A new problem: no walls to bounce off
In the previous guide you solved the heat equation u_t = k u_xx on a finite rod by separation of variables: the two fixed ends acted like walls, only certain wavelengths fit between them, and the answer was a Fourier series whose modes each decayed at their own rate. That whole machine ran on the boundary. Now take the walls away. Let the bar be infinitely long, the whole real line minus-infinity to plus-infinity, with an initial temperature profile u(x, 0) = f(x) and nothing else. Where do the eigenfunctions go? There are no boundary conditions left to quantize them, so the discrete list of allowed wavelengths dissolves into a continuum.
This is no longer an initial-boundary-value problem; it is a pure Cauchy problem — give the data at one time everywhere in space and march forward. The reward for losing the walls is that the problem becomes, in a sense, simpler and more universal: every smooth start spreads the same way, and a single object will solve all of them at once. To find it we ask a focused question. Forget a general f for a moment, and ask what happens to the most concentrated start imaginable: all the heat piled at one point.
The heat kernel: heat from a single spark
Imagine touching the cold infinite bar with a white-hot needle at x = 0 for an instant: a unit of heat deposited at a single point. Mathematically that idealized point source is the Dirac delta, an object that is zero everywhere except at the origin yet has total integral one — not a function in the ordinary sense but a distribution, the response-to-a-point that you should always picture rather than try to graph. The solution of u_t = k u_xx that starts from this delta is the heat kernel, also called the fundamental solution because, as we will see, every other solution is built from it.
1 / x^2 \
K(x, t) = -------------- exp( - --------- ) for t > 0
sqrt(4 pi k t) \ 4 k t /
* a Gaussian (bell curve) in x, for every fixed t
* width grows like sqrt(t); peak height falls like 1 / sqrt(t)
* total area under K equals 1 for all t > 0 (heat conserved)Read the formula like a picture. At a fixed time t, K(x, t) is a Gaussian bell centered at the origin: the e^(-x^2 / (4 k t)) factor is the bell's shape, and the 1 / sqrt(4 pi k t) factor out front is exactly the height needed to keep the total area equal to one. As t grows the bell gets wider (its width scales like sqrt(t)) and shorter (its peak scales like 1 / sqrt(t)), trading height for spread while conserving area. That conserved area is conservation of total heat: the needle deposited one unit of warmth and not a joule leaks away — it just disperses. This is the cleanest possible portrait of diffusion: a spark of heat melting outward into an ever-flatter mound.
From one spark to any fire: solving by convolution
Here is the beautiful payoff, and the reason the kernel is worth all this fuss. The heat equation is linear, so the superposition principle holds: if you know the response to one spark, you know the response to a thousand sparks — just add them up. Now think of any initial profile f(x) as a dense crowd of point sources, one at each location y, each carrying f(y) units of heat. The response to the source at y is just the heat kernel centered at y, namely K(x - y, t), scaled by f(y). Add up (integrate) over all source positions y and you have the temperature everywhere, for all later time.
u(x, t) = integral over all y of K(x - y, t) * f(y) dy
= ( K(., t) * f )(x) <- a convolution
every point of the answer is a kernel-weighted average
of the WHOLE initial profile f, near and farThat operation — slide the kernel across f, multiply, and integrate — is a convolution, and it says something physically vivid: the temperature at point x and time t is a weighted average of the entire starting profile, where the weight is the bell K(x - y, t). Points near x count most (the bell is tallest at its center) and far-away points count a little, but every point counts something. This is exactly why the answer is smooth and why heat from everywhere reaches everywhere — the next two ideas of this guide.
Infinite speed and instant smoothing
Look hard at the convolution and a startling fact falls out. For any t greater than zero, no matter how tiny, the Gaussian K(x - y, t) is strictly positive for every x and every y — a bell curve never actually touches zero. So the integral that builds u(x, t) gives positive weight to f(y) at every location, even a billion meters away. Put a single warm bump anywhere and instantly, for all t > 0, the temperature at every other point on the infinite line is raised by some positive amount. This is infinite propagation speed: the heat equation feels a disturbance everywhere the very moment it happens.
The same positivity-and-averaging that gives infinite speed also gives the kernel's other signature: the smoothing effect. Because u(x, t) is a Gaussian-weighted average of f, every jagged corner, every jump, every kink in the initial data is blurred away the instant t leaves zero. Start with a profile that is merely bounded — even discontinuous, like a step from cold to hot — and for any t > 0 the solution is infinitely differentiable, perfectly smooth, with all its derivatives existing. Diffusion is the great smoother: it cannot create sharp features, only erase them. Hold onto this; it is exactly why the arrow of time only points one way.
One bell, every diffusion: self-similarity and irreversibility
There is a deeper economy hidden in the kernel. Notice that x and t never appear separately — they always travel together as the combination x / sqrt(t). This means the bell at a later time is just the earlier bell stretched horizontally by sqrt(t) and squashed vertically by the same factor; its shape never changes, only its scale. A solution that reproduces itself under such a coordinate rescaling is called self-similar, and the variable x / sqrt(t) is why every diffusion process — heat in a bar, ink in water, the spread of a random walker — wears the same Gaussian face. It is also the fingerprint of diffusion as Brownian motion: the sqrt(t) spreading is precisely how far a randomly jiggling particle wanders in time t.
Now run the movie backward and the kernel betrays the cost of all this smoothing. Forward in time, the bell only ever flattens; sharp information dissolves into the broad average and is gone. To go backward you would have to un-average — to reconstruct a spike from a smooth mound — and the smoothing effect makes that impossible to do stably: two wildly different starts can blur into nearly identical profiles, so the data no longer pins down the past. This is irreversibility, and the backward heat equation is the textbook example of an ill-posed problem. The kernel formula itself refuses to run backward: set t < 0 and the exponent flips sign, the 'Gaussian' blows up instead of decaying, and there is no solution at all.