Skip to content

DYSCO latent dynamics study

Muratore and Mathis introduce DYSCO as a multi-view temporal contrastive learning method for recovering latent trajectories and a structured dynamics model from noisy high-dimensional observations. The paper targets the gap between latent representation learning and explicit system identification: it learns an encoder from observations to latent state and a dynamics model expressed in a predefined functional basis.

Source

  • Paolo Muratore and Mackenzie Weygandt Mathis (2026), Extracting Governing Equations from Latent Dynamics via Multi-View Contrastive Learning, arXiv:2606.13260v1.
  • Source URL: https://arxiv.org/abs/2606.13260.
  • Verified in this repository on 2026-06-26 against the arXiv record and the local v1 PDF.

Technical reading

DYSCO assumes multiple independent noisy views of the same latent dynamical process. The multi-view structure gives the encoder a signal shared across views, while view-specific nuisance noise is not shared. The learned dynamics are parameterized with a structured library of basis functions, so the recovered latent flow can be fed into sparse symbolic-regression procedures.

The theoretical claim is identification of the latent state and deterministic dynamics up to a common affine indeterminacy under asymptotic assumptions. That affine gauge matters for MIF: a recovered coordinate system is not automatically the physical coordinate system, and symbolic coefficients are gauge-dependent unless the downstream recovery step accounts for the affine transform.

The experiments cover several dynamical regimes: Lorenz, Duffing, FitzHugh-Nagumo, Stuart-Landau, and metastable systems. The paper reports recovery under Gaussian observation noise and also studies Poisson observation noise, which is practically relevant for neural and event-count data but does not strictly satisfy every theoretical assumption in the main theorem.

Relevance to SCPN-MIF-CORE

DYSCO is relevant as a future diagnostics-identification pattern, not as a current trigger-runtime component. MIF already has a hard boundary around the verified chamber-side trigger path:

  • DYSCO does not change the verified trigger path.
  • DYSCO is not a production dependency.
  • Any learned model remains advisory and subordinate to the safety certificate, veto, and formally verified trigger fabric.
  • Any future DYSCO-style surrogate must declare its feature boundary, training provenance, uncertainty envelope, and fallback behaviour before it can affect a MIF decision surface.

The closest existing MIF surface is the merge-window predictor. That predictor is already constrained to a closed lock-window feature vector, runtime weights with verified-surrogate provenance, and safety/veto subordination. A DYSCO-style study could inform how future sibling-generated surrogate data are screened before they become predictor training material, but it does not widen the current predictor's inputs.

Boundary notes

The paper does not validate a pulsed magneto-inertial fusion plant, an FPGA trigger, a capacitor-bank controller, or a FUSION-owned plasma solver. It also does not remove the need for source-grounded physics constraints. For MIF, the correct use is narrow:

  1. Treat it as a candidate method for identifying latent dynamics from repeated noisy diagnostic views.
  2. Keep the learned dynamics outside the formally verified safety path.
  3. Require sibling ownership for any physical plasma state reconstruction.
  4. Record affine-gauge handling before interpreting symbolic coefficients as physical equations.

Follow-up criteria

A future DYSCO-derived MIF experiment is in scope only if it satisfies all of the following:

  • The data source is real or sibling-owned simulated diagnostics, not invented examples.
  • The latent dimensionality and basis library are justified before training.
  • The recovered model is benchmarked against held-out trajectories and a source-grounded baseline.
  • The output is advisory only unless a separate formal/safety argument proves the promoted use.
  • The resulting claim states whether it is latent-coordinate, gauge-corrected, or physical-coordinate evidence.