[pmlugato]

AxisExpPolyModel is a per-(process, cell) exponential-polynomial background of any order:

rnorm = exp(lnAmpl_c + sum_k c_k P_k(t))

Here t is the shape-axis coordinate mapped onto [-1, 1] from the axis edges. P_k are polynomials orthonormalised (modified Gram-Schmidt) on the actual bin grid. Order 1 is the AxisExpModel slope. Each higher order adds one direction and nests the order below at zero, so the χ² difference between adjacent orders is a one-parameter likelihood ratio.

Why orthogonalise on the grid. Any degree-N basis spans the same functions, so the fitted optimum is unchanged, but the Hessian is not. On bin centres, nominal Legendre P₂ overlaps the constant at 1e-2. That constant is the free per-cell lnAmpl, so the fit has a nearly flat direction. On the grid the overlaps are below 1e-15 for orders 1–4.

Each order can be grouped over its own subset of the cell axes: for example, an amplitude per cell and a shape shared across cells. All parameters are unbounded.

--paramModel AxisExpPolyModel <channel> <procs> <shape_axis> <cell_axes> <order> [<axes_for_order_1> [<axes_for_order_2> ...]]

We use it for the B+ → J/ψ K+ background in the WRemnants muon-calibration fits. There the measured gain is 2.00× at order 2, 2.03× at order 3, 2.07× at order 4, and nothing at order 5.

Tests

tests/test_axis_exp_poly_model.py (11 tests, passing on main) checks:

• orthonormality of the basis on the grid for orders 1–4;
• parameter count and names;
• a flat background at zero;
• the analytic order-1 form;
• nesting of the orders;
• the degenerate-order and bad-grouping errors;
• a fit that recovers the parameters it was generated from (|Δ| < 1e-6).

🤖 Generated with Claude Code

[github-actions]

Issues Found

No issues found.

Summary

Total issues: 0 critical, 0 important, 0 minor.

Overall Verdict

Status: Patch is correct

Explanation: AxisExpPolyModel correctly implements the ParamModel contract: parameter layout (per-process lnAmpl block plus per-order groupings), channel-ordered reshapes, broadcasting in compute, masked-channel handling, and metadata attributes all match the sibling AxisExpModel and the Fitter's expectations. The Gram-Schmidt basis on the bin grid, degenerate-order validation, and CLI parsing behave as documented, and the tests exercise orthonormality, naming, nesting, error paths, and an end-to-end fit recovery using verified tensorwriter/fitter APIs. No qualifying issues found.

Review details

• Model: cern/Kimi-K3-FLATIRON

Review metrics: 13 turns · 17 tool calls · 5m 11s

• Tokens: 267.1K input · 241.0K cached (90%) · 18.2K output

[cmsmwbot]

CI status — workflow Hodor AI review, run 42661709996: ✅ 1/1 jobs green (slowest/only review, 5m47s). Main CI workflow still in progress at this head.

[cmsmwbot]

CI status — workflow CI, run 42661710587: ✅ 19/19 jobs green (slowest all-unit-tests, 1h20m; run wall time 1h38m at 3a4a09992). Head is green and ready.