Ordinary Fourier partial sums keep some modes unchanged and discard all higher modes. Abel summation makes a different choice: retain every mode, but multiply the th mode by , where . For fixed the weighted series converges absolutely. As approaches one, it recovers boundary data through a positive averaging kernel.
This chapter studies what that recovery means. Continuous data are recovered uniformly; integrable data are recovered at their Lebesgue points; finite jumps are recovered at their midpoint. None of these facts alone proves convergence of the ordinary partial sums. We will also prove a converse under a small-coefficient condition and explain exactly where that additional information enters.
Motivation
The factor suppresses large frequencies smoothly. Positivity of the resulting kernel prevents values outside the range of bounded real data, while concentration near zero produces boundary recovery. These properties come from a concrete formula, not from a formal limit of each coefficient.
There are two limiting operations to distinguish. First, for a fixed , sum the infinite weighted series. Then let . Sending to one in each term before summing returns the original convergence problem and is not justified by the weighted convergence.
We use the circle with normalized measure . Thus
for finite . Functions and series may be complex-valued. One-sided boundary limits are always interpreted periodically.
Definitions
Definition 1.7.1
(Poisson kernel and Abel average)
For , define
For the integral exists for every , since is bounded. The average is unchanged by altering on a null set. In the multiplier notation, the constant-mode factor is always one, including at .
Definition 1.7.2
(Lebesgue boundary value)
An angle has Lebesgue value for if
This is a condition on local averages. It need not imply pointwise convergence of as , and the assigned value at the single point is irrelevant.
For a scalar series , an Abel sum means that converges for and tends to as . All scalar coefficient sequences used below are bounded, which ensures absolute convergence for each fixed .
Theorem/proposition
Theorem 1.7.3
(Positive Abel approximation)
For every ,
absolutely and uniformly in for fixed . The kernel is nonnegative, has normalized integral one, and concentrates at zero as . The operators are contractions on every , . For , in whenever ; for continuous periodic , convergence is uniform.
Theorem 1.7.4
(Radial limits at Lebesgue points and jumps)
If has Lebesgue value at , then . If instead a chosen representative has finite one-sided limits and there, then
The jump conclusion does not require piecewise differentiability. It follows from the even kernel and the one-sided limits.
Theorem 1.7.5
(A small-coefficient converse)
Suppose a scalar series has Abel sum and . Then its ordinary partial sums converge to .
Consequently, if is continuous and periodic and
then its ordinary symmetric Fourier partial sums converge uniformly to . The small-coefficient condition is sufficient; it is not a necessary condition for ordinary convergence.
Proof sketch or proof idea
The geometric series gives the kernel and the constants
Proof
(Expansion, positivity, contraction and norm recovery)
For , summing the two geometric series gives
The series converges absolutely and uniformly in . Its constant coefficient is one, so , and the closed formula proves positivity. Integrating the series against is valid because . The change of variable in each coefficient gives the displayed Abel expansion. Also , so the weighted expansion is uniformly absolutely convergent. Differentiating in is legitimate at fixed , since every is finite.
Put . For and ,
The kernel mass away from zero therefore tends to zero. Positivity and normalization give the contraction by Jensen and Tonelli for finite , and by the essential bound for .
For continuous periodic , write the difference as the average of . Choose by uniform continuity, making that difference small for uniformly in . The remaining part is bounded by times the vanishing tail mass. This proves uniform convergence.
For , translations are continuous in . Use the standard measure-theoretic density of continuous periodic functions in finite , obtained from simple-function approximation and Lebesgue regularity; that density theorem is an input here. Translation is an isometry, so the uniform-continuity result extends by a triangle inequality. Minkowski then gives
Split at a small fixed , using translation continuity near zero and elsewhere. Concentration proves convergence.
There is a concise harmonic interpretation. If , the Abel expansion is . On every smaller disk the series and its derivatives converge uniformly, so its real and imaginary parts are harmonic. The present chapter concerns Fourier summation and boundary recovery; a full disk Dirichlet-problem treatment requires additional work.
Lebesgue values require an averaged local estimate
Proof
(Radial limits at Lebesgue points and at jumps)
Fix and its Lebesgue value . Define
The function is absolutely continuous, , and . Almost everywhere,
Together with evenness of the kernel, this identity will turn the absolute error integral into an integral against . The kernel is even and decreases on : direct differentiation gives
Given , choose such that for . For fixed , integration by parts gives
The lower endpoint term vanishes because and is bounded for this fixed . On , the corresponding integral is at most , which tends to zero. Finally,
The near part is at most after normalization, and the far part vanishes. This proves the radial limit. The statement is conditional on the displayed Lebesgue-value property; using the Lebesgue differentiation theorem would make it an almost-everywhere conclusion, but that measure-theoretic theorem is not proved here.
