Log-Signature (Signature Cumulant)

Updated 22 February 2026

Log-signature is the logarithm of a path's iterated integral signature, offering a compressed, coordinate-independent encoding of geometric features.
It is computed via the Magnus expansion and BCH formula, placing the result in a free Lie algebra for efficient, hierarchical representation.
Applications include generative modeling, time-series analysis, and writer identification, leveraging robustness to noise and reparameterization invariance.

A log-signature—also known as the signature cumulant—is the logarithm, in the completed tensor algebra, of the (iterated-integral) signature of a continuous path. The log-signature provides a compressed, coordinate-independent, and Lie-algebraic encoding of the geometric features of a path. It is fundamentally linked to the algebraic and analytic structure of rough paths, stochastic processes, and high-dimensional sequential data, acting simultaneously as a canonical non-commutative generalization of classical cumulants and as a central object in practical algorithms for generative modeling, inference, and dimension reduction.

1. Algebraic Structure and Definition

The signature of a path $\gamma: [0,T]\to\mathbb{R}^d$ is the sequence of all its iterated integrals, viewed as a group-like element in the completed tensor algebra $T((\mathbb{R}^d)) = \prod_{n=0}^\infty (\mathbb{R}^d)^{\otimes n}$ . Formally,

$S(\gamma) = \left(1,\; X^1,\; X^2,\; \dots \right) \qquad X^n = \int_{0<t_1<\dots<t_n<T} d\gamma_{t_1}\otimes\cdots\otimes d\gamma_{t_n}.$

The log-signature, or signature cumulant, is defined as the non-commutative (tensor algebra) logarithm of the signature,

$\log S(\gamma) = \sum_{n=1}^\infty L^n, \qquad L^n\in (\mathbb{R}^d)^{\otimes n},$

where the sum converges in the appropriate topology and each $L^n$ is an explicit Lie polynomial in the increments of $\gamma$ . By Chen–Ree's theorem, $\log S(\gamma)$ lies in the free Lie algebra $\mathcal{L}((\mathbb{R}^d))$ and can be expanded as

$\log S(\gamma) = \sum_{w\in\mathcal{H}} c_w \, w,$

with $\mathcal{H}$ a Hall (or Lyndon) basis for the Lie algebra and $c_w$ the signature cumulants.

2. Magnus Expansion and Lie Series

The algebraic structure of the log-signature is made explicit via the Magnus expansion, a universal Lie-series which recursively expresses the logarithm of the signature in terms of the path derivative and its nested commutators: $\log S(\gamma) = \int \gamma' - \frac{1}{2} \int [\gamma',\gamma'] + \frac{1}{6} \int [[\gamma',\gamma'],\gamma'] + \cdots,$ with the $n$ th homogeneous piece

$\kappa_n(\gamma) = \int_{0<t_1<\cdots<t_n<T} [\gamma'(t_1), [\ldots, \gamma'(t_n)]] dt_1\cdots dt_n,$

generalizing the cumulant structure of scalar random variables to the non-commutative, tensor-algebraic case (Friz et al., 9 Sep 2025, Friz et al., 2024, Friz et al., 2021). The log-signature thus provides an intrinsic, hierarchically ordered, and compact description of the path's geometry, with higher-order terms encoding complex geometric correlations such as areas and iterated areas.

3. Computational and Algorithmic Methods

Efficient algorithms for computing log-signatures exploit:

The truncated tensor algebra $T^n(\mathbb{R}^d)$ at order $n$ and the corresponding projection onto a Hall or Lyndon basis for the free Lie algebra (dimension $\ll d^n$ ) (Reizenstein, 2017, Barancikova et al., 2024, Lai et al., 2019).
Baker–Campbell–Hausdorff (BCH) formula for concatenating segments: For $\Delta_1, \Delta_2\in \mathbb{R}^d$ , the log-signature of their concatenation is

$\log(\exp(\Delta_1)\exp(\Delta_2)) = \Delta_1 + \Delta_2 + \frac{1}{2}[\Delta_1, \Delta_2] + \frac{1}{12}[\Delta_1,[\Delta_1, \Delta_2]] - \frac{1}{12}[\Delta_2,[\Delta_1,\Delta_2]] + \cdots,$

with all terms in the chosen Lie basis (Reizenstein, 2017).

