Daily Summary for 2026-10-22

Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.

Ada: It's the twenty-second of October, twenty twenty-six, and this is the day's research.

Emmy: 3 new papers came out today.

Ada: I'm Ada.

Emmy: And I'm Emmy, here with Kolya.

Ada: We'll take the day in one pass, then pull out the papers we're staying with.

Emmy: Start us off with Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain.

The day's research: Ada: We need better ways to learn complex system dynamics when observations are partial or noisy.

Emmy: Understanding the true underlying process needs more than just looking at data distribution.

Kolya: How do we integrate different types of supervision effectively?

Ada: Specifically, we should combine soft law constraints with path-dependent supervision.

Emmy: This combination captures both the general shape and temporal structure of evolution.

Kolya: Does this move beyond simply fitting empirical distributions?

Ada: Yes, it allows us to penalize deviations while enforcing structural rules along paths.

Emmy: That unified framework leads to a discrete path-dependent McKean-Vlasov control problem.

Kolya: What method do we use to solve that control problem then?

Ada: We solve it using a forward-backward stochastic differential equation, or FBSDE.

Emmy: The path information influences the process at specific observation times within the FBSDE.

Ada: The core challenge is deriving a reduced FBSDE for quadratic cost and constant diffusion.

Emmy: That work also establishes local well-posedness for the Maximum Mean Discrepancy path-law formulation.

Kolya: And it develops a sample-based neural solver using causal representations.

Ada: This solver has been tested on synthetic transport, latent face transport, and human motion data.

Emmy: A key finding is that a two-state AMASS experiment captures both low-to-high and high-to-low transitions under path supervision.

Kolya: Marginal supervision alone could not distinguish these transitions, showing the framework learns complex temporal dynamics where standard distribution matching fails.

Ada: We need to discuss the application of Learning Generative Dynamics with Soft Path-Law Constraints: A McKean-Vlasov FBSDE Approach.

Emmy: That paper introduces a unified framework for learning stochastic dynamics from distributional observations and discretely observed trajectories using soft law constraints with path-dependent supervision.

Kolya: It combines soft law constraints with path-dependent supervision to learn dynamics from distributional observations and discrete trajectories.

Ada: Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain explores learning and optimal control under robust MDPs using the constant gain setting for average reward problems.

Emmy: That paper explores learning optimal control under robust MDPs specifically using the constant-gain setting for average reward problems.

Kolya: We also have Explicit Steady-State Approximations for Parallel Server Systems with Heterogeneous Servers which studies steady-state performance of parallel server systems under an immediate routing architecture.

Ada: That paper analyzes the weighted workload task allocation policy to study steady-state performance in parallel server systems.

Emmy: The papers we review today are Learning Generative Dynamics with Soft Path-Law Constraints: A McKean-Vlasov FBSDE Approach, which introduces a unified framework for learning stochastic dynamics from distributional observations and discretely observed trajectories by combining soft law constraints with path-dependent supervision. Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain explores learning and optimal control under robust Markov decision processes using the constant gain setting for average reward problems.

Kolya: Learning Generative Dynamics with Soft Path-Law Constraints: A McKean-Vlasov FBSDE Approach is worth hearing because it introduces a unified framework for learning stochastic dynamics from distributional observations and discretely observed trajectories by combining soft law constraints with path-dependent supervision. Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain is worth hearing because it explores learning and optimal control under robust Markov decision processes using the constant gain setting for average reward problems.

Ada: That covers the research today.

Emmy: These papers are Learning Generative Dynamics with Soft Path-Law Constraints: A McKean-Vlasov FBSDE Approach and Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain. We introduce a unified framework for learning stochastic dynamics from distributional observations and discretely observed trajectories by combining soft law constraints with path-dependent supervision, and this paper explores learning and optimal control under robust Markov decision processes using the constant gain setting for average reward problems.

Kolya: That concludes today's review.

