Motivation
A vibrating string and a cooling rod have infinitely many degrees of freedom: specifying the state means giving a function, not just a finite vector. Yet their linear equations can often be decomposed into independent scalar equations. The useful coordinates are sines and cosines, whose derivatives reproduce the same functions up to a constant. Fourier series make that decomposition precise. The infinite limit matters: solving each coefficient equation is not enough until the series and the derivatives used in the differential equation are justified.
A continuous Fourier series is distinct from a discrete Fourier transform. The former describes a function on an interval through infinitely many coefficients; the latter transforms a finite list of samples. Sampling and truncation connect them, but can also introduce aliasing. We first develop the continuous problem on ; changing length requires rescaling the frequencies.
Definitions
Definition 2.2.1
(Fourier coefficients and partial sums)
For an integrable -periodic complex-valued function , define
We use the inner product on . Functions equal almost everywhere represent the same element. For real , .
The exponentials are orthonormal: integrating gives one when and zero otherwise. Thus . The real form is , where and for real and positive . The constant is , a common source of factors of two.
Main results
Theorem 2.2.2
(Orthogonal projection and Parseval)
For , is its unique best approximation among trigonometric polynomials using frequencies . Moreover in , and
The last two assertions use completeness of the trigonometric system in .
Proof
(Orthogonal splitting plus density)
Let . Orthogonality gives . Consequently
where uses the normalized inner product. This proves the unique minimizer and Bessel's inequality . Completeness says trigonometric polynomials are dense: for every some such lies within of . For all sufficiently large , the best-approximation property then gives . Taking the limit in the orthogonal identity with proves Parseval. The density theorem is an essential input; orthogonality alone yields only Bessel's inequality.
Theorem 2.2.3
(A pointwise convergence criterion)
Suppose the periodic extension of has finitely many pieces on a period, and on each open piece both and extend continuously to its closed subinterval. Then at every its symmetric Fourier partial sums converge to
In particular they converge to at continuity points. Endpoint values refer to the periodic extension.
Proof
(Dirichlet kernel and the Riemann–Lebesgue lemma)
The Dirichlet kernel identity gives , with . Put . Pairing and and using gives
The quotient is integrable: near zero the numerator is by the one-sided piecewise hypothesis, and away from zero it is bounded on each piece. The Riemann–Lebesgue lemma makes the oscillatory integral tend to zero. This proof shows why the two one-sided limits enter symmetrically; it does not assert pointwise convergence for every function.
Proof X-ray
Where the one-sided regularity is spent
The proof has three moves. Unit mass of the Dirichlet kernel lets us subtract the candidate limit inside the integral. Pairing with uses evenness of , and it is this pairing that produces the average of the two one-sided limits rather than either one. The key move is to divide by , which vanishes at : the quotient is integrable only because the piecewise hypothesis makes the numerator near . That is the single place where local regularity is used. Once the quotient is in , the Riemann–Lebesgue lemma closes the argument without any rate. Continuity alone does not give an numerator, and indeed continuity alone is not enough for pointwise convergence. This argument is different in kind from the projection proof, which controls only the integral of the squared error and says nothing at a single point.
Worked examples
Worked example 2.2.4
(The sawtooth and its jump)
Let for , extended periodically. Oddness makes all cosine coefficients zero. Integration by parts yields
Thus for . At the periodic one-sided limits are and , so the series converges to zero, whatever value is assigned to . Parseval gives , hence . The normalization is important: is the combined energy of the pair .
Worked example 2.2.5
(A coefficient from projection)
Take . Orthogonality gives , , , and ; the other coefficients vanish. Parseval gives mean square . This equals the direct integral because all cross terms have zero mean. A finite Fourier polynomial has no convergence issue, but it is an excellent check of sign conventions.
Worked example 2.2.6
(Heat flow on a rod)
Let for , , with and . If , separation gives
Both boundary values vanish; differentiating verifies the PDE and setting returns the initial function. The higher frequency decays nine times as fast in its exponential rate. It can dominate the initial shape but becomes relatively less important as time passes.
Methods and reasoning
For an interval with homogeneous Dirichlet conditions the eigenfunctions are and eigenvalues of are . A general initial condition has sine coefficients
The sine system is complete in , so the expression at represents in . For each , Cauchy–Schwarz bounds the absolute sum of the terms after any fixed number of or derivatives, because exponential decay dominates every power of . These derivative series converge uniformly on , justifying termwise differentiation and the PDE there. Parseval and dominated convergence in the coefficient sum show in as . Pointwise agreement at requires extra assumptions, and initial endpoint values inconsistent with the boundary data cannot be satisfied classically at the corners.
The choice of basis encodes boundary conditions. A cosine basis fits homogeneous Neumann conditions, including a constant mode that does not decay under heat flow. A periodic basis contains positive and negative complex frequencies; its zero mode is the mean. For Poisson's equation on a circle, has a periodic solution only if the mean of is zero; the solution is determined only up to a constant. Dividing the zero Fourier coefficient by is not allowed. These are properties of the differential operator, not numerical defects.
Smoothness and coefficient decay
If is periodic and absolutely continuous, with and matching endpoint values, integration by parts gives . Thus for nonzero . More derivatives and matching periodic traces permit further integrations by parts. If the traces fail to match, the boundary term survives; a smooth function on the closed interval may have a jump after periodic extension. Coefficient decay is a statement about the periodic function actually being expanded.
