Classification of diffusion processes in dimension $d$ via the Carleman approach with applications to models involving additive, multiplicative or square-root noises

This paper applies the Carleman approach to classify diffusion processes in dimensions $d=1$ and $d=2$ with polynomial drift and diffusion terms (including additive, multiplicative, and square-root noises) by analyzing the block structure of the resulting infinite-dimensional linear system governing the moments, thereby identifying models with diagonal or triangular spectral decompositions that facilitate the study of their dynamics and steady states.

The Gist

Imagine you are trying to predict the future of a chaotic system, like a swarm of bees, a stock market, or the growth of bacteria in a petri dish. These systems are governed by diffusion processes: things that move randomly but are also pushed by underlying forces (like gravity, wind, or economic trends).

The problem is that these systems are non-linear. This means that a small change today can lead to a massive, unpredictable explosion tomorrow. Calculating the average behavior of such a system is usually a mathematical nightmare.

This paper introduces a clever trick called the Carleman Approach to solve this problem. Here is the breakdown in simple terms:

1. The Core Idea: Turning a Messy Knot into a Straight Line

Imagine you have a tangled ball of yarn (the non-linear system). It's hard to pull apart. The Carleman approach says: "Don't try to untangle the yarn directly. Instead, imagine every possible knot, loop, and twist as a separate, distinct string."

By doing this, you replace one messy, tangled equation with an infinite library of simple, straight lines (a linear system).

• The Old Way: Trying to solve the messy, twisting path of a single particle.
• The Carleman Way: Tracking the "average height," the "average spread," the "average wobble," and every other statistical detail of the crowd all at once, but treating them as if they were simple, straight lines that don't interact in a messy way.

2. The "Carleman Matrix": The Master Control Panel

Once you've turned the messy system into this library of straight lines, you can organize them into a giant spreadsheet called the Carleman Matrix.

Think of this matrix as a traffic controller for the system's statistics.

• It tells you how the "average" moves.
• It tells you how the "spread" (variance) moves.
• It tells you how the "skew" (asymmetry) moves.

The genius of this paper is realizing that this giant spreadsheet isn't just a random mess. It has a natural structure. It can be chopped up into blocks based on the "total degree" of the statistics (how complex the calculation is).

3. The Three Types of Traffic Patterns

The author classifies these systems based on how their "traffic controller" (the matrix) is arranged. This is like looking at the layout of a city's road network:

• The Diagonal City (Geometric Brownian Motion):

  • Analogy: Imagine a city where every street runs perfectly parallel. No intersections.
  • What it means: The different statistics (average, spread, etc.) don't talk to each other. They evolve independently. This is the simplest case, like a stock price that just drifts up or down randomly (Geometric Brownian Motion). It's easy to solve.

• The Staircase City (Pearson Diffusions):

  • Analogy: Imagine a city where you can only drive forward or turn left, but never right. You can go from a high floor to a lower floor, but not the other way around.
  • What it means: The complex statistics depend on the simpler ones, but the simple ones don't depend on the complex ones. It's like a staircase. You can solve the bottom step first, then use that answer to solve the next step up, and so on.
  • Real-world examples: This covers many famous models like the Ornstein-Uhlenbeck process (a particle bouncing back to a center point) and models with power-law tails (where rare, huge events happen more often than you'd expect, like in earthquakes or financial crashes).

• The Upside-Down Staircase:

  • Analogy: The reverse of the staircase. You can only go up, never down.
  • What it means: This happens in specific models where the forces push the system toward infinity in a very specific way.

4. The "Noise" Factor

The paper focuses on systems with different types of "noise" (randomness):

• Additive Noise: Like rain falling on a car. The rain is the same intensity regardless of how fast the car is going.
• Multiplicative Noise: Like wind pushing a sailboat. The faster the boat goes, the harder the wind pushes it. The randomness scales with the system.
• Square-Root Noise: A special type of randomness often found in biology (population growth) where the noise gets weaker as the population gets smaller, preventing the population from going negative.

The author shows that for many of these real-world scenarios, the "Traffic Controller" (the matrix) falls into one of the neat patterns (Diagonal or Staircase), making them solvable.

5. Why This Matters: The "Power Law" Secret

One of the most exciting findings is about steady states (what the system looks like after a long time).

In many of these "Staircase" models, the system settles into a state where rare, huge events are much more common than in a normal bell curve.

• Analogy: In a normal city, most people are average height. In a "Power Law" city, there are a few giants and a few dwarfs, and the distribution follows a specific mathematical rule.
• The paper identifies exactly when and why these systems develop these heavy tails. It tells us that if the "traffic controller" has a specific triangular shape, the system will naturally produce these power-law distributions (like the Kesten, Fisher-Snedecor, or Student distributions).

