Motivation
A sine mode solves the heat equation on an interval because it fits the endpoint conditions. On the whole space there are no endpoints to select those modes. Instead, we ask a different question: if an initial temperature profile is prescribed everywhere, how does each small part of that profile contribute to a later temperature? The heat kernel gives the contribution as a translated Gaussian. Adding those contributions is an integral rather than a discrete sum.
The Gaussian does two jobs at once. Its integral is one, so a constant temperature stays constant. Its width tends to zero as time decreases to zero, so averaging with it recovers a continuous initial profile.
Definitions
Fix a diffusivity and a spatial dimension . Write and . The whole-space initial value problem, also called a Cauchy problem, is
A classical solution has continuous second spatial derivatives and a continuous first time derivative for positive time, and satisfies the equation pointwise. Initial values are imposed by continuous extension to time zero. The notation means bounded uniformly continuous functions, with norm .
Definition 3.1.1
(Heat kernel and its averaging operator)
For , define
The convolution is absolutely convergent for bounded Lebesgue-measurable . We set when discussing the extension to initial time.
The Gaussian integral, applied in each coordinate, gives , and the substitution shows that its width is proportional to . Since as , the kernel is not a function-valued solution with a finite initial profile; it is an averaging family concentrating near the origin.
For a nonempty bounded open set and , the parabolic boundary of the cylinder consists of the initial base and the lateral boundary ; the final-time face is excluded.
Theorem/Proposition
Theorem 3.1.2
(A bounded whole-space solution selected by its initial profile)
For every real , is smooth for positive time, solves , and satisfies
Its extension is continuous on . It is the unique classical solution with those initial values that is bounded on every finite slab .
Theorem 3.1.3
(Weak parabolic maximum principle)
Let be nonempty, bounded and open, and let . Suppose a real function is continuous on , belongs to in , and satisfies there. Then its maximum on the closed cylinder is at most its maximum on the parabolic boundary. Consequently, two classical heat solutions on the cylinder are ordered everywhere whenever their initial and lateral boundary data are ordered.
This is a weak maximum principle: it gives a bound, rather than a classification of all cases of equality. It requires continuity on the closed cylinder, so incompatible corner data are not included.
Theorem 3.1.4
(Positive averaging, conservation, and composition)
For real bounded Lebesgue-measurable data, the operators preserve order and satisfy and for positive . For such ,
If also , then and . If is nonnegative, continuous, and not identically zero, then at every point for every positive time.
Mass conservation is an integral statement and therefore needs integrable data: the constant function one is preserved but has infinite total integral. Strict positivity concerns a nonzero nonnegative profile; signed data may cancel inside the integral.
Proof sketch or proof idea
Proof
(Kernel identity, initial limit, maximum principle and bounded uniqueness)
First differentiate the kernel itself:
Every further derivative is a polynomial in the spatial variable times a Gaussian, with powers of time in its coefficients. On a compact positive-time interval, these derivatives have integrable dominating functions. Locally in the observation point the same bounds control their translates. Since is bounded, differentiation under the integral is valid there. Thus is smooth for , and the kernel identity gives .
For the initial limit, write the convolution with . For every ,
where . Uniform continuity makes the first term small by choosing small. For that fixed choice, scaling moves the second integral into the tail of one integrable Gaussian, so it tends to zero. The estimate is independent of , proving uniform attainment of the initial profile. Together with continuity of , it also gives joint continuity as space and time approach a point of the initial base.
To prove the maximum principle, first replace by , so . If its maximum exceeded the maximum of on the parabolic boundary, it would occur at a point with interior spatial coordinate and positive time. There its spatial Hessian is negative semidefinite, hence . Its time derivative is zero at an interior time maximum, or is nonnegative as a left derivative at final time. Therefore , a contradiction. The resulting bound on , followed by , proves the assertion. Apply it to a difference of two solutions to get comparison.
Whole-space uniqueness needs one extra step because there is no finite lateral boundary. Let be the difference of two candidate solutions with identical initial data, bounded on a finite slab. For use the barrier
If on the slab, choose so large that . Then is negative on the lateral boundary for all times. It is already negative at initial time. The bounded-cylinder maximum principle gives inside that ball. For each fixed point and each , such a ball can be chosen to contain the point. Letting decrease to zero gives ; applying the same argument to proves equality. The slab bound is what makes one radius control every time on the lateral boundary.
The only new difficulty on the whole space is at infinity.
Proof X-ray
Why the boundary at infinity cannot simply be ignored
The bounded-cylinder proof finds a maximum on a compact set. A bounded function on the whole space need not attain its supremum, so that step cannot be reused verbatim. Subtracting the small barrier creates controlled values on a sufficiently distant sphere; slab boundedness is what lets one radius work for every time. The finite-cylinder theorem then applies, and the barrier is removed at each fixed point. The barrier grows only quadratically, which is why the argument covers bounded solutions only; the section on Gaussian growth replaces it by a faster-growing one.
Proof
(Positivity, conservation and composition)
The remaining operator properties come from the integral. Positivity and unit integral give order preservation and the supremum contraction. For integrable , Fubini's theorem is justified by ; integration in then proves conservation. A continuous nonnegative nonzero function is bounded below by a positive number on some ball, and the kernel is strictly positive on that ball, proving strict positivity everywhere.
Finally, complete the square in the Gaussian convolution. The exponent identity is
The normalized Gaussian integral gives . Absolute integrability of the two kernels against bounded permits reassociation by Fubini, proving the composition law. Evolving for one duration and then another is therefore the same evolution as using their sum.
Worked examples
Worked example 3.1.5
(A Gaussian profile spreads while keeping its mass)
Take with . Completing the square in the convolution gives
Indeed, the exponent before integration is
The first factor integrates to the stated amplitude. The central height decreases and the width increases, but a change of variables shows that the total integral remains .
Worked example 3.1.6
(A bounded mode with no finite total mass)
Let for a fixed nonzero vector . The function
solves the equation by direct differentiation, is bounded on every finite slab, and attains the initial data uniformly. Uniqueness therefore identifies it with , without evaluating an oscillatory Gaussian integral. The data are not integrable, so the finite-mass statement is inapplicable, while the supremum estimate remains valid.
Worked example 3.1.7
(Separating a data error from its later decay)
In one dimension compare initial profiles and , where is a positive integer. Linearity and the mode calculation give
The general contraction estimate bounds this by . The exact mode calculation says more: a high-frequency perturbation decays especially rapidly.
Counterexample mode
A discontinuity is smoothed, not recovered uniformly
In one dimension take for and zero for . Its convolution is
At this equals for every positive time, while the assigned initial value is zero. Uniform convergence is impossible in any case: a uniform limit of continuous functions is continuous. The initial-attainment theorem assumes uniformly continuous data for exactly this reason.
Common mistakes
Common mistake
(Changing the normalization of the kernel)
Multiplying the Gaussian by an arbitrary constant still solves the differential equation for positive time, but fails to preserve constant initial data. The factor in front is fixed by , not by the PDE alone.
Common mistake
(Expecting a finite propagation speed)
The Gaussian is positive at every spatial point for positive time, so any nonzero nonnegative continuous compactly supported initial profile becomes positive everywhere immediately. For instance, is at least on , giving a strictly positive contribution to the convolution at every observation point. Positivity far away can be extremely small; it is still nonzero.
Common mistake
(Reversing the contraction inequality)
The forward estimate cannot be inverted into a backward stability bound. The solutions have initial norm one and final norms , so terminal data arbitrarily close to zero come from initial data at fixed distance from zero. The map from terminal data back to initial data is not continuous in the supremum norm.
Uniqueness for solutions with Gaussian growth
The bounded-class theorem leaves a real gap. On the line, satisfies with , and by the exercise below on polynomial solutions it is also the kernel average . Yet is unbounded on every slab, so the theorem above does not say that is the only classical solution with these data. The same objection applies to every nonconstant polynomial profile.
Nor can the growth condition simply be deleted. Tychonoff (1935) constructed, for the one-dimensional heat equation, a solution on that is smooth for positive time, continuous up to , has zero initial values, and is not identically zero. The theorem below shows that no single bound can hold for such a solution on a finite slab where it is nonzero.
Theorem 3.1.8
(Uniqueness and comparison in a Gaussian-growth class)
Let , and let be a real function that is continuous on , belongs to on , and satisfies there. Suppose there are constants and with
If for every , then on the whole slab. In particular, if for every , then .
Two classical solutions with bounds and have a difference bounded by . Hence, on a finite slab, the Cauchy problem has at most one solution of at most Gaussian growth, and ordered initial data give ordered solutions within that class. When the theorem reduces to the bounded-class result.
Proof
Proof
(A growing backward Gaussian barrier)
The bounded-class proof subtracted . This barrier grows only quadratically, so it cannot dominate on large spheres when . We need a heat solution that grows faster than at infinity. The particular growing Gaussian barrier constructed below becomes singular after a finite time, so the proof first treats a short slab and then repeats the argument.
Step 1: a short slab. Assume first that . Choose so small that still holds, and write
Fix a point and a number . With , which lies in for , define
Since and ,
so . Up to a constant factor, is the Gaussian kernel formula evaluated at the negative time , where the decaying exponent becomes a growing one.
Put . Then is continuous on the closed slab, belongs to for positive time, and satisfies . For consider the cylinder . At we have . On the lateral boundary we have . Both factors of decrease in , and . Hence
The difference of the exponents is
which tends to as . So for large enough, depending on and , on the entire parabolic boundary of the cylinder. The weak parabolic maximum principle gives throughout the cylinder, in particular at its axis:
The left side does not depend on . Letting gives , and was arbitrary.
Step 2: an arbitrary slab. If , the restriction is vacuous and Step 1 already proves the theorem. If , put , so . Step 1 on gives there. If , the translated function on is again continuous on its closed slab, belongs to for positive time, solves the heat equation, satisfies the same bound , and has initial values . Step 1 applies again. After at most such steps the whole interval is covered. Applying the result to proves the statement for zero initial data.
