Phase Sync Playground

What's actually being computed here

The paper reduces power-system voltage-angle estimation to phase synchronization: recover unit-modulus phases x ∈ ℂⁿ from noisy pairwise angle-difference measurements on the grid graph. The estimator solves min ‖Cx‖² s.t. |x₁| = ⋯ = |xₙ| = 1, equivalent to maximizing x*Hx with H = −C*C — a Hermitian "connection matrix" built straight from the graph and the noisy measurements. This demo builds the classical connection-Laplacian specialization of that same matrix (paper §III–IV, Lemma 1–4): for every edge (i,j) present in the grid, it draws a noisy phase measurement z_ij = exp(i·(θᵢ − θⱼ) + noise) and accumulates it into an n×n complex Hermitian matrix. Every slider drag rebuilds this matrix and diagonalizes it for real — via a from-scratch complex Hermitian eigensolver (a real-symmetric Jacobi rotation on the 2n×2n real embedding of the matrix, since a Hermitian eigenproblem of size n is exactly a real-symmetric eigenproblem of size 2n). The top eigenvector of H, projected onto the unit circle per-entry, is the spectral estimate θ_si (paper eq. 19–20) — drawn here as amber arrows next to the true angles.

The duality gap is the paper's actual dual certificate (eq. 31–32): S(x) := C*C − Re[ddiag(C*C · xxᵀ)],   δ(x) := −n·λ_min(S(x)) computed here from a second real eigensolve of S(x) each frame. The paper proves 0 ≤ f(x) − f(xwls) ≤ δ(x), so δ(x) ≈ 0 is a genuine certificate of global optimality — not a cosmetic threshold. In the noiseless observable case the paper shows this gap is exactly zero and recovery is exact "without local refinement"; you can watch that happen live by dragging noise to 0.

The real result, in plain language

• Reduction: once voltage magnitudes are known accurately, estimating phase angles from noisy power-flow measurements is exactly a phase synchronization problem (paper, following Guzel & Zhang [8]).
• Governing quantity: everything depends on one normalized noise level η := (K_s‖ε_s‖_∞,obs + K_u‖ε_u‖_∞) / ρ (eq. 25), where ρ is the classical observability margin — the smallest singular value of the measurement Jacobian, eq. 10 — and Ks, Ku are network/measurement constants. This demo's slider-driven η = ε/ρ is a simplified single-noise-source version of exactly that ratio.
• Below threshold: Theorem 5 proves spectral initialization recovers the true angles to RMSE ≤ η radians, and Theorem 6 shows the WLS estimator is unique up to global phase with RMSE ≤ ½η.
• Noiseless case (η = 0): spectral initialization exactly recovers the ground truth in one shot, and the spectral certificate is exact with zero duality gap — no local refinement needed.
• Edge density ↔ observability: more/better-placed measurements raise ρ (the graph's algebraic "rigidity" against noise), directly lowering η and pushing δ(x) toward zero — this is why the density slider in this demo moves the same threshold curve.
• Paper's own experiments (standard IEEE benchmark grids) show the spectral method staying effective beyond the conservative regime the theorems formally cover — reproduced qualitatively here by the gap between the dashed η* line and where δ(x) actually blows up.