Inferring the dynamics of underdamped stochastic systems

The paper presents a principled framework called Underdamped Langevin Inference (ULI) that robustly infers the dynamics of underdamped stochastic systems from noisy, discrete-time experimental data, successfully applying to complex biological systems like migrating cells and animal flocks.

The Gist

Imagine you are watching a chaotic dance floor. There are hundreds of people (particles, cells, or animals) moving around. Sometimes they move smoothly, but often they bump into each other, get pushed by invisible hands, or stumble due to the crowd. You want to figure out the rules of the dance: Why do they move that way? What forces are pushing them? What is the "noise" or chaos making them stumble?

In physics, this is called inferring the dynamics of a system. For a long time, scientists could only figure out these rules if the dancers were moving very slowly and heavily (like walking through molasses). This is called the "overdamped" regime, and we have good tools for it.

But many real-world systems—like a bird flocking, a cell crawling, or a fish swimming—are light and fast. They have momentum. They don't stop instantly when they hit a wall; they bounce or slide a bit. In physics, we call this underdamped.

The problem is: We didn't have a good way to figure out the rules for these fast, bouncy systems.

Here is the story of how the authors of this paper solved that puzzle, explained simply.

The Problem: The "Blurry Photo" Effect

Imagine you are trying to guess how fast a car is going and how hard the driver is braking, but you only have a series of blurry photos taken every second.

1. The Discreteness Problem: To know how hard the driver is braking (acceleration), you need to look at how the speed changes between photos. But because the photos are taken at distinct moments (not a smooth video), the math gets messy. If you try to calculate the "braking force" using standard math, you get a wrong answer. It's like trying to measure the slope of a hill by looking at two points that are far apart; you miss the bumps in between.
2. The Measurement Error Problem: Real cameras aren't perfect. Your photos might be slightly blurry, or the GPS might be off by a few inches. In slow systems, a little blur doesn't matter much. But in fast, bouncy systems, trying to calculate acceleration from blurry data is like trying to hear a whisper in a hurricane. The tiny errors get amplified into massive, wrong conclusions.

For years, scientists tried to fix this, but their methods either required knowing the answer beforehand or failed completely when the data was noisy.

The Solution: "Underdamped Langevin Inference" (ULI)

The authors created a new method called ULI. Think of it as a super-smart detective that knows exactly how to handle blurry photos and fast-moving suspects.

Here is how ULI works, using a few analogies:

1. The "Smooth Guess" (Basis Functions)

Instead of trying to guess the rules for every single point on the dance floor, ULI assumes the rules follow a general pattern (like a polynomial curve or a wave). It's like saying, "The dance moves probably follow a smooth rhythm, not random chaos." This helps the detective ignore the tiny, meaningless jitter and focus on the big picture.

2. The "Bias Correction" (Fixing the Math)

The authors realized that the standard way of calculating acceleration from discrete steps creates a specific, predictable error (a "bias"). It's like a scale that always adds 5 pounds to your weight.

• Old Method: Weighs you, sees 150 lbs, and says, "You weigh 150 lbs." (Wrong, because the scale is broken).
• ULI Method: Weighs you, sees 150 lbs, knows the scale adds 5 lbs, and subtracts it. "You actually weigh 145 lbs."
  They derived a special mathematical formula that automatically subtracts this error, even when the data is sampled at discrete times.

3. The "Noise Filter" (Handling Blurry Photos)

This is the magic trick. When the data is blurry (measurement error), the standard math explodes. ULI uses a clever trick: instead of looking at just two points to guess the speed, it looks at three or four points and averages them in a very specific way.

• Imagine trying to guess the speed of a runner. If you look at where they were 1 second ago and 2 seconds ago, a tiny mistake in measuring their position makes your speed guess wild.
• ULI looks at the position 2 seconds ago, 1 second ago, now, and 1 second in the future. By averaging these in a specific "symmetric" pattern, the random errors cancel each other out, leaving the true speed visible.

What Did They Prove?

