Theory Digest — Apr 23, 2026

1. Motivation (2–3 sentences): Existing weighted network community detection methods rely on Poisson models that fail for overdispersed data and make MCMC difficult with covariates. Real networks often have missing edges that could be predicted using external information like pairwise covariates.

2. Mathematical setup: Let A be an n × n adjacency matrix where Aij ∈ Z≥0 represents weighted edges, with Aij = Aji and Aii = 0. Each node i has latent community label zi ∈ {1,…,K}. The ZINB-SBM model is:

   \[K - 1 \sim q_K\] \[z_i | \pi, K \stackrel{i.i.d.}{\sim} \text{Categorical}(\pi)\] \[\pi | \gamma, K \sim \text{Dirichlet}_K(\gamma/K, \ldots, \gamma/K)\] \[A_{ij} | z_i = l, z_j = m, p_{lm}, \psi_{lm}, r_{lm} \stackrel{\text{ind}}{\sim} \text{ZINB}(p_{lm}, \psi_{lm}, r_{lm})\]

   Assumptions:

   1. Network is undirected with no self-loops
   2. Community assignments are exchangeable
   3. Within-block edge weights follow ZINB distribution
   4. Zero-inflation accounts for missing interactions

3. Toy example: When n = 4, K = 2, with nodes 1,2 in community 1 and nodes 3,4 in community 2, edge A12 follows ZINB(p11, ψ11, r11) while A13 follows ZINB(p12, ψ12, r12).

4. Formal objective: Estimate community labels z and block parameters to maximize:

   \[p(z, \theta | A) \propto p(A | z, \theta) p(z | \pi) p(\theta)\]

The ZINB-SBM uses zero-inflated negative binomial distributions with latent variable representation:

\[A_{ij} = w_{ij}(1 - x_{ij})\] \[w_{ij} | z_i = l, z_j = m, \psi_{lm} \sim \text{NB}(\psi_{lm}, r_{lm})\] \[x_{ij} | z_i = l, z_j = m, p_{lm} \sim \text{Bernoulli}(p_{lm})\]

For CZINB-SBM with covariates yij ∈ Rq, the method uses logistic regression:

\[\psi_{ij} = \frac{\exp(y_{ij}^T \beta_{z_i z_j, 1})}{1 + \exp(y_{ij}^T \beta_{z_i z_j, 1})}\] \[p_{ij} = \frac{\exp(y_{ij}^T \beta_{z_i z_j, 2})}{1 + \exp(y_{ij}^T \beta_{z_i z_j, 2})}\]

The algorithm uses:

1. Extended telescoping sampler for DMFM inference
2. Pólya-Gamma data augmentation for logistic regression

3. Partially collapsed Gibbs sampling with block parameters marginalized

   Toy example application: For the 4-node case, if y13 = (genetic_distance, geographic_distance)T, then ψ13 depends on β12,1 coefficients specific to the (1,2) block pair, allowing community-specific covariate effects.

Step 1 — Prior work:

• Mariadassou et al. (2010): Bayesian SBM with Poisson regression for covariates, used variational inference
• Lu (2023): ZINB-SBM without covariates, fixed K via information criteria
• Lu et al. (2025): Zero-inflated Poisson SBM with node attributes in prior

Step 2 — Delta: This paper adds (1) block-specific covariate effects via logistic regression in ZINB likelihood, (2) Pólya-Gamma augmentation enabling Gibbs sampling for regression coefficients, (3) DMFM framework for automatic K selection.

Step 3 — Theory-specific assessment:

• Main contribution is methodological rather than theorem-proving
• Pólya-Gamma technique is routine application of Polson et al. (2013)
• No theoretical analysis of convergence rates, consistency, or optimality
• DMFM application follows standard framework from Frühwirth-Schnatter et al. (2021)

The key technical insight is enabling efficient MCMC for ZINB regression via data augmentation, whereas Poisson regression requires variational methods.

Verdict: INCREMENTAL — solid methodological combination of existing techniques without theoretical novelty.

