Global universality via discrete-time signatures

Imagine you are trying to teach a computer to understand a story. But this isn't a story made of words; it's a story made of movement. Think of a stock price chart, a robot's arm moving through space, or the path of a drunk person walking home. These are all "paths" or "trajectories."

The big challenge is: How do you describe the entire history of a moving object in a way a computer can use to make predictions?

This paper, titled "Global Universality via Discrete-Time Signatures," by Mihriban Ceylan and David J. Prömel, offers a brilliant new way to solve this. Here is the breakdown in simple terms.

1. The Problem: The "Infinite" Story

In the real world, data doesn't come in perfect, smooth, continuous lines. It comes in snapshots. You check your bank account every morning. You record a GPS signal every second. You have a list of points, not a smooth curve.

Mathematicians have long known a powerful tool called the "Signature" (borrowed from a concept called Rough Path Theory).

The Analogy: Imagine a signature is like a super-compact summary of a journey. It doesn't just say "I went from A to B." It captures the twists, turns, loops, and interactions of the journey. It's like a DNA test for a path.
The Magic: If you know the Signature of a path, you can mathematically reconstruct almost any function of that path. It's a "universal language" for movement.

The Catch:

Real data is messy: We only have snapshots (discrete time), not the smooth movie.
The "Compact" Trap: Old math rules said, "This magic only works if the path stays within a small, contained box." But in the real world (like stock markets or Brownian motion), paths can wander off to infinity. They don't stay in a box.

2. The Solution: "Discrete-Time Signatures"

The authors say: "Let's stop trying to force the messy, infinite world into a small box. Let's build a new theory specifically for the snapshots we actually have."

They propose a two-step process:

Connect the Dots: Take your scattered data points and draw straight lines between them. This creates a "Piecewise Linear Path." It's a jagged line, but it preserves the order and the jumps of your data.
Calculate the Signature: Compute the "Signature" of this jagged line.

3. The Big Breakthrough: "Global Universality"

The paper proves a massive theorem: You can approximate any complex function of a path using just these jagged-line signatures.

The Old Way: "We can approximate any story, but only if the story happens inside a small, safe room."
The New Way: "We can approximate any story, even if the character runs off to the edge of the universe, as long as we have the right mathematical 'weight' to handle the distance."

They use a concept called Weighted Spaces.

The Analogy: Imagine you are trying to describe a journey. If the journey is short, it's easy. If the journey is huge (like a rocket to Mars), it's harder to describe. The authors introduce a "weight" that gets heavier the further you go. This allows them to mathematically control the "infinite" paths without losing accuracy.

4. Why This Matters (The "Brownian Motion" Test)

To prove their theory works in the real world, they tested it on Brownian Motion.

What is it? It's the random, jittery movement of a pollen grain in water (or a stock price). It's the ultimate example of a path that wanders everywhere and never stays in a "box."
The Result: They proved that if you take a Brownian motion, chop it up into time steps, connect the dots with straight lines, and calculate the signature, you can approximate anything related to that movement.
- You can predict option prices.
- You can solve complex differential equations (the math behind physics and finance).
- You can train AI to understand time-series data better.

5. The "So What?" for Everyday Life

Why should you care?

For Finance: It means we can build better models to price complex financial derivatives (insurance for stocks) using only the data we actually have (daily prices), without needing impossible, perfect continuous data.
For AI & Machine Learning: It gives AI a new, powerful "feature" to look at. Instead of just feeding an AI raw numbers, you feed it the "Signature" of the movement. The paper proves this is mathematically guaranteed to work for any pattern, no matter how wild the data gets.
For Robotics: It helps robots understand their own movement history more efficiently, allowing them to learn from discrete sensor data rather than needing perfect, smooth sensors.

Summary Metaphor

Imagine you are trying to guess the plot of a movie, but you only have a few still photos (discrete data).

Old Math: Said, "We can only guess the plot if the movie takes place in a small living room."
This Paper: Says, "No! If we connect the photos with straight lines and look at the shape of those lines (the Signature), we can guess the plot of a movie that takes place anywhere in the galaxy, even if the characters run off into space."

They have built a universal translator that turns jagged, real-world data into a language that computers can understand perfectly, no matter how chaotic the data gets.

Here is a detailed technical summary of the paper "Global Universality via Discrete-Time Signatures" by Mihriban Ceylan and David J. Prömel.

1. Problem Statement

The paper addresses the challenge of approximating complex functionals that depend on the entire trajectory of a data stream (path-dependent functionals). While the signature of a path (an infinite collection of iterated integrals) is known to be a universal feature set for continuous path functionals on compact sets, two major practical limitations exist:

Computational Intractability: For continuous stochastic processes like Brownian motion, iterated integrals cannot be computed in closed form. Practical applications rely on discrete-time observations, necessitating the use of piecewise linear interpolations to reconstruct continuous paths.
Lack of Global Approximation: Classical universal approximation theorems (UAT) for signatures are typically restricted to compact subsets of path space. However, in probabilistic settings (e.g., Brownian motion), sample paths do not lie in fixed compact sets with positive probability. Existing global results for continuous rough paths do not directly apply to the discrete-time setting due to topological and structural differences.

The core problem is to establish global universal approximation theorems for linear functionals of signatures derived from discrete-time observations (specifically piecewise linear interpolations) in $L^p$ -spaces and weighted function spaces, without the restriction of compactness.

2. Methodology

The authors develop a framework that bridges discrete-time data with rough path theory using the following methodological steps:

Discrete-Time Framework: Instead of working with continuous rough paths directly, the authors define the space of paths $\hat{C}^\pi$ as piecewise linear interpolations of discrete observations. They consider the time-extended path $\hat{X}^\pi_t = (t, X^\pi_t)$ to handle non-anticipative functionals.
Weighted Function Spaces: To overcome the non-compactness of path spaces, the authors utilize weighted function spaces ( $B_\psi$ $B_{ψ}$ ). They define an admissible weight function $\psi(\hat{X}^\pi) = \exp(\beta \|\hat{X}^\pi\|_\alpha^\gamma)$ $ψ (\hat{X}^{π}) = exp (β ∥ \hat{X}^{π} ∥_{α}^{γ})$ , where $\|\cdot\|_\alpha$ $∥ \cdot ∥_{α}$ is the Hölder norm.
- They prove that the space of piecewise linear paths equipped with a weaker topology (lower Hölder exponent) forms a weighted space where the weight function is admissible (pre-images of bounded sets are compact).
Weighted Stone–Weierstrass Theorem: The authors apply a generalized Stone–Weierstrass theorem (adapted from Ceylan, Sauer, and Tresser [CST26]) to show that the linear span of signature coordinates is dense in these weighted spaces.
Integrability Conditions: To extend the results from weighted spaces to $L^p$ -spaces, they impose an exponential moment condition on the weight function. Specifically, they require that the $p$ -th moment of the weight function is finite under the probability measure of the underlying process.
Probabilistic Verification: The authors verify that piecewise linear interpolations of Brownian motion satisfy this exponential moment condition. They leverage the fact that the signature of Brownian motion has Gaussian tails, ensuring the necessary integrability for $L^p$ convergence.

3. Key Contributions

A. Theoretical Advances

Global UAT for Discrete-Time Paths: The paper establishes that linear functionals of signatures of piecewise linear paths are dense in weighted spaces $B_\psi(\hat{C}^\pi)$ and $B_\psi(\Lambda^\pi_T)$ (stopped paths). This extends classical UATs from compact sets to global settings.
$L^p$ -Approximation Theorems: Under the condition that the weight function has finite $p$ -th moments, the authors prove that any $L^p$ -functional can be approximated arbitrarily well by linear combinations of signature coordinates.
Discrete vs. Continuous Distinction: The authors explicitly clarify that their discrete-time results are not immediate consequences of continuous-time rough path results (e.g., [CP25]). The weighted structure is not preserved under the canonical embedding of discrete paths into continuous rough path spaces, necessitating a separate discrete-time analysis.

B. Probabilistic Applications

Brownian Motion Integrability: The paper proves that the law of piecewise linearly interpolated Brownian motion satisfies the required exponential moment condition. This validates the use of signature-based approximations for stochastic processes in an $L^p$ setting.
Approximation of SDEs and Random ODEs:
- Random ODEs: Solutions to ODEs driven by piecewise linear Brownian motion can be approximated by linear functionals of the signature.
- Stochastic Differential Equations (SDEs): By combining the discrete-time approximation with the stability of the Brownian signature (Lemma 4.4), the authors show that solutions to SDEs driven by true Brownian motion can be approximated by linear functionals of the discrete-time signature of the interpolated path.

4. Key Results

Proposition 3.1 & 3.4: The linear span of signature coordinates is dense in the weighted Banach spaces $B_\psi(\hat{C}^\pi)$ and $B_\psi(\Lambda^\pi_T)$ .
Theorem 3.5 ( $L^p$ -Convergence): If the weight function $\psi$ satisfies $\int \psi^p d\nu < \infty$ , then for any $f \in L^p$ , there exists a linear signature functional $f_\ell$ such that $\|f - f_\ell\|_{L^p} < \epsilon$ .
Corollary 4.1 & 4.2: For Brownian motion $W$ , measurable and continuous path functionals (and non-anticipative functionals) can be approximated in $L^p$ by linear functionals of the signature of its piecewise linear interpolation $W^{\pi_n}$ .
Lemma 4.4: Establishes the $L^p$ -convergence of linear functionals evaluated on the true Brownian signature to those evaluated on the piecewise linear interpolated signature as the mesh size goes to zero.
Corollary 4.6: Solutions to SDEs driven by Brownian motion can be approximated in $L^p$ by linear functionals of the signatures of piecewise linearly interpolated Brownian paths.

5. Significance

Bridging Theory and Practice: The paper provides a rigorous theoretical foundation for the widespread practical use of signatures in machine learning and quantitative finance, where data is inherently discrete. It justifies why training models on discrete signatures works even when the underlying theory assumes continuous paths.
Global Approximation: By moving beyond compact sets, the results are directly applicable to probabilistic modeling where paths can exhibit unbounded growth or large deviations, a common scenario in financial markets and physical systems.
Computational Feasibility: The results confirm that one does not need to compute complex continuous iterated integrals; instead, computing signatures on discrete, piecewise linear interpolations is sufficient to approximate a vast class of path-dependent functionals and stochastic differential equation solutions with arbitrary precision in the $L^p$ sense.
Foundation for Machine Learning: This work strengthens the theoretical basis for "Signature Methods" in deep learning (e.g., for time-series classification, forecasting, and generative modeling), ensuring that these methods possess universal approximation capabilities even when applied to real-world, discretely sampled data streams.