---
title: 'Signature Kernels: Theory, Computation & Applications'
url: https://www.emergentmind.com/topics/signature-kernels
type: topic
---

# Signature Kernels: Theory, Computation & Applications

Signature kernels are positive definite kernels for sequential data obtained by taking inner products of path signatures, either directly in tensor algebra or after lifting paths through a static kernel into an RKHS. They link rough path theory, kernel methods, and sequential learning: algebraically, they encode ordered iterated-integral structure; analytically, they admit dynamic-programming and Goursat-PDE formulations; statistically, they induce RKHSs, MMDs, and strictly proper scoring rules on path space; computationally, they support GPU-parallelized solvers and large-scale implementations [2305.04625], [1601.08169].

## 1. Algebraic construction and pathwise feature maps

For a bounded-variation path $x:[0,T]\to\mathbb{R}^d$ with $x(0)=0$, the signature map produces the graded tensor sequence
$$
S(x) = (S_0(x), S_1(x), S_2(x), \ldots),
$$
where $S_0(x)=1$ and
$$
S_m(x)=\int_{0<t_1<\cdots<t_m<T} dx_{t_1}\otimes\cdots\otimes dx_{t_m},\qquad m\ge 1.
$$
On each tensor level $(\mathbb{R}^d)^{\otimes m}$ one uses the Hilbert–Schmidt inner product, and the full tensor algebra is equipped with the levelwise sum of these inner products. The basic unweighted signature kernel is then
$$
k(x,y)=\sum_{m=0}^{\infty}\langle S_m(x),S_m(y)\rangle.
$$
This realizes the signature as a feature map into a Hilbert space and makes the kernel positive definite by construction [2305.04625].

A more general and practically important construction begins with a positive definite static kernel $k$ on the state space, with RKHS $H$. The path $x$ is lifted to the $H$-valued path $k_x(t)=k(x(t),\cdot)$, and one defines
$$
k_{\mathrm{Sig},:M}(x,y)=\langle S_{\le M}(k_x),S_{\le M}(k_y)\rangle,
$$
or, when the full series converges, the corresponding untruncated kernel. This “kernelization before taking signatures” is central in modern applications because it allows nonlinear static similarity and pathwise ordering to be combined within a single RKHS formalism [2305.04625].

For discrete sequences, the same idea appears as a sequentialization of a static kernel. Given a sequence $x=(x_1,\ldots,x_n)$ and first differences $\Delta x_i=x_{i+1}-x_i$, the discrete level-$p$ signature feature is
$$
S^p(x)=\sum_{1\le i_1<\cdots<i_p\le n-1}\Delta x_{i_1}\otimes\cdots\otimes\Delta x_{i_p}.
$$
More generally, any static PSD kernel can be sequentialized via ordered tensor features or, equivalently, via the difference Gram matrix
$$
k^\Delta_{ij}=k(x_{i+1},y_{j+1})+k(x_i,y_j)-k(x_{i+1},y_j)-k(x_i,y_{j+1}),
$$
yielding kernels for time series, sequences of graphs, or strings; string kernels arise as a special case, and alignment kernels are closely related after a modification that restores positive definiteness [1601.08169].

A fundamental structural property is invariance under strictly monotone time reparameterization, and more generally under tree-like equivalence. This is often advantageous for path geometry, but it also means that timing information can be suppressed. When timing should remain identifiable, time augmentation $x(t)\mapsto (t,x(t))$ is used; the survey literature states that this removes reparameterization invariance and restores injectivity up to the standard signature identifiability framework [2305.04625].

## 2. Goursat PDEs, dynamic programming, and numerical computation

For continuously differentiable paths, the signature kernel admits a PDE characterization that functions as a kernel trick. If
$$
k_{x,y}(s,t)=\langle S(x)_s,S(y)_t\rangle,
$$
then
$$
\frac{\partial^2}{\partial s\,\partial t}k_{x,y}(s,t)=\langle \dot x_s,\dot y_t\rangle\,k_{x,y}(s,t),
\qquad
k_{x,y}(u,\cdot)=k_{x,y}(\cdot,v)=1.
$$
Equivalently,
$$
k_{x,y}(s,t)=1+\int_u^s\int_v^t k_{x,y}(p,q)\,\langle \dot x_p,\dot y_q\rangle\,dp\,dq.
$$
This is a hyperbolic Goursat problem whose coefficient depends only on local path increments. It avoids explicit construction of signatures and gives an untruncated kernel computation route with the same raw complexity as earlier truncated-kernel tricks, while being naturally suited to GPU antidiagonal parallelization [2006.14794].

In the lifted setting, the same PDE holds in the RKHS:
$$
\frac{\partial^2}{\partial s\,\partial t}k_{X,Y}(s,t)=\langle \dot X_s,\dot Y_t\rangle_H\,k_{X,Y}(s,t),
$$
and, for sampled data, the coefficient can be approximated directly from the static kernel by finite differences. For piecewise linear paths, the PDE reduces on each grid cell to a constant-coefficient problem, which leads to explicit finite-difference updates and antidiagonal dynamic programs [2006.14794].

A complementary computational route uses exact discrete-time recursions for truncated kernels. For finite truncation level $M$, a Horner-type dynamic program computes $k_{\mathrm{Sig},:M}(x,y)$ exactly in $O(M^3L^2)$ time for sequence length $L$, and this remains one of the standard references for tractable finite-level computation [2305.04625].

The limitations of first-order PDE solvers on highly oscillatory inputs have motivated several high-accuracy variants. “Log-PDE Methods for Rough Signature Kernels” derive a coupled PDE system whose coefficients involve higher-order iterated integrals of smooth rough paths, giving a higher-order approximation to rough signature kernels together with quantitative error bounds [2404.02926]. “Numerical Schemes for Signature Kernels” introduces polynomial approximation and interpolation schemes based on local analytic propagation across Goursat rectangles; the paper reports improvements of several orders of magnitude in mean absolute percentage error compared to traditional finite differences, without increasing computational complexity [2502.08470]. “Scalable Signature Kernel Computations for Long Time Series via Local Neumann Series Expansions” replaces global grids by dynamically truncated local power series propagated tile by tile, with stated advantages in adjustable accuracy, reduced memory, and scalability to paths with up to half a million points or more on a single GPU [2502.20392].

## 3. RKHS theory, universality, characteristicness, and scoring rules

Because signature kernels are inner products in tensor Hilbert spaces, they induce RKHSs on path space. This gives immediate access to representer theorems, kernel mean embeddings, Gaussian process constructions, and MMD-based statistics [2305.04625].

On compact sets of bounded-variation paths, linear functionals of signatures are dense in continuous functionals once time augmentation is used to avoid the path-identification ambiguities discussed above. In the same regime, expected signatures characterize probability measures, and convergence of expected signature levels implies weak convergence. The survey literature therefore presents universality and characteristicness as central theoretical guarantees of the signature kernel framework on compact domains [2305.04625].

On non-compact domains, the situation is subtler. The same survey explicitly states that plain signatures and moments fail to be characteristic and robust there, and introduces a normalization $\lambda$ producing a robust signature kernel
$$
k^{\mathrm{robust}}(x,y)=\langle \lambda(S(\varphi(x))),\lambda(S(\varphi(y)))\rangle.
$$
This robust variant is presented as guaranteeing universality in the strict topology, characteristicness to signed measures, and $B$-robustness in the sense of Hampel/Huber [2305.04625]. A common misconception is therefore that characteristicness is automatic for every path domain; the cited theory instead distinguishes compact from non-compact settings and plain from normalized signatures.

The MMD associated with a signature kernel takes the standard RKHS form
$$
d_k(\mu,\nu)=\sqrt{\mathbb{E}[k(X,X')]-2\mathbb{E}[k(X,Y)]+\mathbb{E}[k(Y,Y')]}.
$$
This viewpoint becomes especially important in probabilistic modeling. In “Non-adversarial training of Neural SDEs with signature kernel scores,” the signature kernel is shown to be characteristic on compact path sets, and the induced score
$$
\phi_{\mathrm{sig}}(\mathbb{P},y)
=
\mathbb{E}_{x,x'\sim \mathbb{P}}[k_{\mathrm{sig}}(x,x')]
-
2\,\mathbb{E}_{x\sim\mathbb{P}}[k_{\mathrm{sig}}(x,y)]
$$
is proved strictly proper relative to $\mathcal{P}(\mathcal{K})$. The same paper states that minimizing its expectation is equivalent to minimizing squared MMD, and gives the unbiased U-statistic estimator
$$
\widehat{\phi}_{\mathrm{sig}}(\mathbb{P},y)
=
\frac{1}{m(m-1)}\sum_{i\neq j}k_{\mathrm{sig}}(x^i,x^j)
-\frac{2}{m}\sum_{i=1}^m k_{\mathrm{sig}}(x^i,y),
$$
which is unbiased and consistent [2305.16274].

