Separation of variables

The formula

Start with a linear PDE. The one-dimensional heat equation in a rod is the standard teaching case:

∂u/∂t = α · ∂²u/∂x²

Reading: the rate a point heats up in time equals a diffusivity α times how much the temperature curves in space. Curvature drives flow.

Now guess the solution factors — space stuff times time stuff:

u(x,t) = X(x) · T(t)

Substitute and divide by X·T. Everything on the left depends only on t, everything on the right only on x. Two things equal for all x and all t can only be a constant. Call it −λ:

T'(t) / (α · T(t))  =  X''(x) / X(x)  =  −λ

Reading: one PDE in two variables just became two ODEs in one variable each, tied together by a single number λ — the separation constant, which turns out to be an eigenvalue.

For a rod of length L held at zero on both ends, the space equation X'' + λX = 0 with X(0)=X(L)=0 only has nonzero solutions for a discrete list:

λ_n = (n·π / L)²        X_n(x) = sin(n·π·x / L),   n = 1, 2, 3, …

Reading: only certain shapes fit between the two clamped ends. These are the modes. Each one decays in time like T_n(t) = exp(−α·λ_n·t).

Add them all up, weighted to match the starting temperature f(x):

u(x,t) = Σ  b_n · sin(n·π·x / L) · exp(−α·λ_n·t)
         n

b_n = (2/L) · ∫₀ᴸ f(x)·sin(n·π·x / L) dx

Reading: the answer is a Fourier series whose every term fades at its own rate. High modes (n large) die fast because λ_n grows like n².

Where you meet it

Heat-soak on a test article. A rod, fin, or bracket clamped in a thermal chamber with fixed end temperatures relaxes toward steady state exactly as the series above. The n=1 mode has the longest time constant τ₁ = 1/(α·λ₁) = L²/(α·π²), so after a couple of τ₁ the whole thing is a single decaying sine and everything else is noise. That is why "wait five time constants" works on a soak — you are waiting out the slowest mode.
Modal survey on a vibration stand. The wave equation for a string, beam, or panel separates the same way. The space eigenfunctions are the mode shapes; the eigenvalues set the natural frequencies you chase with a shaker and accelerometers. When a review board asks "what's the first mode," they are asking for X_1 and its frequency — the lowest λ.
Waveguide and cavity fields. RF engineers separate the Helmholtz equation in a rectangular guide into X·Y·Z. The eigenvalues become cutoff frequencies; the eigenfunctions are the TE and TM field patterns. The mode chart on the bench is a separation-of-variables solution in disguise.
Transient conduction hand-check. Before anyone trusts the FEA thermal run, a senior engineer will drop the geometry to a 1-D slab, run the first two terms of this series on a napkin, and confirm the sim's time constant lands in the right decade. If it doesn't, the mesh or the boundary condition is wrong, not the napkin.

How it works

The whole method rests on two loadbearing facts, and both are places people get burned.

It needs linearity. The reason you can add up modes is superposition — the sum of solutions is a solution. Put a u² or a temperature-dependent conductivity into the equation and superposition dies. The series is meaningless for a nonlinear PDE. Separation of variables is a linear-problem tool, full stop.

It needs homogeneous boundary conditions. The eigenfunctions come out clean only when the boundaries are zero (or zero-flux). The classic mistake: the ends of the rod are held at 100 °C, not 0, and someone tries to expand the actual temperature directly in sin(nπx/L). It fits badly and the series fights you. The fix is standard — split off a steady-state piece v(x) that carries the nonzero boundary values, solve the homogeneous transient for what's left, then add them back. u = v + (series). Skip that step and your "solution" won't satisfy the boundaries.

The eigenfunctions are orthogonal, and that's the whole magic. For the clamped rod, ∫₀ᴸ sin(mπx/L)·sin(nπx/L) dx = 0 whenever m ≠ n — numerically it's flat zero, not "small." That orthogonality is what lets you pick off each b_n independently with the integral above, the same way a Fourier analyzer pulls one bin out of a signal without the others leaking in. This isn't luck. The operator d²/dx² with those boundary conditions is self-adjoint (a Sturm–Liouville problem), and self-adjoint operators hand you a complete orthogonal basis for free.

Convergence and the honest limits. Truncating the series at a few terms is fine for smooth data and later times, because the fast modes are already dead — at t = 0.1·L²/α in the unit case, the exact center temperature is 0.47449 and the single first mode gives 0.47455, a fraction of a percent off. But at t = 0 with a discontinuous start (a rod suddenly at uniform 1 with zero-clamped ends), the partial sum overshoots and rings near the jump — Gibbs phenomenon. A 400-term sum lands the center at 0.9984, not 1.0000, and no number of terms kills the overshoot at the corner. So the method is exact in the limit but slow to converge exactly where the initial data is ugly. Know which regime you're in.

Where it simply won't separate. The geometry has to line up with a coordinate system the equation separates in — Cartesian for a box, cylindrical (Bessel functions show up as the eigenfunctions) for a pipe, spherical for a ball. An odd-shaped bracket in Cartesian coordinates does not separate, and that is exactly the situation FEA exists to handle. Separation of variables is the sharp analytic tool for clean geometry and the sanity check for everything else.

History

The phrase came out of the Bernoulli–Leibniz calculus circle before anyone was thinking about heat. Johann Bernoulli used "separation of variables" in a 1694 letter to Leibniz, and worked out the method in print in the Acta Eruditorum that same year ^[1]. He and his brother Jacob were deep in Leibniz's calculus by the early 1690s — Johann was lecturing on the differential calculus in 1691 and integrating differential equations on curve problems like the catenary and the caustics ^[2]. That was the ODE version — the same move of getting all the x on one side.

Pushing it into partial differential equations was the mid-1700s vibrating-string fight. Jean le Rond d'Alembert wrote down the wave equation in 1747; Euler and Daniel Bernoulli piled in over the next few years, with Daniel Bernoulli arguing that a plucked string's motion was a sum of simple sinusoidal modes — physically, separation of variables plus superposition, though the mathematicians of the day brawled for decades over whether an arbitrary shape could really be written as such a sum ^[3]^[4].

The person who made it a method was Joseph Fourier, working on how heat spreads through a solid. He presented a memoir on heat propagation to the Paris Institute on 21 December 1807, but Lagrange blocked its publication — he distrusted the free way Fourier tossed arbitrary functions into trigonometric series ^[5]^[6]. Fourier reworked it, won the Academy's 1811 prize on heat diffusion (still dinged for a lack of rigor), and finally published the full treatment as Théorie analytique de la chaleur in 1822, after Lagrange had died ^[5]^[6]. That book is where separating the heat equation, expanding the initial condition in sines and cosines, and summing the decaying modes all appear as one working technique — the reason the whole thing is still called the Fourier method. The rigor Lagrange wanted arrived later, when Sturm and Liouville (around 1836) proved that this class of eigenvalue problems always delivers a real, orthogonal, complete set of modes — the theorem that makes the guess honest.

Related tools

/tools/heat-conduction
/tools/vibration-natural-freq
/tools/beam-deflection
/tools/lc-resonance
/tools/frequency-period

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →