One Matrix, Two Worlds
The inverse of the Poisson chain's stiffness matrix is the covariance of a Brownian bridge — not approximately, not in a limit, but entry-for-entry in rational arithmetic. What that identity means, where it came from, and what it buys.
A short history of the same object
The formula $\min(s,t)-st$ was discovered at least three times, by communities that did not read each other.
Physics
1828. George Green — a Nottingham miller's son, almost entirely self-taught — publishes by private subscription (about fifty copies) his Essay on the Application of Mathematical Analysis to the Theories of Electricity and Magnetism. Its idea: solve a boundary-value problem once, for a unit point source, and every other solution follows by superposition. The impulse response is the Green's function.
1845. William Thomson rediscovers the forgotten essay and circulates it; the tool becomes the backbone of potential theory and, a century later, of numerical PDE.
Probability
1905 / 1923. Einstein explains Brown's jittering pollen; Wiener's Differential-Space makes the wandering path a rigorous measure: Brownian motion, with $\mathrm{Cov}(W_s,W_t)=\min(s,t)$.
1933–1952. Kolmogorov's empirical-distribution statistic needs the walk pinned at both ends. Lévy studies the tied-down process; Doob's 1949 heuristic and Donsker's 1952 invariance principle make it canonical: the Brownian bridge, $B_t = W_t - tW_1$.
Synthesis
1948–49. Feynman and Kac connect PDEs to path averages: solutions of heat-type equations are expectations over random paths.
1937. (Out of order, and usually forgotten:) Gantmacher & Krein prove inverses of tridiagonal matrices are "single-pair" matrices $u_i v_j$ — the discrete semiseparable fact below.
1974–2011. Besag's conditionally-specified auto-normal fields, Rue & Held's GMRF program, and the Lindgren–Rue–Lindström SPDE link finish the merger: difference operators are precision matrices; their inverses are covariances.
A tent and a pinned walk
Physics. Clamp a rod to $0^\circ$ at both ends and inject one unit of heat at $y$: $-u''=\delta(x-y)$, $u(0)=u(1)=0$. Away from the source there is nothing to bend the profile, so $u$ is piecewise linear: $Cx$ rising on the left, $D(1-x)$ falling on the right. Continuity at $y$ gives $Cy=D(1-y)$; the unit source forces the slope to jump by exactly $-1$, so $C+D=1$. Hence $C=1-y$, $D=y$:
$G(x,y)=\begin{cases}x(1-y), & x\le y\\ y(1-x), & x\ge y\end{cases} \;=\;\min(x,y)-xy.$
Everything in that derivation is visible to the naked eye — and the continuum comes first. Drag the source below. The tent is the continuous object; no grid has been mentioned yet. Only after the tent feels obvious should you press discretize — and watch the matrix agree with the continuum to the last digit.
Probability. Pin the walk: $B_t=W_t-tW_1$. Expand with bilinearity and $\mathrm{Cov}(W_a,W_b)=\min(a,b)$:
$\mathrm{Cov}(B_s,B_t)=\min(s,t)-ts-st+st\cdot 1=\min(s,t)-st.$
That one-liner is the official record, and it is too fast: it asserts the pinning recipe $B_t = W_t - tW_1$ and hides four covariance terms inside "expand". Slow it down twice.
Why $tW_1$? Because pinning is regression
Why subtract $tW_1$ and not something else? Because $tW_1$ is exactly what the endpoint knows about the path. The best linear predictor of $W_t$ from the endpoint $W_1$ is $\beta W_1$ with the least-squares coefficient
$\beta \;=\; \dfrac{\mathrm{Cov}(W_t,W_1)}{\mathrm{Var}(W_1)} \;=\; \dfrac{\min(t,1)}{1} \;=\; t,$
and for Gaussians the linear least-squares predictor is the conditional expectation: $\mathbb E[W_t\mid W_1]=tW_1$. So the bridge is not a trick — it is a residual: $B_t = W_t - \mathbb E[W_t\mid W_1]$, the part of the walk the endpoint cannot explain. Pinning is conditioning is regression — the same dictionary entry as report 09's clamping-is-conditioning, and the same residual that report 15 builds preconditioners out of. Watch it happen:
The four terms, one at a time
Now the "expand with bilinearity" step, term by term. Covariance is bilinear — it distributes over sums like multiplication — so the pinned covariance opens into four pieces:
Same formula. The deterministic influence function of a clamped rod and the covariance of a doubly-pinned random walk are one object — clamping is conditioning. (The matrix version is exact too: with $P=I-t\,e_m^\top$ acting on the discrete walk covariance $\min(t_i,t_j)$, the identity $P\,M\,P^\top=\min(t_i,t_j)-t_it_j$ holds in rational arithmetic — check #7 below.)
