TLS: survey-scale fast path (batch phase-binned kernel + exact refinement)

[johnh2o2]

What

Rewrites Transit Least Squares for survey-scale throughput (N ≫ 1 light curves). The new batch API tls_search_batch() and kernel tls_fast.cu become the default for tls_search / tls_search_gpu / tls_transit (opt out with use_fast=False). The legacy per-point kernel is untouched and retained.

The old kernel did two full O(ndata) passes per (duration, t0) trial and capped light curves at ~3,500 points. Both limits are gone.

Results (end-to-end survey wall time, 100% injected-transit recovery in every regime)

Regime (ndata, periods)	A5000	Ada	V100
TESS FFI (1.3K, 2.5K)	1.25 ms/LC (~800 LC/s)	3.05	1.42
K2 90-d (4.3K, 9.7K)	3.1	6.4	3.9
TESS 2-min (19.7K, 2.5K)	2.8	5.1	3.1
TESS 1-yr (17K, 42K)	18.4	27.3	16.5
Kepler 4-yr (65K, 172K)	168	198	146

Fidelity — not lossy for detection (measured)

The key question: does the coarse epoch grid + refinement sacrifice detectability? Measured with scripts/tls_fidelity_experiment.py — same injected light curves, same Ofir grid, cuvarbase (default t0=3 and reference-matched t0=33) vs the real transitleastsquares, with SDE recomputed by the identical statistic on both methods' χ² spectra (their SR→SDE normalizations differ):

Signal	cuvarbase t0=3 (default)	cuvarbase t0=33 (matched)	recovery
tess-ffi depth 0.005 (strong)	0.99× ref	1.03×	12/12
tess-ffi depth 0.002 (marginal)	0.97× ref	1.01×	10/10
k2 depth 0.004 (narrow)	0.98× ref	1.01×	6/6

Default is within 1–3% of the reference SDE (safe direction — slight under-report), matched t0=33 within 1%, 100% recovery including marginal and narrow transits. SDE is a period-space contrast that normalizes out epoch-grid density, so the coarse grid trades reported t0/parameter precision (restored by the refinement pass) for speed, not detectability. Matched fidelity costs a measured 6–15× (Kepler-4yr 8.4× → 1.48 s/LC).

Speed & cost

Throughput is the market-independent invariant: apples-to-apples on the same machine (same LCs/grid, single GPU vs all CPU cores), ~1,000–3,000× faster than the reference at matched SDE fidelity (more at the default grid; ratio is CPU-dependent). Versus GTLS (the only other GPU TLS, CuPy/RTX 4090, arXiv:2607.00348), a Kepler-class config runs ~22× (matched) to ~190× (default) faster on a cheaper A5000.

Dollar figures are pricing-dependent and were not the headline — an earlier revision over-pinned them by pairing a spot-price GPU with an AWS on-demand CPU. analysis/TLS_COST_ANALYSIS.md has the honest, caveated breakdown (hundreds-to-thousands× cheaper per LC than CPU TLS under any reasonable pricing).

How

• Phase-binned batch kernel: one block per (light curve, period), fold once into shared-memory bins, scan trials against bin-averaged integrated-template tables (S1=∫T, S2=∫T²) with closed-form chi² (chi2_0 − num²/den). Trial cost independent of ndata.
• Period banding: grid split by required bin count so long-period searches don't pay the finest band's cost.
• Float-float fold: ~1e-8 phase error at 4-year baselines with no 1/64-rate double math on consumer GPUs.
• Exact top-K refinement kernel re-fits the best candidate periods on a finer local (duration, t0) grid. Refinement sharpens reported parameters only — SDE/FAP come from the uniform coarse spectrum so the statistic's scale stays consistent (refining only the peak would inflate SDE → false positives).
• Cancellation-free score output; chi² reconstructed in float64 host-side.
• Support fixes: SDE median-detrend window capped at 91 (reference convention; minutes → 0.1 s at 190k periods); duration_grid_keplerian vectorized (1.1 s → 40 ms); threaded per-LC stats; 64-bit offsets; input validation (qmax<1, power-of-two block_size, non-negative refine_top_k).

Validation

• 68/68 GPU tests on RTX 4000 Ada (sm_89), including golden cross-validation vs the reference transitleastsquares package; core suite green on A5000 (sm_86) and V100 (sm_70).
• Fidelity experiment above (SDE parity) + raw numbers in benchmarks/results/tls_survey_jul2026/.
• 25-agent adversarial review; 19 confirmed findings fixed or explicitly deferred.
• Block-size heuristic swept and tuned per compute capability across all three GPU families (no single-GPU over-fit).

Caveats

• Module stays EXPERIMENTAL. Detection SDE is at parity with the reference (above), but that is three regimes, not a full injection–recovery completeness campaign across a (period, depth, ndata) grid (item D3) — the sign-off needed before dropping the flag.
• Small unrelated infra fix: scripts/setup-remote.sh now honors RUNPOD_REMOTE_DIR (some pods have broken /workspace volumes).

🤖 Generated with Claude Code

[johnh2o2]

Follow-up: measured fidelity (is the coarse-grid + refinement lossy?)

Pushed c4d10ff — an apples-to-apples fidelity experiment (scripts/tls_fidelity_experiment.py). It runs the same injected light curves through cuvarbase (default t0_oversample=3 and reference-matched 33) and the reference transitleastsquares on one shared Ofir grid, and recomputes SDE with the identical statistic on both methods' χ² spectra (their SR→SDE normalizations differ, so the statistic is held fixed and only spectrum fidelity varies).

The fast path is not lossy in detectability. On the identical SDE statistic (RTX A5000):

Signal	cuvarbase t0=3 (default)	cuvarbase t0=33 (matched)	reference	recovery
tess-ffi depth 0.005 (strong)	0.99×	1.03×	14.61	12/12
tess-ffi depth 0.002 (marginal)	0.97×	1.01×	12.37	10/10
k2 depth 0.004 (narrow)	0.98×	1.01×	25.82	6/6

Default is within 1–3% of the reference SDE (safe direction — slight under-report); matched t0=33 is within 1%. SDE is a period-space contrast that normalizes out epoch-grid density, so the coarse grid trades reported t0/parameter precision (restored by refinement) for speed, not detectability. Matched fidelity costs a measured 6–15× (Kepler-4yr 8.4× → 1.48 s/LC).

Also corrected the earlier cost claims in analysis/TLS_COST_ANALYSIS.md: the six-figure "cheaper" ratios over-pinned the multiplier by pairing a spot-price GPU with an AWS on-demand CPU. Throughput is the market-independent invariant (thousands× vs CPU at matched SDE fidelity; ~22–190× vs GTLS, arXiv:2607.00348, on cheaper hardware).