Junoh Kang

Probabilistic Signature Inversion: Learning Conditional Distributions from Truncated Signatures

Junoh Kang

Abstract

How can we invert a truncated signature? The full signature determines a path up to tree-like equivalence, motivating the single-path reconstruction viewpoint prevalent in existing inversion methods. However, finite truncation generally leaves multiple compatible paths. This ambiguity motivates a probabilistic view of signature inversion: learning the conditional distribution of paths given their truncated signatures under an underlying path law.

Viewing signature inversion as a probabilistic problem, we quantify its intrinsic uncertainty using posterior-sampling reconstruction error: the expected discrepancy between independent samples from the conditional path distribution. For Gaussian process families, conditioning on linear statistics yields tractable theoretical benchmarks, including a closed-form expression for log-GBM. These benchmarks provide upper bounds on the corresponding errors under richer signature conditioning.

We then use signature-conditioned flow matching to learn a conditional path sampler. We evaluate whether its samples match tractable reference posteriors, reproduce the conditioning signatures, and preserve the distributional and temporal properties of the underlying paths. Experiments cover synthetic stochastic processes and financial time series.

This talk is based on joint work with Kiseop Lee and Bohyung Han: arXiv:2606.15332.

Our speaker

Junoh Kang is a Ph.D. candidate in Electrical and Computer Engineering at Seoul National University, advised by Prof. Bohyung Han. His research focuses on generative modeling, with current applications to financial market simulation. He previously worked on efficient sampling for diffusion models. He received his B.S. in Statistics and Mathematics from Seoul National University.

To become a member of the Rough Path Interest Group, register here for free.