Where does the square root come from?
One honest question remains before the pictures: the tent, the gold parabola $x(1-x)$, and the diagonal $\Sigma_{ii}/h$ are all variances — they live in the field's units squared. A band you draw against sample paths must be in the field's own units, so you take the square root: $\sigma(x)=\sqrt{x(1-x)}$. That is the whole mystery: the $\sqrt{\phantom{x}}$ converts a variance curve into a width. And $\pm2\sigma(x)$ is a pointwise statement — at each fixed $x$ it covers about 95% of samples ($2\Phi(2)-1=95.45\%$; a seeded 200,000-bridge Monte Carlo in the checks script measures 95.48%). It is not a corridor for whole paths: only 53.0% of bridges stay inside the band at every one of the 63 interior samples.
Five nodes, twenty-five fractions, zero error
Take $n=5$ interior nodes, $h=\tfrac16$, $A=\mathrm{tridiag}(-1,2,-1)/h^2$ — diagonal $72$, off-diagonal $-36$. Inverting in exact rational arithmetic:
Why is there no discretization error at all? Because the continuum Green's function is piecewise linear, and the three-point stencil annihilates linear functions exactly. Run $[-36,72,-36]$ along a tent: off the kink it reads a straight line and returns $0$; at the kink, $-u_{i-1}+2u_i-u_{i+1} = h(\text{slope}_L-\text{slope}_R) = h\,(C+D) = h$, so dividing by $h^2$ gives $1/h$ — and the quadrature weight in $\Sigma = h\,G$ restores exactly $1$. $A\Sigma=I$ as a matrix of rational identities.
Sparse conditional, dense marginal
Read the two panels above as statistics. Zeros of the precision matrix are conditional independencies: given its two neighbors, node 3 learns nothing from the rest of the rod — the row of $A$ literally says $\mathbb E[u_i\mid u_{-i}]=\tfrac12(u_{i-1}+u_{i+1})+\tfrac{h^2}{2}b_i$, the mean-value property as a regression. That is the Markov property, and it is why $A$ has $O(n)$ entries. The covariance has no zeros at all: heat injected anywhere warms everything — marginal dependence is global even though conditional structure is nearest-neighbor. One matrix pair, and the whole "sparse precision / dense covariance" doctrine of GMRFs is on the table.
The same tent in three languages
Nothing in §2's derivation was actually thermal. Three professions solve $-u''=\delta(x-y)$, $u(0)=u(1)=0$ every day, each with its own words for the tent, its slopes, and its kink. The dictionary is exact — every row below is a machine-checked identity, not an analogy.
⚡ Resistance
A wire of resistance 1 Ω per unit length, both ends grounded. Inject one ampere at $y$: the voltage profile is the tent, and the peak is the effective resistance to ground, $R_{\rm eff}(y)=y\parallel(1-y)=y(1-y)$ — two resistors in parallel.
The current divider: $1-y$ of the ampere exits the left wall, $y$ the right — the wall at distance $d$ takes share $1-d$, so the closer ground takes the larger share. Verified for all 31 source positions; see mathematica/electrical_networks.wls and the free-vs-grounded row-sum story: grounding deletes two rows of the free path Laplacian, and the surviving end rows have row sum 1 — those are the wires to ground.
🔥 Thermal
A rod with walls held at $0^\circ$ (heat baths). Inject one watt at $y$: the temperature is the tent, the flanks carry constant flux $1-y$ (left) and $y$ (right), and the kink jumps by exactly $-1$ — the injected watt, read off pointwise from $-u''=\delta$.
