Research
My research lives at the intersection of causal inference, probabilistic graphical models, and continuous optimization. The unifying question: when can causal structure be recovered from data — efficiently, and with guarantees — and when can it not?
§ 01
/ 01
Characterizing when causal structure can be uniquely recovered from observational and interventional data, and turning those identifiability conditions into efficient discovery algorithms for identifiable directed acyclic graphs.
/ 02
Continuous-optimization formulations of structure learning that replace combinatorial search over DAGs with differentiable acyclicity constraints, enabling GPU-accelerated estimation on high-dimensional, mixed-distribution data.
/ 03
Variance bounds and lowest-variance estimators for the non- and partially-identifiable regimes — quantifying what can still be learned about causal structure when it is not uniquely determined.
/ 04
Intervention-selection policies that maximize identifiability gain per experiment, reducing the number and cost of interventions needed to recover causal structure.
§ 02
Peer-reviewed work and preprints in causal inference and wireless communications. For the most up-to-date list see my Google Scholar profile.
arXiv preprint
Submitted to JMLR · 2026
S. Mishra, U. Mitra
Proves that every edge between an ordinal node and a one-parameter exponential-family node in a linear parametric causal model is identifiable from the joint distribution alone, with converses showing the three-category and three-support-point conditions are necessary; submitted to the Journal of Machine Learning Research.
arXiv: 2609.17942Asilomar 2026
Accepted for oral presentation
S. Mishra, U. Mitra
Studies causal discovery for equal-variance linear-Gaussian DAGs, tuning ridge-regression regularization by Stein’s unbiased risk estimate (SURE) to recover structure from observational data.
arXiv: 2608.17132Asilomar 2026
Accepted for oral presentation
S. Mishra, Y. Wang, C. K. Johnson, U. Mitra
Motivated by epidemiological data that mixes ordinal, count and continuous measurements, introduces a structured statistical model, proves identifiability of ordinal–exponential-family edge directions, and gives an exhaustive search and a masked DAGMA procedure for recovering mixed DAGs.
arXiv: 2609.20676ICASSP 2026
Barcelona, pp. 6196–6200
C. Peng, S. Mishra, U. Mitra
Proposes a learning-based framework for selecting soft interventions that improves causal-discovery efficiency and reduces experimental cost.
doi: 10.1109/ICASSP55912.2026.11460954IEEE Transactions on Green Communications and Networking
Vol. 10, pp. 1433–1445, 2026
S. Mishra, S. P. Dash, G. C. Alexandropoulos
Investigates one- and two-sided ASK modulations in noncoherent SISO systems assisted by an RIS, proposing novel energy- and sign-based receiver structures.
doi: 10.1109/TGCN.2025.3633182IEEE Wireless Communications Letters
Vol. 15, pp. 300–304, 2026
S. Mishra, S. P. Dash
Considers RIS-aided wireless communication with one-sided ASK and an optimal noncoherent maximum-likelihood detection rule.
doi: 10.1109/LWC.2025.3624154§ 03
Electronic System Design Lab, IIT Bhubaneswar
Collaborate
I’m always glad to chat about causal discovery, identifiability, or scalable structure-learning methods.
sambitmi@usc.edu