This scoring-rule perspective clarifies another recurrent misunderstanding: signature kernels are not only similarity measures for supervised learning. They also define statistically meaningful discrepancies between laws on path space, including losses suitable for non-adversarial generative modeling and forecast verification [2305.16274].

## 4. Generalized, rough, infinite-dimensional, and asymptotic variants

The standard unweighted kernel is only one member of a larger family. “General Signature Kernels” introduces level-weighted kernels
$$
K_\phi^{\gamma,\sigma}(s,t)
=
\sum_{k=0}^{\infty}\phi(k)\,\langle S^{(k)}(\gamma)_{0,s},S^{(k)}(\sigma)_{0,t}\rangle_k,
$$
under a summability condition on $\phi$. These kernels satisfy the shifted integral identity
$$
K_\phi(s,t)=\phi(0)+\int_0^s\int_0^t K_{\phi_{+1}}(u,v)\,\langle d\gamma_u,d\sigma_v\rangle,
$$
and, when $\phi(k)$ is a moment sequence or more generally an integral transform, can be represented as averages of scaled PDE solutions. The same paper establishes a new connection with hyperbolic development; for $\phi(k)=\Gamma(k/2+1)$,
$$
\big\langle \mathbb{E}[S(\circ B)_{0,s}],\,S(\gamma)_{0,t}\big\rangle_\phi
=
\cosh\big(\rho_{\sqrt{s/2}\,\gamma}(t)\big),
$$
where $\rho$ is the hyperbolic distance associated with the development of $\gamma$ [2107.00447].

Rough-path extensions replace classical derivatives or bounded-variation integrals by rough integrations. “The Signature Kernel is the solution of a Goursat PDE” already extends the analysis from $C^1$ and bounded-variation paths to geometric rough paths through a rough integral equation analogous to the Goursat PDE [2006.14794]. “Log-PDE Methods for Rough Signature Kernels” then derives higher-order approximations whose coefficients depend on truncated log-signatures, explicitly targeting highly oscillatory or rough inputs [2404.02926]. “Signature Kernel and Schwinger-Dyson Kernel Equations as Two-Parameter Rough Differential Equations” places the signature kernel equation in a two-parameter rough differential equation framework, proving well-posedness and stability and recovering PDE formulations in the smooth rough-path regime [2605.08844].

The kernel also extends beyond finite-dimensional state spaces. Because the PDE evaluation depends only on inner products, “Non-adversarial training of Neural SDEs with signature kernel scores” states that the same construction extends verbatim to Hilbert-valued paths, including $L^2(D)$-valued spatiotemporal signals and RKHS-valued paths. This is used there to define kernel scores for Neural SPDE generators and mesh-free generation of limit order book dynamics [2305.16274].

Two asymptotic constructions show that signature kernels also arise as limits of other models. “Neural signature kernels as infinite-width-depth-limits of controlled ResNets” proves that controlled ResNets converge, under infinite width and depth, to Gaussian processes indexed by paths whose kernels satisfy activation-dependent PDEs; in the homogeneous identity-activation case, the limiting kernel agrees with the classical signature kernel [2303.17671]. “Free probability, path developments and signature kernels as universal scaling limits” shows that randomized developments into $\mathrm{GL}(N)$ converge to the classical signature kernel, while unitary developments lead to a related Schwinger–Dyson kernel characterized by a quadratic functional equation [2402.12311].

## 5. Applications across statistics, finance, PDEs, and forecasting

Signature kernels were first developed as general sequence kernels, and early empirical studies already placed them among strong performers for multivariate time-series learning. The PDE-based untruncated kernel in “The Signature Kernel is the solution of a Goursat PDE” is evaluated on UEA classification datasets with SVC and on Bitcoin-price regression with SVR, where it is reported to be consistently among the top two kernels on the listed UEA tasks and to outperform RBF, GAK, and truncated-signature baselines in the reported Bitcoin MAPE experiment [2006.14794].

Because the induced RKHS lives on path space, signature kernels can be used to solve path-dependent PDEs by collocation. “A path-dependent PDE solver based on signature kernels” approximates PPDE solutions by minimal-norm elements of the signature RKHS constrained at collocation paths, proves consistency in the linear case, and presents numerical examples in option pricing under rough volatility [2403.11738].

Financial derivatives under rough volatility supply a second application area. “Pricing American options under rough volatility using deep-signatures and signature-kernels” uses signature-kernel RKHS regression for primal continuation values and RKHS martingale parameterizations for dual bounds in rough Bergomi and rough Heston models. The paper reports that kernel methods give strong lower bounds and are more stable than deep-signature methods for small training sets, while deep-signature dual methods produce stronger upper bounds [2501.06758].

