arXiv++ Combinatorics

Machine learning explorations can make significant inroads into solving difficult problems in pure mathematics. One advantage of this approach is that mathematical datasets do not suffer from noise, but a challenge is the amount of data required to train these models and that this data can be computationally expensive to generate. Key challenges further comprise difficulty in a posteriori interpretation of statistical models and the implementation of deep and abstract mathematical problems. We propose a method for scalable tasks, by which models trained on simpler versions of a task can then generalize to the full task. Specifically, we demonstrate that a transformer neural-network trained on predicting permutations from words formed by general transpositions in the symmetric group $S_{10}$ can generalize to the symmetric group $S_{25}$ with near 100\% accuracy. We also show that $S_{10}$ generalizes to $S_{16}$ with similar performance if we only use adjacent transpositions. We employ identity augmentation as a key tool to manage variable word lengths, and partitioned windows for training on adjacent transpositions. Finally we compare variations of the method used and discuss potential challenges with extending the method to other tasks.

Starting from the Izergin-Korepin 19-vertex model in the quadrant, we introduce two families of rational multivariate functions $F_S$ and $G_S$; these are in direct analogy with functions introduced by Borodin in the context of the higher-spin 6-vertex model in the quadrant. We prove that $F_S(x_1,\dots,x_N;z)$ and $G_S(y_1,\dots,y_M;z)$ are symmetric functions in their alphabets $(x_1,\dots,x_N)$ and $(y_1,\dots,y_M)$, and pair together to yield a Cauchy identity. Both properties are consequences of the Yang-Baxter equation of the model. We show that, in an appropriate limit of the spectral parameters $z$, $F_S$ tends to a stable symmetric function denoted $H_S$. This leads to a simplified version of the Cauchy identity with a fully factorized kernel, and suggests self-duality of the functions $H_S$. We obtain a symmetrization formula for the function $F_S(x_1,\dots,x_N;z)$, which exhibits its symmetry in $(x_1,\dots,x_N)$. In contrast to the 6-vertex model, where $F^{6{\rm V}}_S(x_1,\dots,x_N;z)$ is cast as a sum over the symmetric group $\mathfrak{S}_N$, the symmetrization formula in the 19-vertex model is over a larger set of objects that we define; we call these objects 2-permutations. As a byproduct of the proof of our symmetrization formula, we obtain explicit formulas for the monodromy matrix elements of the 19-vertex model in a basis that renders them totally spatially symmetric.

We study the stochastic six-vertex model in half-space with generic integrable boundary weights, and define two families of multivariate rational symmetric functions. Using commutation relations between double-row operators, we prove a skew Cauchy identity of these functions. In a certain degeneration of the right-hand side of the Cauchy identity we obtain the partition function of the six-vertex model in a half-quadrant, and give a Pfaffian formula for this quantity. The Pfaffian is a direct generalization of a formula obtained by Kuperberg in his work on symmetry classes of alternating-sign matrices. One of our families of symmetric functions admits an integral (sum over residues) formula, and we use this to conjecture an orthogonality property of the dual family. We conclude by studying the reduction of our integral formula to transition probabilities of the (initially empty) asymmetric simple exclusion process on the half-line.

We derive combinatorial formulae for the modified Macdonald polynomial $H_λ(x;q,t)$ using coloured paths on a square lattice with quasi-cylindrical boundary conditions. The derivation is based on an integrable model associated to the quantum group of $U_{q}(\wide{sl}_{n+1})$.

We prove that the normalisation of the stationary state of the multi-species asymmetric simple exclusion process (mASEP) is a specialisation of a Koornwinder polynomial. As a corollary we obtain that the normalisation of mASEP factorises as a product over multiple copies of the two-species ASEP.

We introduce a new family of symmetric multivariate polynomials, whose coefficients are meromorphic functions of two parameters $(q,t)$ and polynomial in a further two parameters $(u,v)$. We evaluate these polynomials explicitly as a matrix product. At $u=v=0$ they reduce to Macdonald polynomials, while at $q=0$, $u=v=s$ they recover a family of inhomogeneous symmetric functions originally introduced by Borodin.