The main technical approach uses Pólya-Gamma data augmentation. For the negative binomial component:

\[\frac{\exp(r\eta_{ij,1})}{(1 + \exp(\eta_{ij,1}))^{w_{ij}+r}} = 2^{-(w_{ij}+r)} \exp(\kappa_{ij,1}\eta_{ij,1}) \int_0^{\infty} \exp\left(-\omega_{ij,1}\eta_{ij,1}^2/2\right) p(\omega_{ij,1} | w_{ij} + r, 0) d\omega_{ij,1}\]

where κij,1 = (r - wij)/2 and ωij,1 ~ PG(wij + r, ηij,1).

For the zero-inflation component:

\[\frac{\exp(\eta_{ij,2})^{x_{ij}}}{1 + \exp(\eta_{ij,2})} = 2^{-1} \exp(\kappa_{ij,2}\eta_{ij,2}) \int_0^{\infty} \exp\left(-\omega_{ij,2}\eta_{ij,2}^2/2\right) p(\omega_{ij,2} | 1, 0) d\omega_{ij,2}\]

where κij,2 = xij - 1/2 and ωij,2 ~ PG(1, ηij,2).

The key steps are:

1. Introduce Pólya-Gamma variables to linearize logistic terms
2. Marginalize block parameters for label updates
3. Use conjugacy to sample regression coefficients from normal distributions

4. Apply extended telescoping sampler for DMFM inference

   The technical insight is that Pólya-Gamma augmentation transforms the non-conjugate logistic regression into a conjugate linear Gaussian model conditional on the latent variables.

Simulation study: Compared ZINB-SBM vs ZIP-SBM on synthetic networks with n=150, K=3. Two scenarios:

1. highly overdispersed ZINB data

2. ZIP data. Evaluated using variation of information distance, number of communities recovered.

   Real data: fungusTreeNetwork (51 tree species, pairwise interactions with 3 covariates: genetic, taxonomic, geographic distance). Compared CZINB-SBM vs CP-SBM (Poisson regression baseline). Missing link prediction: 20% edges randomly masked, evaluated via AUC and RMSE over 50 replications.

   Key results: ZINB-SBM perfectly recovered communities in both scenarios, while ZIP-SBM overestimated K in overdispersed case (K̂=8.32 vs true K=3). CZINB-SBM achieved AUC=0.963 vs CP-SBM’s 0.768, and RMSE=2.083 vs 4.525 for missing link prediction.

Limitations:

1. Global overdispersion parameter r rather than block-specific - TECHNICAL (avoids Pólya-Gamma density evaluation but reduces flexibility)
2. Undirected networks only - NATURAL (standard assumption in SBM literature)
3. No theoretical analysis of convergence or consistency - RESTRICTIVE (limits theoretical understanding)
4. MCMC scalability unclear for large networks - RESTRICTIVE (could limit practical applicability)

5. DMFM advantages over standard MFM not rigorously demonstrated - TECHNICAL (empirical claim needs more support)

   Open problems:

6. Develop variational algorithms for scaling to large networks while preserving uncertainty quantification
7. Extend framework to directed networks and incorporate temporal dynamics

Motivation: Masked diffusion large language models (dLLMs) are a promising alternative to autoregressive models, but fine-tuning them with reinforcement learning faces a fundamental challenge: objectives require sequence-level marginal likelihoods that are intractable due to exponentially many unmasking trajectories.

Mathematical setup: Let $V$ be a vocabulary and $m$ the mask token. The state space is $\mathcal{X} = (V \cup {m})^L$ for length-$L$ sequences. A continuous-time Markov chain (CTMC) transports from fully masked prior $\rho_0$ to data distribution $\rho_1$ via rate matrices:

\[R_t(x, x[x_i \leftarrow v]) = \lambda(t) \cdot P[x_1^i = v \mid x_t = x]\]

where $\lambda(t) = \dot{\alpha}(t)/(1-\alpha(t))$ is the hazard rate. The unmasking posterior is:

\[\pi^*(v \mid x_t, i) := P[x_1^i = v \mid x_t]\]

Assumptions:

1. Access to pretrained MDM with unmasking posterior $\pi_0$
2. Reward function $r : V^L \to \mathbb{R}$

3. Tilted target distribution exists: $\rho_{1,a}(x) \propto \rho_1(x) e^{ar(x)}$

   Toy example: When $L=2$, $V={0,1}$, starting from $(m,m)$, the MDM gradually unmasks positions. With reward $r(x_1,x_2) = x_1 + x_2$, tilting favors sequences with more 1’s.