At $y=0.3125$: slopes $0.6875$ / $-0.3125$, peak $0.2148$, wall exits $0.6875+0.3125=1$ W. All fourteen tent facts (superposition included: a $+2$ W and a $-1.5$ W source stack linearly) are the green_tents_checks.py suite — 14/14 PASS.
🌉 Mechanical
A taut string at unit tension, ends pinned. Apply a transverse unit point load at $y$: the deflection is the tent — the influence line of structural engineering (Müller-Breslau), and Maxwell reciprocity $G(x,y)=G(y,x)$ says: deflection at $x$ from a load at $y$ equals deflection at $y$ from a load at $x$.
The internal shear force is the derivative: piecewise constant, $+(1-y)$ then $-y$, jumping by the load at $y$, with support reactions $(1-y)+y=1$ carrying it. The shear diagram is the flux diagram is the current diagram — force balance and Kirchhoff's law are the same slope-jump identity.
| the tent says | electrical | thermal | structural | statistical |
|---|---|---|---|---|
| u(x) = G(x,y) | voltage at x | temperature at x | deflection at x (influence line) | Cov(B_x, B_y) |
| peak = y(1−y) | R_eff to ground, y ∥ (1−y) | thermal resistance to the baths | compliance under the load | Var(B_y) |
| −u′(x) | current in the wire | heat flux | shear force V(x) | ∂ₓ Cov — jumps by −1 at y |
| slope jump −1 at y | injected ampere (Kirchhoff) | injected watt | applied point load (force balance) | row y of AΣ = I |
| flank split 1−y / y | current divider | wall-exit flux split | support reactions R_L, R_R | regression weights on the two pins |
The rank-one triangle
Look again at the closed form: on the upper triangle, $\Sigma_{ij}=(h\,x_i)\cdot(1-x_j)$ — an outer product of two vectors. Gantmacher & Krein called such inverses "single-pair matrices" in 1937: every off-diagonal block of $\Sigma$, however large, has rank one.
This is semiseparability, and it is the structural reason "dense" does not mean "expensive" here: the same compressibility, generalized to blockwise low rank, is the engine of hierarchical matrices (Hackbusch), HSS solvers, and the fast multipole method (Greengard–Rokhlin) — apparently-dense inverse operators applied in near-linear time. It is also the mirror image of §4: a matrix has a rank-one-triangle inverse exactly when it is tridiagonal. Semiseparable covariance and Markov precision are the same property, seen from the two sides of the inverse.
Preconditioning is decorrelation — and CG is the right customer
A preconditioner $M=CC^\top$ is a surrogate Gaussian; applying $C^{-1}$ re-expresses the problem in coordinates the surrogate believes are white, and the solver then faces the true covariance of those pseudo-innovations. Perfect model $\Rightarrow$ identity $\Rightarrow$ one step. The interesting question is what happens with a partial model — and here gradient descent and CG part ways.
Steepest descent contracts the error by $(\kappa-1)/(\kappa+1)$ per step: a function of the two extreme eigenvalues only. Cluster 99% of the spectrum at $1$ and it does not care; the leftover outliers set its pace forever. CG, after $k$ steps, has applied the best degree-$k$ polynomial with $p(0)=1$ across the whole spectrum — so it terminates in as many iterations as there are distinct eigenvalue clusters. A cluster is nearly free; an outlier costs about one iteration.
The cleanest laboratory is domain decomposition on the chain itself: block-Jacobi with $p$ blocks models each segment exactly and ignores the $p-1$ interface couplings. The model error has rank $2(p-1)$, so the preconditioned spectrum is a cluster at $1$ plus exactly $2(p-1)$ outliers. Choose your assassin:
That is the honest statement of the folklore: preconditioning for gradient descent is a search for a spherical bowl — anything less leaves it hostage. Preconditioning for CG is a search for a clustered spectrum: decorrelate what is cheap to decorrelate, and hand the solver a short list of named exceptions. Domain decomposition, deflation, incomplete factorizations, Nyström — all of them are strategies for choosing which few directions to leave wrong.