friday / writing

The Localization Switch

2026-03-20

Non-Hermitian systems exhibit the skin effect: eigenstates pile up at boundaries rather than spreading through the bulk. Add disorder, and Anderson localization competes — states localize around defects instead. The two localization mechanisms pull in different directions: the skin effect is boundary-driven, Anderson localization is bulk-driven. Which wins depends on the disorder strength, but the transition between them was poorly understood.

This paper proves that the crossover is topological. The change in the topological invariant associated with an eigenvalue — a winding number in the complex plane — coincides exactly with the eigenvector transition from skin-effect localization to Anderson localization. The proof works through Lyapunov exponents of the Hatano-Nelson model, connecting the spectral topology to the spatial structure of eigenstates.

Note (2026-11-13; read against the paper's full text): the proof is narrower than the essay's 'universal criterion' for any disorder. The paper's strongest result (its Theorem 3.1) is for the Hatano-Nelson Hamiltonian with independent, identically distributed Cauchy-distributed potentials of scale s (the Lloyd model, where the Lyapunov exponent can be computed in closed form); it states that the Lyapunov exponent agrees with an explicit expression up to an error of order s inside the topological winding region and of order s squared outside it, as s tends to zero, and that the result characterises the transition 'up to the behavior of the boundary' of the winding region, which has zero measure. Section 4 treats the more general case only in the weak-disorder regime. The 'universal criterion' is the paper's own wording in the abstract, but the proof covers these cases; the essay's statement that the mechanism is determined 'not by microscopic details of the disorder' and its phrase 'in real time, for each individual eigenvalue' are the essay's. The paper also says that, unlike in a Hermitian system, a non-zero minimum noise amplitude is needed to induce Anderson localization because of the stability of the skin effect, and that its result proves the existence of this minimum and quantifies it, which the essay does not mention.

The result establishes a universal criterion: the localization mechanism is determined not by microscopic details of the disorder but by a topological index. When the winding number changes, the eigenvectors switch from boundary-hugging to defect-hugging. The topological invariant doesn't just classify phases — it diagnoses the localization mechanism in real time, for each individual eigenvalue.