The Dictionary

Mechanics on the left, statistics on the right. Exact whenever energies are quadratic.

MechanicsStatistics
mass / spring stiffness $k$ precision $1/\sigma^2$
potential energy $\tfrac12 k (z-x)^2$ negative log-likelihood
force $k(x-z)$ score $\tfrac{d}{dz}\log p(z)$
equilibrium posterior mean (= mode)
centre of mass Kalman update $m' = \dfrac{P^{-1}m + R^{-1}y}{P^{-1}+R^{-1}}$ demo
masses in parallel add independent precisions add demo
springs in series: reduced mass $\dfrac{m_1 m_2}{m_1+m_2}$ variances add through a noisy channel; the Kalman prediction step demo
rigid rod on vertical springs least squares regression demo
rod on frictionless sliding collars PCA (total least squares) demo
bead chain: measurement + neighbour springs Kalman smoother; Whittaker graduation demo
masses re-weighted by error correlations GLS; bid-ask bounce as anti-redundant evidence demo
bead between corner posts; repulsive spring minimum-variance portfolio; short selling demo
one body, mass shrinking then regrowing over time the Kalman filter, tick by tick demo
pinning a node in a spring network conditioning; Schur complement of a graphical model demo
constant-tension pulley vs. Hookean spring Lasso vs. Ridge; soft-thresholding demo
thermal jitter posterior sampling (Langevin)

The correspondence is exact only for Gaussians: non-Gaussian likelihoods are non-Hookean springs — equilibrium still finds the mode, but subsystems can no longer be replaced by a single equivalent body.