Motivated by causal inference and physical computing systems, we consider the problem of learning mixture of two bounded-support exponential-family distributions with a shared interaction parameter matrix and mixture component-specific linear parameter vectors. The unknown sequence of component labels belongs to a known structured class. We propose an estimator based on a convex loss that avoids the log-partition function; for each fixed label assignment, estimation is a convex optimization problem. We show that the conditional population loss uniquely identifies the interaction matrix and the linear parameter attached to each sample. For a global empirical minimizer, we establish finite-sample bounds on the interaction error in Frobenius norm and the average squared linear-parameter error. The bounds depend on the logarithm of the number of admissible label assignments, and, for fixed dimension and model bounds, imply consistency when this complexity grows sublinearly with sample size. Additional occupancy and separation conditions yield component and label recovery.
Figures & tables
Fig. 1: Mixture learning setup: Each sample k∈[n] comprises the pair (Z(k),Y(k)) , where Z(k) is unobserved and Y(k) is observed. In our model the label sequence is fixed or random within a known set SZ(n) . Conditional on the full label sequence, the observations are independent, with the law of Y(k) determined by Z(k) and unknown exponential-family parameters. Marginalizing out Z(k) can induce additional dependencies among the observed variables {Ys(k)}s∈[t] , resulting in a denser graphical model and complicating learning from the observed distribution.
Fig. 2: Causal Inference : Independent units generate treatment-outcome panels from exponential-family: latent unit-level confounders u(k) perturb the linear parameter γ(u(k)) , while the quadratic interaction parameter Θ remains shared across units. Learning the parameters for conditional distribution fY∣U,A supports counterfactual evaluation under alternative actions a [ 13 ] .
Fig. 3: Quantum Computing : Repeated anneals generate independent readouts σ(k) from low-temperature noisy Gibbs distributions: Latent read-to-read noise s(k) perturbs the local fields γ(s(k)) , while the pairwise interaction parameter Θ encoding qubit interaction remains shared, as modeled in Fig. 4 . Jointly learning the field and coupling Hamiltonians for given inputs hin,J by learning the parameters of exponential family p(σ∣s) is key to device calibration in quantum annealing [ 14 ] .
Fig. 4: Structured two-component model. Sample k has linear parameter γ⋆(k)=γz⋆(k)⋆ and interaction matrix Θ⋆ . The estimator recovers the shared matrix and sample-specific parameters; recovery of both component vectors and labels requires the conditions in Corollary 1 .