A Gaussian average removes fine oscillations without changing the total mass of an integrable function. As the averaging scale shrinks, it should recover the original function. These two statements involve different limits: smoothing concerns a fixed positive time, while recovery concerns time tending to zero. Fourier analysis lets us prove both and see why they are compatible even for discontinuous initial data.
We use Lebesgue measure on the real line and the symmetric angular-frequency transform
The preceding transform notes construct its unitary extension to . We use that extension when an ordinary Fourier integral is unavailable. Functions may be complex-valued, and the spaces are over ; the averaging kernels themselves are real and nonnegative. All equalities between elements mean equality almost everywhere.
Motivation
Consider an initial temperature with a jump. At a positive time, diffusion produces a smooth profile. Asking that this smooth profile approach the initial data uniformly would be too much: a uniform limit of continuous functions is continuous. Convergence in an integral norm can still hold, because a thin transition region contributes little to that norm.
There is a second distinction. A smooth curve for every positive time need not have a finite velocity at time zero in the chosen norm. The heat equation formally says that this initial velocity is a second derivative. The precise statement will identify exactly which initial data possess that velocity, without assuming in advance that they have classical derivatives.
Definitions
Definition 2.4.1
(Translations and concentrated probability kernels)
Write . A family , indexed by , is a probability approximate identity if , , and, for every ,
Its average is , where the integral is defined. The parameter need not be the physical width; only the stated concentration property is required.
Definition 2.4.2
(The heat kernels and their operators)
For , set
and set . The term semigroup means for nonnegative , together with . It does not assert that a bounded inverse exists.
The Gaussian integral gives . Substituting shows that its mass outside tends to zero. Thus these kernels form a probability approximate identity.
Theorem/proposition
Theorem 2.4.3
(Approximation in Lp and in the uniform norm)
Let be a probability approximate identity. For and ,
If is bounded and uniformly continuous, convergence holds in the uniform norm as well. The assertion does not extend to every function.
Theorem 2.4.4
(Heat flow at positive time)
The operators are contractions on every , , and form a semigroup. For they are strongly continuous: as . For , including , the convolution has a smooth representative for which solves
For every nonnegative integer , there is a finite constant such that
with when .
Theorem 2.4.5
(The exact generator and the energy identity)
On , define as the set of for which has an limit as , and call the limit . Then
For every , belongs to for , is differentiable there as an curve, and . For ,
The spatial derivative in this formula is the classical derivative of the positive-time smooth representative and is also an function.
Proof sketch or proof idea
First prove continuity of translations
Proof
(Interval indicators, density and isometry)
For a bounded interval , the symmetric difference of and has measure at most . Therefore
Finite linear combinations of interval indicators have the same translation-continuity property by the triangle inequality. These combinations are dense in . Here is the measure-theoretic content of that claim. Truncate an function both in space and in magnitude, approximate its bounded values by finitely many levels, and obtain a simple function whose level sets have finite measure. Lebesgue regularity approximates each such set in measure by a finite union of bounded intervals. The th power of the error between indicators is exactly the measure of their symmetric difference. The resulting interval combinations approximate the original function in .
Given , choose one such combination with . Translations are isometries, so
First send to zero, then send to zero. This proves translation continuity for every function in the stated finite range. No pointwise continuity of was used.
Concentration turns translation continuity into approximation
Proof
(Jensen, Minkowski and a near–far split)
For , Jensen's inequality for the probability measure , followed by Tonelli, gives . It also ensures that the defining average is absolutely convergent for almost every . The contraction follows directly from the essential bound.
Minkowski's integral inequality gives
Fix . Translation continuity supplies for which the inner norm is below whenever . On the rest of the line it is at most . Hence the right side is bounded by
The second term tends to zero. This proves norm convergence. For bounded uniformly continuous , the same split works with and proves uniform convergence.
Proof X-ray
Why the order of choices matters
Each hypothesis on the kernel is used at a definite point. Positivity and unit mass make a probability measure, which is what Jensen's inequality and Minkowski's integral inequality need; unit mass also lets be written as before subtracting. Concentration is used only once, on the far region. The near region is controlled by translation continuity, a property of the fixed function , so must be chosen before . That property is where enters: for the jump one has for every , and in the argument fails together with the conclusion.
Gaussian composition and smoothness
Proof
(Completing the square and differentiating under the integral)
Completing the square in the convolution integral gives
so integration of the remaining Gaussian gives for . For a bounded compactly supported , Fubini therefore proves . For , density of such functions and the contraction bounds extend the identity to . For bounded , Fubini applies directly because the double integral of the absolute integrand is at most . Cases with a zero time use .
Every derivative of is a polynomial times a Gaussian. If is conjugate to , including the endpoint exponents, scaling gives
Hölder's inequality now gives the stated pointwise bound for convolution with . These derivatives really are derivatives of . For any , and nonnegative integers , the explicit Gaussian derivatives give constants and an integer such that
The right side belongs to every , including the endpoint spaces, so Hölder pairs it with . The mean value theorem gives the same type of bound for difference quotients on slightly enlarged compact intervals. Dominated convergence now justifies differentiation under the integral and continuity of all the mixed derivatives. A direct calculation gives , proving the equation and joint smoothness for .
For , strong continuity at other times follows from the semigroup and contraction: for , . Left continuity at follows from . Both estimates use the already proved convergence at zero.
The multiplier identifies the initial velocity
Proof
(Multiplier bounds and dominated convergence)
The Gaussian transform is . The convolution theorem, first on Schwartz functions and then by continuity, gives
The square root of in that convolution theorem cancels the Gaussian transform's prefactor. Forgetting either factor would incorrectly rescale the positive-time multiplier, whose small-time limit must agree with the identity by strong continuity.
For , . If , the Fourier transform of the difference quotient converges pointwise to and is bounded in absolute value by . Dominated convergence for the squared error and Plancherel give convergence in .
Conversely, suppose the difference quotients converge in to . Their transforms converge in to . Choose a sequence of times decreasing to zero and a subsequence along which that convergence holds almost everywhere. For completeness, such a subsequence exists by making the sum of its squared errors finite and applying Tonelli. At each of those almost-everywhere points, the scalar difference quotient also converges to . Thus almost everywhere. Since , the asserted domain condition is necessary. This proves the exact characterization.
For fixed , is bounded, so . Differentiating the multiplier in , using bounds on a compact positive-time interval, gives . The same reasoning with identifies the spatial derivative with the classical smooth derivative obtained above; it may first be checked on Schwartz approximations in . The differentiated convolutions converge in by the bounded multiplier and uniformly by
Their uniform limit is the classical derivative already constructed, and their limit has the stated multiplier; a subsequence identifies the two almost everywhere.
Concept lensAnalytic
Smooth at positive time does not mean differentiable at time zero
The Gaussian suppresses arbitrarily high frequencies when . At time zero, an initial velocity governed by a second spatial derivative requires the original frequency profile to survive multiplication by in . Smoothness after averaging is automatic; membership of the initial datum in is an additional condition.
Energy is lost through the spatial derivative
Proof
(Differentiate the energy, then let the start time tend to zero)
At positive times, differentiate the squared norm through its Fourier integral:
The differentiated multiplier is bounded uniformly on any compact interval away from zero, so this step is justified even when . Integrate from to . As , strong continuity gives , and monotone convergence applies to the nonnegative time integral. This proves the full energy identity. In particular, the dissipation integral is finite even when the spatial derivative norm becomes unbounded near time zero.
Dominated convergence in the multiplier also gives as : the only frequency not damped is , a set of Lebesgue measure zero. This is a statement about each fixed . There is no decay rate uniform over the unit ball of , because unit energy may be concentrated arbitrarily close to zero frequency.
Worked examples
Worked example 2.4.6
(A Gaussian widens and loses squared norm)
For , choose . The semigroup identity gives , so the width increases while the integral remains one. Direct integration gives
Consequently , exactly as the energy theorem predicts. Preserving total mass does not preserve the squared norm: spreading the same mass lowers its concentration.
Worked example 2.4.7
(A box is smoothed before it has an initial velocity)
Let . At positive time,
This is smooth and tends to one at interior points, zero outside the closed interval, and one half at the two endpoints. The endpoint values assigned to do not affect its class. The norm convergence follows separately from the approximate-identity theorem; the pointwise endpoint limit is irrelevant to that norm statement.
Its transform is , with the continuous value at zero. But is not square integrable: on a fixed positive-length part of every period, is bounded below, and the squared integrand grows like . Hence . Its heat flow has no initial velocity, although it is smooth at every positive time.
Worked example 2.4.8
(Recovering a past temperature is unstable)
Fix . The multiplier is strictly positive, so is injective on : equal later states come from equal initial states. Nevertheless, its inverse on its range is unbounded.
Choose by . Plancherel gives , while
A sequence of later temperatures tends to zero while their earlier temperatures remain a unit distance from zero. Uniqueness therefore does not provide continuous recovery. This conclusion concerns the norm and requires no claim about numerical rounding or a particular reconstruction algorithm.
Counterexample mode
An L-infinity jump prevents strong continuity
For , symmetry gives . More precisely,
For every , the essential supremum of is : the difference approaches that value from either side of zero and is no larger elsewhere. Thus never tends to zero. Altering the value at the single jump point cannot fix this, because for every the error exceeds on a set of positive measure near zero. The bounded-uniformly-continuous hypothesis in the uniform theorem has real content.
Common mistakes
Common mistake
(Expecting pointwise values from an theorem)
An approximate identity need not converge pointwise at every point, and an convergence theorem does not specify a value at a jump. Pointwise recovery at general Lebesgue points requires an additional argument. Conversely, the failure of uniform convergence at a jump does not contradict convergence in a finite norm.
Common mistake
(Strict loss is not a smaller operator norm)
Contractivity, strict loss for a particular input, and the operator norm are different assertions. Heat flow decreases the norm of every nonzero input at each positive time, because the multiplier is strictly below one almost everywhere and a nonzero Fourier profile has positive mass on a set of positive measure. Yet its operator norm is still one: inputs concentrated near zero frequency have arbitrarily small relative loss. Likewise, strong continuity at zero does not mean that tends to zero; the exercises compute this operator norm.
Common mistake
(Reading classically for all data)
The symbol is justified classically for Schwartz functions by the Fourier differentiation identity. The spectral domain theorem is more precise for general data: it defines the correct second-order operator without confusing an almost-everywhere classical derivative with a weak or norm-defined derivative. The full identification of this domain using two weak derivatives in a Sobolev space is deferred.
Summary
Translation continuity and concentration prove approximation in every finite norm. Gaussian composition makes the averages a semigroup, and Gaussian derivatives give immediate smoothness. Plancherel identifies both the exact initial-velocity condition and the positive-time energy dissipation. These facts distinguish recovery in norm, smoothing, and differentiability at the initial time.
Exercises
Quick check
(For , let . Prove that it is a probability approximate identity and deduce in every finite space containing . Give a bound on its mass outside .)
Solution · Answer
Substitution gives . For , , so its tail mass is at most and tends to zero. Positivity completes the approximate-identity hypotheses. The theorem now gives convergence for each and each . This argument needs neither the Fourier transform of nor a pointwise inversion theorem.
Quick check
(Prove the sharp bound for . Explain how to approximate the sharp constant and why no nonzero function attains equality.)
Solution · Answer
The derivative multiplier has absolute value for . Its maximum occurs at and equals . Plancherel proves the bound. Choose unit Fourier profiles supported in shrinking intervals around that maximizing frequency; continuity of the multiplier makes their norm ratios tend to the maximum. Equality for a nonzero input would require its Fourier mass to be supported where the multiplier equals its maximum, namely the two points . An function supported on a null set is zero, so the sharp constant is not attained by a nonzero input.
Quick check
(For , define by . Determine exactly when , including the threshold case.)
Solution · Answer
The initial condition holds because the squared profile behaves like at infinity. The generator condition is . Near zero the integrand is bounded; at infinity it behaves like . The integral is finite exactly when , or . At the tail is comparable to and diverges. Thus positive-time smoothing holds throughout , while an initial velocity requires the stricter threshold.
Quick check
(Prove that the bounded operator has operator norm one for every . Explain why this does not contradict strong continuity.)
Solution · Answer
The multiplier has absolute value at most one, giving the upper bound. For the unit vectors with , the reverse bound is . Hence the operator norm is one. Strong continuity fixes first and then sends to zero; operator-norm convergence would require one estimate uniform over all unit vectors, including profiles moving to arbitrarily high frequencies.
Quick check
(For , define and, for , define its Fourier transform by)
Show that , that for , and that as an identity. Also find its right derivative at zero.
Solution · Answer
Since , Plancherel gives the norm bound and continuity at zero. Also is bounded, so and . The scalar derivative is bounded by one. Dominated convergence for the squared difference-quotient error therefore proves . This equals the transform of . Finally, pointwise and is bounded by one, so in . The right derivative at zero is . There is also a direct time-integral interpretation: . This integral is the limit of Riemann sums of the continuous curve . The bounded Fourier operator commutes with those limits, and integrating gives exactly . The value records the removable scalar limit; changing that single frequency would not change the element.