The differential expression alone does not determine a boundary-value problem. The same expression has different eigenvalues under Dirichlet, Neumann and Robin conditions. Its energy can be positive, have a zero mode or admit a negative eigenvalue. The endpoint terms explain these differences.
An expansion in eigenfunctions introduces a second issue. Orthogonality tells us how to compute coefficients in a finite approximation. It does not prove that the chosen modes can recover every target, nor does convergence of the function series justify differentiating it twice. These distinctions matter when an expansion is used to solve an equation rather than merely to draw an approximation.
Motivation
An initial-value problem supplies two pieces of data at one point. A separated boundary problem instead supplies one condition at each end. Once the left condition is imposed, a homogeneous second-order equation has only one remaining scalar freedom. The right condition either removes that freedom or is satisfied by a nonzero homogeneous solution. The latter case produces an eigenvalue and changes the solvability of a forced equation.
Integration by parts connects this elementary shooting picture to orthogonality and energy. It is essential to carry the boundary bracket through the calculation. A vanished Green bracket makes two functions compatible with the same symmetry identity. It does not imply that the individual endpoint contribution to the energy is zero. The two calculations use the same tool but retain different boundary information.
Definitions
Fix a finite interval , . Throughout, and are real, with and on the whole closed interval. Write
We consider classical solutions , allowing complex values. These regularity assumptions make the equation, endpoint traces and integrations below meaningful. Infinite intervals, a vanishing leading coefficient and singular weights are outside this regular setting.
Definition 2.2.1
(Separated real boundary conditions)
Choose real pairs and , neither equal to , and impose
Each endpoint has its own condition. An eigenfunction is a nonzero classical solution of satisfying both conditions. Its parameter is an eigenvalue.
If is zero, that endpoint is Dirichlet. At a non-Dirichlet endpoint it is convenient to use outward Robin notation:
Thus and . The different signs reflect the two outward directions. A Robin coefficient zero gives Neumann data. A Dirichlet endpoint is handled separately, not by substituting a finite value for its Robin coefficient.
Concept lensAnalytic
Eigenvalues as zeros of a shooting function
For each complex , let solve with and . This initial vector is nonzero and satisfies the left condition, and the solutions of the homogeneous equation that satisfy the left condition are exactly the multiples of . Hence
and in that case spans the eigenspace. The eigenvalue problem becomes one scalar equation in . For on with two Dirichlet ends, gives and , so , read as when . Then vanishes exactly at , . Writing the solution in this form, rather than as , avoids the spurious root met in the first exercise.
Definition 2.2.2
(Weighted inner product and energy)
Set and . Let be the set of Robin endpoints. For an admissible function define
Dirichlet endpoints contribute zero to the sum because the function value is zero. The weight belongs to the inner product; it does not multiply the derivative term in this energy.
Theorem/proposition
Theorem 2.2.3
(Green identity, reality and separated simplicity)
For two classical functions satisfying the same separated real boundary conditions,
Every eigenvalue is real. Eigenfunctions at distinct eigenvalues are weighted orthogonal, and each eigenspace has complex dimension one.
The identity states symmetry on the classical domain used here. It is not, by itself, a proof of the full Hilbert-space spectral theorem. The simplicity conclusion uses separated conditions; coupled periodic conditions have a different multiplicity behavior.
Theorem 2.2.4
(Robin energy and positivity)
Every eigenfunction satisfies
If and every Robin coefficient is nonnegative, then . Under these sign conditions, zero is excluded if at least one endpoint is Dirichlet, at least one Robin coefficient is positive, or is not identically zero. With two Neumann ends and , nonzero constants are zero eigenfunctions.
Theorem 2.2.5
(Forced solvability and resonance)
Let be continuous and real. For the same separated conditions, consider
If is not an eigenvalue, there is exactly one classical solution. If is an eigenvalue with eigenfunction , a solution exists exactly when . When it exists, all solutions are .
Theorem 2.2.6
(What finite orthogonal projection proves)
For any finite weighted orthonormal eigenfunction set , put and . For ,
is the unique best approximation in their span. Convergence of to every requires the additional assertion that the full chosen sequence is complete, meaning its span is dense in .
Here consists of measurable functions with finite weighted squared integral, identified when equal almost everywhere. Continuous targets already suffice for the classical forced problems in this note.
Theorem 2.2.7
(A fully justified Dirichlet solution expansion)
For continuous on and , the unique solution of , , is
This series converges absolutely and uniformly. No claim about twice termwise differentiation is included in this convergence statement.
Proof sketch or proof idea
The boundary bracket and the energy are different calculations
Proof
(Green's identity, reality, simplicity and the energy formula)
Two integrations by parts give Green's identity. At a Dirichlet endpoint both function values vanish. At a Robin endpoint, both flux derivatives are the same real multiple of the corresponding function value, so there. For an eigenfunction the identity with gives . Its norm is positive, proving reality. With two eigenfunctions it then gives , proving orthogonality when the eigenvalues differ.