Summary

This paper is a user manual for chaos. It takes complex, non-linear systems that involve random noise and shows us how to reorganize them into a giant, structured spreadsheet. By looking at the shape of this spreadsheet (is it diagonal? is it a staircase?), we can instantly know:

1. Is the system solvable?
2. Will it settle down to a stable state?
3. Will it produce rare, massive events (power laws)?

It's like having a decoder ring that turns the chaotic noise of the universe into a clear, readable map.

Technical Summary

1. Problem Statement

The paper addresses the challenge of analyzing the dynamics of stochastic differential equations (SDEs) in $d$ dimensions, specifically focusing on the time evolution of statistical moments and correlations. While the Carleman embedding method is well-established for deterministic non-linear systems (converting them into infinite-dimensional linear systems), its application to stochastic processes has been limited.

The specific problem tackled is the classification of diffusion processes where the drift (forces) and diffusion coefficients are polynomials. The goal is to determine under which conditions the infinite-dimensional linear system governing the moments $\mathbb{E}[x_1^{n_1}(t) \dots x_d^{n_d}(t)]$ admits a simple spectral decomposition (e.g., diagonal, triangular, or block-triangular structures). Such structures allow for explicit solutions or iterative calculations of moments, which are crucial for understanding steady states, power-law tails, and large deviation properties.

2. Methodology

The author employs the Carleman Embedding technique, extended to stochastic systems, combined with a rigorous block-decomposition analysis.

• Carleman Embedding for SDEs:

  • The system is defined by $d$ SDEs with polynomial Ito forces $F_j(\vec{x})$ and polynomial diffusion matrix elements $D_{ji}(\vec{x})$.
  • The dynamics of the moments $m_t(n_1, \dots, n_d) = \mathbb{E}[\prod x_i^{n_i}]$ are governed by a linear system $\partial_t |m_t\rangle = \mathcal{M} |m_t\rangle$, where $\mathcal{M}$ is the infinite-dimensional Carleman matrix.
  • The generator of the process, $\mathcal{L}$, acts on monomials. Since the coefficients are polynomials of degree up to 2, the action of $\mathcal{L}$ on a monomial of degree $n$ produces terms of degrees $n-2, n-1, n,$ and $n+1$.

• Block Decomposition:

  • The Carleman matrix $\mathcal{M}$ is decomposed into blocks based on the global degree $n = \sum n_i$.
  • $\mathcal{M} = \mathcal{M}^{[-2]} + \mathcal{M}^{[-1]} + \mathcal{M}^{[0]} + \mathcal{M}^{[1]}$, where the superscript indicates the change in global degree ($\Delta n$).
  • Key Insight: The structure of these blocks depends on the specific types of noise (additive, multiplicative, square-root) and the polynomial order of the forces.
    • $\mathcal{M}^{[-2]}$ arises from constant diffusion (additive noise).
    • $\mathcal{M}^{[-1]}$ arises from linear diffusion (square-root noise) or constant forces.
    • $\mathcal{M}^{[0]}$ arises from linear forces and quadratic diffusion (multiplicative noise).
    • $\mathcal{M}^{[1]}$ arises from quadratic forces.

• Spectral Analysis:

  • The paper analyzes when $\mathcal{M}$ becomes block-diagonal, block-lower-triangular, or block-upper-triangular.
  • Triangular matrices allow for iterative solution of moments (solving lower degrees first).
  • Diagonal matrices imply independent evolution of moments or simple exponential growth/decay.
  • The eigenvalues of the full matrix are often determined solely by the diagonal blocks $\mathcal{M}^{[0]}$ in these simplified cases.

3. Key Contributions

A. General Framework for Dimension $d$

The paper establishes a unified formalism for $d$-dimensional diffusions with polynomial coefficients. It introduces a "fully-diagonal" assumption for the diffusion matrix (where noise acts independently on each coordinate) and derives the explicit matrix elements of the Carleman blocks in terms of the polynomial coefficients.