For a jump, pair the two sides instead:
The expression in brackets tends to as . Its difference from that constant is uniformly small near zero and integrable away from zero. The same concentration split applies. Since , the constant contributes . A genuine jump, with , is never a Lebesgue point for any finite value : its local average absolute deviation tends to . The midpoint therefore comes from the separate pairing argument.
Proof X-ray
Where evenness, monotonicity and the cumulative bound enter
Evenness of folds the error integral onto , where the two one-sided deviations combine into . The Lebesgue-value hypothesis enters only once, as on ; the integrand itself need not be small at any particular nearby point. The key move is the integration by parts, which transfers the derivative from to the kernel. The sign on is exactly what allows the cumulative bound to be substituted inside that integral, and removes the lower boundary term.
The argument closes in two ways. Integrating by parts once more turns into , which cancels the boundary term and leaves the kernel mass, at most for every . The far part vanishes by concentration, because . Without monotonicity the substitution step fails, and a decreasing majorant of the kernel would be needed instead. At a genuine jump the hypothesis itself fails, which is why the jump case uses the pairing argument.
Ordinary convergence always gives Abel convergence
Proof
(Ordinary convergence implies Abel convergence)
Let converge to . Then is bounded, as is , and summation by parts gives
The weights on the right are nonnegative and sum to one. For any , the tail values satisfy after a fixed index. The total contribution of the finitely many earlier errors tends to zero with , while the tail contributes at most . Hence the Abel limit is .
Apply this to the paired Fourier terms and for . If the ordinary symmetric partial sums converge at , their limit is the Abel limit there. This implication runs in one direction without extra information.
The Tauberian estimate supplies a converse
Proof
(The Tauberian comparison at radius )
Assume now that and that the Abel limit is . For an integer , choose and compare
Given , choose with for . For , the finitely many terms tend to zero as . In the remaining first sum, , so its absolute value is at most . The tail is bounded by
Thus the difference tends to zero, since was arbitrary. The Abel expression tends to , proving ordinary convergence.
For the Fourier corollary, regard the paired terms as continuous functions. The hypothesis gives . Every bound in the preceding comparison therefore holds uniformly in . The Abel sums already tend uniformly to , so uniformly.
Concept lensStructural
A summation method and a converse need different information
Abel convergence averages the sequence of ordinary partial sums. Averaging can hide their oscillation. The condition controls the discrepancy between a partial sum at and the Abel average at radius , preventing that discrepancy from surviving the limit.
Worked examples
Worked example 1.7.6
(Individual modes retain different weights)
For ,
The mean is unchanged, while the other modes have different damping. With normalized circle measure, orthogonality gives . Also . The energy and uniform estimates follow from different features: orthogonality for the first, bounded trigonometric modes for the second.
Worked example 1.7.7
(A square-wave Abel profile has no overshoot)
Let , with value zero at the jumps. Its sine-series coefficients are for odd and zero for even . Thus for positive odd its complex coefficients are and . For , integrating the geometric series gives
Use the logarithm with value zero at . The ratio on the right has positive real part when ; its argument is therefore the ordinary arctangent of its imaginary-to-real ratio. Taking imaginary parts gives
Away from the jumps it tends to , namely the plateau or ; at each jump it equals zero. All its values lie between minus one and one, also a consequence of positivity of the averaging kernel.
There is an explicit shrinking transition profile. Put and . Then
locally uniformly in : the arctangent's argument is uniformly on bounded -intervals. This profile differs from the oscillatory ordinary-partial-sum profile studied in the Gibbs chapter; its derivation uses a different summation method.
Worked example 1.7.8
(The small-coefficient condition is not necessary)
For , the ordinary series converges by the alternating-series criterion. Its Abel sums equal , by integrating the geometric series, and tend to . The Abelian implication therefore identifies the ordinary sum as .
Here does not tend to zero. The Tauberian condition is a sufficient extra hypothesis for a general converse, not a test that every convergent series must satisfy.
Counterexample mode
Abel convergence alone does not recover ordinary convergence
The scalar series has Abel sums , while its ordinary partial sums alternate between one and zero.