   Formal objective: Learn unmasking posterior $\pi_\theta$ that induces terminal distribution:

   \[\rho_{1,a+h}(x) \propto \rho_{1,a}(x) e^{hr(x)}\]

Method: Discrete Tilt Matching (DTM) recasts fine-tuning as incremental posterior matching under reward tilting, avoiding likelihood evaluation.

Key steps:

1. Esscher transform: Express tilted posterior via conditional reweighting:

   \[\pi_{a+h}(v \mid x_t, i) = \frac{E_a[e^{hr(x_1)} \mathbf{1}\{x_1^i = v\} \mid x_t]}{E_a[e^{hr(x_1)} \mid x_t]}\]

2. Tractable characterization: The tilted posterior satisfies:

   \[E_a[e^{hr(x_1)} \pi_{a+h}(v \mid x, i) \mid x_t] = E_a[e^{hr(x_1)} \mathbf{1}\{x_1^i = v\} \mid x_t]\]

3. Control variate: Replace one-hot target with lower-variance:

   \[T_c(v \mid x_t, i, x_1) := c\pi_a(v \mid x_t, i) + \frac{(e^{hr(x_1)} - c)\mathbf{1}\{v = x_1^i\}}{e^{hr(x_1)}}\]

4. DTM objective:

   \[L^{c\text{-DTM}}_{a \to a+h}(\theta) = \int_0^1 \lambda(t) E_a\left[\sum_{i: x_t^i = m} e^{hr(x_1)} \cdot \text{CE}(T_c(\cdot \mid x_t, i, x_1) \| \pi_\theta(\cdot \mid x_t, i))\right] dt\]

   Toy example application: For $L=2$ toy case, DTM would weight training samples $(x_1,x_2)$ by $e^{h(x_1+x_2)}$ and train local posteriors on partially masked states like $(0,m)$ or $(m,1)$.

Prior work:

1. “Tilt Matching for scalable sampling and fine-tuning” (Potaptchik et al., 2026): Introduced incremental reward tilting for continuous diffusion models
2. “d1: Scaling reasoning in diffusion large language models via reinforcement learning” (Zhao et al., 2025): Applied RL to dLLMs using biased likelihood surrogates

3. “SPG: Sandwiched policy gradient for masked diffusion language models” (Wang et al., 2025): Developed bound-sandwiching strategies for dLLM fine-tuning

   Delta: This paper extends tilt matching from continuous to discrete space, deriving a likelihood-free cross-entropy objective with explicit minimizer guarantees for masked diffusion models.

   Theory-specific assessment:

   • Main theorem: The derivation of DTM with explicit minimizer is technically sound but follows predictably from Esscher transform principles
   • Proof technique: Routine application of conditional change-of-measure identities; no genuinely new mathematical insight
   • Bound tightness: KL control guarantee (Proposition 3.4) is standard; no lower bounds provided or discussed

   The technical contribution is competent but incremental - adapting known continuous techniques to discrete setting with expected modifications for cross-entropy losses and masking structure.

   Verdict: INCREMENTAL — Solid but predictable extension of existing tilt matching to discrete masked diffusion models.

