Three traits hiding in one formula
By now you can solve the heat equation u_t = k u_xx three different ways: by separating variables into a decaying Fourier series on an interval, by convolving with the Gaussian heat kernel on the whole line, and by recognising its steady state as a harmonic limit. This guide does something different. Instead of producing one more solution, it steps back and asks: what do all of these solutions have in common? What is the personality of diffusion, the set of traits that every solution wears no matter how you found it?
There are three, and they are not independent decorations — they are three views of the same machinery. The heat equation smooths: feed it the jaggedest data you like and an instant later it is infinitely differentiable. It propagates at infinite speed: a disturbance anywhere is felt everywhere at once. And it is irreversible: you cannot, in general, run it backward to recover the past. Smoothing, infinite speed, and irreversibility are the signature of every parabolic equation, and once you see how they fall out of one fact, the heat equation stops being a formula and becomes a story.
Smoothing: roughness is killed in an instant
Picture a temperature profile with a sharp corner, even a jump — a square pulse of heat with two cliff edges. Switch on the heat equation and the corners do not soften gradually over minutes; the very instant time becomes positive, the profile is perfectly smooth, with derivatives of every order. This is the smoothing effect, and it is the most surprising thing diffusion does for free.
Two of your three solution methods explain it, and they agree. Through the heat kernel: the solution u(x,t) = integral of G(x - y, t) f(y) dy is a convolution of your data f against the Gaussian G, and G is infinitely smooth (indeed real-analytic). Convolution inherits the smoothness of the smoother factor, so even if f merely has jumps, u is C-infinity in x for every t > 0. Through Fourier: rough data is rough precisely because it is built from large high-frequency modes, and the factor e^(-k n^2 t) crushes the mode of frequency n. For any t > 0 the n^2 in the exponent annihilates the high harmonics almost completely, and what is left — a function with no fast wiggles — is smooth. Same trait, two windows onto it.
This is one of the deepest dividing lines between PDE types, so hold it against its opposite. The wave equation u_tt = c^2 u_xx does not smooth: a kink in the initial shape of a plucked string stays a kink and rides along a characteristic forever — singularities are transported, not erased. That difference of how singularities behave is exactly the parabolic-versus-hyperbolic split. Diffusion forgives roughness instantly; waves remember it forever.
Infinite speed: every point feels it at once
Light a match at one end of a very long, cold metal bar. Common sense says the far end stays cold for a while. The heat equation says something startling: the far end warms up immediately — by an utterly tiny, undetectable amount, but strictly above zero, the very instant after the match is lit. This is infinite propagation speed, and it follows from one property of the heat kernel.
Look at the Gaussian G(x,t) = (1 / sqrt(4 pi k t)) e^(-x^2 / (4 k t)). For any t > 0 however small, and any x however far, G(x, t) is strictly positive — the exponential of a real number is never zero. So in u(x,t) = integral of G(x - y, t) f(y) dy, if the initial heat f is positive anywhere, the integral is positive everywhere at every later instant. There is no wavefront, no last point the heat has reached beyond which it is exactly zero. The heat equation has no speed limit.
Set this beside the wave equation, the model of finite speed. A signal in u_tt = c^2 u_xx travels at the definite speed c; the region it has not yet reached is exactly, provably undisturbed, fenced off by the equation's characteristics. The wave equation has a sharp domain of dependence; the heat equation has none. Parabolic equations spread instantly; hyperbolic ones obey a finite speed limit — this is one of the cleanest practical consequences of an equation's type.
Be honest about what this means physically: nothing in nature truly moves infinitely fast, so this is the one place the heat equation tells a small lie. It is the price of Fourier's law assuming the flux responds to the gradient instantly. The model survives because the heat that races ahead is exponentially tiny — the Gaussian tail e^(-x^2/(4 k t)) is microscopic at large x, far below anything you could measure until the sensible diffusive time x^2/k has passed. If you genuinely need a finite speed, you must change the model (the hyperbolic telegrapher's equation adds a relaxation time and restores it).