The distinction also occurs within Fourier analysis: the localization and Lebesgue-constants chapter uses logarithmic operator-norm growth and uniform boundedness to prove that a continuous periodic function can have unbounded ordinary partial sums at a point. Its Abel averages nevertheless converge uniformly by the positive-kernel theorem above. The scalar example demonstrates the logical failure directly; the continuous Fourier example requires that separate existence proof.
Common mistakes
Common mistake
(Setting term by term)
Each damped series converges for , but this does not permit putting in every term before summing: that returns the ordinary series, whose convergence is the original question. The limit is taken after the sum. Likewise, the midpoint value at a jump is determined by the two one-sided limits, not by an arbitrary value assigned at the jump itself.
Common mistake
(Dropping the absolute value in the Lebesgue condition)
The Lebesgue-point proof uses an absolute local-average condition. A symmetric average without the absolute value can cancel positive and negative errors and does not provide the estimate for used here.
Common mistake
(Expecting an rate for Lipschitz data)
The scale describes the central concentration of the Poisson kernel, but its tails affect approximation rates. Even Lipschitz boundary data need not have uniform error bounded by a constant times ; the exercises exhibit the logarithmic factor.
Finally, ordinary convergence, uniform Abel convergence, finite- norm convergence and radial convergence at a jump have different hypotheses and conclusions. No non-tangential boundary theorem is proved here.
Yes: Abel summability together with a bounded sequence n times a_n already forces ordinary convergence, but the proof is much harder.
A natural next question
Can the small-coefficient condition be relaxed to boundedness?
Yes: Abel summability together with a bounded sequence n times a_n already forces ordinary convergence, but the proof is much harder.
Littlewood's Tauberian theorem states: if has Abel sum and for all , then converges to . The comparison at radius no longer suffices, because both error terms are then only bounded by , not small; the proof needs a finer approximation argument outside this chapter. The exercise with shows that alone is not enough.
Summary
The Poisson kernel realizes Abel sums as positive concentrated averages. Uniform continuity and translation continuity give norm recovery; a cumulative-deviation argument gives recovery at Lebesgue points; evenness gives jump midpoints. A small-coefficient Tauberian estimate can then recover ordinary convergence, which Abel convergence alone cannot provide.
Exercises
Quick check
(For , prove for , and derive . Show .)
Solution · Answer
The coefficient of is , by the absolutely convergent expansion. Applying a second average multiplies it by . The two averages have the same absolutely convergent Fourier expansion, so the semigroup identity holds everywhere directly. At , the operator is the mean projection. Parseval gives the energy identity. For , , so summing after removing the constant coefficient gives the final estimate. The constant mode is the only one not attenuated.
Quick check
(Suppose is Lipschitz on the circle with constant . For and , prove . Use for .)
Solution · Answer
The error is at most . The given cosine bound, and imply . Thus the error is at most . For , , so the logarithm is at most . This proves the stated bound. The logarithm arises from integrating over displacements larger than the central scale .
Quick check
(For the continuous periodic function on , prove for . Deduce that a uniform bound for all Lipschitz data is impossible.)
Solution · Answer
Since and , . On , this is at least . Therefore . The result follows. The function is 1-Lipschitz for circular distance and has . Dividing its error lower bound by gives an unbounded quantity, ruling out an estimate even for this one Lipschitz function.
Quick check
(Put and for . Prove and that the series has Abel sum zero but no ordinary sum. You may first estimate by comparing it with an integral.)
Solution
The ordinary partial sums are , while . Let . Comparison on each unit interval gives . The derivative is integrable at zero. Substituting in the integral gives , . Hence .
Two summation-by-parts identities now give . Cauchy–Schwarz yields . Thus the Abel sums have absolute value at most and tend to zero.
There is no ordinary limit. Choose integers nearest to and to . The corresponding square roots differ from those phases by quantities tending to zero, so the partial sums have subsequential limits one and minus one. Therefore alone cannot repair the Abel converse.
Quick check
(For the square wave in the worked example, prove for every , where the norm is the essential supremum. Prove also for .)
Solution
For , the explicit arctangent has the sign of and absolute value below one. The error is below one off the jumps but approaches one on sets of positive measure near each jump. Its essential supremum is therefore one; the same value holds at .
Parseval and the odd sine coefficients give . For , use and set . Enlarging the sum to all positive integers bounds it by . Here . The claimed bound follows. The finite norm tends to zero although the essential-supremum error stays fixed.