Why the boundary changes the image
A one-dimensional gray-level profile may describe a row through an image. Heat diffusion suppresses rapid changes in that profile, but the equation in the interior does not say what happens at the ends. An insulated end blocks flux; an end held at zero is in contact with an ideal dark reservoir. The first choice retains the average brightness. The second can exchange intensity with its surroundings. Here we make those statements for functions on an interval, rather than for a finite pixel vector.
Definition 9.2.1
(Neumann and Dirichlet heat problems on an interval)
Fix a length and diffusivity . We seek for , , satisfying
Neumann means ; homogeneous Dirichlet means .
The zero condition is an idealized boundary value, not a claim about the edge of every actual image. The initial datum below is real and belongs to ; two functions equal almost everywhere represent the same datum. In particular, its values at the two endpoints are not part of the datum.
The prerequisite used here is Fourier completeness and Parseval's identity on an interval: the normalized cosines, including the constant, and separately the normalized sines form orthonormal bases of . We use these established results, rather than proving Fourier completeness. The heat construction, convergence, boundary laws and uniqueness are proved below. No general-domain or nonlinear existence theorem is being assumed.
How the two sets of modes arise
Try a separated solution . A nonzero separated solution leads to and . For Neumann conditions, solving this elementary eigenvalue problem gives the constant eigenfunction for and for , . There is no negative eigenvalue: integrating against gives because the boundary term vanishes. For Dirichlet conditions the eigenfunctions are , , with the same positive eigenvalues; the constant is excluded. The boundary condition has therefore changed the allowed frequencies and removed the zero mode.
Definition 9.2.2
(Mean, cosine, and sine coefficients)
For and , the coefficients used below are
Orthogonality explains these factors. For example, for , while the squared norm of the constant is . The unequal factors in the formulas are essential. The prerequisite quoted in the introduction is used in the following form.
Theorem 9.2.3
(Fourier completeness and Parseval's identity on an interval)
The functions and , , form an orthonormal basis of , and so do the functions , . In terms of the coefficients above, every satisfies
This result is quoted rather than proved here; everything after it is derived on the page.
The two explicit solutions
Theorem 9.2.4
(Interval heat flow from square-integrable data)
For define, with ,
For positive time, both functions are smooth in and up to the spatial endpoints and satisfy the heat equation and their stated boundary condition. As , they converge to in . They are unique among solutions that are up to the spatial endpoints for positive time and satisfy in as .
For positive time let , and . Both models obey
For Neumann, and
For Dirichlet, and
Thus the Neumann profile converges to its initial mean, while the Dirichlet profile tends to zero. The mass flux in the latter formula may have either sign for signed data.
Why the infinite calculation is valid
Proof
(Smoothing, initial reconstruction, and uniqueness)
Parseval makes both coefficient sequences square summable. Fix and nonnegative integers . Differentiating a mode times in space and times in time introduces at most a constant times . Cauchy–Schwarz bounds the sum of absolute values of the resulting terms, uniformly for and , by
where denotes or . The Gaussian factor defeats every fixed power. Uniform convergence of each differentiated series on such a strip justifies termwise differentiation, the PDE and the endpoint conditions. This is positive-time smoothing; it does not mean that an arbitrary datum had classical derivatives at time zero.
The completeness prerequisite identifies with the limit of its sine or cosine partial sums. Parseval then gives
Each summand tends to zero and is bounded by its summable coefficient square, so dominated convergence for series proves initial reconstruction in . This is stronger than merely writing a formal separated series, but it is not a claim of uniform convergence at the endpoints.
For uniqueness, take another solution in the stated class and project it onto each corresponding normalized eigenfunction. At positive times integration by parts twice moves the second derivative onto the eigenfunction; the boundary terms vanish because both the solution and eigenfunction obey the same boundary condition. Each coefficient solves , including for Neumann. Continuity in at zero fixes its initial value. Completeness then identifies the other solution with the series in at each time; their continuity in space upgrades equality almost everywhere to equality everywhere.
Proof
(Energy, mass, and decay laws)
For , integration by parts gives . The product vanishes because for Neumann, or for Dirichlet. For , the corresponding boundary term is : under Dirichlet, differentiating the fixed boundary value in time gives ; under Neumann, . Either identity can also be checked by differentiating the exponentially convergent coefficient sums. Mass has a different boundary term, , which is zero for Neumann but need not vanish for Dirichlet.
The cosine expansion of and sine expansion of contain only eigenvalues at least . Parseval gives the displayed decay bounds term by term. The mean is conserved because all nonconstant cosine modes integrate to zero, or equivalently because the no-flux boundary makes .
Proof X-ray
Where positive time and the boundary condition enter
The construction never differentiates . Square integrability enters only through Parseval, which makes the coefficients square summable; Cauchy–Schwarz pairs that with the second factor in the displayed bound. The key move is to work on a strip with , where the Gaussian factor beats every power . At that factor is gone and the second sum diverges, so no uniform estimate is available at time zero. This is why the initial value is recovered by a different argument: in the error series each summand is bounded by its coefficient square, which is a summable majorant independent of .
Uniqueness uses the boundary condition exactly once. Against an eigenfunction , two integrations by parts leave , which vanishes because and obey the same condition. The regularity class up to the endpoints is there to make that integration by parts legitimate. A competitor obeying a different boundary condition would keep these terms, and its coefficients would no longer solve the homogeneous equation .
These are exact continuum statements, not a proof that any particular discretization converges to them.
