Confidently Wrong: Why Ignoring Binaries Biases IMF Inference at Large Sample Sizes

The gist The stellar initial mass function (IMF) high-mass slope α is routinely measured by fitting single-star models to photometric samples that contain 20–90% unresolved binaries, which introduces a systematic negative bias on α that becomes confidently wrong at large sample sizes where the bias exceeds the reported uncertainty <ref:2603.15779#pg2>.

Systematic Bias and Crossover

The paper investigates how ignoring binary stars biases IMF inference at large sample sizes, finding that this bias is constant with sample size N while statistical uncertainties shrink as 1/√N <ref:2603.15779#pg3>. The bias is bracketed between two limiting observation operators: mass-addition (mobs = m1 + m2), which yields a mass-addition bias of 0.054–0.086 with a crossover to confidently wrong at Ncross ∼ 5,000–10,000, and luminosity-addition (mobs = L−1(L1 + L2)), which shows a lower bias of 4–5× smaller at 75,000–150,000.

Observation Operators and Bias Magnitude

The study brackets the real-world bias between two idealized observation operators. The mass-addition operator is described as the worst case: the observer naively assigns the total system mass as the mass of a single star (mobs = m1(1 + q)). Conversely, luminosity-addition is an idealized photometric case based on the ZAMS mass–luminosity relation, representing an idealized lower-bias photometric case for unevolved populations. The luminosity-addition bias is 4–5× smaller than mass-addition because the steep ZAMS mass–luminosity relation strongly suppresses the secondary’s contribution to the inferred mass.

Binary-Aware Inference Framework

The paper employs a binary-aware mixture likelihood to recover the true slope, which marginalizes over the Moe & Di Stefano (2017) binary population model. This model accounts for both single stars and unresolved binaries using an effective binary probability fb(m1) and a mass-ratio distribution p(q m1). The inference is performed using Hamiltonian Monte Carlo (HMC; Betancourt 2018) with the No-U-Turn Sampler (NUTS; Hoffman & Gelman 2014).

Implications for Future Surveys

The crossover sample sizes reach survey-relevant scales, indicating that current and upcoming surveys will deliver N = 104–106 resolved stars per rich cluster, where single-star IMF measurements become bias-dominated unless binaries are modeled. The paper concludes that binary-aware inference is likely necessary to avoid binary-driven systematic bias in the large-N single-star-fitting regime.

Model Robustness and Limitations

The study confirms that the bias is robust across independent realizations, as the interquartile range (IQR) of the mass-addition bias across seeds is ≲ 0.007 at N = 100,000. Limited misspecification tests show that perturbing the assumed binary fraction by ±20% or substituting a flat mass-ratio distribution shifts the recovered α by only ≤ 0.02, far smaller than the naive bias of 0.054–0.086. However, limitations include an underestimation of contamination from higher-order multiples and a need for survey-specific forward models to account for selection effects and observational noise.

This paper provides a benchmark framework for the large-N single-star-fitting regime; survey-specific applications require forward models incorporating photometric noise, completeness, and stellar evolution. The natural next step is to extend this framework from idealized mass-based inference to realistic photometric pipelines that operate on synthetic color–magnitude diagrams, incorporating photometric noise, completeness, and stellar evolution. Both packages will be described in future methods papers and released as open-source software. The author used Claude Code (Anthropic) and Codex (OpenAI) to assist with analysis scripting, numerical verification, and prose editing. The author reviewed and verified all code, outputs, and text, and takes full responsibility for the integrity and accuracy of the work. The analysis code and HMC posterior samples (∼870 runs) are available from the author upon reasonable request. The progenax package is planned for public release as open-source software. The natural next step is to extend this framework from idealized mass-based inference to realistic photometric pipelines that operate on synthetic color–magnitude diagrams, incorporating photometric noise, completeness, and stellar evolution. Both packages will be described in future methods papers and released as open-source software. The author thanks Eric Sandquist and Kristen Dage for discussions about LSST-era star cluster science, and Mike Grudi´c for compiling the published IMF slope measurements used in Figures 1 and 8. This work made use of progenax (Rosen, in prep.), JAX (Bradbury et al. 2018), NumPyro (Phan et al. 2019; Bingham et al. 2019), matplotlib (Hunter 2007), and NumPy (Harris et al. 2020). The AI Assistance Statement states that the author used Claude Code (Anthropic) and Codex (OpenAI) to assist with analysis scripting, numerical verification, and prose editing. All scientific content, methodology, and interpretations are the author’s own. The author reviewed and verified all code, outputs, and text, and takes full responsibility for the integrity and accuracy of the work. The data availability states that the analysis code and HMC posterior samples (∼870 runs) are available from the author upon reasonable request. The progenax package is planned for public release as open-source software. The natural next step is to extend this framework from idealized mass-based inference to realistic photometric pipelines that operate on synthetic color–magnitude diagrams, incorporating photometric noise, completeness, and stellar evolution. Both packages will be described in future methods papers and released as open-source software. The author thanks Eric Sandquist and Kristen Dage for discussions about LSST-era star cluster science, and Mike Grudi´c for compiling the published IMF slope measurements used in Figures 1 and 8. This work made use of progenax (Rosen, in prep.), JAX (Bradbury et al. 2018), NumPyro (Phan et al. 2019; Bingham et al. 2019), matplotlib (Hunter 2007), and NumPy (Harris et al. 2020). The AI Assistance Statement states that the author used Claude Code (Anthropic) and Codex (OpenAI) to assist with analysis scripting, numerical verification, and prose editing. All scientific content, methodology, and interpretations are the author’s own. The author reviewed and verified all code, outputs, and text, and takes full responsibility for the integrity and accuracy of the work <ref:2644.

1. The AI can perform binary-aware IMF inference by implementing a mixture likelihood that marginalizes over a period-marginalized Moe & Di Stefano (2017) binary population model, which recovers the true slope in the synthetic tests presented here.

2. The system can predict when a single-star IMF measurement is confidently wrong by calculating the crossover sample size Ncross based on the observation operator (mass-addition or luminosity-addition), as noted: The crossover occurs at sample sizes reached by modern surveys in these benchmark tests.

3. The AI can provide an order-of-magnitude benchmark for binary systematics, specifically determining that Mass-addition reaches the bias-dominated regime by N ∼ 5,000–10,000 and Luminosity-addition reaches it by N ∼ 75,000–150,000.

The stellar initial mass function (IMF) high-mass slope α is routinely measured by fitting single-star models to photometric samples that contain unresolved binaries. This practice introduces a systematic negative bias on α that is constant with sample size N. Posterior credible intervals shrink as 1/sqrt N, so at sufficiently large N the bias exceeds the full 95% credible-interval width and the true value falls outside the credible interval - a regime we call "confidently wrong." Building on earlier studies of the bias magnitude, we quantify the sample size at which this bias exceeds the statistical uncertainty. We quantify this bias using two idealized observation operators: mass-addition (m obs=m 1+m 2), a formal upper bound on unresolved-system mass overestimation, and luminosity-addition (m obs=L-1(L 1+L 2)), an idealized lower-bias photometric case based on the ZAMS mass-luminosity relation. Across four astrophysical environments spanning α=1.60 - 2.30, we find: (1) mass-addition bias of 0.054-0.086 with crossover to confidently wrong at N cross about5, 000 - 10, 000; (2) luminosity-addition bias of 0.011-0.021 with N cross about75, 000 - 150, 000; and (3) a binary-aware mixture likelihood that marginalizes over the Moe & Di Stefano (2017) binary population model recovers the true slope in the synthetic tests presented here. Published single-star IMF slopes can therefore plausibly carry systematic errors of order 0.01-0.09 if unresolved binaries are not modeled, comparable to or exceeding reported uncertainties in some regimes. Since current and upcoming surveys (Gaia, JWST, Roman, LSST) will deliver large resolved-star samples, binary-aware inference is likely necessary to avoid binary-driven systematic bias in the large- N single-star-fitting regime.