The authors tested their detective (ULI) on three very different "dance floors":

1. A Single Cell: They looked at a cancer cell migrating. Even though the cell's path was short and noisy, ULI figured out the exact "deterministic flow" (the cell's internal drive to move) and the random noise pushing it around.
2. A Complex Oscillator: They simulated a system that moves in loops (like a heartbeat or a pendulum with friction). ULI correctly identified the non-linear rules governing the loops, even when the data was noisy.
3. A Flock of Birds: This was the big test. They simulated a flock of 27 birds. The birds were pushing each other, aligning their directions, and avoiding collisions. This is a massive, high-dimensional puzzle. ULI successfully figured out the rules of attraction (staying together) and alignment (flying in the same direction) just by watching the flock move.

Why Does This Matter?

Before this paper, if you had data on fast-moving, noisy systems (like cells, animals, or even financial markets), you couldn't reliably figure out the laws governing them. You could only guess.

Now, with ULI, scientists can:

• Understand how individual cells make decisions.
• Predict how animal groups will react to threats.
• Model complex physical systems that were previously too "noisy" to study.

In short: The authors built a mathematical "noise-canceling headphone" for physics. It allows us to hear the true melody of the universe, even when the world around it is screaming with static and chaos.

Technical Summary

1. Problem Statement

The paper addresses the challenge of inferring the governing equations of motion for complex physical systems that exhibit underdamped stochastic dynamics. Unlike overdamped systems (where inertia is negligible and dynamics are first-order), underdamped systems (e.g., migrating cells, animal flocks, dust in plasma) are governed by second-order Langevin equations where velocity and acceleration play critical roles.

Key Challenges Identified:

• Discretization Bias: Experimental data is sampled at discrete time intervals ($\Delta t$). Standard estimators for force and noise, which work for overdamped systems, fail for underdamped systems. When calculating acceleration as a discrete second derivative, the product of fluctuating terms in the velocity and acceleration estimators creates a systematic bias of order $O(\Delta t^0)$. This bias does not vanish even as $\Delta t \to 0$.
• Measurement Errors: Real-world data contains time-uncorrelated measurement noise ($\eta$). In underdamped inference, where the signal is differentiated twice, measurement errors lead to a divergent bias of order $O(\Delta t^{-3})$. This makes standard inference impossible with realistic data unless the measurement error is vanishingly small.
• High Dimensionality: Many systems (e.g., animal flocks) involve high-dimensional state spaces, making inference computationally difficult without exploiting symmetries.

2. Methodology: Underdamped Langevin Inference (ULI)

The authors propose a principled framework called Underdamped Langevin Inference (ULI) to derive unbiased estimators for the force field $F_\mu(x, v)$ and noise amplitude $\sigma_{\mu\nu}(x, v)$ from discrete, noisy trajectories.

A. Mathematical Framework

The system is modeled by the Itô stochastic differential equation:
$$\dot{x}_\mu = v_\mu$$
$$\dot{v}_\mu = F_\mu(x, v) + \sigma_{\mu\nu}(x, v)\xi_\nu(t)$$
where $\xi$ is Gaussian white noise. The goal is to infer $F$ and $\sigma$ from an observed trajectory $y(t) = x(t) + \eta(t)$.

The method approximates the force and noise fields as linear combinations of basis functions $b_\alpha(x, v)$ (e.g., polynomials, Fourier modes):
$$F_\mu(x, v) \approx F_{\mu\alpha} \hat{c}_\alpha(x, v)$$
$$\sigma^2_{\mu\nu}(x, v) \approx \sigma^2_{\mu\nu\alpha} \hat{c}_\alpha(x, v)$$
where $\hat{c}_\alpha$ are empirical orthonormal basis functions derived from the trajectory.

B. Derivation of Unbiased Estimators

The core innovation lies in rigorously deriving correction terms to eliminate biases caused by discretization and measurement noise.

