Motivation
The whole-line heat kernel accounts for an initial temperature at every spatial point. A rod occupying only the positive half-line has an endpoint, and the equation alone does not specify what happens there. Holding the endpoint at zero temperature removes heat; insulating it prevents heat from passing through. These two experiments require different kernels, even though the interior differential equation is identical.
Reflection turns the endpoint condition into a symmetry of a whole-line problem. An odd reflected profile enforces zero temperature; an even reflected profile enforces zero spatial derivative. A moving endpoint temperature needs an additional construction. Its contribution is an integral over earlier boundary times, with a kernel different from the initial-data Gaussian.
We derive these formulas and check their limits, rather than treating substitution into the interior equation as a complete solution. In particular, initial and boundary temperatures meet at one corner. If their prescribed values disagree there, no jointly continuous solution can attain both.
Definitions
Fix . On , , consider
Dirichlet data prescribe . Homogeneous Neumann data prescribe , an insulated endpoint. Here denotes real bounded uniformly continuous functions. For Dirichlet problems our solution class is continuous on the closed quadrant, classical in the open quadrant, and bounded on every finite time slab. For Neumann uniqueness we additionally require to extend continuously to at positive times. No derivative continuity at the initial corner is implicit.
Definition 3.2.1
(Reflected initial-data kernels)
Write
These integrals are absolutely convergent for bounded measurable data. The subscripts distinguish endpoint conditions, not two different interior equations.
We use the Gaussian tail functions
The Gaussian integral gives as . These functions will measure how the initial interior and the endpoint share influence on a later observation.
Theorem/Proposition
Theorem 3.2.2
(Reflection constructs the homogeneous boundary solutions)
For , is smooth up to the endpoint for positive time, solves the heat equation, satisfies , and tends uniformly to as . If also , the same conclusions hold for , with replacing the derivative condition. Both satisfy the supremum bound by .
For , both kernels are strictly positive. Their integrals in the initial location are
Theorem 3.2.3
(A continuous endpoint temperature has its own kernel)
Let , , and assume the corner compatibility . For , define
Then is smooth in the open quadrant and extends continuously to the closed quadrant with initial data and endpoint data . In fact, uniformly. It is the unique solution in the stated Dirichlet class. On each finite slab,
Theorem 3.2.4
(Bounded-class uniqueness, order and stability)
Dirichlet solutions in the stated class are unique for their data. For two compatible data pairs, their constructed solutions satisfy
Ordered initial and endpoint data give ordered solutions. Homogeneous Neumann solutions are unique in the stated Neumann class and preserve order, with supremum contraction in the initial data.
Proof sketch or proof idea
Proof
(Reflection, smoothing and kernel masses)
Extend evenly to the line to obtain the Neumann construction. If , extend it oddly for Dirichlet. These extensions are bounded and uniformly continuous: near zero, the odd case uses the zero endpoint value. Splitting their Gaussian convolutions into positive and negative spatial halves gives exactly and . Evenness of the convolution makes its derivative zero at the endpoint; oddness makes its value zero there.
Gaussian derivatives are integrable uniformly on compact positive-time intervals, so differentiation under the integral proves smoothness and the equation. In particular, the Neumann convolution is smooth on the whole line after even extension, so all derivatives required in its boundary solution class extend continuously to the endpoint at positive times. Uniform initial convergence follows from the whole-line approximate-identity estimate: split the Gaussian integral into a small neighborhood controlled by uniform continuity and a tail whose mass tends to zero. Restricting to the half-line preserves that estimate. This verifies the initial condition, which symmetry alone would not establish.
Because for positive , the difference defining is positive. Changes of variables in the two Gaussian integrals give the displayed masses. Thus the Dirichlet kernel has mass less than one at positive time, whereas the insulated kernel has mass one. The missing Dirichlet mass is not a normalization mistake.
The signs in the two kernels have a structural meaning.
Concept lensStructural
Reflection encodes a boundary condition
The reflected source is a mathematical device outside the physical half-line. Subtraction cancels the temperature at the endpoint; addition cancels its derivative. The signs follow from the quantity prescribed there. Neither reflected source is an extra physical initial datum that may be independently chosen.
