Daily Summary for 2026-10-21

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

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

Emmy: 6 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 Testing for genetic interactions in complex disease with distance correlation.

The day's research: Ada: We need to distinguish causal structures with observed data alone.

Emmy: Not every statistical relationship implies one cause.

Kolya: We explore the selected-marginalized directed graph concept from AKMPC.

Ada: Two graphs are indistinguishable if they share the same smDG.

Emmy: This holds even with arbitrary interventions on observed variables.

Kolya: It offers a new criterion for conditional independence relative to passive observations.

Ada: Moving down, we test genetic interactions in complex diseases using distance correlation.

Emmy: Distance correlation finds epistasis from case-control data of SNPs.

Kolya: This measure captures general statistical independence beyond just linear kinds.

Ada: Simulations suggest it has comparable or better power for schizophrenia genetics data.

Ada: We have work on understanding genetic interactions in complex disease with distance correlation papers/2012.05285.

Emmy: That paper uses distance correlation to detect epistasis from case-control data of single-nucleotide polymorphisms.

Kolya: It highlights the derivation of the explicit asymptotic null distribution for computational speed.

Ada: Does that distribution derivation offer any constraints on sample size?

Emmy: The explicit form allows us to speed up computations significantly when analyzing large datasets.

Kolya: What about the ergodic central limit theorems for subgeometric Markov chains papers/2512.18255?

Ada: That work establishes necessary conditions based on discrete-time L-drift conditions to verify central limit theorems.

Emmy: Those conditions provide insights into convergence rates and tail behavior for ergodic averages of Markov chains on general state spaces.

Kolya: Are the L-drift conditions restrictive enough for practical application?

Ada: The work focuses on necessary conditions, which sets the framework for what is verifiable.

Lucky paper: 2012.05285: Ada: So, we are looking at Testing for genetic interactions in complex disease with distance correlation.

Emmy: It uses distance correlation as a new tool to find epistasis from case-control data of SNPs to understand things like schizophrenia susceptibility.

Kolya: The paper defines distance correlation as the quotient of variance and the product of standard deviations, which is improved over Pearson’s correlation squared because it stays between zero and one.

Ada: So, if we have two variables, X and Y, and their distance correlation is zero, what does that actually mean statistically?

Kolya: It means they are independent in a specific metric space; distance correlation being zero is equivalent to X and Y being statistically independent.

Emmy: But the authors are testing the independence of the joint sample Zi and Zj, which is a bit different from just looking at two variables.

Ada: They set up a null hypothesis H0ij where dcov(Zi, Zj) equals zero, and H1ij where it's not equal to zero.

Kolya: To approximate the null distribution of this test statistic dcov(

Zi, Zj), they use finite marginal spaces. [Emmy: That means they only look at a few coefficients of the quadratic form that give the asymptotic null distribution being non-null.

Ada: And how do they decide which specific permutation tests to use to get that null distribution approximation?

Kolya: The most logical approach is resorting to permutation tests because of this structure.

Emmy: In their simulations, they ran R equals one thousand replicates for all cases in the testing for genetic interactions in complex disease with distance correlation study.

Ada: And what did those simulations show about the power of their test?

Kolya: The power is very satisfactory, and it increases as you get further away from that null hypothesis.

Emmy: Experiment II seems to indicate they are detecting a genuine signal at the genetically-regulated gene expression level in their schizophrenia database interrogation.

Ada: They also claim to outperform the BOOST method, which they call the gold standard, in their tested scenarios.

Kolya: The authors state that distance correlation characterizes independence in certain metric spaces, which they then apply to the case of three-point marginal spaces.

Lucky paper: 2603.17080: Ada: Let's look at the paper on A quadratic Grassmann manifold optimization problem arising from quantum embedding methods.

Emmy: So, we're dealing with minimizing a quadratic function J(P) over the Grassmann manifold Gr(m, R M).

Kolya: The core of this is minimizing J(P) equals Tr(BP inverse over two Tr(AP AP), where A and B are symmetric matrices in R M by M.

Ada: That means we're looking for the best P that satisfies P squared equals P and has a trace equal to m.

Emmy: The authors define the tangent space at a point P as TPM equals

X, P: , where X is an antisymmetric matrix.

Kolya: And they compute the Riemannian gradient of this function on M using gradMJ(P) equals

[G(P), P: , P], where G(P) equals B minus A times A inverse.

Ada: That looks like a lot of matrix algebra just to find the direction we should move on the manifold.

Emmy: Then they decompose the Riemannian Hessian HessMJ(P) into two parts, omega(P) and K(P).

Kolya: Omega is defined as TPM times X seven times

[G(P), X: , P], and K is TPM times X seven times minus

[AXA, P: , P].

Ada: It's interesting how they separate the curvature effects into those two terms.

Emmy: The big part of the paper then moves to convexification, introducing an optimization problem Je(D) over the convex hull K.

Kolya: They define Je(D) equals Tr(CD) plus one over four

A, D: squared, where C equals B minus one over two A squared.

Ada: So they're trading the non-convex manifold for a convex problem defined over CH(M).

Emmy: The paper shows that J and Je have the same critical points on M, which is important because it means they share some solutions.

Kolya: Theorem four states that if H star satisfies the gap condition mu m less than mu m plus one, then D star is a global minimizer of the original non-convex problem.

Ada: So the gap condition on those eigenvalues of B dictates whether we can trust the solution found in Je.

Emmy: They also looked at special cases where A and B commute, which gives an explicit solution based on eigenmodes of matrix C.

Kolya: Perturbation analysis is done too, examining what happens when the commutator

A, B: is small and leading to real-analytic solutions for P epsilon under the gap condition.

Ada: That suggests a nice smooth behavior around certain points if those eigenvalue gaps hold up.

Emmy: They also analyzed the DMET bath-construction problem by decomposing A and B into blocks derived from a larger matrix gamma.

Kolya: Full disentanglement in that context requires m greater than or equal to the dimension of X minus l, where X is the smallest gamma-invariant subspace of R L containing Hfrag.

Ada: That constraint on m seems like a practical hurdle for applying this to real systems.

Emmy: Numerically, they detail three approaches: Riemannian optimization on the Stiefel manifold N, Self-Consistent Field SCF algorithms, and solving the convexified problem with the Optimal Damping Algorithm ODA.

Kolya: The ODA method provides a global minimum of J on M if H star is gapped.

Ada: But they noted that SCF algorithms can only converge to local minimizers of J on M, which is a limitation.

Emmy: The numerical tests on the benzene molecule from the DMET bath construction show that SCF always converges to a local minimum, but it doesn't guarantee convergence to the global minimum.

Kolya: The results suggest that Je provides a good starting point for Riemannian optimization methods to reach the assumed global minimum of J on M.