B. Classification of Soluble Models

The author classifies models based on the noise types per coordinate (Additive, Multiplicative, Square-root) and the resulting structure of the Carleman matrix:

1. Block-Diagonal ($\mathcal{M} = \mathcal{M}^{[0]}$): Occurs for Geometric Brownian Motion (multiplicative noise only, linear forces). The moments evolve independently.
2. Block-Lower-Triangular ($\mathcal{M} = \mathcal{M}^{[-2]} + \mathcal{M}^{[-1]} + \mathcal{M}^{[0]}$): Occurs for Pearson diffusions (e.g., Ornstein-Uhlenbeck, Square-Root, Kesten, Fisher-Snedecor, Student processes). Moments of degree $n$ depend only on moments of degree $n, n-1, n-2$.
3. Block-Upper-Triangular ($\mathcal{M} = \mathcal{M}^{[0]} + \mathcal{M}^{[1]}$): Occurs for models with multiplicative noise and quadratic forces (e.g., Lotka-Volterra with noise).

C. Dimension $d=1$ Results

The paper recovers and generalizes known 1D results:

• Geometric Brownian Motion: Diagonal Carleman matrix.
• Pearson Family: Lower-triangular matrix. The eigenvalues are $E_n = n f^{[1]} + n(n-1)D^{[2]}$.
• Power-Law Tails: The paper explicitly links the convergence of moments to the sign of the eigenvalues. If the eigenvalues become positive for high orders $n$, the steady state exhibits power-law tails ($\rho(x) \sim x^{-(1+\mu)}$). The critical exponent $\mu$ is derived from the condition $E_\mu = 0$.
• Specific distributions identified: Gamma (Square-root noise), Inverse-Gamma (Kesten, multiplicative noise), Fisher-Snedecor (mixed square-root/multiplicative), and Student (mixed additive/multiplicative).

D. Dimension $d=2$ Results

The analysis extends to 2D interacting systems:

• Independent Geometric Brownian Motions: The matrix is block-diagonal; eigenvalues are sums of 1D eigenvalues.
• Interacting Systems (Lotka-Volterra, Lorenz-type):
  • The paper introduces a change of variables to the ratio $R(t) = x_2(t)/x_1(t)$ and the first coordinate $x_1(t)$.
  • It is shown that the ratio $R(t)$ often converges to a steady state (equilibrium or non-equilibrium) independent of the scale of $x_1$.
  • The finite-time Lyapunov exponents $\lambda_1(T)$ and $\lambda_2(T)$ are analyzed. In interacting cases, they share a common asymptotic value determined by the steady-state average of the ratio dynamics.
  • The Scaled Cumulant Generating Function (SCGF) for the moments is derived, linking the Carleman eigenvalues to large deviation theory.

E. Extension to $d > 2$

Appendix D outlines the strategy for $d > 2$ by reducing the system to $d-1$ ratios and one scale variable. While the steady state for the ratios becomes a complex $(d-1)$-dimensional problem, the structure of the Carleman matrix and the definition of the Lyapunov exponents remain consistent.

4. Key Results

1. Spectral Simplicity: The paper identifies that the "simplest" spectral decompositions occur when the Carleman matrix is triangular. This allows the eigenvalues of the infinite system to be read directly from the diagonal blocks, which depend only on the linear parts of the forces and the quadratic parts of the diffusion.
2. Moment Convergence Criteria: A rigorous criterion is established for the existence of steady-state moments. For a process with power-law tails, moments of order $n$ exist if and only if $n < \mu$, where $\mu$ is the root of the eigenvalue equation $E_n = 0$.
3. Universality of Power-Law Exponents: In interacting 2D systems (e.g., Kesten, Fisher-Snedecor, Student), the interaction forces a common power-law exponent $\mu$ for all variables, even if they would have different exponents in isolation.
4. Large Deviation Connection: The eigenvalues of the Carleman matrix are identified as the Scaled Cumulant Generating Functions (SCGF) for the finite-time Lyapunov exponents, providing a direct bridge between moment dynamics and large deviation theory.

5. Significance

• Unification: The paper provides a unified algebraic framework (Carleman embedding) to treat a wide variety of stochastic models (Ornstein-Uhlenbeck, Kesten, Fisher-Snedecor, etc.) that are usually studied individually.
• Analytical Tractability: By identifying the block-triangular structures, the paper offers a systematic way to solve for moments in models that are otherwise difficult to handle analytically, especially in higher dimensions.
• Physical Insight: The analysis of the ratio dynamics in 2D systems reveals that interactions often synchronize the scaling behavior (Lyapunov exponents) of the system components, leading to universal power-law tails.
• Methodological Bridge: It connects the classical Carleman method (deterministic) with modern stochastic analysis and large deviation theory, offering tools relevant to non-equilibrium statistical mechanics, population dynamics, and financial mathematics.

In conclusion, this work serves as a comprehensive guide for applying the Carleman approach to stochastic diffusion processes, classifying them by their noise structure and demonstrating how this classification dictates the solvability of their moment dynamics and the nature of their steady states.