Proof
(Boundary kernel, endpoint trace and the corner)
For boundary forcing put . Direct differentiation gives
The lower endpoint in the last integral contributes zero for fixed . The spatial derivative is with respect to the source coordinate , using the smooth kernel formula at its endpoint, not differentiating the boundary data. This identifies the boundary kernel with the correctly signed derivative of the Dirichlet kernel. It also proves its total positive-time mass is one when time is integrated to infinity.
The function solves the heat equation in . On a region with bounded away from zero, it and all its derivatives vanish faster than every power as . Differentiating the integral for therefore produces no upper-endpoint term, even when is merely continuous. The resulting derivatives are locally dominated by integrable kernels. Hence is smooth inside and solves the equation.
To prove its boundary trace, write the integral using . For a fixed positive , its total kernel mass tends to one as . For , the mass with is
On the remaining short time interval, continuity controls . This proves . For joint continuity near , restrict time to a compact neighborhood of and choose . Uniform continuity of the boundary history controls the short-time part uniformly in this neighborhood. The tail is at most , and the missing mass tends to zero uniformly since time stays bounded away from zero. Thus the same estimate allows time to vary, giving the joint boundary trace. The initial-data term tends to zero locally uniformly at that boundary for positive time.
There remains the corner, where each contribution can individually have a mismatch. Let . The mass identities give
Now has a uniformly continuous odd extension, while
Uniform initial convergence follows, including at the endpoint. Positivity of the two kernels and their complementary masses prove the supremum estimate, order and stability. The initial-data and boundary contributions form one bounded average together; bounding them separately would lose the sharper maximum estimate.
The three limits in this proof are of different kinds.
Proof X-ray
Compatibility makes two corner contributions fit together
The existence proof for the forced problem takes three separate limits, and compatibility is used in only one of them. Interior smoothness uses the fact that and all its derivatives vanish faster than any power as when stays away from zero; this is why mere continuity of suffices and no upper-endpoint term appears. The endpoint trace at uses only that as while the mass with tends to zero, so acts as an approximate identity in time. The corner is the non-routine step: there and need not converge separately, since while . The identity , that is , allows one to subtract the common value ; then has a uniformly continuous odd extension and is small for small . If , no single constant can be subtracted, and the incompatible example below shows that joint continuity really fails.
Proof
(Growing barrier and even reflection)
For uniqueness let be the difference of two Dirichlet solutions, bounded on a fixed finite slab and zero on its initial and endpoint boundaries. Set , so . For , choose large enough that on . On the bounded rectangle, a positive maximum cannot occur at positive interior space: there the time derivative is nonnegative, including its one-sided value at final time, and the spatial second derivative is nonpositive, contradicting . Thus . Let at each fixed point and repeat for .
For Neumann uniqueness, reflect the difference evenly. The zero endpoint derivative and the stipulated positive-time regularity make this extension across zero and a whole-line heat solution. It has zero initial data and is slab-bounded. The same growing-barrier argument on proves it is zero. This argument uses derivative regularity at positive boundary times, not an unstated derivative condition at time zero.
Worked examples
Worked example 3.2.5
(A localized warm profile is felt everywhere inside)
Let be nonnegative, compactly supported, not identically zero, and satisfy . With a zero-temperature endpoint, . Continuity makes positive on some interior interval; strict positivity of there proves for every . There is immediate interior influence at arbitrarily distant points. This differs from finite-speed wave propagation and does not mean the endpoint condition was omitted: still holds exactly.
Worked example 3.2.6
(An endpoint whose temperature rises linearly)
For constant initial temperature and endpoint temperature , write , where
Indeed, integrating the boundary convolution by parts gives . The explicit expression has time derivative and tends to zero initially at fixed positive , verifying the antiderivative. Its endpoint value is , and . Also . For , the negative derivative means heat enters the positive half-line from the hotter endpoint. The solution is bounded on each finite time slab although the boundary temperature is not bounded over all time.
Worked example 3.2.7
(The endpoint selects the parity of a mode)
For , initial data with zero Dirichlet boundary give . Initial data with insulated boundary give . Direct differentiation checks the equation and the appropriate boundary condition, and bounded-class uniqueness identifies these functions with the reflected-kernel constructions. Interchanging sine and cosine interchanges a zero value with a zero derivative; it does not preserve the prescribed experiment. Neither profile has finite absolute integral on the half-line.