Energy and computational truncation
For a heat solution with square-integrable initial data, orthogonality gives
Every term is nonincreasing, so total energy cannot grow. If only the first terms are retained, the squared error is the corresponding tail sum. At positive time this tail is bounded by . This is a truncation estimate for exact coefficients, not a bound for quadrature, roundoff, or an aliased FFT computation of those coefficients. Those additional errors require separate accounting.
Orthogonality as an error certificate
The projection property gives more than a way to calculate coefficients. For any trigonometric polynomial in the retained frequencies, the squared error splits exactly into the error of plus the squared distance between and . Thus no other choice of those coefficients can produce a smaller error. It does not follow that gives the smallest maximum error or the smallest error at each individual point. Different error criteria are different optimization problems. In a computation, estimated coefficients also have their own error: the projection identity refers to the exact integrals defining the coefficients.
From a boundary condition to a sine series
An initial temperature need not itself vanish at the endpoints in the formulation. For example, take for . Direct integration yields , so only odd modes appear. At every positive time the heat series vanishes at both ends, but its initial limit is the constant one in the interior. There is no contradiction: endpoint values occupy a set of measure zero and do not affect the limit. The solution cannot be continuous at either space-time corner if it is required to take the incompatible values one initially and zero on the boundary. This example makes the distinction between an initial condition in a norm and a classical pointwise condition concrete.
For a nonzero prescribed boundary temperature, one useful strategy is to subtract a function that matches those boundary values before expanding the remainder. For fixed endpoint temperatures and , take the steady linear profile . If the PDE is the homogeneous heat equation, satisfies the same PDE with zero boundary values and initial data . Solve for by its sine coefficients, then add . For time-dependent boundary data the lifting has a time derivative and the resulting equation generally has a forcing term. Forgetting that term changes the problem.
Common misconceptions
Common mistake
(Expecting the assigned value at a jump)
At a jump the partial sums approach the midpoint of the one-sided limits, not a chosen endpoint value. For the sawtooth at every term vanishes, so the series sums to whatever value is given.
Common mistake
(Expecting the overshoot near a jump to disappear)
Near a jump an overshoot of the partial sums persists as grows, even though the interval where it occurs shrinks. This is compatible with convergence, because the squared error over a shrinking interval tends to zero; it rules out uniform convergence near the jump.
Common mistake
(Dividing the zero mode by zero)
For periodic , the coefficient equation is . For it reads : a solution exists only if has mean zero, and then is free. Writing is not a formula.
Termwise differentiation needs its own justification. The heat equation's exponential factors provide it at positive time, but not at for arbitrary data.
Counterexample mode
Termwise differentiation of the sawtooth series
Claim: a Fourier series converging to may be differentiated term by term to obtain . The sawtooth on has there, but differentiating term by term gives . At its terms are , which do not tend to zero, so the series diverges. The step that fails is : the integration by parts leaves the boundary term , which is nonzero because the periodic extension jumps by . The repaired statement requires periodic and absolutely continuous with ; then the differentiated series is the Fourier series of .
Quick check
Quick check
(What is the long-time limit of periodic heat flow?)
Look at which Fourier mode does not decay.
Solution · Answer
Each mode is multiplied by , so the nonzero modes decay while remains. The limit in is the initial spatial mean ; it is zero only when the initial mean is zero.
Exercises
Quick check
(Compute the Fourier coefficients and mean square of .)
Solution · Answer
and , with all others zero. Parseval gives mean square , which agrees with the direct value .
Quick check
(Let on and on , periodically extended. Find its sine coefficients and the values to which the series converges at and .)
Solution · Answer
Oddness gives , while . The one-sided limits at both jump locations are and , so the series converges to zero at each.
Quick check
(Solve the Dirichlet heat problem on for initial data . Compute its squared norm at time .)
Solution · Answer
. Orthogonality and give .
Quick check
(For periodic , find all twice differentiable periodic solutions. Why is there no periodic solution for ?)
Solution · Answer
A particular solution is , since . The difference of any two solutions has zero second derivative; periodicity eliminates its linear term, so all solutions are . Integrating over a period would give zero on the left and on the right, a contradiction.
Further study
The argument invokes completeness and the Riemann–Lebesgue lemma. Distributional transforms, multidimensional domains, nonhomogeneous boundaries, Gibbs estimates and stable coefficient computation deserve separate treatment. The sine expansion is one boundary-value model, not a complete account of spectral methods.
Yes: the averages of the partial sums converge uniformly for every continuous periodic function, because their kernel is nonnegative.
A natural next question
Can averaging rescue convergence for a merely continuous function?
Yes: the averages of the partial sums converge uniformly for every continuous periodic function, because their kernel is nonnegative.
Continuity alone does not make converge. The Cesàro means are convolutions with the Fejér kernel
which is nonnegative, has mean one and concentrates at . These properties give uniformly for every continuous periodic (Fejér's theorem). Positivity also keeps between the infimum and supremum of , so the averages show no overshoot at a jump. The Dirichlet kernel changes sign, and its mean absolute value grows like a multiple of .