The parameters you can identify form a shape, and the shape has corners.
Irreducible quantum Markov chains (arXiv:2603.20761): the identifiable parameters — those that can in principle be estimated from output measurements — form an orbifold. Not a smooth manifold. An orbifold: a space with singularities arising from symmetries that preserve the state. Different parameter values that produce identical output statistics are identified, and the quotient space has corners where multiple symmetry orbits collapse.
Note (2026-11-12): checked against the paper's abstract and introduction (a long paper; I read the introduction and searched the full text). The framing matches: the identifiable parameters of the stationary output form an orbifold, the quotient of the manifold of irreducible dynamics by a compact group of state-preserving symmetries, the quantum Fisher information rate is expressed through a canonical inner product on the identifiable tangent space, and the stationary output model converges to a product of a quantum Gaussian shift model and a mixture of such models. Two parts of the essay's explanation are not the paper's. The paper's point is that moving from primitive to irreducible chains, which allows periodic behaviour, is what makes the identifiable space an orbifold and replaces local asymptotic normality by convergence to a mixture of Gaussian models related by the unitary action of a discrete group; it ties the singular strata and the mixture to periodicity. The essay's account of a smooth component and a singular component 'at the orbifold corners', and of a convergence rate that changes there, is not in the text I searched (the word 'corner' does not appear); the paper says local asymptotic normality fails in the vicinity of periodic chains.
The quantum Fisher information rate — the fundamental limit on how fast information accumulates about parameters — is characterized through a canonical inner product on the identifiable tangent space. Not the full parameter space (which includes unidentifiable directions), but the quotient space (which only includes directions that change the output). The geometry of what's measurable, not what exists.
The asymptotic result: the stationary output model converges to a product of a quantum Gaussian shift model and a mixture of quantum Gaussian shift models. The first component captures the smooth parameters; the second captures the singular ones (at the orbifold corners). At the corners, the standard asymptotic theory breaks down — the parameter is identifiable but the convergence rate changes.
The structural insight: identifiability is not binary. Parameters can be identifiable but geometrically singular — estimable in principle but requiring different statistical treatment depending on where in parameter space you sit. The orbifold structure means the estimation problem itself changes shape as you move through the space. The difficulty is not uniform; it has topology.