Main proof strategy:

1. Esscher conditional reweighting: Core identity relating tilted and base measures:

   \[\frac{dP_{a+h}}{dP_a} = \frac{e^{hr(x_1)}}{E_a[e^{hr(x_1)}]}\]

2. Key conditional identity: For tilted posterior characterization:

   \[E_{a+h}[G \mid x_t] = \frac{E_a[G e^{hr(x_1)} \mid x_t]}{E_a[e^{hr(x_1)} \mid x_t]}\]

3. Weighted cross-entropy minimization: The DTM objective minimizer is found by pointwise optimization. For fixed $(t,x_t,i)$:

   \[L_{t,i}(\pi; x_t) = E_a[e^{hr(x_1)} \text{CE}(T_c(\cdot \mid x_t,i,x_1) \| \pi(\cdot \mid x_t,i)) \mid x_t]\]

   Using unbiasedness condition $E_a[e^{hr(x_1)} T_c(v \mid x_t,i,x_1) \mid x_t] = E_a[e^{hr(x_1)} \mathbf{1}{x_1^i = v} \mid x_t]$:

   \[L_{t,i}(\pi; x_t) = E_a[e^{hr(x_1)} \mid x_t] \cdot \text{CE}(\pi_{a+h}(\cdot \mid x_t,i) \| \pi(\cdot \mid x_t,i))\]

4. KL bound derivation: Uses standard CTMC path-space KL formula:

   \[\text{KL}(P \| \tilde{P}) = \int_0^1 \lambda(t) E_P\left[\sum_{i: X_t^i = m} \text{KL}(\pi(\cdot \mid X_t,i) \| \tilde{\pi}(\cdot \mid X_t,i))\right] dt\]

   The proofs are routine applications of classical exponential tilting theory with no novel technical insights.

Datasets:

• Synthetic maze planning (41×41 grid)
• GSM8K mathematical reasoning (grade school math)
• MATH500 (competition mathematics subset)
• Countdown (arithmetic planning game)
• Sudoku (constraint satisfaction)

Baselines: LLaDA-8B-Instruct base model, LLaDA-1.5, d1, wd1, UniGRPO, SPG (all recent RL methods for dLLM fine-tuning)

Key numbers:

• Sudoku: 99.2% accuracy (vs 94.0% best baseline, 27.7% base)
• Countdown: 81.6% accuracy (vs 70.7% best baseline, 16.8% base)
• GSM8K: 81.6% accuracy (competitive with baselines, 77.2% base)
• MATH500: 36.0% accuracy (below 40.0% best baseline, 32.4% base)

Ablations: Annealing step size $h \in {1.5,3,6,12}$, control variate $c \in {0,1}$, generation lengths 256 vs 512. Wallclock comparison shows DTM converges faster than SPG on Sudoku.

Limitations:

1. TECHNICAL: Requires reward evaluations on full sequences during training, limiting scalability to very long sequences
2. RESTRICTIVE: Performance gaps on complex mathematical reasoning (MATH500, GSM8K) suggest state-level training may not capture long-horizon coherent reasoning well
3. TECHNICAL: Annealing schedule and step size $h$ require task-specific tuning; no principled selection method provided

4. NATURAL: Limited to additive reward structures $\rho_{1,a}(x) \propto \rho_1(x) e^{ar(x)}$; excludes more complex reward dependencies

   Open problems:

5. Develop principled methods for automatic annealing schedule selection that avoid mode collapse while ensuring convergence
6. Extend DTM to handle non-exponential reward tilting and more complex preference structures beyond simple scalar rewards

Motivation: Treatment effect distributions (TED) reveal who benefits or is harmed by interventions, going beyond average treatment effects. However, TED is not identified without assumptions on the joint distribution of potential outcomes. Existing approaches either impose strong rank preservation or derive wide partial identification bounds.

Mathematical setup: Let $(Y_0, Y_1)$ denote potential outcomes under control and treatment. The fundamental problem is that only one outcome per unit is observed, so the joint distribution $\pi \in \Pi(\mu, \nu)$ is unidentified, where $\mu = \mathcal{L}(Y_0)$ and $\nu = \mathcal{L}(Y_1)$ are marginal laws.