Ada: We introduce Learning Generative Dynamics with Soft Path-Law Constraints: A McKean-Vlasov FBSDE Approach because it introduces a unified framework for learning stochastic dynamics from distributional observations and discretely observed trajectories by combining soft law constraints with path-dependent supervision. Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain is worth hearing because it explores learning and optimal control under robust Markov decision processes using the constant gain setting for average reward problems.

Emmy: This is our final review.

Kolya: We introduce Learning Generative Dynamics with Soft Path-Law Constraints: A McKean-Vlasov FBSDE Approach because it introduces a unified framework for learning stochastic dynamics from distributional observations and discretely observed trajectories by combining soft law constraints with path-dependent supervision. Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain is worth hearing because it explores learning and optimal control under robust Markov decision processes using the constant gain setting for average reward problems.

Ada: This is the end.

Emmy: These are Learning Generative Dynamics with Soft Path-Law Constraints: A McKean-Vlasov FBSDE Approach and Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain. We introduce a unified framework for learning stochastic dynamics from distributional observations and discretely observed trajectories by combining soft law constraints with path-dependent supervision, and this paper explores learning and optimal control under robust Markov decision processes using the constant gain setting for average reward problems.

Kolya: The episode is over.

Ada: This is the end.

Lucky paper: 2509.14203: Ada: Okay, we're diving into Bellman Optimality of Average-Reward Robust Markov Decision Processes with a Constant Gain. So, the main focus here is on moving beyond finite-horizon models when dealing with average reward problems and uncertainty from an adversary.

Emmy: The paper looks at the constant-gain setting for these robust MDPs, which is a specific way of analyzing long-run average reward problems under uncertainty about the controller's information.

Kolya: I'm interested in how they define this robust Bellman equation and what it means for the optimal average reward alpha.

Ada: They define a solution as one where u*(s) is the supremum over all possible strategies phi and all possible adversary paths ps of the expected reward minus alpha star plus the value of u* at time one.

Emmy: And Theorem one is pretty strong because it says if you find a solution (u*, alpha*) to that equation, then alpha* must equal the optimal average reward alpha(mu,, KH) regardless of the specific parameters used in the definition.

Kolya: That's interesting because it links the solution directly to the true average reward measure without needing extra assumptions on communication structures initially.

Ada: The paper then develops sufficient conditions for existence, and Theorem five states that if the controller is weakly communicating and Q and P S are compact for every state s in set S, then a solution to that equation exists.

Emmy: Weak communication seems like a necessary condition here because it dictates how much information the controller has about what the adversary might do next.

Kolya: What exactly does weak communication mean in this context, and how does that relate to the existence of these solutions?

Ada: Definition three defines weak communication as meaning that for every stationary controller policy, the MDP faced by the adversary is weakly communicating. It’s about ensuring some level of control influence remains even when facing an adversary.

Emmy: This moves us into the area where we see results about asymmetric information, where things get more complicated than simple optimal control might suggest.

Kolya: I noticed Theorem eight discusses cases where the adversary is weakly communicating but stationary optimality might not hold for the controller in a specific instance if alpha* and alpha' don't coincide.

Ada: That result implies that even under certain structural assumptions like compactness, we can't always assume a stationary policy will be optimal when there are information asymmetries between the controller and the adversary.

Emmy: It suggests that when you have this kind of robustness, you might need something more dynamic than a fixed policy to achieve the best long-run average reward.

Kolya: And then they show that if solutions exist for both equations, but their gains alpha* and alpha' are different, then the gain alpha' corresponds exactly to the inf-sup control value.

Ada: That means that in those asymmetric cases, even though a stationary policy might not be optimal for the controller, there is still a concrete value—the inf-sup control value—that you can calculate using that other solution.

Emmy: So the paper provides a way to compute this robust optimal control value even when traditional stationary optimality breaks down due to information gaps.

Kolya: This work really expands the dynamic programming theory for average-reward robust MDPs by establishing these conditions for solutions and characterizing the control value under those conditions.

Ada: It’s a solid foundation for doing robust decision making in operational environments where you can't perfectly anticipate everything from an adversary.