Determinantal equations conjecture for hidden Markov model varieties

Let Σ\Sigma be the alphabet, let gn,d\mathbf{g}_{n,d} denote the parameterization of the relevant hidden Markov model, and let p=(p(v))vΣnp=(p(v))_{v\in\Sigma^n} be a string function. For words vi,wjv_i,w_j with wjvin|w_jv_i|\le n, form the matrix whose (i,j)(i,j)-entry is p(wjvi)p(w_jv_i). Determinantal equations conjecture. Let n2d1n\ge 2d-1. Then

(p(v))vΣnimage(gn,d)(p(v))_{v\in\Sigma^n}\in\overline{\operatorname{image}(\mathbf{g}_{n,d})}

if and only if

det [p(wjvi)]1i,jd+1=0\text{\rm det~}[p(w_jv_i)]_{1\le i,j\le d+1}=0

for all choices of words v1,,vd+1,w1,,wd+1v_1,\ldots,v_{d+1},w_1,\ldots,w_{d+1} such that wjvin|w_jv_i|\le n. This gives determinantal equations for the Zariski closure of the hidden Markov model parameterization, addressing the problem of describing that variety by its invariants.

Sources & referencesView supporting material

Primary source

Alexander Schoenhuth, “Equations for hidden Markov models”, arXiv:0901.3749 (2009).

Progress summary

Refreshed
Open

The conjecture remains unresolved: the original 2009 paper states it, but no later proof, counterexample, or verification was found.

The determinantal-equations conjecture, stated as Conjecture 3.5 in a 2009 paper, proposes that for n2d1n \ge 2d-1 the vanishing of all specified (d+1)×(d+1)(d+1) \times (d+1) determinants exactly characterizes the Zariski closure of the hidden Markov model parameterization. The source presents this as a conjecture, not a theorem.

Current status (as of August 2026): The conjecture is unresolved; no publicly retrieved source establishes a proof, counterexample, or subsequent resolution.

Sources

Solutions 0

No solutions have been posted yet.