💰 quantitative finance

Signature SDEs from an affine and polynomial perspective

This paper demonstrates that signature stochastic differential equations can be characterized as affine and polynomial processes on the extended tensor algebra, enabling the derivation of explicit formulas for their Fourier-Laplace transforms and expected values via converging power series solutions to Riccati and linear ODEs, thereby providing a universal framework for path-dependent Itô-diffusions with analytically tractable laws.

Original authors: Christa Cuchiero, Sara Svaluto-Ferro, Josef Teichmann

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Christa Cuchiero, Sara Svaluto-Ferro, Josef Teichmann

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 you are trying to predict the future path of a wandering drunkard (a mathematical concept called a "Brownian motion" or random walk) who is walking through a city. Usually, to predict where they will end up, you only need to know where they are right now. But what if the city has strange rules where the drunkard's next step depends not just on their current location, but on their entire history of where they've been? Maybe they turn left if they've walked in a circle before, or speed up if they've zig-zagged a lot.

This is the problem of path dependence. It's incredibly hard to solve because the "state" of the system isn't just a number; it's the whole story of the journey.

This paper introduces a clever mathematical trick to solve these complex, history-dependent problems. Here is the breakdown using simple analogies:

1. The "Signature": A Unique Fingerprint for Paths

The authors use a tool called the Signature. Think of the signature of a path as a unique mathematical fingerprint.

Just as a fingerprint captures the unique ridges and swirls of a finger, the signature captures the unique twists, turns, and loops of a path.
It works by taking all the little steps the path took and combining them in a specific way (iterated integrals).
The Magic: The signature contains all the information about the path. If you know the signature, you know the path. Even better, the signature turns complex, non-linear history into a set of numbers that behave very nicely mathematically.

2. The "Linearization": Turning a Curve into a Straight Line

The core problem in finance and physics is that these history-dependent paths are usually non-linear and messy. They are like a tangled ball of yarn.

The authors show that if you look at the signature of the path instead of the path itself, the messy, tangled ball of yarn suddenly becomes a straight line (or a very simple, predictable curve).
They call these "Signature SDEs" (Stochastic Differential Equations). By switching to the signature, they transform a chaotic, history-dependent system into a system that follows simple linear rules.

3. The "Affine and Polynomial" Tools: The Calculator

Once the problem is turned into a straight line (linear), the authors use two powerful mathematical "calculators" to solve it:

Affine Processes: These are systems where the future average behavior depends linearly on the current state.
Polynomial Processes: These are systems where the future variance (spread) depends on a simple polynomial (like $x^2$ ) of the current state.

The paper proves that Signature SDEs are actually just these simple "Affine" and "Polynomial" processes in disguise. This is a huge deal because we already have perfect formulas for how to calculate the future of these simple processes.

4. The Solution: Riccati Equations (The Recipe)

Because the problem is now "simple," the authors can write down exact formulas to predict the future.

They don't just guess; they solve specific differential equations (called Riccati equations and Linear ODEs) that act like a recipe.
If you feed the current "signature" into this recipe, it spits out the exact probability of where the path will be, or the expected value of any function of that path.
They show that these recipes can be written as power series (infinite sums), which can be calculated on a computer.

5. Why This Matters (According to the Paper)

Universality: The authors claim this method is "universal." It means that almost any random process driven by Brownian motion (even those with complex, history-dependent rules) can be viewed through this lens.
Explicit Formulas: Instead of running thousands of computer simulations (Monte Carlo) to get an approximation, this method gives you a direct formula.
Real-World Examples: The paper tests this on things like:
- Geometric Brownian Motion: A standard model for stock prices.
- Jacobi Diffusion: A model often used for interest rates or probabilities that stay between 0 and 1.
- Counter-examples: They also show where the math gets tricky (at "degeneracy points" where the rules break down), proving they know exactly where their method works and where it doesn't.

Summary Analogy

Imagine you are trying to predict the weather.

Old Way: You look at the current temperature, wind, and humidity, but you also have to remember every storm that happened last week because the atmosphere has a "memory." This is too complex to calculate directly.
This Paper's Way: You take the entire weather history and compress it into a single "Weather Signature" code. You discover that if you look at this code, the weather rules become simple and linear (like a straight line on a graph). You then use a simple calculator (the Riccati equation) to predict the future weather based on that code.

The paper essentially says: "Don't fight the complexity of the path's history. Encode the history into a signature, and the complexity vanishes, leaving you with a simple, solvable math problem."

Technical Summary: Signature SDEs from an Affine and Polynomial Perspective

Problem Statement
The paper addresses the challenge of characterizing the laws of stochastic processes with path-dependent coefficients, specifically within the class of Itô-diffusions. While affine and polynomial processes are well-understood in finite dimensions due to their tractability (explicit formulas for Fourier-Laplace transforms and moments), many modern models in finance and physics (e.g., rough volatility, neural SDEs) exhibit path dependence that falls outside these standard classes. The authors aim to establish a framework where such general path-dependent processes can be "linearized" into infinite-dimensional affine or polynomial processes, thereby enabling explicit computation of their characteristic functions and expected values.

