Optimal robust detection statistics for pulsar timing arrays

arXiv:2509.06489 · astro-ph.IM, gr-qc · Submitted 2025-09-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: I'm Vera.

Jocelyn: And I'm Jocelyn, here with Subrahmanyan.

Vera: We're looking at the paper: "Optimal robust detection statistics for pulsar timing arrays".

Title and authors: Vera: So this paper, "Optimal robust detection statistics for pulsar timing arrays," is about finding a better way to test for that nano-Hz stochastic gravitational wave background using pulsar timing arrays.

Jocelyn: It’s essentially a critique of the current standard method and a proposal for an improved one.

Subrahmanyan: The core idea here seems to be making the detection statistic more robust against errors we make when we model the timing noise from each individual pulsar.

Vera: Right. They are saying that because the standard statistic doesn't maximize our detection probability at a fixed false alarm probability, it’s not perfectly optimal under our assumed noise and signal models.

Jocelyn: And they call this new NPMV statistic because it tries to be as close as possible to the theoretical Neyman-Pearson optimal statistic while only using cross-correlated data.

Subrahmanyan: That sounds like a clever compromise, trying to get the best of both worlds: sensitivity and robustness against noise modeling errors.

Vera: They derived this by minimizing the variance of the difference between its statistic and the true Neyman-Pearson statistic, but with one big constraint: it only uses cross-correlations.

Jocelyn: So, they are trading some complexity for a specific type of robustness that keeps it away from having to deal with autocorrelations.

Subrahmanyan: That constraint about not using autocorrelation is interesting because that’s where a lot of the uncertainty in pulsar timing noise usually comes from, isn't it?

The paper's summary: Vera: Now they summarize what they did, and the main point is that this NPMV statistic significantly outperforms the conventional cross-correlation detection statistic in simulations.

Jocelyn: They show that when we operate at a specific threshold, their method increases the detection probability by about forty-seven percent compared to the standard DFCC statistic.

Subrahmanyan: That forty-seven percent gain is substantial because it means we’re catching more of the actual signal when we are trying to be conservative with our false alarm rate.

Vera: They compare different statistics, looking at area under the ROC curve, and they show that the NPMV statistic has a better performance in realistic cases.

Jocelyn: The paper also shows specific numbers for the detection probability at a five sigma false alarm probability of two point nine times ten to the minus seven.

Subrahmanyan: So, when we look at those results, the standard DFCC gets about a twenty-seven percent detection probability, but NPMV gets around forty percent. That’s a pretty big difference in what we can claim is real.

Vera: They also showed that the NP statistic itself has an area under the ROC curve of zero point nine eight five, while the standard DFCC is at zero point nine six seven.

Jocelyn: So, they are showing that their approach doesn't just look good in a toy model; it’s better when you apply it to what we see in real PTA data.

Subrahmanyan: The paper then suggests replacing the standard "optimal" cross-correlation statistic with NPMV because it’s more sensitive at the same false alarm probability and handles noise modeling errors better.

The paper's improvements: Vera: So what are these actual improvements they suggest? It really boils down to making our analysis more robust against those pulsar timing residual noise modeling errors.

Jocelyn: They say that by minimizing the variance of its difference with the NP statistic under the constraint of only using cross-correlations, they get a quadratic detection statistic that’s both in between NP and DFCC performance-wise.

Subrahmanyan: That quadratic form is mathematically elegant because it achieves this balance without needing to include those auto-correlations that introduce so much modeling uncertainty.

Vera: They actually provide a Python implementation of the NPMV filter and statistic in the enterprise extensions package, which is a big deal for practical application.

Jocelyn: That means researchers can take their existing analyses based on DFCC and redo them using this new NPMV statistic without having to write everything from scratch.

Subrahmanyan: The paper mentions that the related NPCC statistic is guaranteed to perform even better than NPMV, but it can only be obtained numerically, and their investigations show that in practice, it only provides modest improvements over NPMV.

Vera: So the practical improvement they are pushing is that you get a more reliable detection claim because you're not as sensitive to the specific way you model pulsar noise errors.

Conclusion: Jocelyn: So to wrap up, the paper "Optimal robust detection statistics for pulsar timing arrays" argues that we should move away from the traditional cross-correlation statistic and adopt NPMV.

Vera: They argue it’s more sensitive at a given false alarm probability, and crucially, it’s robust against uncertainties in pulsar timing residual noise modeling.

Subrahmanyan: From my side, I see this as a necessary step toward better signal detection because it directly addresses the modeling errors that plague our current analyses when trying to find those nano-Hz gravitational waves.

Jocelyn: It sounds like they are providing a more reliable tool for interpreting the data we collect from PTAs, and they’ve made it accessible with that Python implementation.

Vera: Yeah, so the main point is that for those working on PTA data right now, switching to NPMV could give them a forty-seven percent boost in detection probability at that five sigma level.

Subrahmanyan: It’s about making sure our detection claims are built on a more solid statistical foundation when we are looking for the hints of gravitational waves in the universe.

Jocelyn: That’s what it is. We'll keep an eye on how this NPMV statistic gets used by the community and what kind of follow-up observations might be needed to settle these claims.

Max-Planck-Institut fur Gravitationsphysik (Albert-Einstein-Institut) · Leibniz Universitat Hannover · Department of Physics and Astronomy, University of Texas Rio Grande Valley

astro-ph.IM, gr-qc

Submitted: 2025-09-08

Updated: 2026-11-17

Comments: 23 pages, 8 figures

Code: https://github.com/nanograv/15yr

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 84/100

The gist: Pulsar timing arrays (PTAs) seek to detect a nano-Hz stochastic gravitational-wave background (GWB) by searching for the characteristic Hellings and Downs angular pattern of timing residual…

Key concepts

Optimal Cross-Correlation Detection Statistic (DFCC)
This is the conventional statistic used by PTA collaborations to assess false alarm probabilities and p-values. It is considered 'optimal' but is not strictly Neyman-Pearson optimal because it can be affected by uncertainties in modeling pulsar timing noise.
Neyman-Pearson (NP) Statistic
The theoretical gold standard for detection statistics, maximizing the probability of correctly detecting a signal at a fixed false alarm probability. The NPMV statistic is designed to be as close to this ideal NP statistic as possible while only using cross-correlated data.
Neyman-Pearson-Minimum-Variance (NPMV) Statistic
A new quadratic detection statistic that minimizes the variance of its difference from the NP statistic, constrained to use only cross-correlations. This makes it robust against noise modeling errors and more sensitive to gravitational wave signals than DFCC.
False Alarm Probability (FAP)
The probability of observing a result as strong as or stronger than the one measured, assuming no actual gravitational wave signal is present. The analysis uses a specific FAP threshold (e.g., 2.9 x 10^-7) to define the '5-sigma' detection threshold.

Terminology

Summary

Pulsar timing arrays (PTAs) seek to detect a nano-Hz stochastic gravitational-wave background (GWB) by searching for the characteristic Hellings and Downs angular pattern of timing residual correlations The NPMV statistic increases the detection probability by 47% when operating at a 5-sigma FAP of 2. p. 0

Where the paper says it

Romano2 1Max-Planck-Institut fur Gravitationsphysik (Albert-Einstein-Institut), Callinstraße 38, D-30167 Hannover, Germany ¨ Leibniz Universitat Hannover, D-30167 Hannover, Germany ¨ 2Department of Physics and Astronomy, University of Texas Rio Grande Valley, One West University Boulevard, Brownsville, TX 78520, USA ABSTRACT Pulsar timing arrays (PTAs) seek to detect a nano-Hz stochastic gravitational-wave background (GWB) by searching for the characteristic Hellings and Downs angular pattern of timing residual correlations. So far, the evidence remains below the conventional 5-𝜎 threshold, as assessed using the literature-standard “optimal cross-correlation detection statistic”.

9 × 10−7 compared to the literature-standard “optimal” cross-correlation detection statistic

The Problem with Existing Statistics

The conventional “optimal cross-correlation detection statistic” (DFCC) is not Neyman-Pearson (NP) optimal because it does not maximize the detection probability at a fixed false alarm probability (FAP), and it is susceptible to uncertainties in modeling pulsar timing noise The NP-optimal detection statistic incorporates autocorrelations, making it more vulnerable to uncertainties in the modeling of pulsar timing noise The authors argue that the NP statistic is not robust because it depends on both autocorrelation and cross-correlations, whereas the DFCC statistic only depends upon pulsar timing residual cross-correlations

The Proposed NPMV Statistic

The paper introduces the “Neyman-Pearson-Minimum-Variance” (NPMV) statistic as a compromise between the NP and DFCC statistics, being “as close as possible” to the NP statistic while using only cross-correlated data Specifically, it minimizes the variance of its difference with the NP statistic subject to the constraint that it only uses cross-correlations This leads to a quadratic detection statistic whose performance lies in between NP and DFCC, but which is robust because it does not use auto-correlations

Derivation and Robustness

The NPMV filter is derived by minimizing the variance of the difference between its statistic and the NP statistic under the null hypothesis subject to the constraint that Diag Q = 0, meaning it does not use autocorrelations For block-diagonal N, this leads to a simple form where QNPMV has diagonal elements equal to zero and off-diagonal elements equal to QNP for non-diagonal indices This is obtained by assuming N is block-diagonal, which is a standard assumption for noise covariance matrices

Performance Comparison

The NPMV statistic significantly outperforms the conventional DFCC statistic in simulations, increasing the detection probability by ∼47% when operating at the 5-sigma detection threshold FAP of 2.9 × 10−7 ROC curves comparing NP, NPCC, NPMV, and DFCC show that the NPMV statistic has better performance (larger area under the ROC curve) than the standard DFCC statistic in realistic cases

Conclusion

The paper concludes that the NPMV statistic should replace the traditional “optimal” cross-correlation detection statistic (DFCC) because it is robust against pulsar noise modeling errors and is more sensitive at a given FAP The NPMV filter can be implemented in the enterprise extensions package, allowing existing analyses based on DFCC to be redone using the NPMV detection statistic

Exact values of the noise parameters can be found at:

https://github.com/nanograv/15yr stochastic analysis/blob/main/tutorials/data/optstat ml gamma4p33. p. 8

Where the paper says it

5 Exact values of the noise parameters can be found at: https: //github. com/nanograv/15yr stochastic analysis/blob/main/tutorials/ data/optstat ml gamma4p33.

p. 8
Where the paper says it

5 Exact values of the noise parameters can be found at: https: //github. com/nanograv/15yr stochastic analysis/blob/main/tutorials/ data/optstat ml gamma4p33.

json

False-alarm probability (FAP)

0 0.2 0.4 0.6 0.8 1

Detection probability

Area under curve: NP AUC = 0.985 NPCC AUC = 0. p. 9

Where the paper says it

8 1 Detection probability Area under curve: NP AUC = 0. 985 NPCC AUC = 0.