At the left endpoint, admissible initial vectors form a one-dimensional line. A nonzero solution cannot have the zero vector there, by initial-value uniqueness. Thus two eigenfunctions at the same eigenvalue have proportional initial data and are proportional everywhere. This proves geometric simplicity without asserting completeness or enumerating the spectrum.
For energy, only one integration by parts is used:
The Robin relations give and , yielding the two positive-sign endpoint contributions in . Their coefficients may still be negative. Under the stated nonnegative assumptions, zero energy forces , hence constant. A Dirichlet value, a positive endpoint coefficient or a nonzero nonnegative potential then forces that constant to be zero. This is incompatible with an eigenfunction.
Proof X-ray
A zero boundary bracket does not remove Robin energy
The Green bracket compares two functions and cancels because they satisfy the same endpoint relation. The energy boundary term involves one function and generally survives as . Symmetry gives real eigenvalues; positivity additionally depends on the signs of these terms and of the potential. The same integration formula should not be shortened in two different ways without checking what remains.
A two-point alternative proved by initial-value methods
Proof
(Shooting from the left endpoint)
Fix . Let solve the forced equation with , and let solve the homogeneous equation with , . Regular linear initial-value theory gives these solutions on the whole finite interval. The second initial vector is nonzero and satisfies the left boundary condition. Every solution satisfying the left condition is therefore .
The right boundary equation is . If , it determines exactly one . If , is an eigenfunction and the equation instead requires . It remains to identify this last condition with orthogonality of the forcing.
Green's identity, applied before imposing the right condition on , gives
The left bracket is zero because both initial data of are zero. The real initial data and real coefficients let us take real. At the right endpoint, if , the bracket equals , and by uniqueness. If , it equals , again with nonzero factor. Thus the remaining bracket vanishes exactly when . This proves necessity and sufficiency of the compatibility condition, not merely its necessity. The homogeneous freedom gives the stated family of solutions.
Proof X-ray
Turning an endpoint condition into an integral condition
Everything before Green's identity is linear algebra on a one-parameter family: the left condition leaves the line , and the right condition is one scalar equation in whose coefficient is the shooting function. The real work is translating the resonant condition , a statement at one endpoint, into , a statement about the whole interval. Green's identity does this because has zero data at , which kills the left bracket. The right bracket is a nonzero multiple of , and uniqueness supplies the nonzero factor or .
Coefficients do not supply completeness
Proof
(Projection identity and Bessel's inequality)
For the projection theorem, is orthogonal to the finite span. Expanding for a vector in that span gives the sum of the squared residual and . This proves optimality and the identity. Increasing gives Bessel's inequality, but an unrepresented orthogonal direction can remain. For example, the even-indexed sine modes are an orthogonal eigenfunction family for the Dirichlet problem, yet every coefficient of against that family is zero.
If completeness is separately known, approximate a target by a finite linear combination, then take large enough to contain all its modes. Best approximation makes the projection error no larger than that chosen error. This proves convergence and Parseval equality. Neither statement determines isolated endpoint values or provides uniform convergence of derivatives. A finite Rayleigh quotient likewise does not, without further arguments, prove a minimax characterization of all eigenvalues.
Proving the explicit expansion without differentiating a series
Proof
(Solve classically, then identify the series)
First use forced solvability to obtain the classical Dirichlet solution . The explicit homogeneous calculation gives eigenvalues , so the assumed is nonresonant. Green's identity with gives .
Cauchy–Schwarz gives , while . Since grows quadratically, the proposed series converges absolutely and uniformly to a continuous function vanishing at both endpoints. Uniform convergence permits termwise integration, so has exactly the same sine coefficients as .
Here is the needed uniqueness fact with its convergence argument. A continuous function on vanishing at the endpoints has a continuous odd -periodic extension . If all its sine coefficients are zero, all Fourier coefficients of are zero. The averages of its Fourier partial sums are convolutions with
The finite sum defines and shows and . The averaged partial sum is the normalized convolution . For and , the kernel is bounded by . In the convolution error, uniform continuity controls for ; boundedness and the displayed estimate make the remaining integral tend to zero. Thus these averages converge uniformly to . All the averages here are zero, forcing .
Apply this to . It proves the expansion. We first solved the classical equation and then identified its series; we never assumed that two differentiations of a uniformly convergent series are legitimate. At resonance , the same reasoning, together with the compatibility theorem, gives a solution exactly when , with the th coefficient arbitrary and all other coefficients .
Worked examples
Worked example 2.2.8
(A change of variable reveals the weight)
On , consider with . Put and . Then and , so the equation becomes on . Therefore
The normalization uses . Unweighted orthogonality is genuinely different: the integral of the unnormalized first two modes against is
Worked example 2.2.9
(Real separated conditions can admit negative energy)
Let , on , and choose , , where . The function satisfies both outward Robin conditions and . Thus is an eigenvalue. Its energy is
The eigenvalue is real, as symmetry predicts. Nonnegativity fails because one Robin coefficient is negative, not because the Green identity has failed.
