Fit returns an all-NaN state with a converged flag (flag 2)

[matthewholman]

Corrected. The original report claimed sequence prediction was ~100x slower than the fit. That measurement was invalid: the fit it was compared against had returned an all-NaN state, so the prediction code was being asked to integrate relative to an undefined epoch. Predicting on valid fits is fine -- 288 objects went through _predict without trouble. There is no performance bug here. Apologies for the noise.

The real problem is what produced that state.

Fitting 4,172 optical observations of (13377) over a 1989-2026 arc, seeded from its MPCORB elements, returns:

flag 2, csq 1.65e9, ndof 8338          (reduced chi-square ~197,000)
epochMJD_TDB   nan
x, y, z        nan
xdot,ydot,zdot nan
cov_0_0        nan

Flag 2 is FLAG_CSQ_TOO_LARGE, "converged; reduced chi-square above threshold", and it is a member of CONVERGED_FLAGS. So a fit whose entire state vector is NaN is reported to callers as converged. Anything that branches on CONVERGED_FLAGS -- or that reads the state without testing it -- consumes it silently. Feeding it to predict is what sent the integrator off, which is how I found it.

Two things worth separating: whether the fit should diverge here at all (a catalogue seed on a well-observed numbered object), and whether a NaN state should ever leave the fitter wearing a converged flag. The second seems worth a guard regardless of the first.

[matthewholman]

Closing: there is no defect here. Both versions of this report were wrong, and the error was mine throughout.

_orbitfit writes the state columns only when res.flag == 0:

success = res.flag == 0
... (tuple(res.state[i] for i in range(6)) if success else (np.nan,) * 6)
    ((res.epoch - 2400000.5) if success else np.nan)
    ("BCART_EQ" if success else "NONE")

So a fit rejected on chi-square reports a NaN state deliberately, and sets FORMAT to "NONE" to say so. The fitter's own state is finite; nothing is corrupted.

What happened: I wrote a harness that accepted flags 0, 2 and 6 as usable, on the assumption that flag 2 meant "converged, chi-square high, residuals still meaningful". It does not -- the state is withheld. I then fed those NaN states to predict, which integrated relative to an undefined epoch and ran for minutes, which is what I originally reported as a prediction performance bug. Predicting from valid fits is fine.

The converged vs accepted distinction is exactly what #527 introduces, so this was already understood.