a world where things move in a predictable, straight line until something sudden and random knocks them off course. Think of a car cruising down a highway at a steady speed, only to be jolted by a pothole or a sudden detour. In the realm of science, this is how many natural systems behave: they follow a smooth, deterministic path until a random event triggers a jump. This is the world of Piecewise-Deterministic Markov Processes (PDMPs). Scientists use these models to understand everything from how data flows across the internet to how bacteria grow and divide. The "jump rate" is the secret heartbeat of these systems; it tells us how likely a jump is to happen at any given moment. If we can estimate this rate accurately, we can predict when a cell will split, when a network might get congested, or when a machine might fail. But here's the catch: calculating this rate is like trying to guess the exact moment a dice will land on a six just by watching the shadow it casts. It's incredibly tricky, and different scientists have been trying to solve it with different tools, often speaking different mathematical languages that make it hard to compare who is doing the best job.

This paper is the ultimate referee match for those different tools. The authors, Romain Azaïs and Solune Denis, decided to stop the confusion by building a unified playing field where they could test the top contenders side-by-side. They didn't just look at the math on paper; they ran thousands of computer simulations using a classic model called the "TCP process" (which mimics how internet traffic behaves) and even tested their methods on real-world data: the growth and division of Escherichia coli bacteria. Their main finding is a bit of a plot twist: there is no single "super-estimator" that wins every time. Depending on the specific conditions and where you are looking in the system, different methods shine. One method might be the champion in the early stages of growth, while another takes the lead later on. The paper suggests that while some methods are more complex and sensitive to tuning, a specific, slightly simpler approach often offers the most reliable balance for one-dimensional systems. Ultimately, the study proves that choosing the right tool depends entirely on the specific landscape you are trying to map, and that no single strategy is universally superior.

The Story of the Jumping Process

To understand what the authors did, let's first meet our main character: the PDMP. Imagine a ball rolling down a smooth, straight ramp. This is the "deterministic" part; if you know where the ball started and how fast it's going, you can predict exactly where it will be in ten seconds. But now, imagine that at random moments, a mischievous hand reaches out and tosses the ball to a new spot on the ramp. This is the "jump." The ball then resumes rolling from its new spot. The "jump rate" is simply the probability of that hand reaching out at any specific moment.

Why do we care? Because nature loves this pattern.

In biology: A cell grows steadily (rolling down the ramp) until it reaches a certain size and suddenly splits in two (the jump).
In technology: Data packets travel smoothly until a network gets too crowded, causing a sudden drop in speed or a reset (the jump).
In insurance: A company collects premiums steadily until a massive claim hits (the jump).

The challenge for scientists is that they can't always see the "hand" tossing the ball. They only see the ball's position over time. They need to work backward to figure out the rules of the game: How likely is a jump to happen right now, given where the ball is? This is called non-parametric estimation. "Non-parametric" means they aren't assuming the jump rate follows a simple, pre-written formula (like a straight line or a curve). Instead, they want to let the data tell them the shape of the jump rate, no matter how weird or wiggly it is.

The Great Estimator Showdown

For years, researchers have developed different mathematical "recipes" to guess this jump rate. Some recipes use the time between jumps, others look at where the ball lands after a jump, and some try to count how often the ball visits certain spots. The problem was that these recipes were written in such different ways that comparing them was like comparing apples, oranges, and a very confused pineapple. You couldn't tell which one was actually better.

Azaïs and Denis decided to put them all in the same kitchen. They standardized the ingredients and the cooking methods so they could taste-test them fairly. They focused on three main strategies:

The "Post-Jump" Detective (Estimator $\hat{\lambda}^\diamond$ ): This method looks at where the ball lands after a jump and how often it lands there. It uses a formula that relates the jump rate to the density of these landing spots.
The "Pre-Jump" Sleuth (Estimator $\hat{\lambda}^\clubsuit$ ): This method is a bit more specialized. It looks at where the ball was right before it jumped. It assumes the jump happens in a very specific, predictable way (like a ball always splitting exactly in half).
The "Oracle" Navigator (Estimator $\hat{\lambda}^\spadesuit$ ): This is the fancy, high-tech method. It tries to find the "perfect" starting point and time to make the best guess. It's like having a map that tells you exactly where to stand to see the jump most clearly. However, in the real world, we don't have this perfect map, so we have to guess the best spot, which makes it harder to use.

The Simulation Lab: The TCP Model

To test these detectives, the authors first used a computer simulation based on the TCP model. This is a famous model for internet traffic where data grows linearly and then gets cut down (fragmented) when it gets too big. They set the "fragmentation" to be deterministic (the data always gets cut by a fixed percentage, like 40%).

They ran the simulation with different amounts of data (1,000 jumps vs. 10,000 jumps) and watched how well each estimator performed. Here is what they found:

No Clear Winner: The most surprising result was that no single method was the best everywhere. In some parts of the state space (the range of possible sizes), the "Post-Jump" detective was better. In other parts, the "Oracle" navigator took the lead.
The Trade-off: The "Oracle" method ( $\hat{\lambda}^\spadesuit$ ) had a theoretical advantage in some areas, but it was very sensitive to the "smoothing parameters" (the knobs you turn to adjust how much you blur the data to make it readable). If you turned the knobs slightly wrong, the results got messy.
The Simple Hero: The "Post-Jump" detective ( $\hat{\lambda}^\diamond$ ) turned out to be the most robust. It didn't require knowing the exact rules of the jump (like the "Pre-Jump" method did) and wasn't as finicky with the knobs as the "Oracle."

They also tested "adaptive" versions of the methods, where the computer automatically adjusts its own knobs. They found that even with these fancy upgrades, the fundamental difference between the methods remained: the formula you use to capture the jump rate matters more than the specific math trick you use to estimate the underlying distributions.

The Real World: Bacteria in a Petri Dish

To make sure their computer results weren't just a digital mirage, they took their methods to the real world. They analyzed data from Escherichia coli bacteria. In this experiment, scientists measured the size of individual bacteria over time as they grew and divided.

The Setup: The bacteria grow exponentially (like a ball rolling down a ramp that gets steeper). When they divide, they split into two. The "jump" is the division event.
The Challenge: Unlike the clean TCP model, real bacteria don't always split exactly in half. Sometimes one daughter cell is bigger than the other. This meant the "Pre-Jump" method (which assumed perfect splitting) was out of the running.
The Results: They compared the "Post-Jump" method and the "Oracle" method.
- At 25°C and 27°C, the "Post-Jump" method ( $\hat{\lambda}^\diamond$ ) produced a model that perfectly matched the real distribution of cell sizes. The "Oracle" method, however, predicted too many small cells, suggesting it was overestimating the jump rate.
- At 37°C (a warmer temperature), the situation flipped. The "Oracle" method actually fit the data better than the "Post-Jump" method.

This confirmed their earlier finding: context is king. The "best" method depends on the temperature, the amount of data available, and the specific behavior of the system.

The Takeaway

So, what does this all mean for a curious teenager?

There is no magic bullet. In science, especially when dealing with complex, random systems, there is rarely one "best" tool that works for every job.
Simplicity often wins. While the fancy "Oracle" method looks cool and has great theoretical properties, the simpler "Post-Jump" method was often more reliable and easier to use, especially when data was scarce.
The map matters more than the compass. The authors found that the formula used to connect the data to the jump rate was more important than the specific math technique used to estimate the numbers. Whether you used a simple kernel or a complex adaptive projection, the results followed the same pattern.

The paper concludes that for one-dimensional systems (like a single line of growth), the "Post-Jump" estimator ( $\hat{\lambda}^\diamond$ ) is likely the safest and most practical choice for scientists. However, the authors suggest that the future might lie in combining these methods—using the strengths of each one depending on where you are in the system. It's a reminder that in the messy, beautiful world of nature, the best solution is often a flexible one, ready to adapt to the unexpected jumps life throws at us.

Technical Summary: Asymptotic Analysis and Practical Evaluation of Jump Rate Estimators in Piecewise-Deterministic Markov Processes

Problem Statement
Piecewise-deterministic Markov processes (PDMPs) provide a framework for modeling systems with deterministic trajectories interrupted by random jumps. A critical challenge in the statistical analysis of PDMPs is the non-parametric estimation of the local jump rate, $\lambda$ . While various non-parametric methods have been developed in recent literature, they rely on distinct theoretical frameworks, observation schemes, and estimator constructions (e.g., recursive vs. non-recursive kernels, different bandwidth selections, and varying assumptions on transition kernels). This heterogeneity complicates direct theoretical comparisons of convergence rates and asymptotic variances. Furthermore, it remains unclear whether a single method uniformly outperforms others across different regions of the state space or under different model specifications.

Methodology
The authors propose a unified framework to standardize and consolidate state-of-the-art approaches for estimating the jump rate in one-dimensional PDMPs. The study assumes the observation of the embedded Markov chain (post-jump locations and inter-jump times) over a long time horizon, with a known deterministic flow $\Phi$ .

The paper focuses on three primary strategies for capturing the jump rate, all formulated as quotient estimators involving invariant measures:

Strategy $\clubsuit$ (based on Formula 8): Derived from the work of Krell (2016) and Krell & Schmisser (2021), this approach assumes deterministic fragmentation ( $Q(\cdot|x) = \delta_{h(x)}$ ). It estimates $\lambda(x)$ using the ratio of the invariant density of post-jump locations composed with the fragmentation function, $\mu(h(x))$ , and an empirical probability term. The estimator is denoted $\hat{\lambda}^\clubsuit_n$ .
Strategy $\diamondsuit$ (based on Formula 4): Generalizing the approach of Fujii (2013) and Krell & Schmisser (2021), this method does not assume deterministic transitions. It estimates $\lambda(x)$ using the ratio of the invariant density of pre-jump locations, $\mu_-(x)$ , and an empirical probability term. The estimator is denoted $\hat{\lambda}^\diamondsuit_n$ .
Strategy $\spadesuit$ (based on Formula 3): Based on the work of Azaïs & Muller-Gueudin (2016), this approach estimates $\lambda(\Phi(t|\xi))$ as a ratio of the joint invariant density of (post-jump location, inter-jump time) to the marginal survival function. It involves an "oracle" selection of the argument $(\xi, t)$ that maximizes the denominator to minimize variance, denoted $\check{\lambda}^\spadesuit_n$ , and a practical version $\hat{\lambda}^\spadesuit_n$ using estimated arguments.