975 NPMV AUC = 0. p. 9
Where the paper says it

985 NPCC AUC = 0. 975 NPMV AUC = 0.

975 DFCC AUC = 0. p. 9
Where the paper says it

975 NPMV AUC = 0. 975 DFCC AUC = 0.

967

At “5-σ” FAP of 2.9×10−7:

NP detection probability 18.7% NPCC detection probability 16. p. 7

Where the paper says it

The first is NP, which (by construction) has the best performance. Next is the (numerically-constructed) NPCC filter, which maximizes the detection probability at fixed FAP, subject to the constraint that it uses only cross-correlations between pulsars.

1% NPMV detection probability 16. p. 0
Where the paper says it

1 × 10−15 and spectral index 𝛾 = 13/3. Compared to the literature-standard “optimal” cross-correlation detection statistic, the NPMV statistic increases the detection probability by 47% when operating at a 5-𝜎 FAP of 𝛼 = 2.

1% DFCC detection probability 12. p. 1
Where the paper says it

This corresponds to filters Q which vanish on the diagonal. At a given FAP, we will see that the detection probability for this “NeymanPearson-Minimum-Variance” (NPMV) statistic is higher than that of the conventional DFCC statistic.

8%

Figure 2.

ROC curves of different detection statistics, for the toy model of Sec. p. 9

Where the paper says it

8% Figure 2. ROC curves of different detection statistics, for the toy model of Sec.

p. 9
Where the paper says it

8% Figure 2. ROC curves of different detection statistics, for the toy model of Sec.

7.1 Left panel: linear scale Legend gives AUC “area under the ROC curve” for the different statistics (larger is better) Right panel: same as left panel, but for logarithmic axes Legend gives detection probability for the different statistics at the 5-sigma discovery threshold FAP of 2. p. 9
Where the paper says it

Legend gives AUC “area under the ROC curve” for the different statistics (larger is better). Right panel: same as left panel, but for logarithmic axes.

9 × 10−7 (leftmost plot boundary)

Figure 3.

ROC curves for the DFCC, NP, and NPMV detection statistics, for the NANOGrav 15-year PTA fixed γ = 13/3 GWB and pulsar noise model (Agazie et al. p. 9

Where the paper says it

7% Figure 3. ROC curves for the DFCC, NP, and NPMV detection statistics, for the NANOGrav 15-year PTA fixed 𝛾 = 13/3 GWB and pulsar noise model (Agazie et al.

p. 9
Where the paper says it

7% Figure 3. ROC curves for the DFCC, NP, and NPMV detection statistics, for the NANOGrav 15-year PTA fixed 𝛾 = 13/3 GWB and pulsar noise model (Agazie et al.

2023a) At the discovery threshold FAP of 2. p. 8
Where the paper says it

These plots of detection probability versus FAP are shown in Fig. At the discovery 5-𝜎 threshold FAP of 2.

p. 8
Where the paper says it

These plots of detection probability versus FAP are shown in Fig. At the discovery 5-𝜎 threshold FAP of 2.

9 × 10−7 (leftmost plot boundary) the NPMV statistic increases the detection probability by ∼ 47% compared to the currently-standard “optimal” cross-correlation detection statistic DFCC

Acknowledgments

We thank Wang-Wei Yu for helpful discussions J.D.R. acknowledges financial support from NSF Physics Frontier Center Award PFC-2020265 and start-up funds from the University of Texas Rio Grande Valley J. p. 9

Where the paper says it

ACKNOWLEDGMENTS We thank Wang-Wei Yu for helpful discussions. acknowledges financial support from NSF Physics Frontier Center Award PFC-2020265 and start-up funds from the University of Texas Rio Grande Valley.

D.R. performed part of this work at the Aspen Center for Physics, which is supported by National Science Foundation grant PHY-2210452

Data Availability Statement

No open or closed data were used in this work We make use of pulsar positions and model parameters as described in publications in the literature (Agazie et al. 2023a) All figures can be regenerated using code available online at https://github. p. 9

Where the paper says it

We make use of pulsar positions and model parameters as described in publications in the literature (Agazie et al. All figures can be regenerated using code available online at https: //github.

com/vhaasteren/robust-ds-figures

References

Afzal, A., Agazie, G. p. 0

Where the paper says it

Keywords: gravitational waves — methods: data analysis — methods: statistical — pulsars: general — stars: neutron 1. INTRODUCTION Four pulsar timing array (PTA) collaborations have reported evidence for a nano-Hz gravitational-wave background (GWB) (Agazie et al.

, Anumarlapudi, A. p. 9
Where the paper says it

, Agazie, G. , Anumarlapudi, A.

p. 9
Where the paper says it

, Agazie, G. , Anumarlapudi, A.

, et al. 2023 The Astrophysical Journal Letters 951 L11 doi: 10. p. 9
Where the paper says it

, Anumarlapudi, A. 2023, The Astrophysical Journal Letters, 951, L11, doi: 10.

3847/2041-8213/acdc91

Agazie, G., Anumarlapudi, A., Archibald, A. p. 9

Where the paper says it

, Anumarlapudi, A. , Archibald, A.

M., et al. 2023a The Astrophysical Journal Letters 951 L8 doi: 10. p. 9
Where the paper says it

, Archibald, A. 2023a, The Astrophysical Journal Letters, 951, L8, doi: 10.

3847/2041-8213/acdac6

—. 2023b The Astrophysical Journal Letters 951 L50 doi: 10. p. 9

Where the paper says it

3847/2041-8213/acdac6 —. 2023b, The Astrophysical Journal Letters, 951, L50, doi: 10.

3847/2041-8213/ace18a

Allen, B. 2023 Physical Review D 107 043018 doi: 10. p. 9

Where the paper says it

20510 Allen, B. 2023, Physical Review D, 107, 043018, doi: 10.

1103/PhysRevD. p. 9
Where the paper says it

2023, Physical Review D, 107, 043018, doi: 10. 1103/PhysRevD.

p. 9
Where the paper says it

2023, Physical Review D, 107, 043018, doi: 10. 1103/PhysRevD.

p. 9
Where the paper says it

2023, Physical Review D, 107, 043018, doi: 10. 1103/PhysRevD.

p. 9
Where the paper says it

2023, Physical Review D, 107, 043018, doi: 10. 1103/PhysRevD.

p. 9
Where the paper says it

2023, Physical Review D, 107, 043018, doi: 10. 1103/PhysRevD.

p. 9
Where the paper says it

2023, Physical Review D, 107, 043018, doi: 10. 1103/PhysRevD.

p. 9
Where the paper says it

2023, Physical Review D, 107, 043018, doi: 10. 1103/PhysRevD.

p. 9
Where the paper says it

2023, Physical Review D, 107, 043018, doi: 10. 1103/PhysRevD.

107.043018

Allen, B., Dhurandhar, S. p. 9

Where the paper says it

043018 Allen, B. , Dhurandhar, S.

, Gupta, Y. p. 9
Where the paper says it

, Dhurandhar, S. , Gupta, Y.

, et al. 2023 The International Pulsar Timing Array Checklist for the Detection of Nanohertz Gravitational Waves arXiv doi: 10. p. 9
Where the paper says it

, Gupta, Y. 2023, The International Pulsar Timing Array Checklist for the Detection of Nanohertz Gravitational Waves, arXiv, doi: 10.

48550/arXiv. p. 9
Where the paper says it

2024, The NANOGrav 15 Yr Data Set: Posterior Predictive Checks for Gravitational-Wave Detection with Pulsar Timing Arrays, arXiv, doi: 10. 48550/arXiv.

p. 9
Where the paper says it

2024, The NANOGrav 15 Yr Data Set: Posterior Predictive Checks for Gravitational-Wave Detection with Pulsar Timing Arrays, arXiv, doi: 10. 48550/arXiv.

p. 9
Where the paper says it

2024, The NANOGrav 15 Yr Data Set: Posterior Predictive Checks for Gravitational-Wave Detection with Pulsar Timing Arrays, arXiv, doi: 10. 48550/arXiv.

2304.04767

Allen, B., & Romano, J. p. 0

Where the paper says it

One potential source for this GWB is a population of supermassive black-hole binaries at the centers bruce. romano@utrgv.

D. 2023 Physical Review D 108 043026 doi: 10. p. 9
Where the paper says it

, & Romano, J. 2023, Physical Review D, 108, 043026, doi: 10.

1103/PhysRevD.108.043026

—. 2024 Optimal Reconstruction of the Hellings and Downs Correlation arXiv doi: 10. p. 9

Where the paper says it

043026 —. 2024, Optimal Reconstruction of the Hellings and Downs Correlation, arXiv, doi: 10.

48550/arXiv.2407.10968

Anholm, M., Ballmer, S. p. 9

Where the paper says it

10968 Anholm, M. , Ballmer, S.

, Creighton, J. p. 9
Where the paper says it

, Ballmer, S. , Creighton, J.

p. 9
Where the paper says it

, Ballmer, S. , Creighton, J.

D. E., Price, L. p. 9
Where the paper says it

, Creighton, J. , Price, L.

R., & Siemens, X. p. 9
Where the paper says it

, Price, L. , & Siemens, X.

2009 Physical Review D 79 084030 doi: 10. p. 9
Where the paper says it

, & Siemens, X. 2009, Physical Review D, 79, 084030, doi: 10.

1103/PhysRevD.79.084030

Antoniadis, J., Arumugam, P. p. 9

Where the paper says it

084030 Antoniadis, J. , Arumugam, P.

, Arumugam, S. p. 9
Where the paper says it

084030 Antoniadis, J. , Arumugam, P.

, et al. 2023 Astronomy & Astrophysics 678 A49 doi: 10. p. 9
Where the paper says it

, Arumugam, S. 2023, Astronomy & Astrophysics, 678, A49, doi: 10.

1051/0004-6361/202346842

—. 2024a Astronomy & Astrophysics 685 A94 doi: 10. p. 9

Where the paper says it

, Arumugam, S. 2023, Astronomy & Astrophysics, 678, A49, doi: 10.

1051/0004-6361/202347433

—. 2024b Astronomy & Astrophysics 690 A118 doi: 10. p. 9

Where the paper says it

, Arumugam, S. 2023, Astronomy & Astrophysics, 678, A49, doi: 10.

1051/0004-6361/202348568

Chamberlin, S. J., Creighton, J. D. E., Siemens, X. p. 9

Where the paper says it

, Price, L. , & Siemens, X.

p. 9
Where the paper says it

, Price, L. , & Siemens, X.

p. 9
Where the paper says it

, Price, L. , & Siemens, X.

, et al. 2015 Physical Review D 91 044048 doi: 10. p. 9
Where the paper says it

20510 Allen, B. 2023, Physical Review D, 107, 043018, doi: 10.

1103/PhysRevD.91.044048

Cox, D. 1962 Renewal Theory 1st edn Methuen’s Monographs on Applied Probability and Statistics (London: Methuen)

Das, A., & Geisler, W. p. 5

Where the paper says it

A discussion of their cumulative distribution function (CDF), and their application to PTAs (for example, computing 𝑝- values) may be found in Hazboun et al. This also contains citations to the more general literature on this topic, for example Das & Geisler (2021).

S. 2021 Journal of Vision 21 1 doi: 10. p. 9
Where the paper says it

, Anumarlapudi, A. 2023, The Astrophysical Journal Letters, 951, L11, doi: 10.

1167/jov.21.10.1

Ellis, J. A., Siemens, X., & van Haasteren, R. p. 0

Where the paper says it

Draft version September 9, 2025 Typeset using LATEX twocolumn style in AASTeX631 Optimal robust detection statistics for pulsar timing arrays Rutger van Haasteren, 1 Bruce Allen, 1 and Joseph D. Romano2 1Max-Planck-Institut fur Gravitationsphysik (Albert-Einstein-Institut), Callinstraße 38, D-30167 Hannover, Germany ¨ Leibniz Universitat Hannover, D-30167 Hannover, Germany ¨ 2Department of Physics and Astronomy, University of Texas Rio Grande Valley, One West University Boulevard, Brownsville, TX 78520, USA ABSTRACT Pulsar timing arrays (PTAs) seek to detect a nano-Hz stochastic gravitational-wave background (GWB) by searching for the characteristic Hellings and Downs angular pattern of timing residual correlations.

2013 The Astrophysical Journal 769 63 doi: 10. p. 9
Where the paper says it

, Anumarlapudi, A. 2023, The Astrophysical Journal Letters, 951, L11, doi: 10.

88/0004-637X/769/1/63

Ellis, J. A., Vallisneri, M. p. 1

Where the paper says it

For a PTA similar to the published 1 In practice, for composite hypotheses (models with free parameters), these statistics are still used (Agazie et al. 2024; Vallisneri et al.

, Taylor, S. p. 2
Where the paper says it

To encourage its adoption, we provide a Python implementation of the NPMV filter and statistic in the enterprise extensions package (Ellis et al. 2020; Taylor et al.

p. 2
Where the paper says it

To encourage its adoption, we provide a Python implementation of the NPMV filter and statistic in the enterprise extensions package (Ellis et al. 2020; Taylor et al.

R., & Baker, P. T. 2020 ENTERPRISE: Enhanced Numerical Toolbox Enabling a Robust PulsaR Inference SuitE Zenodo doi: 10. p. 2
Where the paper says it

2 At a given FAP, the NPMV statistic achieves higher detection probability while remaining robust to pulsar noise modeling errors. To encourage its adoption, we provide a Python implementation of the NPMV filter and statistic in the enterprise extensions package (Ellis et al.

5281/zenodo.4059815

Hazboun, J. S., Meyers, P. M., Romano, J. p. 0

Where the paper says it

One potential source for this GWB is a population of supermassive black-hole binaries at the centers bruce. romano@utrgv.

D., Siemens, X., & Archibald, A. p. 9
Where the paper says it

, Anumarlapudi, A. , Archibald, A.

M. 2023 Phys Rev D 108 104050 doi: 10. p. 9
Where the paper says it

, & Romano, J. 2023, Physical Review D, 108, 043026, doi: 10.

1103/PhysRevD.108.104050

Hellings, R. W., & Downs, G. p. 0

Where the paper says it

Draft version September 9, 2025 Typeset using LATEX twocolumn style in AASTeX631 Optimal robust detection statistics for pulsar timing arrays Rutger van Haasteren, 1 Bruce Allen, 1 and Joseph D. Romano2 1Max-Planck-Institut fur Gravitationsphysik (Albert-Einstein-Institut), Callinstraße 38, D-30167 Hannover, Germany ¨ Leibniz Universitat Hannover, D-30167 Hannover, Germany ¨ 2Department of Physics and Astronomy, University of Texas Rio Grande Valley, One West University Boulevard, Brownsville, TX 78520, USA ABSTRACT Pulsar timing arrays (PTAs) seek to detect a nano-Hz stochastic gravitational-wave background (GWB) by searching for the characteristic Hellings and Downs angular pattern of timing residual correlations.

S. 1983 The Astrophysical Journal 265 L39 doi: 10. p. 9
Where the paper says it

, Anumarlapudi, A. 2023, The Astrophysical Journal Letters, 951, L11, doi: 10.

1086/183954

Horn, R. A., & Johnson, C. p. 5

Where the paper says it

The general case has a real data space, a normal distribution defined by a real positive-definite symmetric covariance matrix, and a statistic (1) defined by a real symmetric matrix 4 Since N is Hermitian and positive-definite, this follows from the Schur product theorem (see Assertion (c) of (7. 3) in Horn & Johnson (2012), with A = N −1 and B = (N −1 ) ∗ ).

R. 2012 Matrix Analysis 2nd edn Cambridge University Press doi: 10. p. 5
Where the paper says it

The general case has a real data space, a normal distribution defined by a real positive-definite symmetric covariance matrix, and a statistic (1) defined by a real symmetric matrix 4 Since N is Hermitian and positive-definite, this follows from the Schur product theorem (see Assertion (c) of (7. 3) in Horn & Johnson (2012), with A = N −1 and B = (N −1 ) ∗ ).

1017/CBO9781139020411

Isserlis, L. 1918 Biometrika 12 134 doi: 10. p. 5

Where the paper says it

6 van Haasteren, Allen, Romano Expanding the variance of 𝚫 in terms of the data 𝑧, we find: 𝑉(Q) = |𝐷(𝑧 | 𝚫)|2 𝐻0 − |⟨𝐷(𝑧 | 𝚫)⟩𝐻0 | 2 = ∑︁ 𝑎,𝑏,𝑐,𝑑 Δ𝑎𝑏Δ ∗ 𝑐𝑑 𝑧 ∗ 𝑎 𝑧𝑏𝑧𝑐 𝑧 ∗ 𝑑 𝐻0 − 𝑧 ∗ 𝑎 𝑧𝑏 𝐻0 𝑧𝑐 𝑧 ∗ 𝑑 𝐻0 = ∑︁ 𝑎,𝑏,𝑐,𝑑 Δ𝑎𝑏Δ ∗ 𝑐𝑑 𝑧 ∗ 𝑎 𝑧𝑐 𝐻0 𝑧𝑏𝑧 ∗ 𝑑 𝐻0 = ∑︁ 𝑎,𝑏,𝑐,𝑑 Δ𝑎𝑏Δ ∗ 𝑐𝑑𝑁 ∗ 𝑎𝑐𝑁𝑏𝑑 = Tr(𝚫N𝚫N) , (26) where the third equality follows from Isserlis’s theorem (Isserlis 1918), the fourth equality from ⟨𝑧𝑎𝑧 ∗ 𝑏 ⟩𝐻0 ≡ 𝑁𝑎𝑏, and the fifth from the Hermitian property of N and 𝚫. We now use the method of Lagrange multipliers to solve the constrained optimization problem.

1093/biomet/l2.i-2.i-5

Powell, M. J. D. 2009 The BOBYQA algorithm for bound constrained optimization without derivatives Technical Report DAMTP 2009/NA06 University of Cambridge Cambridge http://www. p. 8

Where the paper says it

In practice, we find that from generic initial conditions, many optimizers converge to DFCC rather than to the desired cross-correlation NPCC solution. To overcome this unwanted convergence to DFCC, we employ an ensemble search strategy, combining the stochastic simultaneous perturbation stochastic approximation (SPSA) method (Spall 1998)—whose stochastic updates resemble an annealing process—with subsequent local refinement using the derivative-free BOBYQA algorithm (Powell 2009).

damtp.cam.ac.uk/user/na/NA papers/NA2009 06

Rasmussen, C. E., & Williams, C. p. 4

Where the paper says it

Relating DF and NP filters for realistic PTA data Realistic PTA data consists of multiple samples per pulsar. These data may be modeled using Gaussian processes (Rasmussen & Williams 2006; van Haasteren & Vallisneri 2014).

K. 2006 Gaussian Processes for Machine Learning MIT Press

Reardon, D. J., Zic, A., Shannon, R. M., et al. 2023 The Astrophysical Journal Letters 951 L6 doi: 10. p. 9

Where the paper says it

, Anumarlapudi, A. 2023, The Astrophysical Journal Letters, 951, L11, doi: 10.

3847/2041-8213/acdd02

Romano, J. D., & Cornish, Neil. p. 8

Where the paper says it

(54) Here, 𝐻 ≈ 70 km s−1 Mpc−1 is the present-day Hubble constant, and 𝑓yr ≡ 1/yr [see e. 5 in (Romano & Cornish 2017)].

J. 2017 Living Reviews in Relativity 20 2 doi: 10. p. 3
Where the paper says it

CURN is a conservative choice of 𝐻0. Rather than saying “there is no GWB”, it says “there is a background, but not of the type described by the general theory of relativity (GR)”.

1007/s41114-017-0004-1

Spall, J. 1998 IEEE Transactions on Aerospace and Electronic Systems 34 817 doi: 10. p. 7

Where the paper says it

For low FAP, we only need the 16 𝜆 𝑗 > 0 terms in (50). This is still possible in the precision provided by IEEE 754 doubles, which have a 53 bit mantissa.

1109/7.705889

Taylor, S. R., Baker, P. T., Hazboun, J. p. 5

Where the paper says it

GENERALIZED 𝜒 2 DISTRIBUTION For a multivariate normal distribution, the detection and false-alarm probabilities associated with quadratic statistics (1) are described by generalized 𝜒 2 distributions. A discussion of their cumulative distribution function (CDF), and their application to PTAs (for example, computing 𝑝- values) may be found in Hazboun et al.

S., & Simon, J. 2021 Enterprise extensions Vallisneri, M. p. 2
Where the paper says it

2 At a given FAP, the NPMV statistic achieves higher detection probability while remaining robust to pulsar noise modeling errors. To encourage its adoption, we provide a Python implementation of the NPMV filter and statistic in the enterprise extensions package (Ellis et al.

, Meyers, P. M., Chatziioannou, K., & Chua, A. J. K. 2023 Physical Review D 108 123007 doi: 10. p. 9
Where the paper says it

, & Romano, J. 2023, Physical Review D, 108, 043026, doi: 10.

1103/PhysRevD.108.123007

van Haasteren, R. 25 On the Calculation of P-Values for Quadratic Statistics in Pulsar Timing Arrays arXiv doi: 10. p. 9

Where the paper says it

3847/2041-8213/ace18a —. 2024, The NANOGrav 15 Yr Data Set: Posterior Predictive Checks for Gravitational-Wave Detection with Pulsar Timing Arrays, arXiv, doi: 10.

48550/arXiv.2506.10811

van Haasteren, R., & Vallisneri, M. p. 1

Where the paper says it

For a PTA similar to the published 1 In practice, for composite hypotheses (models with free parameters), these statistics are still used (Agazie et al. 2024; Vallisneri et al.

2014 Physical Review D 90 104012 doi: 10. p. 9
Where the paper says it

20510 Allen, B. 2023, Physical Review D, 107, 043018, doi: 10.

1103/PhysRevD.90.104012

Vigeland, S. J., Islo, K., Taylor, S. R., & Ellis, J. p. 2

Where the paper says it

2 At a given FAP, the NPMV statistic achieves higher detection probability while remaining robust to pulsar noise modeling errors. To encourage its adoption, we provide a Python implementation of the NPMV filter and statistic in the enterprise extensions package (Ellis et al.

A. 2018 Physical Review D 98 044003 doi: 10. p. 9
Where the paper says it

20510 Allen, B. 2023, Physical Review D, 107, 043018, doi: 10.

1103/PhysRevD.98.044003

Xu, H., Chen, S., Guo, Y., et al. 2023 Research in Astronomy and Astrophysics 23 75024 doi: 10. p. 9

Where the paper says it

, Arumugam, S. 2023, Astronomy & Astrophysics, 678, A49, doi: 10.

1088/1674-4527/acdfa5

Figure 1.

In the complex z-plane, the integrand of (45) is analytic apart from n + 1 poles along the imaginary axis We illustrate the 3-dimensional case for which E has one positive and two negative eigenvalues If γ < 0 we close the path of integration with a half-circle in the lower half complex f-plane which encloses all of the poles along the negative imaginary axis, as shown in Fig. p. 6

Where the paper says it

In the complex 𝑧-plane, the integrand of (45) is analytic apart from 𝑛 + 1 poles along the imaginary axis. We illustrate the 3-dimensional case for which E has one positive and two negative eigenvalues.

1 >

ROC curves of different detection statistics, for the toy model of Sec. 7.1 These compare the standard “optimal” cross-correlation deflection statistic DFCC, the Neyman-Pearson-optimal statistic NP, the cross-correlation-only Neyman-Pearson statistic NPCC, and the crosscorrelation-only Neyman-Pearson-Minimum Variance statistic NPMV Left panel: linear scale Legend gives AUC “area under the ROC curve” for the different statistics (larger is better) Right panel: same as left panel, but for logarithmic axes Legend gives detection probability for the different statistics at the 5-sigma discovery threshold FAP of 2. p. 9

Where the paper says it

These compare the standard “optimal” cross-correlation deflection statistic DFCC, the Neyman-Pearson-optimal statistic NP, the cross-correlation-only Neyman-Pearson statistic NPCC, and the crosscorrelation-only Neyman-Pearson-Minimum-Variance statistic NPMV. Left panel: linear scale.

9 × 10−7 leftmost plot boundary >

ROC curves for the DFCC, NP, and NPMV detection statistics, for the NANOGrav 15-year PTA fixed γ = 13/3 GWB and pulsar noise model (Agazie et al. 2023a) At the discovery threshold FAP of 2.9 × 10−7 leftmost plot boundary the NPMV statistic increases the detection probability by ∼ 47% compared to the currently-standard “optimal” cross-correlation detection statistic DFCC

The NPMV filter is implemented in the enterprise extensions package, meaning existing analyses based on the DFCC statistic can be easily redone using the NPMV detection statistic The related NPCC statistic is guaranteed to perform even better than NPMV, but it can only be obtained numerically, and our investigations indicate that in practice it only provides modest improvements over NPMV

5 Exact values of the noise parameters can be found at:

https://github.com/nanograv/15yr stochastic analysis/blob/main/tutorials/data/optstat ml gamma4p33.json

At “5-σ” FAP of 2.9×10−7: NP detection probability 5. p. 0

Where the paper says it

So far, the evidence remains below the conventional 5-𝜎 threshold, as assessed using the literature-standard “optimal cross-correlation detection statistic”. While this quadratic combination of cross-correlated data maximizes the deflection (signal-to-noise ratio), it does not maximize the detection probability at fixed false-alarm probability (FAP), and therefore is not Neyman-Pearson (NP) optimal for the assumed noise and signal models.

1% NPMV detection probability 4. p. 0
Where the paper says it

1 × 10−15 and spectral index 𝛾 = 13/3. Compared to the literature-standard “optimal” cross-correlation detection statistic, the NPMV statistic increases the detection probability by 47% when operating at a 5-𝜎 FAP of 𝛼 = 2.