Worked example 2.2.10
(Resonance is a compatibility question)
For on with zero endpoint values, the forcing is not orthogonal to the eigenfunction . No solution exists. Replace the forcing by ; it is orthogonal to , and all solutions are . Direct substitution confirms the sign and the factor . An arbitrary homogeneous contribution remains because resonance removes uniqueness.
Worked example 2.2.11
(Uniform convergence of the solution does not justify two derivatives)
For with zero Dirichlet values on , direct integration gives . The expansion theorem yields
The series is absolutely and uniformly convergent. Twice formally differentiating gives a sine series for the constant forcing, but every partial sum of that forcing series is zero at both endpoints. It cannot converge uniformly on the closed interval to the constant one. The actual solution is smooth up to the endpoints; the obstruction concerns convergence of the differentiated series, not existence of the derivatives of the solution.
Common mistakes
Common mistake
(Outward signs at the two endpoints)
The outward signs at the endpoints must be written before computing Robin energy. Using the same derivative sign at both ends changes the relation between the boundary coefficients and the quadratic form. A real eigenvalue also need not be positive; the exponential example is an exact counterexample.
Counterexample mode
An orthogonal family of eigenfunctions need not expand every target
Claim: because eigenfunctions at distinct eigenvalues are orthogonal, projections onto any infinite orthonormal family of eigenfunctions converge to the target. Take the Dirichlet problem , , and keep only the even-indexed modes , . They are orthonormal eigenfunctions. For ,
because both frequencies are odd integers. Every projection of is therefore zero, and for every . The repair is the completeness hypothesis in the projection theorem: the span of the full sequence must be dense. The explicit Dirichlet expansion above uses every mode and is justified separately by Fourier uniqueness. The general regular spectral-completeness theorem, the oscillation theorem and eigenvalue asymptotics are not proved here, so that justification must not be transferred to unspecified eigenfunctions.
Common mistake
(Dividing by at resonance)
At resonance, dividing by hides the compatibility condition and the free coefficient. The correct statement is the forced-solvability theorem: a solution exists only when the resonant coefficient of the forcing is zero, and then that coefficient of the solution is arbitrary.
Common mistake
(Passing derivatives through a convergent series)
Away from resonance, a convergent function series still needs an argument before derivatives or the differential operator are passed through the sum. A target function also need not satisfy the homogeneous boundary conditions obeyed by the eigenfunctions. Endpoint mismatches can be invisible in an limit while ruling out uniform convergence on the closed interval, as the last worked example shows.
Summary
Separated real conditions cancel the Green bracket, giving reality, weighted orthogonality and simple eigenspaces. Robin endpoint terms survive in the energy and determine its sign. At an eigenvalue, forced solvability requires orthogonality to the homogeneous mode. Projection computes coefficients; completeness and convergence need further work. In the explicit Dirichlet case, classical solvability, uniform series bounds and Fourier uniqueness provide that work without unjustified termwise differentiation.
Exercises
Quick check
(For on , impose and . For , derive the equation for positive eigenvalues and explain why must be checked separately. What happens at ?)
Solution · Answer
For , energy excludes nonpositive eigenvalues because the left endpoint is Dirichlet. For , the left condition gives and the right gives . Substitution of into this parameter equation gives zero, but the corresponding sine function is identically zero. At , solve the original equation instead: , and the right condition is . It has no nonzero solution for . At , is a genuine zero eigenfunction, and its energy is .
Quick check
(Solve the mixed problem on , with and . Find the normalized eigenfunctions in the correct weighted inner product.)
Solution · Answer
The substitution gives , , . Nonnegative energy and the left Dirichlet condition exclude zero and negative eigenvalues. Thus for , and . Its weighted squared norm is .
Quick check
(For on with , find the compatibility condition. Solve for , and select the solution of mean zero.)
Solution · Answer
The homogeneous zero mode is the constant, so compatibility is . The proposed forcing satisfies it. Integrating gives and . Both derivative conditions hold. Its integral is , so the mean-zero choice is . All other solutions differ by a constant.
Quick check
(Solve on with zero Dirichlet values, derive its sine expansion and explain why its twice-differentiated series cannot converge uniformly to the forcing on the closed interval.)
Solution · Answer
Direct integration gives . The ordinary sine coefficients of are , so the solution expansion is , with uniform absolute convergence. Twice applying the negative second derivative formally recovers the sine series for . All its partial sums are zero at , whereas the forcing equals there. Uniform convergence to the forcing on is therefore impossible.
Quick check
(For periodic conditions , , exhibit a two-dimensional eigenspace of . Identify the step of the separated-simplicity proof that no longer applies.)
Solution · Answer
Both and satisfy the periodic conditions and ; they are independent, giving a two-dimensional eigenspace at eigenvalue one. Periodic conditions couple the two endpoints. They do not restrict initial vectors at the left endpoint to a single line, which was the step used to prove simplicity for separated conditions. The periodic Green bracket still cancels between the two endpoints, so this multiplicity does not contradict symmetry or reality.