arxiv: 2604.06602 · v2 · submitted 2026-04-08 · 🌌 astro-ph.IM · astro-ph.EP

Recognition: no theorem link

UMI: GPU-Accelerated Asymmetric Robust Estimator for Photometric Detrending in Exoplanet Transit Searches

Omar Khan

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:32 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.EP

keywords UMIphotometric detrendingexoplanet transitrobust estimatorGPU accelerationasymmetric M-estimatorTESSKepler

0 comments

The pith

An asymmetric robust estimator recovers shallow exoplanet transit depths more accurately while running 69 times faster on GPU.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces UMI as a modified Tukey bisquare M-estimator that applies stronger penalties to downward deviations than upward ones and computes its noise scale only from above-median residuals. This exploits the fact that planetary transits produce brightness decreases but never increases relative to the stellar continuum. Injection-recovery tests on TESS, Kepler, and K2 data show the largest accuracy gains at 0.1 percent transit depths, with median depth errors falling from 20.5 percent to 15.8 percent on TESS and from 14.6 percent to 4.2 percent on Kepler. The fused HIP/CUDA implementation delivers the speedups needed for processing large photometric catalogs.

Core claim

UMI modifies the standard Tukey bisquare M-estimator with an asymmetric weight function that penalizes downward deviations more aggressively than upward ones and an upper-RMS scale estimator computed from above-median residuals only, ensuring that transit dips never contaminate the noise estimate, and implements the full procedure as a fused HIP/CUDA GPU kernel.

What carries the argument

The asymmetric weight function together with the upper-RMS scale estimator inside the modified Tukey bisquare M-estimator, executed as a single fused GPU kernel.

If this is right

At 0.1 percent transit depth, median depth recovery error drops from 20.5 percent to 15.8 percent on TESS data.
The same error drops from 14.6 percent to 4.2 percent on Kepler data.
Detrending time per star falls from 234 ms to 3.4 ms, a 69-fold speedup.
End-to-end pipeline throughput rises by a factor of 37 relative to the wotan biweight implementation.
The accuracy advantage is strongest for transit depths well above the photometric noise floor.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The directional bias in the weighting could be adapted to other astronomical time series that contain one-sided signals, such as certain eclipsing binaries or flare detections.
The reported GPU kernel speed suggests the method could support real-time detrending in future high-cadence surveys.
Performance on ground-based photometry with stronger correlated noise remains untested and may narrow the reported gains.

Load-bearing premise

Transit signals are always strictly below the continuum and no other downward-going systematics or correlated noise will be mistaken for them.

What would settle it

A controlled injection test on light curves that contain artificial downward spikes unrelated to transits, measuring whether UMI produces systematically worse trend fits or depth biases than a symmetric biweight estimator.

Figures

Figures reproduced from arXiv: 2604.06602 by Omar Khan.

**Figure 1.** Figure 1: Weight function comparison. Left: standard Tukey bisquare, which assigns weight w = 0.83 to a transit dip at u = −0.3 (nearly full weight, causing the trend to absorb the dip). Right: UMI asymmetric bisquare (α = 2.0), which reduces the weight to w = 0.41 for the same dip. The shaded region indicates where transit-dip residuals typically fall. The asymmetric version excludes these points from the trend est… view at source ↗

**Figure 2.** Figure 2: Asymmetry parameter optimization on 10,000 TESS stars. Left: depth recovery error versus asymmetry parameter α at 7 depths. The value α = 2.0 is the inflection point where accuracy improves sharply from 1.0 to 2.0 but plateaus beyond. Right: systematic bias increases linearly with α. The −451 ppm bias at α = 2.0 is below the TESS noise floor (∼1000 ppm). Lower is better for both panels. 3.2. Injection-Reco… view at source ↗

**Figure 3.** Figure 3: Transit depth recovery error across 6 injected depths for UMI (default and aggressive modes) versus 5 comparison methods on 1000 real TESS stars. UMI achieves the lowest error at 0.05 to 0.3% depth, which corresponds to the super-Earth and sub-Neptune regime. The aggressive mode (α = 10, c = 2.5) further improves shallow-depth accuracy at the cost of increased bias. Lower is better. 12 hours. This limitati… view at source ↗

**Figure 4.** Figure 4: Left: per-star detrending speed for 8 methods on 1000 TESS stars. UMI (3.4 ms) is the fastest method with useful accuracy, 69× faster than biweight and 113× faster than Welsch. Right: accuracy at 0.1% transit depth (super-Earth regime). UMI is simultaneously the fastest and most accurate method at this depth. UMI welsch biweight savgol 0 50 100 150 200 250 300 350 400 Planets Won 425 179 160 38 Others comb… view at source ↗

**Figure 5.** Figure 5: Known planet recovery on 802 confirmed exoplanets (81 TESS, 721 Kepler). Left: overall win count, showing that UMI recovers more planets (425) than all other methods combined (377). Right: wins broken down by mission, showing that UMI dominates on Kepler while remaining competitive on TESS. Higher is better [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 7.** Figure 7: UMI consistency across 4 Kepler quarters (Q2, Q5, Q9, Q17; 1000 stars each). The maximum spread is 1.5 percentage points at 0.1% depth, demonstrating stable performance across different observing epochs [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

read the original abstract

We present UMI (Unified Median Iterative), a novel robust location estimator for detrending photometric time series in exoplanet transit surveys. UMI modifies the standard Tukey bisquare M-estimator with two innovations: (1) an asymmetric weight function that penalizes downward deviations (transit dips) more aggressively than upward ones, exploiting the physical constraint that transits are always below the stellar continuum, and (2) an upper-RMS scale estimator computed from above-median residuals only, ensuring that transit dips never contaminate the noise estimate. Implemented as a fused HIP/CUDA GPU kernel, UMI achieves 69x faster detrending (3.4 ms vs 234 ms per star) and 37x faster full pipeline throughput compared to the wotan biweight implementation. Injection-recovery tests across TESS, Kepler, and K2 show that UMI's advantage is concentrated at planet-scale transit depths above the photometric noise floor: at 0.1% transit depth, UMI reduces median depth recovery error from 20.5% to 15.8% on TESS (23% improvement) and from 14.6% to 4.2% on Kepler (71% improvement). At shallower depths approaching the noise floor, all sliding-window methods converge to comparable performance. Validated across 802 confirmed exoplanets from TESS and Kepler, UMI occupies a previously unfilled region of the speed-accuracy tradeoff for transit detrending. The tool is publicly available as pip install torchflat.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

UMI adds asymmetric weighting and upper-only scale estimation to the Tukey bisquare, delivering 69x speed and better 0.1% depth recovery in the reported tests, but the bias risk from other downward signals is the part that still needs checking.

read the letter

UMI modifies the Tukey bisquare M-estimator with two targeted changes: an asymmetric weight function that down-weights negative residuals more than positive ones, and a scale estimate taken only from residuals above the median. These exploit the fact that transits sit below the continuum while most noise does not. The GPU kernel then makes the whole thing fast enough for large surveys. The abstract reports 69x faster per-star detrending and 37x pipeline throughput versus the wotan biweight, which is a practical win if the numbers hold up in the code. At 0.1% injected depth the median recovery error drops from 20.5% to 15.8% on TESS and from 14.6% to 4.2% on Kepler, with the gap closing at shallower depths. That pattern is consistent with the method protecting the continuum estimate from the signal itself. The injection-recovery setup on real survey data avoids the circularity problem that sometimes appears in detrending papers. The tool is also released as a pip package, which lowers the barrier for others to test it. The soft spot is the handling of other downward features. Real TESS and Kepler light curves contain instrumental dips, correlated noise, and stellar variability that also go below the median. If the asymmetric penalty treats those the same way it treats transits, the fitted continuum can shift and the recovered depths can be biased. The abstract gives the performance numbers but does not detail how the test light curves were cleaned or whether mixtures of contaminants were included. A reader would want to see that section to judge whether the gains survive realistic conditions. This work is aimed at teams running transit searches on TESS, Kepler, or K2 data who already use sliding-window robust fits and want both speed and better protection for shallow signals. The concrete metrics and open implementation make it worth a referee's time even if the asymmetry needs more stress-testing.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces UMI (Unified Median Iterative), a GPU-accelerated modification of the Tukey bisquare M-estimator for photometric detrending in exoplanet transit searches. It incorporates an asymmetric weight function that penalizes downward deviations more aggressively than upward ones and an upper-RMS scale estimator computed exclusively from above-median residuals. The central claims are a 69x speedup in detrending (3.4 ms per star) relative to wotan biweight, plus improved transit depth recovery in injection-recovery tests: at 0.1% depth, median depth recovery error drops from 20.5% to 15.8% on TESS and from 14.6% to 4.2% on Kepler, with comparable performance at shallower depths and validation across 802 confirmed exoplanets.