For piecewise linear or polygonal paths, the log-signature can be maintained incrementally under concatenation; widely used libraries such as iisignature provide optimized implementations (Reizenstein, 2017).
In stochastic/semimartingale settings, the expected log-signature is computed via Magnus-type expansion or non-commutative Riccati-type functional equations, sometimes reducing to finite-dimensional PDEs/ODEs in Markov, affine, or Lévy-type models (Friz et al., 9 Sep 2025, Friz et al., 2021).

4. Statistical and Machine Learning Applications

The log-signature enables powerful statistical summarization and inference on path-valued data:

Truncated log-signature coordinates provide a geometrically rich, low-dimensional embedding for sequential data, invariant under time-reparameterization and robust to noise (Lai et al., 2019, Barancikova et al., 2024).
In explicit generative modeling, e.g., SigDiffusions, score-based diffusion processes are trained directly on the space of log-signatures, enabling sample generation and closed-form inversion to recover the underlying path or its coefficients in suitable bases (e.g., Fourier, orthogonal polynomials) (Barancikova et al., 2024).
In writer identification, log-signature embeddings of pathlets extracted from image contours yield compact, discriminative codes for clustering or retrieval, outperforming raw signatures in both computational efficiency and classification performance (Lai et al., 2019).
For stochastic processes, the expectation of the signature (and hence its log) uniquely determines the law under growth conditions (Chevyrev–Lyons theorem), making log-signatures central in statistical modeling of time series, financial processes, and rough signals (Friz et al., 2024, Friz et al., 2021).

5. Analytic Properties: Decay, Convergence, and Rigidity

While signature coefficients decay factorially, log-signature coefficients typically decay only geometrically. The analytic radius of convergence of the log-signature's power series is finite except for straight-line paths, as formalized in the Lyons–Sidorova conjecture and subsequent results:

For rough, tree-reduced paths, log-signature coefficients have strictly smaller radii of convergence than signature coefficients (Boedihardjo et al., 22 Jun 2025).
Infinite radius of convergence for log-signature can only arise for line segments (or paths that are locally linear on all subintervals).
Explicit vanishing (determinantal) identities on iterated integrals arise as geometric constraints in such “entire” cases, implying rigidity of the log-signature transform and reinforcing its discriminative power for generic paths (Boedihardjo et al., 22 Jun 2025).

6. Numerical, Algorithmic, and Basis Considerations

The dimension of the truncated signature up to level $n$ is $\sum_{k=0}^n d^k$ , while the dimension of the truncated log-signature (Hall/Lyndon basis for the free Lie algebra) is much smaller, allowing for more efficient representation and clustering in high dimensions (Lai et al., 2019, Barancikova et al., 2024).
Closed-form inversion formulae permit explicit, basis-adapted reconstruction of paths from log-signature data (Fourier, polynomials), with controlled numerical properties and scalable preprocessing (Barancikova et al., 2024).
Modern library implementations exploit precomputed Hall or Lyndon bases, BCH expansions, and vectorized computation for practical deployment in scientific computing and ML pipelines (Reizenstein, 2017).

7. Domain-Specific Applications and Empirical Findings

Empirical studies demonstrate that log-signature features are highly effective for graph-based retrieval and identification tasks. In offline writer identification, log-signature codebooks coupled with codeword co-occurrence features yield top-1 accuracy exceeding 94% on the IAM dataset, 99% on CVL, and outperform raw signatures and many deep learning baselines (Lai et al., 2019). The structural and computational advantages—principled geometric compressiveness, scale-invariance, and reparametrization invariance—are central to these results, and similar architectures are leveraged in generative time-series models, rough path classification, and stochastic kernel learning (Lai et al., 2019, Barancikova et al., 2024, Friz et al., 2024).

References: