Modifications of the BIC for order selection in finite mixture models

Finite mixture models are ubiquitous in modern statistical modeling, and a recurring practical issue is choosing the model order. In \citet[Sankhyā Series A, \textbf62, pp. 49–66]{keribin2000consistent}, the Bayesian information criterion (BIC) was proved consistent in mixtures, but under strong regularity, including high moments and high-order derivatives of the component density. We introduce the $ν$-BIC and $ε$-BIC, which weight the BIC penalty by negligibly small logarithmic factors immaterial in practice. This minor modification yields consistency under substantially weaker conditions, without differentiability and with mild moment assumptions, and we also give a misspecification result: when the truth lies outside the candidate family, any vanishing-penalty IC eventually selects a Kullback–Leibler optimal order among candidates. Finally, we clarify two limitations of consistent IC-based selection in mixtures: there is no universally minimal BIC-scale penalty within our sufficient conditions, and order consistency can conflict with minimax optimality in Hellinger risk. We illustrate the theory for Gaussian mixtures, non-differentiable Laplace mixtures, heavy-tailed $t$-mixtures, and mixtures of regression models.

💡 Research Summary

This paper revisits the problem of selecting the number of components (order) in finite mixture models from the perspective of information criteria (IC). While the Bayesian Information Criterion (BIC) is widely used because it enjoys consistency—i.e., it selects the true order with probability tending to one as the sample size grows—its classical consistency proofs rely on very strong regularity conditions. Keribin (2000) showed BIC’s consistency for mixtures but required the component density to possess up to five derivatives and for all those derivatives to have finite third moments. Subsequent works (e.g., Gassiat & Van Handel 2012) extended the result to infinite candidate orders but imposed even stronger smoothness and moment assumptions. Consequently, many practically important mixtures—such as Laplace mixtures (non‑differentiable), heavy‑tailed t‑mixtures, or mixtures of regression models—fall outside the scope of existing theory.

The authors propose two modest modifications of the BIC penalty that they call ν‑BIC and ε‑BIC. Both retain the familiar term (½ dim Sₖ log n) but multiply it by a factor that is asymptotically negligible:

• ν‑BIC: penₙ,ν(k) = α(k) n⁻¹ Lₙ^{∘ν}(n) log n,
• ε‑BIC: penₙ,ε(k) = α(k) n⁻¹ (log n)^{1+ε},

where Lₙ(n)=log(e∨n) and Lₙ^{∘ν} denotes ν‑fold composition of Lₙ. The function α(k)=dim(Sₖ)/2 coincides with the usual BIC coefficient. By choosing ν large enough or ε small enough, the extra factor stays below 1.1 for astronomically large sample sizes (e.g., ν=3 yields Lₙ^{∘3}(n)≤1.1 up to n≈5.7×10⁸). Hence, in practice the ν‑BIC or ε‑BIC behaves indistinguishably from the classical BIC, while theoretically it provides a safety margin that relaxes the required assumptions.

The main technical contribution is a set of consistency theorems under dramatically weaker conditions:

1. Compact parameter spaces: The parameter set Sₖ for each candidate order is assumed only to be compact.
2. Glivenko‑Cantelli (GC) class: The class of log‑densities must satisfy a uniform law of large numbers; this is ensured by modest bracketing entropy conditions.
3. Lipschitz component densities: Each component density ϕ(x;θ) needs to be Lipschitz in θ with an envelope having a finite second moment. No differentiability is required.
4. Moment condition: Only a second‑order moment of the envelope is needed; higher‑order moments are unnecessary.

Using these assumptions, the authors develop a uniform concentration bound for the empirical log‑likelihood (Proposition 5) based on bracketing entropy integrals J(δ). The bound shows that the supremum deviation between empirical and population log‑likelihoods over Sₖ is of order Oₚ(√{J(δ)/n}), which is sufficiently small to guarantee that the penalized criterion selects the true order with probability → 1. This yields consistency for both ν‑BIC and ε‑BIC.

A second major result addresses model misspecification. When the true data‑generating density f₀ does not belong to any mixture family under consideration, any IC whose penalty vanishes (penₙ(k)→0) but does not vanish too quickly (√n penₙ(k)→∞) will eventually select the order whose best‑fitting mixture minimizes the Kullback–Leibler divergence to f₀. Thus, the proposed criteria retain a meaningful target even under misspecification.

The paper also discusses two inherent limitations of IC‑based order selection in mixtures:

• No universally minimal BIC‑scale penalty: Within the broad sufficient conditions, many different penalty functions (different ν or ε, or even log‑log factors) guarantee consistency. Hence, consistency alone cannot single out an “optimal” penalty.
• AIC/BIC tension: In a simple Gaussian mixture setting, an AIC‑type criterion attains the parametric minimax Hellinger risk, whereas any order‑consistent criterion (including BIC, ν‑BIC, ε‑BIC) incurs a risk that exceeds the minimax bound by an unbounded factor. This reveals a fundamental trade‑off: insisting on order consistency can sacrifice optimal estimation risk.

The authors illustrate their theory with four concrete families:

• Gaussian mixtures – confirming that the new penalties behave like BIC in practice while allowing non‑regular component configurations.
• Laplace mixtures – demonstrating that non‑differentiable components satisfy the Lipschitz condition and thus enjoy consistency.
• Heavy‑tailed t‑mixtures – showing that only a second moment of the envelope is needed, accommodating infinite higher moments.
• Mixtures of regression models – extending the framework to conditional densities.

Finally, the paper situates its contributions within the broader literature on mixture model selection, noting connections to singular BIC (sBIC), WBIC, complete‑data likelihood approaches, and recent “vanishing‑penalty” frameworks (e.g., Nguyen 2024). It emphasizes that while many alternative methods (hypothesis testing, shrinkage estimators, risk‑based bounds) exist, the ν‑BIC and ε‑BIC provide a theoretically sound, practically indistinguishable modification of the classic BIC that works under far weaker regularity.

In summary, the paper introduces two minimally altered BIC‑type criteria that retain the familiar form of BIC but achieve consistency under substantially weaker assumptions, handle misspecified models, and expose fundamental trade‑offs between order consistency and minimax risk. This advances both the theory and practice of mixture model order selection, especially for non‑smooth or heavy‑tailed component distributions.