Notes on this project's Richardson iteration, for readers of algos.tex#

Why this exists#

evolving_surface/notes/algos.tex (Sec. 5, "Implicit timestepping and the linear solve") specifies the surface diffusion step as backward Euler, (I - Δt Δ_Γ) u^{n+1} = u^n, solved by preconditioned GMRES: the round-sphere Laplacian M = I - Δt Δ_S (diagonal, invertible by eigenvalue) preconditions the full operator, the real-embedding map E puts the half-spectrum complex coefficients into a real vector space, and restarted GMRES iterates to a residual tolerance.

turing-surface (this project) solves the same split operator with a different numerical method: a preconditioned Richardson (fixed-point) iteration, reusing exactly algos.tex's preconditioner M^{-1} but with no Krylov subspace, no orthogonalization, and no adaptive stopping. This note gives the map between the two, in this project's variable names, and why the switch.

Notation map#

The fixed point this project actually iterates#

Same split as algos.tex, lap_g = lap_s + dlap, substituted into backward Euler and rearranged so every occurrence of the unknown is Un. Starting from (I - dt*D*lap_g) Un = B and substituting the split:

(I - dt*D*(lap_s + dlap)) Un = B

Expanding, and moving the dlap term to the right so only the exactly invertible round-sphere part remains on the left:

Un - dt*D*lap_s(Un) = B + dt*D*dlap(Un)

lap_s is diagonal with eigenvalue -l(l+1), and lam holds +l(l+1), so lap_s(Un) = -lam .* Un — the left side becomes Un .* (1 + dt*D*lam), and dividing through gives:

Un = (B + dt*D*dlap(Un)) ./ (1 + dt*D*lam)

B is the explicit-reaction right-hand side — this project's models are IMEX (explicit reaction, implicit diffusion), where algos.tex's worked example is the bare heat equation, so B here is u^n plus a reaction term. Richardson iteration on this fixed point:

Un^(0)   = B ./ (1 + dt*D*lam)                                  [dlap = 0]
Un^(k+1) = (B + dt*D*dlap(Un^(k))) ./ (1 + dt*D*lam)

for k = 0 .. niter-1. models/schnakenberg.m's for k = 1:niter loop is this: Un^(0) is the divide computed just before the loop, and each pass computes Un^(k+1) from Un^(k). It is written as a full re-evaluation rather than an accumulated correction δ = Un^(k+1) - Un^(k) on purpose: at dlap ≡ 0 (the round sphere), every Un^(k) is then bit-for-bit Un^(0), with no cancellation to round differently — a stronger, and cheaper to check, statement than "close to the round-sphere answer."

Convergence, and why it isn't GMRES#

Writing M = I - dt*D*lap_s and A = M - dt*D*dlap, each step is Un^(k+1) = M^{-1}(B + dt*D*dlap(Un^(k))) — a stationary iteration that converges to the exact solution of A·Un = B exactly when the spectral radius of M^{-1}(dt*D*dlap) is below 1: while the geometric correction stays small against what the round-sphere solve already inverts. Unlike GMRES, there is no residual check and no adaptive iteration count: niter is fixed before the run starts, so a shape/timestep/diffusivity combination outside the convergence radius fails silently — the state saturates or diverges over many steps — rather than being caught the way algos.tex's solve_step catches it (its info != 0 return, logged when GMRES fails to reach tol within maxiter).

That tradeoff is deliberate, not an oversight, and it comes from where the two projects run. algos.tex's GMRES needs, every iteration: a dot product across the whole spectral state (Arnoldi orthogonalization) and a residual norm to test against tol — both require reading a scalar back to the host mid-solve. This project's solver instead records one whole timestep as a single GPU command buffer, submitted once, with the entire for k = 1:niter loop unrolled at compile time into a fixed sequence of dispatches — there is no point in that sequence where the host makes a decision, and no path for a data-dependent stopping rule to plug in. (Recompiling — which changing niter triggers — is the only way this project can change how much work a step does; see the README's "for loops, unrolled".) Richardson iteration is the cheapest method that still fits that shape: the same preconditioner as algos.tex, one dlap evaluation per iteration, a fixed and recompile-on-change trip count, in exchange for linear rather than superlinear convergence.

One more difference worth flagging#

algos.tex maps the half-spectrum complex coefficients through a real vector space embedding E (Sec. 6.4) because GMRES needs one flat, real-linear operator to hand to a generic solver. This project never needs E/E^{-1}: its spectral state is already carried as a real "2 x nlm" array — row 0 the real part, row 1 the imaginary — rather than packed complex, so every step here, dlap included, is already ℝ-linear arithmetic on that layout with no embedding or un-embedding step at all.