--- title: 'ACDC: Accumulated Cutoff Discrepancy Criterion' url: https://www.emergentmind.com/topics/accumulated-cutoff-discrepancy-criterion-acdc type: topic --- # ACDC: Accumulated Cutoff Discrepancy Criterion 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 [2602.22062]. ## 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 [2602.22062]. 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 [2602.22062]. 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 \[ P_\star\in m^{(\infty)}:=\bigcup_{K\ge 1} m^{(K)}, \] the closest element to the truth in Kullback–Leibler divergence, typically with no finite \(K\), which leads to overestimating \(K_o\). Robust nonparametric methods are less sensitive to parametric misspecification but can underestimate \(K_o\) and lack generality. ACDC is defined to address this tension through a component-level, mechanism-aware discrepancy [2602.22062]. ## 2. Mechanistic framework and component-level discrepancy The general framework allows observations \(x_n\in\mathcal{X}\) to depend on covariates \(w_n\in\mathcal{W}\). Each observation is represented as a deterministic combination of component contributions, \[ x_n=g^{(K)}(y_{n1},\ldots,y_{nK}),\qquad g^{(K)}:\mathcal{Y}^{\otimes K}\to\mathcal{X}, \] with the “no-contribution” property \[ g^{(K)}(y_1,\ldots,y_{K-1},\emptyset)=g^{(K-1)}(y_1,\ldots,y_{K-1}). \] Each component contribution is \[ y_{nk}= \begin{cases} \emptyset, & \text{if } z_{nk}=\emptyset,\\ f(w_n,z_{nk},\phi_k,\varepsilon_{nk}), & \text{otherwise}, \end{cases} \] where \(z_n=(z_{n1},\ldots,z_{nK})\in\mathcal{Z}^{(K)}\), \(z_n\mid \eta,w_n\sim H^{(K)}_{\eta,w_n}\), and \(\varepsilon_{nk}\sim G\) are independent noise variables [2602.22062]. 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 \(\varepsilon_{nk}\) given observed data, for indices that are “used” \((y_{nk}\neq \emptyset)\). For model parameter \(\theta\) and data distribution \(P_o\), the inferred distribution of the \(k\)-th component noise is \[ \widetilde{G}_k^{(\theta)}=\int \mathcal{L}\big(\varepsilon_k\mid y_k\neq\emptyset,x,w,\theta\big)\,P_o(dx,dw), \] and the component-level discrepancy is \[ D^{(K)}(\theta,k,P_o)=\mathcal{D}\big(\widetilde{G}_k^{(\theta)}\|G\big). \] Here \(\mathcal{D}\) is a divergence or distance between probability measures, with examples including KL divergence, Wasserstein distance, and MMD [2602.22062]. In the plug-in implementation, one fits a point estimator \(\widehat{\theta}^{(K)}\), computes per-sample conditional noise laws \[ \widehat{G}_{nk}^{(K)}=\mathcal{L}\big(\varepsilon_{nk}\mid x_n,w_n,\widehat{\theta}^{(K)}\big), \] usage indicators \(u_{nk}^{(K)}=\mathbf{1}(y_{nk}^{(K)}\neq\emptyset)\), and counts \(N_k^{(K)}=\sum_{n=1}^N u_{nk}^{(K)}\), then aggregates these into an empirical component noise distribution \(\widehat{G}_k^{(K)}\) and evaluates \[ \widehat{D}^{(K,k)}=\mathcal{D}\big(\widehat{G}_k^{(K)}\|G\big) \] [2602.22062]. For mixture models, the paper also gives a component-level realized discrepancy via responsibilities \[ p(k\mid x,\theta)=\eta_k\,\frac{dF_{\phi_k}}{dP_\theta}(x), \qquad \widetilde{F}^{(\theta)}_{ok}= \frac{p(k\mid x,\theta)}{\int p(k\mid y,\theta)\,P_o(dy)}\,P_o, \] followed by \[ D^{(K)}(\theta,k,P_o)=\mathcal{D}\big(\widetilde{F}^{(\theta)}_{ok}\|F_{\phi_k}\big). \] The paper develops ACDC and its theory using the noise-based construction, while this mixture-specific formulation