1. Discretization Correction (Force Field):
   The naive estimator $\langle \hat{a}_\mu \hat{c}_\alpha \rangle$ is biased. By performing an Itô-Taylor expansion, the authors identify a bias term proportional to the derivative of the basis function with respect to velocity.
   The corrected estimator is:
   $$\hat{F}_{\mu\alpha} = \langle \hat{a}_\mu \hat{c}_\alpha(x, \hat{v}) \rangle - \frac{1}{6} \langle \sigma^2_{\mu\nu}(x, \hat{v}) \partial_{v_\nu} \hat{c}_\alpha(x, \hat{v}) \rangle$$
   (Note: The prefactor changes to $1/2$ when using specific symmetric velocity definitions to handle measurement errors).

2. Measurement Error Correction:
   To handle measurement noise $\eta(t)$, the authors derive specific constructions for the conditioning variables (position and velocity) used in the basis functions.

   • Force Estimator: Uses a local average position $\bar{y}(t) = \frac{1}{3}(y_{t-\Delta t} + y_t + y_{t+\Delta t})$ and a symmetric velocity $\hat{v} = \frac{y_{t+\Delta t} - y_{t-\Delta t}}{2\Delta t}$. This choice cancels the leading-order divergent bias terms ($O(\Delta t^{-3})$).
   • Noise Estimator: Requires a more complex construction using four time points and a specific linear combination of increments to cancel bias terms arising from multiplicative noise and measurement errors.

3. Basis Selection via Partial Information:
   To determine the optimal number of basis functions ($n_b$) without overfitting, the authors utilize a "partial information" metric. They calculate the information contribution of each basis function added to the set, allowing for the identification of the most relevant terms in the force field directly from data.

3. Key Contributions

• Theoretical Breakthrough: First rigorous derivation of unbiased estimators for underdamped Langevin equations that converge correctly in the limit of discrete sampling and non-zero measurement error.
• Operational Framework (ULI): A complete, ready-to-use algorithm that corrects for both discretization bias and measurement noise divergence.
• Handling Multiplicative Noise: The method successfully recovers state-dependent (multiplicative) noise amplitudes, which are common in biological and active matter systems.
• Data Efficiency: The framework enables "single-cell profiling," allowing the inference of governing equations from short, individual trajectories rather than requiring large ensembles of averaged data.

4. Results and Validation

The authors validated ULI on several benchmarks:

• Stochastic Damped Harmonic Oscillator:
  • Demonstrated that naive estimators yield a biased friction coefficient (recovering only $2/3$ of the true value), while ULI recovers the exact value.
  • Showed that ULI remains accurate even with significant measurement noise, whereas standard methods fail completely.
• Non-Linear Dynamics (Van der Pol Oscillator):
  • Successfully inferred the non-linear force field $F(x,v) = \kappa(1-x^2)v - x$ and multiplicative noise terms.
  • Validated that the method works with different basis sets (polynomials and Fourier modes) and in high dimensions ($d=1$ to $6$).
• Experimental Data (Migrating Cells):
  • Applied ULI to experimental trajectories of human breast cancer cells (MDA-MB-231) in two-state confinements.
  • Successfully inferred the deterministic flow field from single-cell trajectories, predicting cell behavior consistent with experimental observations.
• Collective Systems (Viscek-like Flocks):
  • Inferred interaction kernels (cohesion and alignment) for a 3D flock of 27 particles.
  • Despite the "curse of dimensionality" (6N degrees of freedom), exploiting particle exchange symmetry allowed accurate recovery of the interaction forces and noise structure. Simulations using the inferred model reproduced the flocking behavior qualitatively.

5. Significance

This work provides a critical tool for the physics of active matter and biological systems.

• Bridging Theory and Experiment: It resolves the long-standing issue of how to extract physical laws from noisy, discrete experimental data in underdamped regimes.
• New Insights into Active Matter: By enabling the inference of non-linear forces and multiplicative noise from single trajectories, it allows researchers to uncover the specific physical mechanisms (e.g., self-propulsion, alignment rules) governing complex systems like cell migration and animal swarms.
• Generalizability: The framework is applicable to any system described by second-order stochastic differential equations, ranging from molecular motors to macroscopic animal groups.

The authors have made a Python package implementing ULI publicly available, facilitating its adoption across scientific disciplines.