Motivation
An initial temperature profile is introduced once and then diffuses. A heat source acts repeatedly: heat deposited late has had less time to spread than heat deposited early. The contribution from each source time must therefore evolve for its own remaining duration. Duhamel's formula adds these differently aged contributions.
This interpretation suggests a formula, but does not justify every derivative of it. The heat kernel becomes concentrated as elapsed time tends to zero. A second spatial derivative of that kernel has a nonintegrable-looking time bound near the most recent source time. We will use sufficient spatial regularity of the source to move those derivatives onto the data instead. The endpoint of the time integral then contributes the current source itself.
The resulting construction proves a classical existence theorem, comparison, stability and a precise mass-input law. It also explains a limitation: initial heat is smoothed for a positive duration, while heat being supplied now may retain the source's limited spatial regularity. We do not assume that every forced solution becomes infinitely differentiable merely because the homogeneous equation does.
Definitions
Fix , a dimension and a finite horizon . All functions are real-valued. Consider
Write for bounded uniformly continuous functions on the whole space. The space consists of functions whose derivatives through order two are bounded and uniformly continuous, with norm given by the sum of their supremum norms. Assume and . Thus the source and its spatial derivatives through order two vary continuously in supremum norm on the closed time interval. This is a sufficient hypothesis, not an optimal source class.
Definition 3.5.1
(Heat evolution and the accumulated source)
For positive elapsed time, let
Define
The time integrals may be understood as uniform limits of Riemann sums in the indicated function norm. The diagonal value at zero elapsed time is the source slice, through the definition of ; it is not a finite pointwise value of the Gaussian kernel.
The Gaussian has unit integral and is positive. Consequently the operators preserve order and contract the supremum norm. Completing the square gives , and the approximate-identity estimate gives in supremum norm for . These basic homogeneous properties will control the source integral as well.
Theorem/Proposition
Theorem 3.5.2
(A sufficient classical Duhamel construction)
Under the stated assumptions, is continuous up to initial time, bounded on the finite slab, and belongs to for positive time. It solves the forced equation and tends uniformly to initially. The source contribution has spatial derivatives through order two continuous up to time zero, and
In particular its initial time derivative is the initial source. If also , the full solution has continuous spatial derivatives through order two and a continuous first time derivative up to initial time, with .
Theorem 3.5.3
(Uniqueness, comparison and accumulated error)
The construction is unique among solutions in that are continuous on and bounded there. For data/source pairs and with solutions ,
Ordered initial data and sources give ordered solutions. The source error is accumulated over its duration; a bound on its instantaneous size alone is not a time-independent error bound.
Theorem 3.5.4
(Restarting the evolution and accounting for total heat)
For with ,
The source keeps its original time argument. If in addition and
then the solution is integrable at every time, with
Total heat is therefore absolutely continuous in time, with derivative equal to the source integral almost everywhere. This finite-mass statement needs the extra integrability assumptions.
Proof sketch or proof idea
Proof
(Transferring derivatives to the source)
For a function in , spatial differentiation of a Gaussian convolution can be transferred to that function, through difference quotients and bounded derivatives. Thus derivatives through order two of the source potential are
The multi-index counts spatial derivatives. Supremum contraction gives integrable bounds uniform in source time. Each continuous source/derivative path has compact range in its function space. Strong continuity of the heat operators is uniform on those compact ranges: use a finite cover and contraction to reduce to finitely many fixed functions. Thus the evolved slices depend jointly continuously in supremum norm on observation time and source time, including the diagonal. This justifies the Riemann integrals and their continuity, including their zero initial values. It avoids estimating a second derivative of the kernel by a constant divided by elapsed time and then integrating that divergent bound.
For a time-independent , the kernel equation and transferred derivatives give the identity
One may first integrate the kernel equation away from zero and then use the approximate-identity limit at zero. Apply this identity to the difference of two source potentials at nearby times. The contribution over the old integration interval, divided by the time increment, tends to the integral of the evolved source Laplacian. Contraction and the uniformly bounded source derivatives justify passing the limit through the integral.