The paper defines rank stickiness via a single parameter $\rho \in (0,1]$. Let $\epsilon(\rho) := (1-\rho)/\rho$. The coupling maximizing $\rho$-rank preservation solves:

\[\max_{\pi \in \Pi(\mu,\nu)} R(\pi) - \epsilon(\rho) \text{KL}(\pi | \mu \otimes \nu)\]

where

\[R(\pi) := \frac{1}{2}\mathbb{E}_{(Y_0,Y_1) \sim \pi} \mathbb{E}_{(Y_0',Y_1') \sim \pi} \langle Y_0 - Y_0', Y_1 - Y_1' \rangle\]

Assumptions:

1. $\mathcal{Y} \subset \mathbb{R}^d$ is compact
2. $\mu, \nu$ are absolutely continuous with strictly positive densities

3. Rank stickiness parameter $\rho$ is known or specified as sensitivity parameter

   Toy example: When $d=1$, $\mu = \text{Uniform}[0,1]$, $\nu = \text{Uniform}[0,1]$, and $\rho = 1$, this reduces to the comonotonic coupling where $Y_1 = Y_0$ almost surely. For $\rho < 1$, rank violations occur with positive probability.

   Formal objective:

   \[\pi_\rho := \arg\max_{\pi \in \Pi(\mu,\nu)} R(\pi) - \epsilon(\rho) \text{KL}(\pi | \mu \otimes \nu)\]

The method constructs the Bregman-Sinkhorn copula via Sinkhorn dual potentials $\phi_\rho, \psi_\rho: \mathcal{Y} \to \mathbb{R}$ satisfying:

\[\phi_\rho(y_0) = \epsilon(\rho) \log \int_\mathcal{Y} \exp\left(\frac{\langle y_0, y_1 \rangle - \psi_\rho(y_1)}{\epsilon(\rho)}\right) d\nu(y_1)\] \[\psi_\rho(y_1) = \epsilon(\rho) \log \int_\mathcal{Y} \exp\left(\frac{\langle y_1, y_0 \rangle - \phi_\rho(y_0)}{\epsilon(\rho)}\right) d\mu(y_0)\]

The copula density is:

\[\frac{d\pi_\rho(y_0, y_1)}{d\mu(y_0) d\nu(y_1)} := \exp\left(\frac{\langle y_0, y_1 \rangle - \phi_\rho(y_0) - \psi_\rho(y_1)}{\epsilon(\rho)}\right)\]

Key property: The conditional distribution $Y_1

Y_0 = y_0$ is an exponential tilt with Bregman divergence exponent:

\[\mathbb{P}(dy_1 | Y_0 = y_0) = \frac{\exp\left(-\frac{B_{\psi_\rho}(y_1 | \nabla\psi_\rho^*(y_0))}{\epsilon(\rho)}\right) d\nu(y_1)}{\int_\mathcal{Y} \exp\left(-\frac{B_{\psi_\rho}(y_1' | \nabla\psi_\rho^*(y_0))}{\epsilon(\rho)}\right) d\nu(y_1')}\]

Application to toy example: For uniform marginals on $[0,1]$ with $\rho = 0.9$, the Sinkhorn system is solved iteratively. The resulting copula exhibits rank violations: $\mathbb{P}(Y_1 \neq u

Y_0 = u) > 0$ for any $u \in (0,1)$, unlike the comonotonic case.

Prior work:

1. Heckman et al. (1997): Introduced rank preservation assumption for point identification, derived Fréchet-Hoeffding bounds
2. Firpo (2007): Showed rank preservation yields quantile-by-quantile TED differences

3. Fan & Park (2010), Kim (2014): Tightened partial identification bounds under shape restrictions

   Delta: This paper introduces rank stickiness parameter $\rho$ allowing continuous relaxation from independence ($\rho \to 0$) to perfect rank preservation ($\rho = 1$). Key innovations:

   • Single scalar parameter achieves nonparametric point identification
   • Bregman-Sinkhorn copula has closed-form conditional moments
   • $\sqrt{n}$-rate convergence despite infinite-dimensional parameter space

   Theory-specific assessment:

   • Main theorem is somewhat predictable as entropic optimal transport regularization is well-studied
   • Proof technique combines Sinkhorn theory with contraction arguments - not fundamentally new but technically solid
   • Bounds are not compared to lower bounds (none appear to be known for this setting)
   • The exponential family structure of conditionals is elegant but follows from standard Sinkhorn theory

   The connection between rank correlation and entropic OT (Proposition 2.3) provides theoretical foundation, but this link has been explored in other contexts. The empirical process theory (Theorems 4.3-4.4) is technically competent but uses standard functional delta method and implicit function theorem techniques.

