Researchers show a simple way to estimate an isolated root (zero) of a polynomial by randomly sampling points on a closed loop that surrounds that root. Using a classical formula from complex analysis, they rewrite a single, simple root as the exact average (expectation) of a complex-valued function taken around a contour. Sampling that function at random angles on the contour gives an unbiased Monte Carlo estimate of the root without changing the polynomial itself.

At a high level the idea uses the Cauchy integral / residue theorem. For a polynomial P(z) with a simple root r inside a closed curve, the root can be written as a contour integral of z times P'(z)/P(z) around the curve. The authors parameterize a convenient circle c + ρ e^{iθ} and define a function g(θ) so that r = E[g(θ)] when θ is uniformly random on [0,2π). In practice one draws M independent angles θ1,...,θM, computes Y_m = g(θ_m) for each, and takes the average r̂ = (1/M)∑Y_m. This r̂ is unbiased and has a quantified variance.

The paper develops finite-sample error bounds for this Monte Carlo estimator. The bounds come from concentration inequalities (Hoeffding-type bounds) and depend explicitly on the geometry of the chosen contour and how far it is separated from other roots. The framework also yields a stochastic version of the argument principle for counting how many roots lie inside a contour, and the authors outline a combined isolation-plus-estimation algorithm. They also report numerical experiments showing how the estimator behaves for different polynomial degrees and compare it with other root-finding ideas.

Important caveats follow directly from the construction. The method requires an isolating contour: a loop that contains exactly one root and no other roots, and it assumes the root is simple (not a repeated root). There must be no zeros on the contour itself. If the contour is close to other roots or to the root being estimated, the variance grows and more samples are needed. Because this is a Monte Carlo estimator, it returns a probabilistic estimate with a small chance of noticeable error; this differs from Las Vegas randomized algorithms that use randomness but still guarantee a correct answer (possibly with variable running time).