There is also a new short interval at its upper end. Its average converges in supremum norm to the current source, by time continuity and the strong initial limit of the heat evolution. The forward difference quotients converge uniformly over allowed base times as their increments decrease: source continuity is uniform on the closed time interval, and the strong kernel limits are uniform on the compact source/derivative ranges. The backward quotient at a positive time is a forward quotient based at . Uniform convergence and continuity of the candidate derivative therefore give the same limit from that side. This proves the displayed time derivative and its continuity. Combining it with the spatial formula proves the PDE. The homogeneous term is smooth for positive time by integrable Gaussian derivatives, and
Together with the homogeneous initial limit this proves uniform attainment. When the initial data also have the two bounded uniformly continuous spatial derivatives, the same identities extend the full equation to time zero.
Two ideas carry this proof.
Proof X-ray
The most recent source time supplies the PDE's right side
The construction rests on two choices. The first is where the derivatives fall: putting on rather than on replaces the kernel bound , which is not integrable at , by the bounded quantity . This is the only use of , and it is why that hypothesis is sufficient rather than optimal. The second is the split of the time increment of . Over the old interval the identity turns the increment into ; over the new interval the average of tends to because strongly. Every earlier slice solves the homogeneous equation, so the right side of the PDE comes entirely from the moving upper limit. Uniformity of both limits over base times, obtained from compactness of the source ranges, is what upgrades a one-sided limit to a derivative. Dropping the endpoint term would turn the accumulated input into a homogeneous solution with zero data, that is, into zero.
Proof
(Positivity, contraction and a growing barrier)
Positivity and contraction immediately prove comparison and the difference estimate for the constructions. To identify them with every solution in the bounded class, subtract two candidates with the same source and data. Their difference solves the homogeneous equation with zero initial values. Use the growing barrier
For any positive barrier coefficient, a sufficiently distant sphere makes the bounded difference minus the barrier negative on its lateral boundary. It is negative at initial time as well. On that bounded cylinder a positive maximum at positive time would have nonnegative time derivative and nonpositive Laplacian, contradicting the strictly negative heat operator. At final time use the left time derivative. Letting the barrier coefficient decrease to zero at each fixed point, then repeating for the negative difference, proves uniqueness. No uniqueness claim is made for unrestricted spatial growth.
Proof
(Splitting the source integral and counting mass)
For the restart identity, split the source integral at the intermediate time and apply the Gaussian composition law to the older contributions. Absolute bounds from contraction permit interchanging the integrations. The initial term and all old source terms become the homogeneous evolution of the current state; the remaining interval supplies the new heat. This is a restart with the correct time-dependent input, not a composition law obtained by resetting that input's clock.
The source is jointly continuous, so the space-time integrands are measurable. Fatou's lemma at converging times shows that the extended spatial norm is lower semicontinuous and hence measurable. Under the extra mass assumptions, Tonelli applied to absolute values gives an integrable bound for the full spatial/source-time convolution. Fubini then justifies the signed iterated integrals and their measurable time integrands. Unit Gaussian mass proves the stated norm and signed-mass formulas. The source integral is time-integrable, so its indefinite integral is absolutely continuous. The derivative assertion is consequently almost everywhere; continuity of the source in supremum norm alone does not assert continuity of its whole-space integral.
The restart identity has a structural reading.
Concept lensStructural
Different slices of the source have different diffusion ages
The same observation time contains old contributions that have spread widely and recent contributions that remain close to their supplied profile. Restarting preserves the current temperature but must also preserve the absolute schedule of future input. This age structure explains both the accumulated error bound and why ongoing rough forcing can limit spatial smoothing.
Worked examples
Worked example 3.5.5
(A persistent spatial mode approaches a forced balance)
Take zero initial data and a time-independent source with . Put . Then
The scalar amplitude satisfies with zero initial value. Unlike an unforced mode, it does not decay to zero: it approaches the stationary response with amplitude equal to the reciprocal decay rate. The source continually replenishes what diffusion removes. These bounded oscillatory data are not absolutely integrable on the whole space.
Worked example 3.5.6
(A source decaying at the mode's own rate)
Keep the same spatial mode, but multiply its source by . Each source-time contribution has the same product of decay factors, so
Directly, . For a source with positive decay rate , the amplitude is the exponential-response quotient, with limit
The extra time factor comes from the two decay rates meeting. It does not cause unbounded growth: exponential decay ultimately wins, unlike sustained resonance in an undamped wave oscillator.