Generative modeling is now a major use case. In “Non-adversarial training of Neural SDEs with signature kernel scores,” signature kernels define a strictly proper score on path space for unconditional and conditional training of Neural SDEs and Neural SPDEs. The paper reports improved stability over adversarial alternatives, no need for Wasserstein-GAN-style Lipschitz constraints, and strong empirical performance on geometric Brownian motion, rough Bergomi, EUR/USD and USD/JPY forex generation, conditional EUR/USD forecasting, and limit-order-book spatiotemporal generation [2305.16274].

Forecast verification and probabilistic weather modeling provide a newer domain. “Signature Kernel Scoring Rule as Spatio-Temporal Diagnostic for Probabilistic Forecasting” frames weather variables as continuous paths, uses path augmentations to obtain strict properness, and reports high discriminative power on WeatherBench 2 scorecards. The same paper trains sliding-window generative neural networks on ERA5 reanalysis data and reports that signature-kernel-based training outperforms climatology for forecast paths of up to fifteen timesteps [2510.19110].

A more specialized but conceptually important application is sparse signature recovery. “Sparse Signature Coefficient Recovery via Kernels” shows that suitable linear combinations of signature kernels can isolate an anagram class of coefficients or even a single high-level iterated integral without computing the full truncated signature, with a use case based on sparse controlled differential equations [2412.08579].

## 6. Implementations, scalability, limitations, and current practice

Several software stacks now implement signature kernels and related approximations.

| Library | Stated focus | Reported capability |
|---|---|---|
| `sigkernel` | PDE kernel with CPU/CUDA support | antidiagonal GPU updates for the Goursat solver [2006.14794] |
| `KSig` | Scikit-Learn-compatible dual and primal methods | exact truncated DP, PDE kernels, random Fourier signature features, tensor sketch [2501.07145] |
| `pySigLib` | PyTorch-compatible CPU/GPU signature and kernel ops | exact reverse-mode differentiation through the finite-difference solver; GPU backward speedups up to $665\times$ over `sigkernel` in the reported benchmark [2509.10613] |
| `polysigkernel` | polynomial PDE solvers | several-orders-of-magnitude MAPE improvements over finite differences without higher complexity [2502.08470] |
| `PowerSig` | local Neumann-series PDE solver | signature kernels for sequences of length $524{,}289$ using under 100 MB of GPU memory [2502.20392] |

The practical bottlenecks remain significant. The survey literature states that full Gram-matrix construction scales as $O(N^2)$ in the number of sequences, and exact truncated computations scale as $O(M^3L^2)$ per pair; PDE approaches avoid explicit signature construction but still have quadratic dependence on sequence length in raw arithmetic count [2305.04625]. KSig accordingly distinguishes dual methods, whose memory remains quadratic in sequence length and sample count, from primal randomized-feature approximations that scale more favorably for large datasets [2501.07145].

Sensitivity to scaling is another recurring issue. The survey paper emphasizes that signature levels scale as $S_m(\lambda x)=\lambda^m S_m(x)$ and that truncation error depends on path variation. The Neural SDE paper likewise stresses time normalization, path scaling, and standardization to prevent numerical overflow or domination by low-order terms in PDE solves and gradient computation [2305.04625], [2305.16274]. A common practical response is to rescale channels, add time augmentation when timing matters, and use RBF lifts or robust normalization when plain signatures are too brittle.

Theoretical assumptions also matter. Much of the core theory is stated for bounded-variation paths or compact path sets; extensions to rough paths, non-compact domains, and infinite-dimensional settings exist, but each requires additional machinery or assumptions [2305.16274], [2404.02926]. It is therefore inaccurate to treat every signature-kernel result as automatically path-space-wide. The current literature instead offers a layered picture: compact BV theory is mature; rough-path, stochastic, and long-sequence numerics are active frontiers.

A plausible implication is that contemporary practice should be method-selective rather than doctrinal. Exact or untruncated PDE kernels are natural when high fidelity on moderate sequence lengths is paramount; polynomial or local-series solvers are preferable for oscillatory or very long sequences; randomized features and primal approximations are attractive when Gram-matrix costs dominate; and robust or augmented kernels are often necessary when timing, non-compactness, or conditioning issues are central. Across these variants, the defining idea remains unchanged: path signatures furnish a universal ordered feature system, and signature kernels make that system computationally and statistically usable at scale.

Source: https://www.emergentmind.com/topics/signature-kernels