Heat flow pt = e−(t/2n)∂² p
t = 0.000
actual roots (zeros of the evolved polynomial)theory predictionstarting roots
Repeated differentiation roots of p(k)
k = 0
roots of p(k)starting rootsGauss–Lucas hull of the starting roots
What is exact and what is illustrative.
- Exact: the heat-flow dots are the true zeros of
e−(t/2n)∂²p. They obey the zero-dynamics ODEdzj/dt = (1/n) Σk≠j 1/(zj−zk), integrated with adaptive RK4. Time is scaled bynso that O(1) motion happens at t = O(1). - Exact: the differentiation dots are the true roots of the k-th derivative. The monomial coefficients of p are far too ill-conditioned for doubles at n in the hundreds (the polynomial is ~e−n/2 inside the support), so the coefficients are built in BigInt fixed-point arithmetic (about n+128 bits), and each p(k) is solved by Aberth–Ehrlich iteration warm-started from the previous derivative's roots.
- Theory overlay (heat): orange marks apply the transport map
z ↦ z + t·m0(z), wherem0(z) = ∫ dμ0(w)/(z−w)is the Cauchy transform. For the disk this is the exact mapz + t·z̄from the paper: circular law to an ellipse with semi-axes 1±t, then a semicircle at t = 1. For the ring, square and clusters I estimate m0 numerically from a 5000-point reference sample with a small smoothing length, so it is a first-order picture that is only trustworthy for small t, not a proof. - Theory overlay (derivatives): only the rigorous Gauss–Lucas hull is drawn. The centroid of the roots never moves under differentiation. The "mass moves toward the origin" result is for rotationally invariant μ0 (try disk or ring) and you can see it directly; this demo does not draw a predicted limiting curve for it.
- The paper's other fact: differentiating o(n) times leaves μ0 unchanged. Drag k to a handful and the cloud barely moves.