Jacobi-Trudi determinants over finite fields (Q1787959)

Fix a finite field \(\mathbb{F}_q\) and an integer partition \(\lambda = (\lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_k)\). It is well-known that \(\lambda\) induces a Schur function \(s_\lambda\) in the ring \(\Lambda\) of symmetric functions. How likely is a uniformly random ring homomorphism \(\Lambda \to \mathbb{F}_q\) to send \(s_\lambda\) to \(0\)? This question is the guiding thread of the paper. To formalize it, one should introduce the complete homogeneous symmetric functions \(h_1, h_2, h_3, \ldots \in \Lambda\), which generate \(\Lambda\) freely as a commutative ring; one also sets \(h_0 = 1\) and \(h_i = 0\) for all negative \(i\). Now, the Jacobi-Trudi formula says that the Schur function \(s_\lambda\) equals the determinant of the \(k \times k\)-matrix \(( h_{\lambda_i - i + j} )\). To pick a uniformly random ring homomorphism \(\Lambda \to \mathbb{F}_q\) means to pick images \(v_1, v_2, v_3, \ldots\) of \(h_1, h_2, h_3, \ldots\) in \(\mathbb{F}_q\) (independently and uniformly); this homomorphism then sends \(s_\lambda\) to the determinant of the \(k \times k\)-matrix \(( v_{\lambda_i - i + j} )\) (with the conventions \(v_0 = 1\) and \(v_i = 0\) for \(i < 0\)). Of course, only finitely many \(v_i\) appear in this matrix, so we only need to pick finitely many elements (e.g., picking \(v_1, v_2, \ldots, v_{\lambda_1 - 1 + k}\) will suffice). The question can now be restated without reference to symmetric functions as follows: Fixing \(\mathbb{F}_q\) and \(\lambda\) as before, we pick \(\lambda_1 - 1 + k\) many elements \(v_1, v_2, \ldots, v_{\lambda_1 - 1 + k}\) of \(\mathbb{F}_q\) uniformly at random, and we set \(v_0 = 1\) and \(v_i = 0\) for \(i < 0\). What is the probability that the \(k \times k\)-matrix \(( v_{\lambda_i - i + j} )\) will have determinant \(0\)? Note that this is a submatrix of the Hankel matrix \(( v_{i+j} )\). For example, if \(\lambda = (3, 3, 2)\), then the matrix is \( \begin{pmatrix} v_3 & v_4 & v_5 \\ v_2 & v_3 & v_4 \\ 1 & v_1 & v_2 \end{pmatrix} \); the probability that its determinant is \(0\) is \((q^4 + q^3 - 2q^2 + q) / q^5\). For other \(\lambda\)'s, the probability can become more complicated; for instance, for \(\lambda = (4, 4, 2, 2)\) it is given by two different rational expressions in \(q\) depending on whether \(q\) is odd or even (Proposition 5.6). The authors conjecture that such an expression (a quasi-polynomial in \(q\) divided by a power of \(q\)) can always be given (Conjecture 5.8). For certain partitions \(\lambda\), the probability takes a particularly simple form. Namely, if \(\lambda\) is a rectangle (i.e., all parts of \(\lambda\) are equal), then the probability is \(1/q\) (Corollary 6.4). The same holds if \(\lambda\) is a hook -- i.e., a partition of the form \((a, 1, 1, \dots, 1)\) (Proposition 3.1) -- and if \(\lambda\) is a staircase -- i.e., a partition of the form \((k, k-1,\dots, 1)\) (Theorem 6.5). In characteristic \(\neq 2\), these three classes of partitions cover all partitions \(\lambda\) for which the probability is \(1/q\) (Theorem 7.9). In general, the probability is always at least \(1/q\) (Corollary 5.2). The authors also conjecture an upper bound: they suspect that the probability is always at most \(1 - \left| \operatorname{GL}_k(\mathbb{F}_q)\right| / q^{k^2}\) (Conjecture 5.10). (This can be restated as saying that the random \(k \times k\)-matrix constructed above is at least as likely to be invertible as a uniformly random \(k \times k\)-matrix.) The paper poses many more questions and answers some of them. For example, when \(\lambda\) is a hook, it is shown not just that the probability of the determinant to be \(0\) is \(1/q\), but rather that the determinant is equidistributed across \(\mathbb{F}_q\) (Proposition 3.1). This is not the case for rectangles \(\lambda\), but nevertheless the probability of the determinant to be any specific value in \(\mathbb{F}_q\) is determined in that case (Proposition 9.9). For a few pairs of partitions \(\lambda\) and \(\mu\), it is shown that the vanishing of the determinant for \(\lambda\) is independent from its vanishing for \(\mu\) (Section 8). We refer the reader to the paper for these and further results. The paper opens up an essentially new field of research, relating classical determinant combinatorics with the study of polynomial evaluations over finite fields. It proposes several inviting conjectures, as well as numerous ``implicit questions'' such as what happens if the Schur functions \(s_\lambda\) are replaced by other symmetric functions (e.g., the monomial symmetric functions). We suspect that it will spawn further work in the coming years. A pre-referee preprint of the paper appeared in [\url{arXiv:1611.00216}].