Counterexample mode
Incompatible temperatures cannot meet continuously
Take initial temperature one and endpoint temperature zero. The interior formula is . It approaches one at each fixed positive as time decreases to zero and is zero at the endpoint for positive time. Along it equals instead. No choice at the corner makes the function jointly continuous there. The formula still describes interior smoothing with separate traces, but it is outside the compatible classical class proved above.
Even with compatible data, the corner regularity stops at the level proved.
Counterexample mode
Uniform initial convergence does not include derivatives
Take insulated initial data . The theorem gives uniform convergence of to . Yet for every positive time, while . Consequently the spatial derivative cannot extend continuously to the initial corner with value . A stronger derivative-attainment theorem would need the additional compatibility and further regularity; it is not a consequence of uniform convergence of temperatures.
Common mistakes
Common mistake
(Giving the Dirichlet image kernel unit mass)
A zero-temperature endpoint absorbs heat: for . Only the insulated kernel has unit mass. The missing mass is not a normalization error to be corrected.
Common mistake
(Zero temperature is not zero flux)
Prescribing is different from prescribing . In the coordinate used here the conductive flux toward increasing is , and the outward normal at the left endpoint points toward decreasing . The heat leaving through the endpoint per unit time is therefore , which vanishes for the insulated problem but not for the absorbing one.
Common mistake
(Convolving the endpoint temperature with the initial Gaussian)
The boundary kernel averages over elapsed times and has dimensions different from the initial Gaussian. Replacing it by misses the normal derivative and gives the wrong trace. Nor may one obtain the trace by substituting into the integrand, which gives zero; the trace was proved by concentration of .
Finally, state the uniqueness class and corner regularity. Slab boundedness controls the distant boundary in the barrier proof; continuity of the temperature at time zero does not promise continuity of its derivatives.
Summary
Odd and even reflection construct different half-line heat evolutions. A prescribed endpoint temperature contributes through a positive time kernel, whose mass complements the absorbing initial kernel. Their shared corner value makes uniform initial attainment possible. Positivity gives order and stability, while a growing barrier or even reflection supplies uniqueness in precise bounded classes.
Exercises
Quick check
(Verify the boundary-kernel derivative formula and the zero-derivative condition for the Neumann image kernel.)
Differentiate in the correct source or observation variable before setting it to zero.
Solution
Since , we obtain
The first identity inserts boundary temperatures; the second expresses insulation. Confusing the source derivative with the observation derivative obscures these separate roles.
Quick check
(For zero initial data and endpoint temperature , write the solution and prove that . What happens as at fixed positive time?)
Use positivity and monotonicity of the boundary data together with the kernel mass.
Solution
The solution is . On this time interval, , hence
At fixed positive time the displayed tail tends to zero as . The theorem supplies the endpoint trace and uniform initial attainment, since the two corner values are zero.
Quick check
(For the linearly heated endpoint with zero initial data and , compute the total half-line heat and compare its derivative with the incoming boundary flux .)
Integrate the positive boundary kernel in space first; Tonelli justifies changing the order.
Solution
A Gaussian substitution gives
Its derivative is , exactly . This is incoming heat at the left endpoint, so the total heat increases. The calculation establishes the balance from convergent integrals rather than silently discarding an infinite-end boundary term.
Quick check
(Suppose is also integrable and nonnegative. Prove the total-mass identities for homogeneous Dirichlet and Neumann evolution, and determine their limits as .)
Use kernel symmetry and the two spatial integrals; avoid assuming an unproved vanishing flux at infinity.
Solution
Tonelli and symmetry give
The first weight decreases to zero as and is bounded by one. Dominated convergence therefore gives total Dirichlet heat tending to zero. The Neumann total remains the initial mass. As , the Dirichlet weight tends to one for every , so its total tends to the initial mass as well. Differentiation at fixed positive time, justified by the integrable data and a bounded derivative weight on nearby times, further gives
If is positive on a set of positive measure, this is strictly negative: heat leaves into the zero-temperature endpoint. The mass identities hold for the kernel formulas even when ; membership in the continuous-corner homogeneous Dirichlet class additionally requires . No finite-mass assertion is being applied to the constant or oscillatory examples.