0% DFCC detection probability 2. p. 1
Where the paper says it

This corresponds to filters Q which vanish on the diagonal. At a given FAP, we will see that the detection probability for this “NeymanPearson-Minimum-Variance” (NPMV) statistic is higher than that of the conventional DFCC statistic.

p. 1
Where the paper says it

This corresponds to filters Q which vanish on the diagonal. At a given FAP, we will see that the detection probability for this “NeymanPearson-Minimum-Variance” (NPMV) statistic is higher than that of the conventional DFCC statistic.

7%

8 Conclusion

This paper introduces a new quadratic detection statistic for pulsar timing arrays (PTA), which we call the “Neyman-Pearson-Minimum-Variance” statistic and denote with NPMV We argue that the NPMV statistic should replace the traditional “optimal” cross-correlation detection statistic, which we denote by DFCC, and which is currently used by all of the PTA collaborations to assess their false alarm probabilities (FAPs) and p-values Like DFCC, the NPMV statistic uses only pulsar crosscorrelations, making it robust against uncertainties in pulsar timing residual noise modeling But the NPMV statistic has a significant advantage over DFCC: at a given FAP, it is more sensitive In simulations that model the current generation of PTAs, NPMV has a 47% higher detection probability than DFCC, when operating at the 5sigma-equivalent FAP of 2.9 × 10−7 The related NPCC statistic is guaranteed to perform even better than NPMV, but it can only be obtained numerically, and our investigations indicate that in practice it only provides modest improvements over NPMV

At “5-σ” FAP of 2.9×10−7: NP detection probability 18. p. 0

Where the paper says it

So far, the evidence remains below the conventional 5-𝜎 threshold, as assessed using the literature-standard “optimal cross-correlation detection statistic”. While this quadratic combination of cross-correlated data maximizes the deflection (signal-to-noise ratio), it does not maximize the detection probability at fixed false-alarm probability (FAP), and therefore is not Neyman-Pearson (NP) optimal for the assumed noise and signal models.

p. 0
Where the paper says it

So far, the evidence remains below the conventional 5-𝜎 threshold, as assessed using the literature-standard “optimal cross-correlation detection statistic”. While this quadratic combination of cross-correlated data maximizes the deflection (signal-to-noise ratio), it does not maximize the detection probability at fixed false-alarm probability (FAP), and therefore is not Neyman-Pearson (NP) optimal for the assumed noise and signal models.

7% NPMV detection probability 4. p. 0
Where the paper says it

1 × 10−15 and spectral index 𝛾 = 13/3. Compared to the literature-standard “optimal” cross-correlation detection statistic, the NPMV statistic increases the detection probability by 47% when operating at a 5-𝜎 FAP of 𝛼 = 2.

0% DFCC detection probability 2.7%

The NPMV filter is implemented in the enterprise extensions package (Ellis et al. 2020; Taylor et al. p. 2

Where the paper says it

To encourage its adoption, we provide a Python implementation of the NPMV filter and statistic in the enterprise extensions package (Ellis et al. 2020; Taylor et al.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Robust statistical inference for pulsar timing arrays can be improved by replacing standard optimal cross-correlation detection statistics with a Neyman-Pearson-Minimum-Variance (NPMV) statistic when analyzing PTA data. This improvement makes the analysis more sensitive to GWB signals while remaining robust against uncertainties in pulsar timing residual noise modeling, as the NPMV statistic is as close as possible to the NP statistic under the constraint that it only uses cross-correlated data.

  2. The improved AI system can perform higher fidelity detection claims by utilizing the NPMV filter and statistic instead of the conventional cross-correlation deflection statistic (DFCC). This leads to a 47% increase in detection probability when operating at a 5 sigma false alarm probability of 2.9 x 10-7 for simulated PTA data matching NANOGrav results.

  3. The system can incorporate advanced noise modeling uncertainties more effectively by utilizing the structure of the NPMV filter, which is derived by minimizing the variance of its difference with the NP statistic subject to using only cross-correlations, thus making it less affected by pulsar noise modeling errors.

Abstract

Pulsar timing arrays (PTAs) seek to detect a nano-Hz stochastic gravitational-wave background (GWB) by searching for the characteristic Hellings and Downs angular pattern of timing residual correlations. So far, the evidence remains below the conventional 5σ threshold, as assessed using the literature-standard "optimal cross-correlation detection statistic". While this quadratic combination of cross-correlated data maximizes the deflection (signal-to-noise ratio), it does not maximize the detection probability at fixed false-alarm probability (FAP), and therefore is not Neyman-Pearson (NP) optimal for the assumed noise and signal models. The NP-optimal detection statistic is a different quadratic form, but is not used because it also incorporates autocorrelations, making it more susceptible to uncertainties in the modeling of pulsar timing noise. Here, we derive the best compromise: a quadratic detection statistic which is as close as possible to the NP-optimal detection statistic (minimizing the variance of its difference with the NP statistic) subject to the constraint that it only uses cross-correlations, so that it is less affected by pulsar noise modeling errors. We study the performance of this new NPMV statistic for a simulated PTA whose noise and (putative) signal match those of the NANOGrav 15-year data release: GWB amplitude A gw = 2.1 times 10-15 and spectral index γ= 13/3. Compared to the literature-standard "optimal" cross-correlation detection statistic, the NPMV statistic increases the detection probability by 47% when operating at a 5σ false-alarm probability of α= 2.9 times 10-7.

Sources

Related papers