Methodology
The core methodology relies on the signature of a continuous semimartingale, which consists of iterated Stratonovich integrals of the process against itself. The authors treat the signature process as taking values in the extended tensor algebra $T((\mathbb{R}^d))$ , specifically on the subset of group-like elements.

Affine and Polynomial Framework on Tensor Algebras:
The authors first generalize the theory of affine and polynomial processes to subsets of the extended tensor algebra.
- Affine Processes: Defined via a martingale problem where the generator applied to exponential functions $\exp(\langle u, x \rangle)$ yields a result proportional to the function itself, with a coefficient determined by a map $R(u)$ . This leads to an affine transform formula where the Fourier-Laplace transform is expressed as $\exp(\langle \psi(T), x_0 \rangle)$ , with $\psi$ solving a tensor-algebra valued Riccati ODE.
- Polynomial Processes: Defined similarly but for linear functions $\langle u, x \rangle$ . The generator maps these to linear combinations, leading to a moment formula where expected values are expressed via a tensor-algebra valued linear ODE.
- Duality: The paper utilizes a duality perspective, showing that the evolution of the dual process (deterministic) is governed by these ODEs, allowing the computation of expectations of the primal stochastic process.
Signature SDEs:
The authors introduce Signature SDEs (Sig-SDEs) of the form:
$dX_t = b(X_t)dt + \sqrt{a(X_t)}dB_t$
where $X_t$ is the signature process, and the coefficients $b$ and $a$ are entire maps (converging infinite linear combinations) of the signature components. Unlike standard setups considering only finite truncations, this framework handles proper entire maps.
Main Theoretical Result:
The central result is that the signature process of a Sig-SDE is itself an affine and polynomial process on the state space of group-like elements.
- The drift and diffusion characteristics of the Sig-SDE determine the vector fields $R$ (for the Riccati ODE) and $L$ (for the linear ODE) in the extended tensor algebra.
- Specifically, $R(u)$ involves quadratic terms in $u$ (due to the diffusion coefficient), while $L(u)$ is linear in $u$ .
- Consequently, the Fourier-Laplace transform and the expected value of entire functions of the signature can be computed by solving these infinite-dimensional ODEs.
One-Dimensional Case and Real-Analyticity:
In Section 5, the authors specialize to one-dimensional diffusions with real-analytic coefficients. Here, the signature process corresponds to the sequence of powers $(1, X_t, X_t^2, \dots)$ .
- They prove that under uniform ellipticity and boundedness conditions, the solution to the Kolmogorov equation is real-analytic in the spatial variable.
- This allows the moment and affine transform formulas to hold for a dense class of functions.
- The paper also provides a counterexample (Section 5.3.5) showing that at degeneracy points of the diffusion coefficient, real-analyticity of the expected value can fail, highlighting the necessity of specific conditions for the formulas to hold globally.

Key Contributions and Results

Universality of Signature SDEs: The paper demonstrates that a broad class of path-dependent Itô-diffusions can be represented as projections of infinite-dimensional affine and polynomial processes. This establishes a form of universality for the affine/polynomial class within the space of Brownian-driven processes.
Explicit Formulas: The authors derive explicit formulas for the Fourier-Laplace transform and expected values of entire functions of the signature. These are given as converging power series (or linear series) where the coefficients are solutions to Riccati and linear ODEs in the extended tensor algebra.
Unification of Approaches: The paper unifies the affine (Riccati) and polynomial (linear) approaches, showing they coincide for signature processes on group-like elements. It also connects these to transport PDEs, offering an alternative to Riccati ODEs that avoids explosion issues in certain cases.
Numerical Illustrations: The theoretical framework is validated through numerical examples, including the computation of the Laplace transform for geometric Brownian motion and the fourth power of Brownian motion. The authors demonstrate that while naive truncation of Riccati ODEs can lead to instability, schemes based on transport equations or specific moment formula implementations yield stable results.

Significance and Claims
The paper claims to provide a "universal" framework for analyzing path-dependent stochastic processes by lifting them to the signature space. By doing so, it allows for the application of powerful affine and polynomial techniques to models that were previously intractable or required complex numerical approximations without explicit structural guarantees.

The authors emphasize that their approach allows for the characterization of laws on path spaces via the signature without requiring exponential moment conditions (by working with Fourier-Laplace transforms). They note that this is particularly relevant for modern models in machine learning (e.g., neural SDEs with entire activation functions) and rough volatility modeling.

The paper remains modest regarding the existence and uniqueness of solutions to the infinite-dimensional ODEs in full generality, noting that while global existence is established for specific cases (e.g., bounded coefficients, uniform ellipticity), the general theory relies on the assumption that the Sig-SDE admits a solution. The numerical results serve to illustrate the practical applicability of the derived formulas rather than to prove general existence theorems for all possible path-dependent coefficients.