ACDC: Accumulated Cutoff Discrepancy Criterion

• ACDC is a robust model selection method that accumulates component-level discrepancies to determine the number of latent processes in misspecified models.
• It uses a truncation cutoff (ρ) to ignore minor discrepancies, balancing the sensitivity of likelihood-based methods with nonparametric robustness.
• Empirical results show that ACDC outperforms traditional criteria like AIC, BIC, and the Elbow method in both synthetic and real-data scenarios.

Accumulated Cutoff Discrepancy Criterion (ACDC) is a robust model selection criterion for discovering a mechanistically meaningful number of latent processes in misspecified latent variable models. In the formulation proposed in “Robust Model Selection for Discovery of Latent Mechanistic Processes,” ACDC accumulates component-wise discrepancies after truncation at a tolerance level $\rho$ and selects the smallest $K$ minimizing the resulting robust loss. Its stated objective is to combine the sensitivity of likelihood-based methods with the robustness of nonparametric ones, especially in settings where likelihood-based criteria overestimate the number of latent processes under misspecification while robust nonparametric procedures can be overly conservative (Li et al., 25 Feb 2026).

1. Problem setting and motivation

The criterion is formulated for latent structures that consist of $K_o$ mechanistic processes, where $K_o$ is unknown. One considers a sequence of model families

$$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$

where $K$ counts processes or components (Li et al., 25 Feb 2026).

A canonical example is a mixture model with

$$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$

latent labels $z_n\in\{1,\ldots,K\}$,

$$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$

This formulation places ACDC in a general family of component-based latent variable models rather than restricting it to a single architecture (Li et al., 25 Feb 2026).

The central motivation is misspecification. Likelihood-based criteria such as AIC, BIC, and marginal likelihood are consistent when $m^{(K_o)}$ is well-specified. Under misspecification, however, they asymptotically select models approaching

$K$0

the closest element to the truth in Kullback–Leibler divergence, typically with no finite $K$1, which leads to overestimating $K$2. Robust nonparametric methods are less sensitive to parametric misspecification but can underestimate $K$3 and lack generality. ACDC is defined to address this tension through a component-level, mechanism-aware discrepancy (Li et al., 25 Feb 2026).

2. Mechanistic framework and component-level discrepancy

The general framework allows observations $K$4 to depend on covariates $K$5. Each observation is represented as a deterministic combination of component contributions,

$K$6

with the “no-contribution” property

$K$7

Each component contribution is

$K$8

where $K$9, $K_o$0, and $K_o$1 are independent noise variables (Li et al., 25 Feb 2026).

The distinctive feature of ACDC is its mechanism-aware component-level discrepancy. Rather than comparing only fitted and observed distributions at the aggregate level, it defines a discrepancy through the conditional distribution of the component noise $K_o$2 given observed data, for indices that are “used” $K_o$3. For model parameter $K_o$4 and data distribution $K_o$5, the inferred distribution of the $K_o$6-th component noise is

$K_o$7

and the component-level discrepancy is

$K_o$8

Here $K_o$9 is a divergence or distance between probability measures, with examples including KL divergence, Wasserstein distance, and MMD (Li et al., 25 Feb 2026).

In the plug-in implementation, one fits a point estimator $K_o$0, computes per-sample conditional noise laws

$K_o$1

usage indicators $K_o$2, and counts $K_o$3, then aggregates these into an empirical component noise distribution $K_o$4 and evaluates

$K_o$5

(Li et al., 25 Feb 2026).

For mixture models, the paper also gives a component-level realized discrepancy via responsibilities

$K_o$6

followed by

$K_o$7

The paper develops ACDC and its theory using the noise-based construction, while this mixture-specific formulation illustrates the same mechanistic principle (Li et al., 25 Feb 2026).

3. Criterion, computation, and cutoff calibration

For a fixed $K_o$8, ACDC defines the robust loss as

$K_o$9

where $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$0 is the cutoff level. The selected number of components is

$$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$1

The outer minimum resolves ties in favor of parsimony (Li et al., 25 Feb 2026).

A notable detail is that ACDC sorts nothing explicitly. It accumulates component-wise discrepancies through a sum after truncation at $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$2. The truncation prevents misspecification within the acceptable tolerance from influencing the loss and thereby prevents overfitting by adding spurious components (Li et al., 25 Feb 2026).

Operationally, the procedure runs over $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$3. For each $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$4, the model is fit to obtain $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$5; the conditional noise laws and usage indicators are computed; the empirical component noise distributions are formed; and the discrepancies $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$6 are evaluated. The criterion is then assembled as $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$7, and $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$8 is selected as the smallest minimizer. The computational cost is dominated by model fitting for each $$m^{(K)}=\{P_\theta:\theta\in\Theta^{(K)}\},\qquad K=1,2,\ldots,$$9 and by evaluation of the divergence $K$0 on the empirical component noise distributions. Per-component effective sample size is $K$1 (Li et al., 25 Feb 2026).

The paper gives several strategies for choosing $K$2. A domain-knowledge calibration uses related datasets with ground-truth labels or in silico mixtures to maximize an accuracy metric such as F-measure over $K$3, then transfers the calibrated $K$4 to new datasets of similar type. A generally applicable unlabeled approach plots $K$5 as a function of $K$6 for each $K$7, with small $K$8 used for visual separation, and chooses the smallest $K$9 at the first “stability interval” where one $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$0 curve remains minimal across a substantial range of $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$1. An automated version defines a minimum stability width $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$2, tracks intervals $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$3 over which a single $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$4 is best, and selects the first interval whose width is at least $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$5 (Li et al., 25 Feb 2026).

Divergence choice is part of the practical specification. The recommended default is KL divergence because of its connection to likelihood and invariance to smooth reparameterizations of $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$6. In high-dimensional settings, KL estimation via $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$7-NN can degrade; the paper suggests exploiting structure or using Sinkhorn, described there as entropy-regularized Wasserstein, for stability. It also provides consistent KL estimators, including a one-sample $$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$8-NN estimator, a bias-corrected variant, and an adaptive-$$m^{(K)}=\left\{P_\theta=\sum_{k=1}^K \eta_k F_{\phi_k}:\theta=(\eta,\phi_1,\ldots,\phi_K)\in\Theta^{(K)}\right\},$$9 version with $z_n\in\{1,\ldots,K\}$0 and $z_n\in\{1,\ldots,K\}$1 (Li et al., 25 Feb 2026).

4. Robust model selection consistency

The theoretical target is a robust model selection consistency property defined through a distribution-level discrepancy $z_n\in\{1,\ldots,K\}$2 and the component-level discrepancy $z_n\in\{1,\ldots,K\}$3. For each $z_n\in\{1,\ldots,K\}$4,

$z_n\in\{1,\ldots,K\}$5

and the worst-case component-wise discrepancy at the true $z_n\in\{1,\ldots,K\}$6 is

$z_n\in\{1,\ldots,K\}$7

A model selection procedure $z_n\in\{1,\ldots,K\}$8 is $z_n\in\{1,\ldots,K\}$9-robustly consistent for $$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$0 and $$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$1 if, whenever the mismatch or “gap condition”

$$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$2

holds, then

$$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$3

(Li et al., 25 Feb 2026).

For mixture models, the paper states that ACDC using the noise-based component discrepancy is $$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$4-robustly consistent provided $$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$5 is one of KL divergence, Wasserstein distance, or MMD. The theorem specifies that for KL divergence, $$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$6 and $$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$7; for Wasserstein distance or MMD, $$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$8 and $$z_n\mid \theta \sim \mathrm{Categorical}(\eta),\qquad x_n\mid z_n=k,\theta\sim F_{\phi_k}.$$9. For probabilistic matrix factorization with $m^{(K_o)}$0, ACDC is $m^{(K_o)}$1-robustly consistent with $m^{(K_o)}$2 and

$m^{(K_o)}$3

(Li et al., 25 Feb 2026).

The proof idea reported in the paper is that $m^{(K_o)}$4, whereas for $m^{(K_o)}$5, $m^{(K_o)}$6 in probability, so the minimum is achieved at $m^{(K_o)}$7 once parsimony resolves ties. The paper presents this as the formal mechanism through which ACDC balances sensitivity and robustness: likelihood-based criteria optimize predictive fit and may add spurious components under misspecification, while ACDC measures component-level mechanistic alignment via noise distributions and ignores discrepancies below tolerance $m^{(K_o)}$8 (Li et al., 25 Feb 2026).

5. Model classes and empirical evaluation

The paper develops ACDC for several latent variable model classes. In unsupervised probabilistic matrix factorization, the setting is $m^{(K_o)}$9 with components $K$00, contributions $K$01 when active, and

$K$02

Classical instances include nonnegative matrix factorization with $K$03 and factor analysis with $K$04. The same noise-based discrepancy extends to supervised PMF variants by conditioning on $K$05. Mixture models appear in unsupervised form through $K$06 and in covariate-dependent form through $K$07 (Li et al., 25 Feb 2026).