   Verdict: INCREMENTAL — Solid theoretical contribution that smoothly interpolates between known extremes, but the core techniques (Sinkhorn iteration, entropic OT) are established and the main innovation is problem formulation rather than proof technique.

The main proof strategies involve:

1. Uniqueness via asymmetric contraction (Theorem 4.2): Uses interpolation argument. For two solutions $(\phi_0, \psi_0)$ and $(\phi_1, \psi_1)$, define $h(y_1) = \psi_1(y_1) - \psi_0(y_1)$. The key inequality is:

   \[\|\phi_1 - \phi_0\|_\infty \leq \alpha_\psi \|\psi_1 - \psi_0\|_\infty\]

   where $\alpha_\psi = 1 - \delta_\psi c_\psi \nu(A_\psi) < 1$ for some $\delta_\psi \in (0,1)$ and open set $A_\psi$ where $

   h(y_1)

   < (1-\delta_\psi)M_\psi$. Combined with $|\psi_1 - \psi_0|_\infty \leq |\phi_1 - \phi_0|_\infty$, this yields contradiction unless $h$ is constant.

2. Central limit theorem (Theorem 4.4): For $\rho \in (0,1)$, applies implicit function theorem in Banach spaces. Key steps: - Establish Fréchet differentiability of Sinkhorn operator $\mathcal{S}$ with respect to marginals - Show linearized operator is bounded linear isomorphism on appropriate function space - Apply empirical process theory to marginal perturbations

3. Rank preservation case (Theorem 4.3): Uses functional delta method for composition $G_n^{-1} \circ F_n$. The key concentration inequality leverages Hadamard differentiability:

   \[\sqrt{n}(G_n^{-1} \circ F_n(y) - G^{-1} \circ F(y)) \Rightarrow \frac{\sqrt{2} B(F(y))}{g(G^{-1}(F(y)))}\]

4. Bregman divergence representation: The exponential tilt structure follows from convex conjugate properties:

   \[\langle y_0, y_1 \rangle - \psi_\rho(y_1) - \psi_\rho^*(y_0) = -B_{\psi_\rho}(y_1 | \nabla\psi_\rho^*(y_0))\]

   The proofs are technically sound but primarily assemble known results from Sinkhorn theory, optimal transport, and empirical process theory.

Numerical simulations on synthetic data with two DGPs:

1. Bregman-Sinkhorn copula with $\rho_1 = 0.99$

2. Gaussian copula with $\rho_2 = 0.90$

   Both exhibit rank violations. Key findings with $n = 2000$:

   • Estimated TED recovers bimodal structure (60% negative, 40% positive treatment effects) despite small positive ATE
   • Comonotonic coupling ($\rho = 1$) underestimates heterogeneity
   • Bregman-Sinkhorn variance estimator for ATE substantially tighter than Fréchet-Hoeffding and Neyman bounds
   • Method shows appropriate sensitivity to stickiness parameter $\rho$

   Purely theoretical contributions would benefit from empirical validation on real datasets from job training programs or similar policy interventions where treatment effect heterogeneity is substantively important.

Limitations:

1. RESTRICTIVE: Requires knowing or specifying rank stickiness parameter $\rho$ - paper treats this as sensitivity analysis but practical guidance for choosing $\rho$ is limited

2. TECHNICAL: Compact support assumption needed for uniqueness proof - likely removable with more work on tail behavior

3. NATURAL: Absolute continuity and positive density assumptions standard in this literature

4. RESTRICTIVE: Extension to observational studies requires unconfoundedness - strong assumption that rules out unmeasured confounding

5. TECHNICAL: Convergence rate analysis limited to bounded domains, may extend to unbounded support with moment conditions

   Open problems:

6. Data-driven selection of optimal $\rho$ value, perhaps via cross-validation on imputed joint likelihood or stability of estimated functionals
7. Formal consistency guarantees for the covariate adjustment algorithm using binary classification to estimate density ratios - current approach lacks theoretical backing