Significance. If the performance claims hold under realistic conditions, UMI would occupy a useful position in the speed-accuracy tradeoff for large-scale transit surveys by exploiting the one-sided nature of transits. The GPU kernel implementation and public pip package are concrete strengths that could enable broader adoption. The reported gains are concentrated at planet-scale depths above the noise floor, which aligns with practical needs in TESS/Kepler/K2 analyses.

major comments (2)

[Abstract and §4] Abstract and §4 (injection-recovery tests): The reported depth-recovery improvements (e.g., 14.6% → 4.2% on Kepler at 0.1% depth) rest on the assumption that the asymmetric weighting and upper-RMS scale can isolate transits without bias from other downward systematics. However, the test description does not specify whether the light curves included realistic correlated noise, instrumental dips, or additional astrophysical signals below the continuum; without such mixtures, the 23–71% error reductions may not generalize to actual survey data.
[§3.1] §3.1 (asymmetric weight function and upper-RMS definition): The claim that 'transit dips never contaminate the noise estimate' follows directly from restricting the RMS to above-median residuals, but the interaction between the asymmetric weights and the location estimate itself is not shown to remain unbiased when multiple downward excursions (transits plus systematics) are present simultaneously. A concrete test with injected transits plus independent downward noise components would be required to support the central performance claims.

minor comments (2)

[Abstract] The abstract states '37x faster full pipeline throughput' without defining the baseline pipeline components or whether the comparison includes I/O and downstream transit search steps.
[§3] Notation for the asymmetry tuning constant is introduced but its default value and sensitivity analysis are not tabulated or plotted.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments highlight important aspects of validation and clarity that we will address in the revised manuscript. We respond point by point below.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (injection-recovery tests): The reported depth-recovery improvements (e.g., 14.6% → 4.2% on Kepler at 0.1% depth) rest on the assumption that the asymmetric weighting and upper-RMS scale can isolate transits without bias from other downward systematics. However, the test description does not specify whether the light curves included realistic correlated noise, instrumental dips, or additional astrophysical signals below the continuum; without such mixtures, the 23–71% error reductions may not generalize to actual survey data.

Authors: The injection-recovery tests were performed on real TESS and Kepler light curves drawn from the survey archives (with transits injected at known depths), which inherently contain realistic correlated noise, instrumental artifacts, and other astrophysical signals. The separate validation across 802 confirmed exoplanets from TESS and Kepler further demonstrates performance on unaltered survey data. We acknowledge that §4 could be more explicit about these data properties. In the revision we will expand the test description to include quantitative noise statistics for the light curves, representative examples of included systematics, and a statement confirming that the tests use real survey data rather than idealized white-noise simulations. revision: yes
Referee: [§3.1] §3.1 (asymmetric weight function and upper-RMS definition): The claim that 'transit dips never contaminate the noise estimate' follows directly from restricting the RMS to above-median residuals, but the interaction between the asymmetric weights and the location estimate itself is not shown to remain unbiased when multiple downward excursions (transits plus systematics) are present simultaneously. A concrete test with injected transits plus independent downward noise components would be required to support the central performance claims.

Authors: We agree that an explicit demonstration of unbiased location estimation under simultaneous downward signals strengthens the central claims. While the upper-RMS definition prevents scale contamination by construction, the coupled effect of asymmetric weights on the iterative location estimate merits direct verification. In the revised manuscript we will add a targeted simulation (new panel or short subsection in §3.1) that injects both planetary transits and independent downward systematics (e.g., additional instrumental dips or flares) into controlled light curves and quantifies the residual bias in the recovered continuum. This will directly address the referee’s request for a concrete test. revision: yes

Circularity Check

0 steps flagged

No circularity: method is explicit modification of known estimator; performance from independent injection tests

full rationale

The paper defines UMI explicitly as a modification of the standard Tukey bisquare M-estimator, adding an asymmetric weight function and an upper-RMS scale estimator computed only from above-median residuals. These are presented as direct innovations without deriving them from the target performance metrics. Validation relies on injection-recovery tests using real TESS, Kepler, and K2 light curves with injected transits, which are external benchmarks independent of the estimator's internal parameters or definitions. No self-citations, self-definitional loops, or fitted inputs renamed as predictions appear in the derivation or claims. The central results (speed and depth-recovery improvements) are therefore not equivalent to the inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The abstract describes modifications to an existing estimator but does not specify exact parameter values or additional axioms beyond the physical constraint of transits.