The empirical study covers both synthetic and real-data settings. In mixture simulations with multivariate skew-normal mixtures, varying $K$08, dimensions $K$09, $K$10 up to $K$11, and varying misspecification and cluster sizes, the baselines were Elbow, Silhouette, and Gap statistic, with parameter estimates by EM and ACDC using KL with $K$12-NN. The reported metrics were mean absolute error (MAE), mean 0–1 loss, and median signed deviation. In low dimension, ACDC was best overall with MAE $K$13, 0–1 loss $K$14, and median $K$15; competitors had MAE at least $K$16 and 0–1 loss at least $K$17. In high dimension, ACDC was approximately tied with Silhouette and much better than Elbow and Gap, with Wilcoxon tests showing significant MAE improvements for ACDC versus Elbow and Gap (Li et al., 25 Feb 2026).

In flow cytometry, twelve datasets were used. On training datasets 1–6, F-measure versus $K$18 curves aligned, and a single $K$19 maximized average F-measure and per-dataset F-measure. On test datasets 7–12, ACDC matched coarsened posterior average accuracy at far lower computational cost, reported as Python ACDC at $K$20h versus Julia coarsened posterior at $K$21h. The paper describes ACDC as comparable to nonparametric methods in this setting, while Elbow performed best on dataset 7, characterized there as a hard case (Li et al., 25 Feb 2026).

For single-cell RNA-seq using Tabula Muris, the experiments used 80 uniform subsets across 20 tissues with ground-truth cell types and Gaussian mixtures on PCA features under clear misspecification. Both KL and unbalanced Sinkhorn worked well; KL had slight overestimation bias with median $K$22, whereas Sinkhorn had slight underestimation with median $K$23. Automated $K$24 selection was close to manual selection, with manual MAE $K$25 versus automated $K$26. Compared with Elbow, Gap, Silhouette, Seurat, and SC3, ACDC achieved the best MAE ($K$27) and lowest 0–1 loss ($K$28), while ARI and AMI were comparable to Seurat; Seurat had median $K$29 bias and SC3 showed gross overestimation (Li et al., 25 Feb 2026).

The PMF applications include mutational signature discovery under Poisson NMF and hyperspectral unmixing under Gaussian factor analysis. In synthetic breast cancer datasets based on COSMIC v2 and PCAWG, under both well-specified and misspecified scenarios, BIC always picked $K$30 and Parallel Analysis consistently underestimated, often selecting $K$31. ACDC selected $K$32 or $K$33, where signatures and exposures had meaningful errors and the decomposition remained interpretable. In the hyperspectral unmixing example with ground truth of five end-members, ACDC selected $K$34, which the paper states maximizes sARI; BIC overfit with $K$35, and Parallel Analysis underfit with $K$36 (Li et al., 25 Feb 2026).

6. Terminological ambiguity, limitations, and related usages

The acronym ACDC is not unique in the arXiv literature. In “Approximate confidence distribution computing,” ACDC denotes Approximate Confidence Distribution Computing, a likelihood-free inferential method that targets a confidence distribution rather than a Bayesian posterior. The data provided for that paper explicitly states that the phrase “Accumulated Cutoff Discrepancy Criterion” does not appear there, and that the method is a frequentist alternative to ABC rather than a model selection criterion for latent mechanistic processes (Thornton et al., 2022).

A second, analytically distinct usage arises in work on Dirichlet forms and heat kernel bounds. In “Energy inequalities for cutoff functions and some applications,” the supplied terminology maps the query term to the pair $K$37 and $K$38, together with an accumulated inequality across chains of annuli. In that setting, the core inequality is

$K$39

and the accumulated form controls the energy of a cutoff $K$40 built from multiple annuli. This suggests a broader overlap in cutoff-based terminology, but it is a distinct analytic construction from the model selection criterion developed in 2026 (Andres et al., 2012).

Within the 2026 formulation, the stated limitations are substantive. ACDC requires a mechanistically meaningful discrepancy at the component level and a modest misspecification regime; if misspecification is very large, robust recovery of $K$41 is generally impossible. KL estimation in high dimension is statistically hard, and calibration of $K$42 is problem-dependent, with principled heuristics and domain-informed strategies but no universal rule. The paper identifies potential extensions to topic models, supervised factor analysis, functional clustering, hierarchical mixtures, time-series models, nonlinear generative models such as VAEs, and hierarchical structures with nested discrepancies and multi-stage ACDC (Li et al., 25 Feb 2026).

In this sense, Accumulated Cutoff Discrepancy Criterion designates a specific robust model selection procedure centered on component-level discrepancy truncation, not a generic family of cutoff-based methods. Its defining features are the robust loss $K$43, the parsimonious selection rule for $K$44, and the use of mechanism-aware discrepancies to recover a mechanistically meaningful number of latent processes under misspecification (Li et al., 25 Feb 2026).