illustrates the same mechanistic principle [2602.22062]. ## 3. Criterion, computation, and cutoff calibration For a fixed \(K\), ACDC defines the robust loss as \[ \mathcal{R}^\rho(x_{1:N},K)=\sum_{k=1}^K \max\bigl(0,\widehat{D}^{(K,k)}-\rho\bigr), \] where \(\rho\ge 0\) is the cutoff level. The selected number of components is \[ \widehat{K}^\rho(x_{1:N})= \min\left\{\operatorname*{argmin}_{K}\mathcal{R}^\rho(x_{1:N},K)\right\}. \] The outer minimum resolves ties in favor of parsimony [2602.22062]. A notable detail is that ACDC sorts nothing explicitly. It accumulates component-wise discrepancies through a sum after truncation at \(\rho\). The truncation prevents misspecification within the acceptable tolerance from influencing the loss and thereby prevents overfitting by adding spurious components [2602.22062]. Operationally, the procedure runs over \(K=1,2,\ldots,K_{\max}\). For each \(K\), the model is fit to obtain \(\widehat{\theta}^{(K)}\); the conditional noise laws and usage indicators are computed; the empirical component noise distributions are formed; and the discrepancies \(\widehat{D}^{(K,k)}\) are evaluated. The criterion is then assembled as \(\mathcal{R}^\rho(x_{1:N},K)\), and \(\widehat{K}^\rho\) is selected as the smallest minimizer. The computational cost is dominated by model fitting for each \(K\) and by evaluation of the divergence \(\mathcal{D}\) on the empirical component noise distributions. Per-component effective sample size is \(N_k^{(K)}\) [2602.22062]. The paper gives several strategies for choosing \(\rho\). 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 \(\rho\), then transfers the calibrated \(\rho\) to new datasets of similar type. A generally applicable unlabeled approach plots \(\mathcal{R}^\rho(x_{1:N},K)+\lambda K\) as a function of \(\rho\) for each \(K\), with small \(\lambda>0\) used for visual separation, and chooses the smallest \(\rho\) at the first “stability interval” where one \(K\) curve remains minimal across a substantial range of \(\rho\). An automated version defines a minimum stability width \(\Delta_{\min}\), tracks intervals \([\rho_{\mathrm{start}},\rho_{\mathrm{end}}]\) over which a single \(K\) is best, and selects the first interval whose width is at least \(\Delta_{\min}\) [2602.22062]. 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 \(\varepsilon\). In high-dimensional settings, KL estimation via \(k\)-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 \(k\)-NN estimator, a bias-corrected variant, and an adaptive-\(k\) version with \(k_N\to\infty\) and \(k_N/N\to 0\) [2602.22062]. ## 4. Robust model selection consistency The theoretical target is a robust model selection consistency property defined through a distribution-level discrepancy \(d\) and the component-level discrepancy \(\mathcal{D}\). For each \(K\), \[ \Theta_\star^{(K)}(P_o):= \operatorname*{argmin}_{\theta\in\Theta^{(K)}} d\big(P_o,P_\theta\big), \] and the worst-case component-wise discrepancy at the true \(K_o\) is \[ \rho(P_o,K_o):= \sup_{\theta\in\Theta_\star^{(K_o)}(P_o)} \max_{k\in\{1,\ldots,K_o\}} D^{(K_o)}(\theta,k,P_o). \] A model selection procedure \(\widehat K(x_{1:N},\rho)\) is \(\kappa\)-robustly consistent for \(\Theta_\star\) and \((d,\mathcal{D})\) if, whenever the mismatch or “gap condition” \[ \inf_{\theta\in\Theta_\star^{(K)}(P_o)} d\big(P_o,P_\theta\big)\ge \kappa\big(\rho(P_o,K_o),K\big) \qquad\text{for all }K