The time-stepping is forced by the barrier, not by the solution.
Proof X-ray
Why the proof advances in time steps
The barrier must exceed on distant spheres throughout the time interval and must still solve the heat equation. The particular positive Gaussian solution with initial profile has a coefficient and becomes singular at . The growth constant is fixed on the whole slab, so the step length is the same at every stage, and finitely many stages suffice.
Worked example 3.1.9
(Polynomial data are now covered)
Return to on the line, with . Because for , we have , and on
Thus lies in the Gaussian-growth class with and . Any classical solution on with initial values and some Gaussian bound must equal . Comparison gives more. If another solution in this class has initial data with , then the constants are solutions in the class. Applying comparison to and to gives on the slab.
Worked example 3.1.10
(Gaussian growth can end in finite time)
Let , , , and . For define
With we have , so . This is a constant multiple of the barrier with , , and replaced by , so . Also . For put . On ,
so is the unique Gaussian-growth solution on each such slab.
At the center, however, as . Consequently, no classical solution with a Gaussian bound exists on a slab with . If were one, the theorem applied on for each would give there. Then as , which contradicts the continuity of at . Uniqueness holds on every finite slab; for this initial profile, existence fails at and beyond .
Continuity at the initial time is part of the hypothesis, not a formality.
Counterexample mode
Pointwise zero initial limits are not enough
On let
This is a constant multiple of . Spatial derivatives of a solution with constant coefficients are again solutions, so . For each fixed , as : trivially when , and because the Gaussian factor beats every power of when . Since , we get , so is bounded on every slab with . Along the curve , however, . Setting thus gives a nonzero solution whose every pointwise initial limit is zero but which is not continuous at . The theorem is not contradicted: its hypothesis is continuity on the closed slab, together with a bound valid down to .
Practice with solutions
Quick check
(On the line, find a polynomial solution with initial values , and prove that it is the only classical solution on a finite slab with at most Gaussian growth.)
Look for a correction that is linear in time; for the growth bound use that a polynomial times is bounded.
Solution · Answer
Try . Then and , so and , with . This agrees with the second-moment computation , because the odd moments of vanish. The function is continuous and tends to zero at infinity, hence is bounded by some constant . So on . If is another solution with these data and a Gaussian bound, the theorem applied to gives .
Quick check
(Let and be classical solutions on a slab, each with its own Gaussian bound, and suppose their continuous initial data satisfy . Prove . Deduce that every Gaussian-growth solution with initial values is at least .)
Combine the two growth constants before applying the theorem.
Solution · Answer
Let and . Then satisfies and . The theorem gives , so . For the deduction, the unique solution with data is by the worked example on polynomial data, and its minimum over at time is , attained at . The lower bound is attained, so it cannot be improved.
Quick check
(In dimensions, with , find the solution with initial values and determine how long a Gaussian-growth solution can exist. Does the lifespan depend on the dimension?)
A product of one-variable heat solutions, one for each coordinate, solves the -dimensional equation.
Solution · Answer
If solves the one-dimensional equation for each , then satisfies . Taking every from the worked example on finite-time growth gives
with . This is a constant multiple of the barrier with and replaced by . On every slab with it has a Gaussian bound, so it is the unique such solution. The argument of that worked example, applied at , shows that no Gaussian-growth solution reaches . The lifespan does not depend on ; only the blow-up rate does.
Summary
The heat kernel combines a differential identity, unit normalization and concentration near the origin; together they give a smooth solution that attains bounded uniformly continuous data uniformly. The maximum principle supplies comparison and bounded-class uniqueness, and a growing barrier extends both to solutions of at most Gaussian growth, within which data such as have solutions only until . Smoothing and forward stability do not imply finite propagation or stable backward recovery.
Exercises
Quick check
(In one dimension, prove that the heat kernel has second moment by integrating its spatial derivative by parts.)
The Gaussian decay removes the boundary term at infinity.
Solution
Since ,
The boundary term vanishes at both infinities. In dimension , the product structure gives by adding the coordinate moments.
Quick check
(Suppose is bounded and globally Lipschitz with constant . Prove the uniform initial error bound .)
Use the second moment and Cauchy–Schwarz with the Gaussian as a unit-mass weight.
Solution · Answer
The Lipschitz inequality gives . Cauchy–Schwarz and unit mass bound the last integral by . The bound is independent of .
Quick check
(On a bounded cylinder, let solve the heat equation and let satisfy . If on the parabolic boundary, prove . Assume the classical regularity required by the maximum principle.)
Subtract a time-dependent comparison barrier from .
Solution
Put . Then , and this function is nonpositive on the parabolic boundary. The maximum principle gives . Apply the same reasoning to to obtain the lower bound.
Quick check
(For the interval solution on , compute its total integral and boundary flux. Explain why this does not contradict whole-space conservation.)
Integrate the equation in space, retaining the endpoint derivatives.
Solution
The mass is . Integrating the PDE gives
which agrees with differentiating . Zero endpoint temperature does not mean zero endpoint flux: the spatial derivatives are nonzero.