The Maximum Principle and Duhamel's Principle

Why we still need more tools

By now you can solve the heat equation in three ways: separation of variables on an interval, the heat kernel on the line, and the qualitative reading of smoothing, infinite speed, and irreversibility. But every one of those leaned on writing the solution down — a Fourier series, a Gaussian convolution. Two honest gaps remain. First: can you say anything firm about a solution without solving for it, just from the data? Second: what happens when there is a heater running inside the bar, so the equation has a source term and the clean formulas no longer apply?

This guide closes both gaps. The maximum principle answers the first: it bounds the solution everywhere using only the initial and boundary data, instantly, with no formula for the solution itself. Duhamel's principle answers the second: it constructs the response to any source by treating the source as a stream of tiny initial pulses and reusing the source-free solver you already trust. Together they are the qualitative and the constructive capstones of the whole rung.

The maximum principle: peaks live on the parabolic boundary

Heat a metal bar, hold its ends fixed, and watch. No interior spot ever spontaneously grows hotter than the bar was at the start or hotter than the ends are being held — the hottest temperature you will ever see was already present, either in the initial profile or on the boundary. Diffusion only flattens; it never manufactures a fresh peak out of nowhere. The parabolic maximum principle states exactly this for the heat equation: over a space-time region, the maximum (and the minimum) of u is attained on its boundary, not in the interior.

The subtle and crucial point is which boundary counts. For a problem on an interval across a time window from t = 0 to t = T, the relevant set — called the parabolic boundary — is the bottom (the initial slice t = 0) together with the sides (the two ends, for all times), but pointedly NOT the top (the final slice t = T). The principle reaches back to the initial data and the side data; it never reaches forward to the final time. That one asymmetry — past matters, future does not — is the irreversibility of diffusion showing up as plain geometry.

       t = T  (final time)   <-- NOT part of the parabolic boundary
        +-----------------+
        |                 |
  x = 0 |   interior:     | x = L   <-- sides ARE part of it
 (side) |   no new max    | (side)
        |   forms here    |
        +-----------------+
       t = 0  (initial slice) <-- bottom IS part of it

   max of u  =  max over {  bottom (t=0)  +  sides (x=0, x=L)  }
            EXCLUDES the top edge t = T

Why it is true, and what it instantly buys you

The proof is a short, beautiful trick worth seeing once. Suppose u solved u_t = k u_xx and had a strict interior maximum at some point. At an interior maximum in space, the profile curves downward, so u_xx <= 0; and at a maximum in time, u is not still rising, so u_t >= 0. But the equation demands u_t = k u_xx with k > 0, which would force a nonnegative quantity to equal a nonpositive one — only possible if both are zero, ruling out a strict peak. To kill the borderline case, look instead at v = u - epsilon t for a tiny epsilon > 0: then v_t - k v_xx = -epsilon < 0, so v can have no interior max at all, its max sits on the parabolic boundary, and letting epsilon go to zero hands the conclusion back to u.

From this one principle, three prizes drop out at once. Uniqueness: two solutions with the same initial and boundary data have a difference w that solves the heat equation with zero parabolic-boundary data, so its max and min are both zero, so w = 0 — the heat initial-boundary-value problem has exactly one solution. Stability: shift the data by at most delta and the solution shifts by at most delta everywhere, which is precisely the continuous dependence that makes the problem well-posed. And a bound for free: the whole interior is bracketed by the largest and smallest values on the parabolic boundary, a genuine a priori estimate you hold before solving anything.

The most useful repackaging is the comparison principle: bigger data gives a bigger solution, pointwise. If one bar starts and is edged hotter than another at every point, it stays hotter everywhere for all time — apply the maximum principle to the difference of the two solutions. A clean special case is positivity preservation: nonnegative initial and boundary data force a nonnegative solution, which is exactly why a temperature (or a probability density, or a concentration) never goes negative under diffusion. Common sense, finally proven.

Duhamel's principle: building the solution with a heater on

Now the second gap. Until now the bar had no internal heat sources — just an initial profile that spreads. But suppose a heater is switched on inside, adding heat continuously over time, so the equation becomes the inhomogeneous u_t = k u_xx + f(x,t), with f the source. The Gaussian convolution and the bare Fourier series were built for the source-free equation; they do not directly cover f. Duhamel's principle is the elegant bookkeeping trick that bridges the gap, and it rests entirely on the one property the heat equation has in abundance: linearity, hence the superposition principle.

The idea is a single vivid picture. Think of the running source not as one continuous push but as an unbroken stream of tiny instantaneous pulses. The heat injected during a brief interval around an instant s behaves like a fresh initial condition f(x,s) ds switched on at time s — and from that moment it simply diffuses forward on its own, obeying the source-free equation for the remaining time t - s. Each pulse forgets it was ever a source and just spreads like ordinary initial data. By superposition, the full forced solution is the sum (integral) of all those independently diffusing pulses.

1. Freeze the source at an instant s: read off the heat it delivers there, f(x,s) ds, and treat that slice as a brand-new initial condition switched on at time s.
2. Let that frozen pulse diffuse freely under the source-free heat equation for the leftover time t - s — using exactly the solver (heat kernel, or eigenfunction series) you already mastered.
3. Add up the contributions over every firing time s from 0 to t. Writing S(tau) for the source-free evolution over time tau, the forced solution is u(x,t) = integral from 0 to t of S(t - s) f(x,s) ds.
4. If the initial data is not zero, add on the separately-evolved initial condition by superposition — Duhamel handles the source, ordinary evolution handles the start, and the two simply sum.

Spelled out with the kernel, the response to the heater is u(x,t) = integral from 0 to t of [ integral of G(x - y, t - s) f(y,s) dy ] ds, where G is the heat kernel. This is the exact PDE analogue of the variation-of-parameters formula you met for ODEs, and the inner object G(x - y, t - s) is the causal Green's function in time: it lives only where t - s > 0, encoding that the temperature now can only depend on heat added in the past, never the future.

Honest limits, and how the two principles fit together

Both tools come with honest fine print. Duhamel's principle is a linearity statement through and through: it works only because the responses to separate pulses superpose. For a nonlinear source — say f depending on u itself, as in a reaction-diffusion model — the clean sum breaks, and you must turn the Duhamel formula into an integral equation and iterate toward a fixed point rather than read off the answer in one stroke. It also quietly assumes the source-free flow is itself well-posed, which is exactly why you can run it forward in time but never backward.

The maximum principle has limits of its own, and they are the same boundary that separates diffusion from waves. It is fundamentally a parabolic-and-elliptic phenomenon, born of the smoothing, order-preserving nature of diffusion. The wave equation obeys no maximum principle at all: its solutions oscillate freely above and below the data, an interior crest can rise higher than anything on the boundary, and the comparison principle fails outright. If you ever catch yourself expecting bigger data to give a bigger solution for a wave, stop — that intuition is diffusion's, not the wave's.