Skip to content

lbfgs-ad from the automatic initial guess stalls at ε_E ≈ 0.1 for chirped SHG pulses #34

Open

Description

@chrisbrahms

What happens

When the input pulse is chirped, lbfgs-ad started from croak's automatic initial guess stops in a local minimum. The retrieved field is about 10% off (ε_E ≈ 0.07–0.12) even though the trace error is lower than that of a transform-limited pulse retrieved the same way. Started from the true pulse instead, the same solver reaches ε_E ≈ 0.02 with a lower trace error. More iterations from the automatic start change nothing.

Seen in a simulation study: SHG FROG traces of 3.5 fs pulses at 1030 nm, simulated by an independent full-field propagation code after 0.5 µm of BBO. The retrieval runs retrieve_from_tracedata(td, algorithm="lbfgs-ad", maxiters=800, fit_tau0=True) with the Type I crystal model and two-beam smearing from the dispersive-SHG branches (#31 and the stacked follow-ups):

pulse start iterations trace error ε_E
TOD +86 fs³ truth 800 5.5e-5 0.020
TOD +86 fs³ automatic 800 7.6e-5 0.107
TOD +86 fs³ automatic 4000 7.6e-5 0.107
GDD stretched 2× truth 800 2.6e-4 0.014
GDD stretched 2× automatic 800 2.7e-4 0.070
GDD stretched 2× automatic 4000 2.7e-4 0.070

Transform-limited pulses on the same grids reach ε_E 0.01–0.03 from the automatic start. The same floor shows up for every chirped and TOD pulse in the study, at 650, 800 and 1030 nm. Most of the error is in the spectral amplitude: the retrieved spectrum has spiky wings on the blue side, where the spectral intensity is 1–10 % of its peak.

Expected

From the automatic start, the solver should reach the truth's basin, or croak should give a way to get there: a better initial guess for chirped pulses, restarts, or a different solver for the first iterations. Since the initial guess is deterministic, more seeds give the same result.

Not yet known

  • Whether this happens with the ideal SHG model and no crystal or smearing. That decides whether it is the start or the crystal model's landscape.
  • Whether other solvers (copra, lbfgs) start well enough from the automatic guess.
  • A reproduction from croak's own synthetic traces. The traces above came from an external simulation, so a minimal reproduction still needs writing.

Drafted by Claude (AI).

Activity

  1. added
    bugSomething isn't working
    priority: mediumReal work that should not wait long
    area: solversRetrievers and their optimisers
    on Nov 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area: solversRetrievers and their optimisersbugSomething isn't workingpriority: mediumReal work that should not wait long

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions