How spectral geometry affects decentralized gradient tracking

The steady-state mean-square disagreement scales as α2σ2Δλ−3\alpha^2\sigma^2\Delta_\lambda^{-3} when the nontrivial eigenvalue approaches $-1$.

The study examines how the spectral geometry of the mixing matrix affects the convergence of decentralized stochastic gradient tracking (GT). This process is used for smooth strongly convex problems.

The research shows that for two agents with identical quadratic objectives, the exact steady-state mean-square disagreement scales as α2σ2Δλ−3\alpha^2\sigma^2\Delta_\lambda^{-3}. This occurs as the nontrivial eigenvalue approaches $-1$, provided the stepsize α\alpha is sufficiently small relative to the gap.

The absolute spectral gap is denoted as Δλ\Delta_\lambda. The gradient-noise variance is denoted as σ2\sigma^2.

For diagonalizable mixing matrices, a stochastic bound averages all nontrivial eigenvalue contributions up to an eigenvector conditioning factor. As a real or complex eigenvalue approaches the unit circle away from 11, its contribution grows as the inverse cube of its distance to the circle.

For symmetric mixing with gaps pp at 11 and ν\nu at $-1$, the weighted sum θ1/p+θ2/ν3\theta_1/p+\theta_2/\nu^3 replaces the prior factor (Δλν2)−1(\Delta_\lambda\nu^2)^{-1}. For NN agents, Nθ1N\theta_1 and Nθ2N\theta_2 count the nonnegative and negative nontrivial eigenvalues, respectively.

A finite-horizon lower bound for the average iterate matches the worst-case factor p−1+ν−3p^{-1}+\nu^{-3} up to universal constants. Numerical experiments confirm this predicted scaling.

Key facts

  • The steady-state mean-square disagreement scales as α2σ2Δλ−3\alpha^2\sigma^2\Delta_\lambda^{-3} as the nontrivial eigenvalue approaches $-1$.
  • The contribution of an eigenvalue grows as the inverse cube of its distance to the unit circle away from 11.
  • Numerical experiments confirm the predicted scaling.

All sections

Edition