To ensure comparability, the authors restrict the analysis to non-recursive kernel estimators for all strategies. They establish new theoretical results regarding the consistency and asymptotic normality of $\hat{\lambda}^\clubsuit_n$ and $\hat{\lambda}^\diamondsuit_n$ under ergodicity and Lipschitz mixing conditions. These results are then compared against the existing asymptotic theory for $\check{\lambda}^\spadesuit_n$ .

Key Contributions

Unified Theoretical Framework: The paper demonstrates that several distinct methods in the literature implicitly rely on the same underlying quotient formulas (specifically Formulas 3, 4, and 8). It provides a rigorous comparison of these strategies within a single statistical setting.
New Asymptotic Results: The authors prove the strong consistency and asymptotic normality for the estimators $\hat{\lambda}^\clubsuit_n$ and $\hat{\lambda}^\diamondsuit_n$ . They derive explicit expressions for the asymptotic variances, allowing for a quantitative comparison of efficiency.
Numerical Validation on TCP Model: Using the Transmission Control Protocol (TCP) model as a representative one-dimensional PDMP, the authors conduct extensive numerical simulations. They compare the theoretical asymptotic variances with empirical estimation errors across different sample sizes and state space regions.
Real-World Application: The methods are applied to real-world single-cell data from Escherichia coli to model cell growth and division dynamics. The study fits a PDMP model to the data and evaluates the estimators by comparing the resulting continuous-time invariant distributions against empirical cell size distributions.

Results

No Uniform Dominance: Theoretical analysis and numerical simulations reveal that no single estimator uniformly outperforms the others. The relative performance depends on the specific region of the state space and the model parameters (e.g., the fragmentation factor $\kappa$ in the TCP model).
Variance Comparison: In the TCP model, the asymptotic variances of the estimators intersect. For instance, the oracle estimator $\check{\lambda}^\spadesuit_n$ (and its practical counterpart) may exhibit lower variance in certain regions (e.g., smaller cell sizes) but higher variance in others compared to $\hat{\lambda}^\diamondsuit_n$ .
Bias and Sensitivity: The estimator $\hat{\lambda}^\spadesuit_n$ (and $\check{\lambda}^\spadesuit_n$ ) is found to be more sensitive to the choice of smoothing parameters (bandwidths) because it requires smoothing in both space and time, unlike the other methods which only require spatial smoothing. This complexity can lead to higher bias in finite samples.
Adaptive Projection: The study extends the comparison to adaptive projection-based estimators for the invariant densities. The results indicate that the trend in estimation error (which estimator performs better where) remains consistent regardless of whether kernel or adaptive projection methods are used to estimate the underlying invariant measures.
Real Data Findings: In the E. coli application, the estimator $\hat{\lambda}^\diamondsuit_n$ (and its adaptive variant) generally provided a better fit to the empirical cell size distribution at lower temperatures (25°C, 27°C), while $\hat{\lambda}^\spadesuit_n$ performed better at 37°C. This suggests that the optimal choice depends on the data characteristics and sample size.

Significance and Claims
The paper claims to provide the first rigorous, both theoretical and numerical, comparison of the strategies used to capture the jump rate in PDMPs. By establishing that the performance of these estimators is not uniform, the authors argue that the choice of method should be guided by the specific characteristics of the problem and the region of interest.

The authors modestly conclude that, given the current state of knowledge and the practical difficulties in implementing the more complex $\hat{\lambda}^\spadesuit_n$ (due to its sensitivity to bandwidths and argument selection), the estimator $\hat{\lambda}^\diamondsuit_n$ (or its adaptive variant) is likely the most suitable choice for one-dimensional PDMPs, despite potentially having higher variance in specific regions. They suggest that a promising future direction involves aggregating these estimators to leverage the strengths of each approach at different points in the state space.

The work also highlights the importance of one-dimensional PDMPs in understanding more complex multidimensional processes, noting that while general non-parametric approaches for multidimensional PDMPs are limited to Formula 3, specific structures in multidimensional models might allow for the derivation of tractable formulas analogous to those used in one dimension.

The Story of the Jumping Process

The Great Estimator Showdown

The Simulation Lab: The TCP Model

The Real World: Bacteria in a Petri Dish

The Takeaway

More like this