free parameters (1)

asymmetry tuning constant
The strength of the asymmetric penalization is likely controlled by a parameter that may be chosen or fitted.

axioms (1)

domain assumption Transits produce only downward deviations in photometric time series
Invoked to justify the asymmetric weight function.

pith-pipeline@v0.9.0 · 5581 in / 1298 out tokens · 98013 ms · 2026-05-10T18:32:14.577749+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 11 canonical work pages · 1 internal anchor

[1]

, keywords =

Akeson, R. L., Chen, X., Ciardi, D., et al. 2013, Publications of the Astronomical Society of the Pacific, 125, 989, doi: 10.1086/672273

work page doi:10.1086/672273 2013
[2]

Kepler Planet-Detection Mission: Introduction and First Results.Science2010,327, 977

Borucki, W. J., Koch, D., Basri, G., et al. 2010, Science, 327, 977, doi: 10.1126/science.1185402

work page doi:10.1126/science.1185402 2010
[3]

Cleveland, W. S. 1979, Journal of the American Statistical Association, 74, 829, doi: 10.1080/01621459.1979.10481038

work page doi:10.1080/01621459.1979.10481038 1979
[4]

AJ , volume =

Foreman-Mackey, D., Agol, E., Ambikasaran, S., & Angus, R. 2017, The Astronomical Journal, 154, 220, doi: 10.3847/1538-3881/aa9332

work page doi:10.3847/1538-3881/aa9332 2017
[5]

Stahel, W. A. 1986, Robust Statistics: The Approach Based on Influence Functions (John Wiley & Sons)

1986
[6]

J., Mulders G

Hippke, M., David, T. J., Mulders, G. D., & Heller, R. 2019, The Astronomical Journal, 158, 143, doi: 10.3847/1538-3881/ab3984 10Khan

work page doi:10.3847/1538-3881/ab3984 2019
[7]

2019, Astronomy & Astrophysics, 623, A39, doi: 10.1051/0004-6361/201834672

Hippke, M., & Heller, R. 2019, Astronomy & Astrophysics, 623, A39, doi: 10.1051/0004-6361/201834672

work page doi:10.1051/0004-6361/201834672 2019
[8]

Hoare, C. A. R. 1961, Communications of the ACM, 4, 321, doi: 10.1145/366622.366647

work page doi:10.1145/366622.366647 1961
[9]

B., Sobeck, C., Haas, M., et al

Howell, S. B., Sobeck, C., Haas, M., et al. 2014, Publications of the Astronomical Society of the Pacific, 126, 398, doi: 10.1086/676406 Kov´ acs, G., Zucker, S., & Mazeh, T. 2002, Astronomy & Astrophysics, 391, 369, doi: 10.1051/0004-6361:20020802 Lightkurve Collaboration. 2018, Lightkurve: Kepler and TESS time series analysis in Python, Astrophysics Sou...

work page doi:10.1086/676406 2014
[10]

2019, in Advances in Neural Information Processing Systems 32, 8024–8035

Paszke, A., Gross, S., Massa, F., et al. 2019, in Advances in Neural Information Processing Systems 32, 8024–8035

2019
[11]

2014, Experimental Astronomy, 38, 249, doi: 10.1007/s10686-014-9383-4

Rauer, H., Catala, C., Aerts, C., et al. 2014, Experimental Astronomy, 38, 249, doi: 10.1007/s10686-014-9383-4

work page doi:10.1007/s10686-014-9383-4 2014
[12]

R., Winn, J

Ricker, G. R., Winn, J. N., Vanderspek, R., et al. 2015, Journal of Astronomical Telescopes, Instruments, and Systems, 1, 014003, doi: 10.1117/1.JATIS.1.1.014003

work page internal anchor Pith review doi:10.1117/1.jatis.1.1.014003 2015
[13]

Journal of the American Statistical Association , author =

Rousseeuw, P. J., & Croux, C. 1993, Journal of the American Statistical Association, 88, 1273, doi: 10.1080/01621459.1993.10476408

work page doi:10.1080/01621459.1993.10476408 1993
[14]

Tukey, J. W. 1977, Exploratory Data Analysis (Addison-Wesley)

1977