Worked example 3.5.7
(A localized continuous heater can have growing total heat)
Take zero initial data and the constant-in-time source , with . Gaussian convolution gives
At the center, in dimensions one and three respectively,
The one-dimensional center grows without bound, while the three-dimensional center has a finite limit. In either dimension total heat is . A bounded limiting local temperature does not imply that total heat has stopped increasing. No steady-state existence theorem is being inferred solely from one observed limit.
Counterexample mode
A zero spatial frequency can accumulate indefinitely
The constant source one with zero initial data gives . In one dimension a stationary response would satisfy , hence be a quadratic plus an affine function, never bounded on the whole line. The nonzero-frequency response formula cannot be used at zero frequency by retaining a reciprocal zero decay rate. Slab boundedness permits growth over longer time horizons.
Bounded forcing can also limit regularity, not only size.
Counterexample mode
Ongoing forcing need not allow unlimited smoothing
In one dimension choose a smooth compact cutoff equal to one near zero and let be that cutoff times . This source belongs to but is not near zero. With zero initial data and constant-in-time source, the solution satisfies
The first term is smooth. If the solution were in space near zero, the identity would make the source there, a contradiction. The constructed solution is classical to the order proved, but need not be spatially smooth to all orders. The time-independent source obstructs higher regularity at every positive time through this identity.
Common mistakes
Common mistake
(Substituting the kernel at zero elapsed time)
The Gaussian at zero elapsed time is not a finite function to be substituted into a source endpoint term. Its operator limit is the identity, as , and that is how the current source enters the time derivative.
Common mistake
(Time continuity of the source does not justify spatial derivatives)
Continuity of in supremum norm does not justify differentiating twice in space. Differentiating the kernel instead gives a bound of order , which is not integrable. The proof above uses two controlled spatial derivatives of explicitly.
Common mistake
(Resetting the source clock at a restart)
The forced evolution is not a homogeneous semigroup on the temperature alone with a freely reset time-dependent source. Retain the original source schedule at restart; in elapsed-time notation use the shifted source at elapsed time .
Finally, distinguish finite-slab bounds, uniform-in-time bounds and finite total mass. Each has different hypotheses, and a localized heater can keep adding mass even when a local temperature stabilizes.
Summary
Duhamel's formula evolves every source slice for its remaining diffusion time. Controlled spatial derivatives justify a classical construction, while the changing upper limit supplies the current source in the PDE. Positivity gives comparison and accumulated stability, and integrability gives the exact heat-input law. Restarting retains absolute source time, and continuing input can prevent the unlimited smoothing familiar from homogeneous heat flow.
Exercises
Quick check
(For zero initial data and source , with and , find the scalar mode response and separate its transient from its periodic part.)
Solve the first-order amplitude equation with zero initial value.
Solution
With , the solution is , where
Substitution gives and . The exponential term is the transient; the other terms give a periodic response with a phase shift. Its limiting amplitude is .
Quick check
(Integrable initial data have total heat , and the source is with , where is smooth and integrable with total integral and has the stated bounded-derivative regularity. Find the total heat at time .)
The mass formula does not require evaluating the full spatial convolution.
Solution
The total heat is . It tends to . This follows from the signed source integral and also applies when the source changes sign in space. The mass limit alone does not determine a limiting temperature profile.
Quick check
(For zero initial data and the spatially constant source , compare the correct restart at time with incorrectly resetting the source to elapsed time zero.)
The correct new source is evaluated at its original absolute time.
Solution
The solution is . The correct restart gives
Resetting the source to elapsed time would instead give , missing the cross term . The homogeneous operator preserves spatial constants; the discrepancy comes entirely from changing the future input.
Quick check
(An approximation is bounded and continuous on the closed slab and classical for positive time. Its initial error is at most and its residual has absolute value at most . Prove that for the constructed solution .)
Use both signs of the difference with a time-linear bound and a small growing spatial barrier.
Solution
Let . Its heat operator is bounded above by the residual bound. Subtract the proposed initial/time bound and a positive multiple of the growing barrier:
Its heat operator is strictly negative. It is strictly negative initially and, on a sufficiently distant finite sphere, negative throughout the slab by boundedness. The maximum argument gives . Let the barrier coefficient decrease to zero at a fixed point and repeat with the negative difference. Thus . Here the exact solution is constructed under this note's data/source assumptions. The comparison proof needs the approximation's stated classical continuity and boundedness, but does not need its residual to satisfy the stronger source-derivative assumptions used for that construction.