Concept lensAnalytic
The same solutions by reflection
Extend to the whole line: evenly across and for Neumann, oddly for Dirichlet, so the extension is -periodic. Heat flow on the line with the Gaussian preserves these symmetries, which encode the boundary conditions. Folding the line back onto gives, for ,
with for Neumann and for Dirichlet. Each cosine or sine mode is mapped to its exponentially damped multiple, so this kernel reproduces the series above. Its Neumann form is a sum of positive Gaussians, so nonnegative data stay nonnegative, a fact the series does not show directly.
Initial values and endpoint compatibility
The formulation deliberately permits a profile whose endpoint behavior conflicts with the boundary imposed for . If is continuous and , a homogeneous Dirichlet solution cannot converge uniformly to on : every positive-time solution has value zero at . Similarly, suppose and are jointly continuous on the closed space-time strip, , and is differentiable at the endpoints. Then (by the fundamental theorem of calculus), so the Neumann condition requires . These are necessary conditions for those stronger notions of attainment, not extra hypotheses for the theorem. Demanding higher derivatives continuous at the space-time corners introduces further compatibility requirements; the theorem makes no such assertion.
For a useful sufficient check of uniform reconstruction of the profile itself, suppose and, in the Dirichlet case, . No endpoint derivative condition is needed for the Neumann case. Integrating the appropriate coefficient twice by parts gives or ; the latter bound includes the finite boundary terms involving and . The Fourier series then converges absolutely and uniformly to a continuous function. Fourier completeness says this function equals almost everywhere, hence everywhere; dominated convergence of the absolutely summable coefficients gives uniform convergence of to . This result concerns , not automatic convergence of all its derivatives.
Worked examples
Worked example 9.2.5
(An insulated ramp tends to its mean)
Take . Integration by parts yields and
Consequently,
The initial mean and limiting profile are . Although conflicts with a continuously attained Neumann derivative, the theorem still provides an initial solution. The preceding criterion even gives uniform convergence of to the ramp. At every positive time its endpoint derivative is zero, however, so these derivatives do not converge to the initial endpoint derivatives. There is no contradiction: convergence of functions in does not imply convergence of their endpoint derivatives.
Worked example 9.2.6
(A zero-boundary arch loses its brightness)
For , the sine coefficients are zero for even and for odd . Thus
The initial mass is . Since is smooth and satisfies the endpoint values, the preceding argument also gives uniform initial convergence. The Dirichlet estimate shows that the entire profile tends to zero in . Formally evaluating the initial boundary flux gives ; for this polynomial, the positive-time flux tends to that value because its first differentiated sine series is absolutely summable. It illustrates outward loss for this nonnegative profile, not a general monotonicity claim for signed data.
Worked example 9.2.7
(A single mode reveals the rate)
For Neumann data , the solution is . Its mass is , while . The conserved constant part prevents from decaying to zero. If the initial data were the Dirichlet mode instead, there would be no constant component and the whole profile would decay at the same amplitude rate.
A counterexample and modeling cautions
Counterexample mode
L2 convergence at time zero does not give uniform convergence
A tempting claim is that if the solution converges to a continuous profile in , it converges uniformly. Take Dirichlet data represented by the constant function . Its sine coefficients are for odd and zero for even , and the theorem gives in . But for every , so for every positive time. The repair is the sufficient condition of the previous section: with in the Dirichlet case.
Conversely, the same constant is a stationary Neumann solution. Thus neither an interior PDE nor the word “smoothing” determines the boundary model or its long-time limit.
Common mistake
(Using one normalization for the constant and the cosines)
The constant mode has squared norm , each cosine has squared norm . Applying the cosine formula with to the Neumann datum gives , and taking that as the constant term produces , with mass instead of . The constant term is the mean , equivalently half of that value.
Common mistake
(Reading mass conservation or positivity from dissipation)
The two energy laws have different meanings: measures squared intensity, whereas measures squared spatial slope. Both decrease here, but that does not imply that total intensity is conserved. Nor does a small error alone control an individual endpoint value. Continuum pixels may also be negative mathematically; claims about positivity require their own proof and are not deduced merely from energy dissipation.
Exercises
Exercise 9.2.8
(Let in the Neumann problem. Find , , and , and verify both dissipation identities.)
Foundation
Solution(3)
- 1
Solve mode by mode
Put . Orthogonality gives and .
- 2
Compute the two quadratic quantities
and .
- 3
Differentiate and compare
, and .
Exercise 9.2.9
(Let in the Dirichlet problem. Compute , its squared norm and mass. Which mode controls the long-time decay?)
Foundation
Solution(3)
- 1
Attach the decay factors
With , .
- 2
Use orthogonality and integrate
. Since and , .
- 3
Identify the slowest mode
The first mode, with eigenvalue , controls the nonzero leading long-time term.
Exercise 9.2.10
(Show directly from the cosine coefficients that an Neumann solution preserves mass. Explain why the same argument cannot be applied to the sine modes.)
Core
Solution(2)
- 1
Integrate the cosine series termwise
Each nonconstant cosine integrates to zero, so .
Because: The positive-time series converges uniformly.
- 2
Compare with the sine modes
Odd sine modes integrate to , so the sine integral need not vanish and its coefficients decay at different rates. Boundary flux, rather than a constant mode, determines mass change.
Exercise 9.2.11
(Consider Dirichlet data . Derive all and explain exactly why convergence at zero does not imply uniform convergence on . Can a solution continuous on the closed space-time strip including have both this initial profile and zero boundary values for positive times?)
Core
Solution(3)
- 1
Compute the coefficients
Direct integration gives .
- 2
Separate the two notions of convergence
Key step
convergence ignores the two endpoint values and allows mismatch near them on an interval whose contribution to squared error tends to zero. Since for while , the uniform error is at least one.
- 3
Test continuity at the corner
Continuity at would demand both limiting values zero and one, a contradiction. No such solution exists.