Skip to content

TODO_DEFERRED.md

Latest commit

 

History

History
61 lines (48 loc) · 2.93 KB

TODO_DEFERRED.md

File metadata and controls

61 lines (48 loc) · 2.93 KB

Deferred TODO

Notes on unrelated items noticed during the v1.0.0 modernization pass. Not blocking v1.0.0; revisit later.

Matrix scaling: robustness parity with SciPy

  • Our hand-rolled scalers in wlsqm/utils/lapackdrivers.pyx (rescale_columns_c, rescale_rows_c, rescale_twopass_c, rescale_ruiz2001_c, rescale_scalgm_c) do not detect singular rows/columns. LAPACK's DGEEQU does, and rescale_dgeequ_c now propagates that as numpy.linalg.LinAlgError. The others silently divide by the row/column norm without checking for zero, which for a pathological input will produce inf/nan scaling factors and poison the downstream solve. Add explicit zero-row/zero-column checks to each scaler and propagate via the existing int return-code channel.
  • Also add SciPy-style input validation at the Python entry points (rescale_*, do_rescale): reject non-finite entries, non-square where required, zero-size arrays, wrong dtype/layout. Right now we rely on the Cython memoryview machinery to reject the most egregious cases, but the error messages are opaque.
  • While at it, revisit wlsqm.fitter.simple.fit_* / wlsqm.fitter.expert.ExpertSolver input validation too — same flavour.

Issue #5: non-contiguous memoryview ValueError

Open on GitHub: ValueError: Buffer and memoryview are not contiguous in the same dimension. Reproduce, then decide: clearer error at the Python entry point, or fix to accept non-contiguous inputs where cheap. Deferred from the main modernization pass.

sudoku_lhs.py extraction

examples/sudoku_lhs.py is a small Latin-hypercube helper that could stand alone as a tiny package. Not urgent.

Documentation / tutorial pass for doc/

doc/wlsqm.pdf and doc/wlsqm_gen.pdf are the original theory documents. Potential arXiv tutorial target — rewrite them with the "surrogate model, not Taylor series" framing, and mention that the same method appears in the literature under several names (MLS, WLSQM, "diffuse approximation").

Windows long width in ExpertSolver.interpolate

ExpertSolver.interpolate() in wlsqm/fitter/expert.pyx backs its I_out return array with a Cython long[::1] view and an allocation of dtype=np.int_. On Linux/macOS 64-bit, C long is 64 bits; on Windows 64-bit (MSVC) it is 32 bits. The function therefore silently produces different element widths across platforms, and cannot address more than 2**31 local models on Windows. Low priority (no one has a WLSQM setup with > 2 billion neighborhoods) but the fix is small: switch both the cdef view type and the numpy dtype to a fixed width, e.g. np.int64_t / np.int64.

Weighting function support

The literature distinguishes MLS variants mainly by the weighting function applied to neighbor points (e.g. distance-based decay). wlsqm currently has only WEIGHT_UNIFORM and WEIGHT_CENTER. Adding a user-supplied radial weight function would be a natural extension and improves robustness against outliers.