Why transform the equation?
On a finite interval, boundary conditions select a discrete family of sine or cosine modes. On the whole line, the analogous frequencies form a continuum. The Fourier transform records the amount of each frequency in the initial profile. Differentiating twice in space multiplies the frequency component by a nonpositive number, so the partial differential equation becomes an ordinary differential equation in time for each fixed frequency. Solving that family of equations suggests a solution; we must still justify inversion and recovery of the initial data.
Consider
For this derivation take in the Schwartz class.
Definition 3.4.1
(Schwartz class )
A function belongs to if is smooth and is bounded on for all nonnegative integers . In particular, and all its derivatives decay faster than any inverse power of .
This convenient assumption is stronger than necessary; it gives convergent integrals and licenses the Fourier identities below without treating distributions. The separate heat-kernel and Cauchy-problem note permits much broader initial profiles.
Convention and the frequency equation
We use the following convention throughout.
Definition 3.4.2
(Fourier transform and its inverse)
For ,
The second formula is the inversion theorem; the transform maps onto itself.
Integrating by parts twice, with boundary terms vanishing by rapid decay, gives
Suppose temporarily that a solution and its derivatives have enough decay and time regularity to transform the PDE and move the time derivative through the integral. Then, for each fixed ,
The scalar ODE has the unique solution
This calculation motivates the candidate. It is not a proof that every classical solution admits an ordinary Fourier transform: a bounded function need not be integrable. Instead, we construct the candidate by inversion and verify it directly.
The factor is one at zero frequency and strictly less than one at every nonzero frequency for ; a larger is damped faster. This is a statement about Fourier components, not about pointwise values of the solution.
Inversion, regularity and the initial trace
The candidate is now defined directly by inversion, and its properties are verified without assuming anything about solutions in advance.
Theorem 3.4.3
(Frequency-side solution of the heat equation)
Let and define
Then is smooth on (with one-sided time derivatives at ), , for every , and as .
Proof
(Differentiation under the inverse integral)
Because is Schwartz, is integrable for every nonnegative integer . On any bounded interval of nonnegative times, differentiating the integrand times in and times in introduces the factor , while the exponential has modulus at most one. An integrable majorant therefore permits differentiation under the integral by dominated convergence. Time derivatives at are understood from the right; on positive times they are ordinary derivatives. In particular,
The inversion theorem gives . Moreover,
by dominated convergence. Thus the constructed solution has the prescribed initial values uniformly, not merely at individual points. The same reasoning gives joint continuity at . For every fixed the multiplier times is Schwartz, so is Schwartz as well.
Every step of this proof leans on the same bound.
Proof X-ray
What the multiplier bound does in the proof
The proof uses one estimate three times. Differentiating times in and times in , letting , and placing back in the Schwartz class each need an integrable majorant independent of , and is one because . That inequality is where and enter. The inversion theorem is used only at , to identify the trace with . Uniform convergence to then follows from the pointwise limit and the same majorant; no property of the Gaussian kernel is needed. For the multiplier is , which grows, and the majorant disappears.
The last sentence is not a technicality.
Counterexample mode
The frequency formula does not run backwards
Claim: the same inverse integral defines a solution for negative times. Take , whose transform is . For the integrand has modulus , which is not integrable once . Correspondingly, the Gaussian formula of the worked example with gives , whose peak tends to infinity as . The hypothesis in the proposition cannot be dropped.
From frequency damping to the Gaussian kernel
For positive times the frequency formula can be rewritten as a spatial convolution.
Theorem 3.4.4
(Gaussian kernel representation)
For and , the solution of the previous proposition is
Proof
(Inverting a Gaussian and exchanging integrals)
For , the Gaussian Fourier transform is
One may verify the constant without a complex contour: if denotes the integral on the left, differentiation under the integral and integration by parts in give ; the real Gaussian integral gives . Solving this ODE proves the formula.
With , denote the resulting Gaussian by . Substitute the definition of into the inverse integral. Since
Fubini's theorem permits interchange of the integrals, and the inner integral is by the Gaussian formula with . The factor belongs to the inverse transform; it is absorbed into the normalized Gaussian.
This computation connects the frequency and spatial descriptions. The other heat-kernel note proves directly that has unit integral and that convolution approaches bounded uniformly continuous data. That extension does not require such data themselves to have an ordinary Fourier transform.
Concept lensAnalytic
Heat flow as an average over random displacements
For , is the density of a normal random variable with mean and variance . If is a standard normal variable, the kernel representation therefore reads
Positivity and are properties of an average, and the initial trace says that the displacement shrinks to zero. The cosine example below is the identity with .
Which uniqueness statement is justified?
Within solutions that are Schwartz in space and sufficiently regular for the preceding transform calculation, the scalar ODE and Fourier inversion give uniqueness. A broader result needs a different argument: the heat-kernel and Cauchy-problem note proves uniqueness among classical solutions continuous to the initial line and bounded on each finite slab , using a maximum-principle barrier. Our constructed solution lies in that class because
Common mistake
(Applying the frequency ODE to an arbitrary solution)
Do not drop the growth condition and infer uniqueness among all classical whole-line solutions from the frequency ODE. An arbitrary candidate need not be Fourier transformable as a function: the bounded solution is not integrable, and the uniqueness argument for the transformable class says nothing about it.
Existence of our candidate, uniqueness in the transformable class, and bounded-slab uniqueness are three related but distinct conclusions.
Worked examples
Worked example 3.4.5
(Track both height and width)
Take with . Its transform is . Frequency damping adds to the positive coefficient of inside the negative exponent. Applying the Gaussian inverse formula gives
At the formula returns . The peak decreases while the spatial width increases. The integral remains : the decrease in height and increase in width balance. This conservation applies here because the profile is integrable; a constant profile is preserved too, but its whole-line integral is infinite.
The decrease of the peak is a statement about the maximum, not about every point.
Common mistake
(Frequency damping is not pointwise decay)
Every nonzero frequency is damped, but values at a fixed point can increase. For and , the Gaussian solution gives , while at it gives . Heat moving outward from the peak raises the tails.
Worked example 3.4.6
(A signed profile: the derivative of a Gaussian)
Take , the derivative of the previous profile. Its transform is , which vanishes at . Since the multiplier commutes with multiplication by , the solution is the -derivative of the Gaussian solution. With ,
The integral is zero at every time, because is odd. The maximum of is attained at and equals . It decays like , faster than the decay of the positive Gaussian: the positive and negative parts cancel as they spread.
Worked example 3.4.7
(A single frequency outside the Schwartz class)
The profile is bounded but not integrable, so is not an ordinary function. The kernel formula still applies. Since is even, the part of integrates to zero, and
The last integral is the characteristic function of a centered normal distribution with variance , evaluated at . Direct differentiation confirms . The single mode is damped by exactly the multiplier at . For the profile is constant and is preserved.
Practice
Quick check
(For initial data , what is the ratio ? Compare it with the corresponding ratio at for .)
Solution · Answer
The ratios are and respectively. The first is smaller for positive . The ratios are well defined here because the Gaussian transform is nonzero at both frequencies.
Quick check
(For Schwartz , express and explain why uniformly as .)
Solution · Answer
The transform is . Inverting the difference and taking absolute values gives the bound , which tends to zero by dominated convergence. Rapid decay supplies the integrable majorant.
Quick check
(Why is it incorrect to assert that tends to the finite value of a delta function as ? How is recovered instead?)
Solution · Answer
diverges. The delta is not an ordinary function with a finite value there. For the stated Schwartz data, the inverse-transform estimate above proves uniformly; concentration of the normalized kernel also explains the same limit spatially.