Daily Summary for 2026-10-23

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

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

Emmy: 8 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 Kinetic theory for Transformers and the lost-in-the-middle phenomenon.

The day's research: Ada: We looked at information movement in large language models today.

Emmy: How does that affect user service when memory is limited?

Kolya: Service-induced congestion in memory-limited serving was the focus.

Ada: That describes server running out of GPU memory during text creation.

Emmy: When that happens, items get removed and output rate drops.

Kolya: This structural instability sets the limits for high-rate output.

Ada: We studied kinetic theory for Transformers today.

Emmy: That links to the lost-in-the-middle phenomenon.

Kolya: How does the particle system explain self-attention dynamics?

Ada: It clarifies token retrieval profiles in prompts.

Emmy: Specifically, it shows a shape with primacy and recency.

Kolya: And what about that interior minimum when conditions are met?

Ada: That provides a solid explanation for how models handle input sequences.

Lucky paper: 2605.09213: Ada: So we're looking at kinetic theory for Transformers now and how that relates to the lost-in-the-middle phenomenon.

Emmy: It’s about viewing the attention dynamics as an interacting particle system, and then proving a mean-field limit result for that system.

Kolya: That model uses a one-dimensional torus T, which is structurally faithful to what they used in previous work.

Ada: And that framework shows how the retrieval observable satisfies a U-shape when the number of tokens N gets large enough.

Emmy: The paper shows that this U-shape has a unique interior minimum, which explains why we see primacy and recency effects in token profiles.

Kolya: The mechanism for this depends on the cross-correlation evolving according to a linear equation with a forcing term determined by A phi.

Ada: That forcing term injects the fluctuation carried by the source token at position sigma zero into the dynamics of that correlation.

Emmy: So, if we look at Figure three we see how particles drift toward the first token and then form a coherent packet near the end.

Kolya: The paper states that for iid uniformly distributed prompts, this soft accuracy σ0 is U-shaped with a unique global minimum in (zero one).

Ada: That confirms that the empirically observed lost-in-the-middle phenomenon has a rigorous explanation through this kinetic theory.

Emmy: It’s interesting how they use Glauber calculus adapted to that triangular causal structure to get those sharp bounds on covariances.

Kolya: The error estimates show the error in the soft accuracy is bounded by terms involving N to the power of minus delta and one p e C t⟨n⟩ C.

Ada: That gives us a concrete idea of how small an error we can expect when scaling up the model size N.

Lucky paper: 2509.14229: Ada: So we're looking at the Spacing Test for Fused Lasso paper now, which deals with detecting changepoints in one-dimensional signals and post-selection inference.

Emmy: It seems like this paper tackles a very specific geometric problem related to how uncertainty is measured after you select some data points.

Kolya: The key finding here is that the onedimensional fused lasso has a surprisingly simple geometry where each changepoint enters in a strictly one-sided fashion, and there are no leave events.

Ada: That means the conservative spacing test from Tibshirani et al., which we usually treat as an approximation, becomes exact for this setting.

Emmy: Exactly. The paper shows that the truncation region reduces to a single lower bound given by the next knot on the LARS path, which is pretty neat.

Kolya: And what about that exact selective p-value? It has a closed form identical to the simple spacing statistic used in the LARS or lasso setting, with no extra computation needed.

Ada: That's significant because it removes the need for tracking many "hit" and "leave" events along the regularization path.

Emmy: So, what about how they prove that this exactness holds? The mechanism behind this is what I'm curious about.

Kolya: It stems from the fact that changepoints enter sequentially without leaving under general position, and the selection event is one-sided.

Ada: So once you condition on having reached step k with entrant j and sk, the relevant headroom for eta top y is exactly from the previous knot down to the next knot.

Emmy: That means no other inequalities are active in one dimension, which simplifies things immensely when you try to apply this idea elsewhere.

Kolya: The paper formalizes this in Theorem two by stating that lambda plus k plus one equals lambda k plus one which shows the lower truncation endpoint is precisely the next knot.

Ada: That relationship between those knots is what establishes the exact selective pivot and gives us that closed form expression for T k.

Emmy: It's interesting how they connect this result to empirical validation, showing that their one-sided representation produces the same selective inference quantities as Hyun et al.'s generalized lasso hit/leave conditioning.

Kolya: They also confirmed that both methods yield identical selective power behavior under a single changepoint, proving the one-sided reduction isn't just an approximation but an exact representation.

Ada: So this work provides a transparent geometric explanation for post-selection inference in the fused lasso, giving us a closed form expression for the exact selective p-value.

Emmy: It bridges modern L1 regularization with traditional sequential segmentation methods, which is useful when you need something computationally simple for applied fields like genomics.

Kolya: And they suggest that exact post-selection inference for change-point problems can be achieved without needing high-dimensional polyhedral algebra, which is a big deal.

Lucky paper: 2603.15934: Ada: Let's talk about Fast Relax-and-Round Unit Commitment with Economic Horizons, that paper from PJM and the energy system stuff.

Emmy: So it deals with unit commitment, which is figuring out which power generators to use for a given time period.

Kolya: What's interesting is how they expand the formulation to include ramping constraints and long-horizon planning. That adds complexity to the math, right?

Ada: Exactly. They introduce decision variables like Pi and Pdelta,i, plus commitment status ui for discretionary generators i.

Emmy: And they show a fast algorithm specifically for committing hydro-generators before traditional units. That's a big part of the paper’s contribution.

Kolya: The paper claims this method can solve unit commitment problems with over twenty thousand generators in approximately ten seconds on commodity hardware. That runtime estimate is pretty impressive.

Ada: It also shows that increasing the planning horizon leads to sizable operational cost savings for the system operator.

Emmy: I see how they handle hydro units differently by decoupling them and using predicted demands to commit them based on remaining periods with highest demand.

Kolya: That specific approach runs in log-linear time, O(n(log n + T log T)), which is much better than what we usually see.

Ada: The simulation results show that the RRUC method produces numerically indistinguishable solutions from an exact algorithm, even though RRUC is twice as fast on the largest problem tested.

Emmy: They also found a strong positive correlation between marginal cost and demand in their analysis of demand-marginal cost pairs.

Kolya: That correlation is what motivates them to use hydro generators to flatten load or peak shave by smoothing out Dt = Dt - Pn j=one uj,tκj.

Ada: The conclusion points to Algorithm two being able to commit one thousand seven hundred ninety-two generators for an entire year in twenty seconds.

Emmy: They also mentioned future work involving incorporating transmission constraints and using GPUs for acceleration of the solver.

Kolya: It’s a scientific study on developing a fast and accurate unit commitment method that handles ramping constraints and long-horizon planning, which is important because it allows operators to make horizon-aware economic decisions for large systems with hydro units.

Lucky paper: 2510.07324: Ada: We’re moving into something pretty deep with this paper, Geodesics in the Deep Linear Network.

Emmy: It looks like it starts with a phenomenological model of training dynamics for neural networks, which is the DLN.

Kolya: The DLN consists of an upstairs parameter space of N by d matrices, and a downstairs observable space linked by this product map phi.

Ada: So, the core idea is tracking how these models train through this geometry.

Emmy: And they are deriving a general system of ordinary differential equations for geodesics in that DLN geometry.

Kolya: They found horizontal straight lines in an invariant balanced manifold that stay geodesics under Riemannian submersion.

Ada: That sounds like a very specific geometric property they are proving about the network dynamics.

Emmy: What is the relationship between this balanced manifold and how we define things upstairs and downstairs?

Kolya: In that balanced manifold M, all matrices W p have the same singular values for every p from one to N.

Ada: If they have the same singular values, what does that imply about the structure of those matrices?

Kolya: It means the right singular vectors of matrix W p+one align perfectly with the left singular vectors of matrix W p for all steps from one to N-one.

Emmy: That alignment sounds like a very strong constraint on how the layers interact.

Kolya: Exactly, and they use that structure to parameterize any matrix X using only, the singular values, and the matrices Q N down to Q zero.

Ada: So, this mapping from matrix space to parameter space is bijective in this balanced manifold.

Kolya: Right, and they define a metric g N on the downstairs space where the squared norm of a tangent vector Z is defined by Tr Z T A N,X-one X(Z).

Emmy: That looks like they are defining how distances are measured in that observable space.

Kolya: And because of Theorem one point one, the product map phi acts as a Riemannian submersion from the manifold M to this downstairs space with metric g N.

Ada: So, if you have a path in the upstairs parameter space that stays within that balanced manifold, it translates to a path of the same length downstairs.

Emmy: That’s interesting because it connects the dynamics of training directly to geometric properties.

Kolya: The main results involve ODEs for general geodesics and then explicit formulas when we have specific relationships between the singular vectors of our endpoints, A and B.

Ada: When they give us those explicit formulas, what is the condition they need to meet for that simple straight line in the upstairs space to translate into a geodesic?

Kolya: They need there to exist some matrix Q such that both endpoints are related by U =UQ and V =VQ.

Emmy: That condition involving Q seems like the key constraint for finding these explicit paths.

Kolya: And they provide a specific formula for the geodesic in the downstairs space between A and B, which involves terms like (one-t)U 1N + tU 1N Q T.

Ada: That formula looks incredibly complex, but it’s the explicit solution they are aiming for.

Emmy: Does this mean that if we can find those A and B that satisfy the condition with some Q, we know exactly what the path looks like?

Kolya: Yes, Theorem two point three gives them a specific formula for the geodesic between any two endpoints given that condition holds.

Ada: So, if we look at training dynamics, this paper suggests a certain "straight line" behavior exists within these network geometries.

Emmy: What about those open problems they mentioned? What are the next steps for this research direction?

Kolya: The main follow-up work involves solving the general case geodesic equations for arbitrary endpoints and generalizing these results to low rank matrices, say rank less than d.

Lucky paper: 2305.14632: Ada: So we're moving into Supermodular Rank today, which is paper number five of five on this day in mathematics.

Emmy: It sounds like it deals with decomposing functions on a lattice into supermodular parts to help with optimization problems.

Kolya: The core idea is defining the supermodular rank as the smallest number of terms needed to decompose a function into a sum of supermodular functions, where those summands can have different partial orders.

Ada: That sounds complicated because the partial order pi changes between each term in the decomposition.

Emmy: And it says that functions with at most rank r are characterized as being a union of Minkowski sums of at most r supermodular cones.

Kolya: Specifically, Theorem eight shows that the Minkowski sum of supermodular cones L pi(one) + + L pi(m) is a convex polyhedral cone whose facet inequalities are the same for all m cones.

Ada: That means there's a shared boundary definition across all those different partial orders involved.

Emmy: It also gives us some bounds on the rank itself, showing that for two variables, the maximum supermodular rank is two n + one when n is three or greater.

Kolya: And if we look at submodular functions inside the interior of-L(one one), their rank can be two n.

Ada: That gives us a concrete measure for how complex a function on n variables can structurally be.

Emmy: The real payoff seems to be in the optimization algorithms they propose, like R-SPLIT and R-SPLIT RATIO.

Kolya: They use the rank decomposition insight that "a function with elementary submodular rank r can be split into two r submodular pieces."

Ada: That splitting idea is what allows them to improve approximation guarantees for constrained set function maximization.

Emmy: For instance, Corollary sixty-four shows that if a function has elementary submodular rank r + one the approximation ratio for the R-SPLIT GREEDY is (R(alpha, gamma), (one - alpha e-) - O(epsilon)).

Kolya: That formula is interesting because it shows how trading off computational cost and accuracy depends directly on that rank r.

Ada: Experimentally, they tested this across four types: DETERMINANTAL, BAYESIAN, COLUMN, and RANDOM functions.

Emmy: And the results showed that the approximation ratio guarantee can increase quickly as r gets higher for DETERMINANTAL functions.

Kolya: They also found that the R-SPLIT GREEDY algorithm outperforms GREEDY on several datasets for small n, like n = twenty.

Ada: So, even with this decomposition, they still have open questions about figuring out those supermodular cones themselves.

Emmy: It sounds like the theoretical complexity of computing those decompositions is still a major hurdle for practical application.