Irreversibility: the arrow of time
Stir a drop of cream into coffee. It blends into an even tan, and you will never see, however long you wait, the tan coffee spontaneously un-mix back into a sharp drop of cream. Diffusion runs one way only. The irreversibility of diffusion is the mathematical face of that everyday certainty, and it is simply the smoothing effect viewed from the other end of time.
Here is why the two are the same fact. Forward in time, every mode decays like e^(-k n^2 t), so fine detail is steadily destroyed and many different rough pasts all flow toward similar smooth futures. The map from past to future loses information — and a map that loses information cannot be inverted. To go backward you would have to multiply each mode by e^(+k n^2 t), reviving the lost high frequencies by an explosive factor: a change in the present at the 10^(-10) level would correspond to a wild, unbounded change in the inferred past. Smoothing forward is exactly what makes the past unrecoverable.
This is the cautionary tale of the whole subject. The backward heat equation u_t = -k u_xx looks like an innocent request — find the earlier picture that blurred into this one — but it is ill-posed in Hadamard's exact sense: arbitrarily tiny changes in the data produce arbitrarily large changes in the answer, so the solution does not depend continuously on the data. Note the careful wording: ill-posed does not mean no solution exists; for ideal noise-free data a unique one may. It means the answer is useless with any real, imperfect measurement, because the lost detail you are trying to amplify is drowned in noise.
That is also why so many tempting real problems are quietly hopeless in their raw form. Deblurring a photo, thermal back-tracking to find a body's temperature an hour ago, identifying a heat source from later sensor data — all of these are backward diffusion. None can be solved by literally running the equation in reverse; they can only be tamed by regularization, where you replace the exact backward problem with a nearby well-posed one by adding extra assumptions (smoothness bounds, penalty terms) that hold the explosive high frequencies in check. Regularization does not undo the arrow of time; it answers a deliberately modified question.
One picture that ties the knot
The cleanest way to see all three traits at once is to watch a single rough lump of heat evolve and ask what the heat kernel does to it at each stage. The same Gaussian convolution that smooths is the same strictly-positive Gaussian that reaches everywhere is the same information-destroying average that cannot be undone. Read the picture below from a square pulse of heat dropped onto a cold line.
initial f(x): ___|-----|___ (two sharp jumps: a square pulse)
t > 0 (any t): u(x,t) = convolve f with the Gaussian
G(x,t) = e^(-x^2/(4 k t)) / sqrt(4 pi k t)
SMOOTHING corners gone instantly -> C-infinity hump, for every t>0
INFINITE SPEED G(x,t) > 0 for ALL x -> u(x,t) > 0 everywhere at once
IRREVERSIBLE the average loses detail -> a backward step needs e^(+k n^2 t): blows up
later pictures: .-=#####=-. then .-===-. then .-_-. (spreading, fading bell)
width ~ sqrt(4 k t); area = total heat = conservedNotice the two quiet conservation facts riding along in the bottom line. The width grows like sqrt(4 k t) — the x ~ sqrt(k t) diffusive scaling you have seen throughout — so doubling the spread costs four times the time; and the area under the bell stays fixed, the conservation of total heat that makes the heat kernel a unit of heat redistributed but never lost. The very same Gaussian that smooths and reaches everywhere also happens to be the self-similar profile every localized lump of heat eventually relaxes into, having forgotten the details of where it started.
Where this leads next
These three traits are qualitative pictures; the next guide turns one of them into an airtight theorem. Infinite speed and positivity hint that the heat equation cannot manufacture a new hot spot out of nothing — heat can only flow downhill from existing peaks. Made precise, that is the parabolic maximum principle: the maximum of a solution over a space-time region is always attained at the very start or on the boundary, never freshly in the interior. It is the rigorous engine behind two things you have been using on faith — uniqueness and continuous dependence — and so the real reason the forward heat problem is well-posed.
The same final guide also returns to sources. Everything here assumed no heater inside the bar; switch one on and you have the inhomogeneous equation u_t = k u_xx + f. Duhamel's principle handles it by treating the source as a stream of tiny instantaneous pulses, each of which then smooths, spreads, and conserves exactly as a fresh initial condition would. The three traits you have just met are not a tour stop — they are the working grammar of the whole parabolic world, and you will keep reading them off every diffusion equation you meet.