A Fourier series is an infinite series of sine and cosine functions, or equivalently complex exponentials, whose frequencies are integer multiples of a fundamental frequency. Its Fourier coefficients are calculated by integrating a function against these harmonics over one period. The resulting series represents the function in a sense determined by its regularity and the mode of convergence; it need not converge to the function at every point. (dlmf.nist.gov)
Named after Joseph Fourier, whose investigations of heat conduction established its importance, Fourier series connect the geometry of Hilbert spaces with the analysis of periodic phenomena. They provide methods for solving partial differential equations, separating signals into frequency components, and constructing numerical approximations. Their theory encompasses orthogonality, convergence, harmonic energy, discontinuities, and summation methods, while their computational use connects them to the discrete Fourier transform and fast Fourier transform. (mathshistory.st-andrews.ac.uk)
Harmonic Representation and Coefficient Formulas
A function is periodic with period if for every in its domain. Fixing such a period determines the angular frequency and the harmonic frequencies , where . When is the smallest positive period, is the fundamental angular frequency of the function; otherwise it is the fundamental frequency of the chosen expansion. The th harmonic consists of sine and cosine components with period . The zero-frequency component is constant and represents the mean over one period. These definitions generalize the usual -period convention by rescaling the independent variable. (math.mit.edu)
Suppose that is absolutely integrable over one period, meaning . Its trigonometric, or real, form is the formally assigned infinite series
where
The symbol here assigns the series to ; it does not assert equality with its sum. “Real form” refers to the trigonometric functions used: for complex-valued , the coefficients may themselves be complex. (dlmf.nist.gov)
The starting point is arbitrary because each integrand is -periodic, so integration over any complete period gives the same coefficient. In particular,
The factor in the nonconstant coefficients reflects the fact that the integral of a squared sine or cosine over a full period is , whereas the constant function has integral . Distinct harmonics have zero cross-integrals, allowing their coefficients to be isolated separately. For example, with , the function has , , and , with all remaining coefficients zero. (math.mit.edu)
Using the complex exponential, the same expansion takes the form
Here , and denotes all integers, including negative indices and zero. Expanding the exponentials into sines and cosines gives, for ,
Conversely, and . For real-valued , , so positive and negative frequency coefficients are conjugate pairs rather than independent data. (dlmf.nist.gov)
The symmetric partial sum retains both signs of every frequency through order :
Ordinary convergence of this expansion means convergence of as , not an arbitrary ordering of the doubly indexed terms. The coefficient integrals can exist without establishing pointwise convergence to . Even when convergence holds, the sum need not equal an assigned value at a discontinuity: under suitable regularity conditions it equals the average of the two one-sided limits. Defining the coefficients and proving reconstruction are therefore distinct mathematical steps. (dlmf.nist.gov)
Origins in Vibrating Strings and Heat Conduction
In 1747, Jean le Rond d’Alembert formulated a partial differential equation for the transverse motion of a stretched string and described its solution through waves travelling in opposite directions. Leonhard Euler’s investigations of 1748 broadened the discussion of admissible initial shapes: rather than requiring one analytical expression for the entire string, Euler allowed curves whose different portions could be specified differently. Daniel Bernoulli, in his work of 1753, approached the problem through the superposition principle. He argued that a string’s motion could be assembled from simultaneous simple vibrations, with frequencies equal to integer multiples of the fundamental frequency. The disagreement concerned not merely alternative solution methods, but whether every physically admissible initial shape could be represented by an infinite sum of sinusoidal modes. (mathshistory.st-andrews.ac.uk)
For a string fixed at both ends, each sinusoidal standing-wave mode vanishes at the endpoints; its amplitude specifies how strongly that mode contributes to the displacement. Bernoulli’s proposal therefore required an arbitrary initial curve between the endpoints to be expressible as a sine series. A plucked string, whose initial profile has a corner, made the difficulty especially apparent: how could sums of smooth trigonometric functions reproduce a curve that was not smooth? Eighteenth-century arguments often transferred properties of finite expressions to infinite series without adequately distinguishing their limiting behavior. The vibrating-string controversy consequently became a debate about what counted as a function, as well as about the equations of mechanics. (bibliotheque.imo.universite-paris-saclay.fr)
Joseph Fourier submitted his heat-conduction memoir to the Institut de France on December 21, 1807. His central problem was to determine temperatures inside a body from its geometry, thermal properties, and prescribed initial or boundary temperatures. Trigonometric expansions supplied the connection between those prescribed data and elementary solutions of the heat equation. The memoir was not published in full at the time; a revised submission won the Academy’s heat-theory competition in 1812, although the judges criticized aspects of its rigor and generality. Fourier developed its mathematical content into Théorie analytique de la chaleur, published in 1822. Unlike the earlier string debate, this work made trigonometric representation a systematic computational instrument across problems involving rectangular bodies, rings, and other geometries. (fr.wikisource.org)
The 1822 book introduced its first trigonometric expansion through an equilibrium-temperature problem: a semi-infinite rectangular strip had its long sides maintained at zero temperature and its end maintained at a constant nonzero temperature. Fourier combined elementary solutions that decayed into the strip, choosing their coefficients so that the sum reproduced the constant temperature along the end. Elsewhere, prescribed initial temperatures determined the amplitudes of modes that decayed with time. His integral formulas for coefficients gave a general procedure rather than a separate algebraic construction for each temperature profile. Fourier also supplied convergence arguments for particular examples, but his broad claims about arbitrary functions exceeded what those arguments established. (bibliotheque.imo.universite-paris-saclay.fr)
Peter Gustav Lejeune Dirichlet’s 1829 paper shifted attention from constructing coefficients to proving that the resulting partial sums approached the generating function under stated conditions. His method expressed partial sums as integrals and controlled their limiting behavior, establishing a model for rigorous convergence analysis. In an 1837 exposition, he explicitly treated a function as a correspondence assigning a unique value to each argument, without requiring a single analytical formula. Bernhard Riemann extended the investigation in his 1854 habilitation memoir, published posthumously in 1867. To study trigonometric series beyond the earlier regularity assumptions, he formulated the definition now called the Riemann integral and examined functions with infinitely many discontinuities. Questions about trigonometric representation thus directly shaped both the definition of integration and the separation of a function from any particular formula representing it. (mathshistory.st-andrews.ac.uk)
Orthogonality, Completeness, and Mean-Square Approximation
For complex-valued, -periodic functions, the space consists of square-integrable functions identified when they agree almost everywhere. It is a Hilbert space with the normalized inner product and associated norm
Writing and , direct integration gives : distinct harmonics are orthogonal, and each has norm one. Consequently,
is the coordinate of the orthogonal projection of onto the one-dimensional subspace generated by . In real notation, the corresponding normalized system is , for ; the factors account for the coefficient normalization in the usual sine–cosine formulas. (math.mit.edu)
Orthogonality alone does not ensure that a system represents every function in the space. The exponential system is also complete: its finite linear combinations, the trigonometric polynomials, are dense in . Equivalently, a function orthogonal to every must be zero as an element. Thus the exponentials form an orthonormal basis in the Hilbert-space sense. Unlike an algebraic basis, which represents each vector by a finite combination, this basis generally requires an infinite expansion interpreted as a limit in norm. One route to completeness approximates continuous periodic functions uniformly by trigonometric polynomials and uses the density of continuous periodic functions in . (math.mit.edu)
Let and . Since is orthogonal to , every satisfies
Hence uniquely minimizes the mean-square error among all trigonometric polynomials using these frequencies. Completeness implies
This conclusion applies to every square-integrable periodic function, without requiring continuity or differentiability. It concerns the integrated squared discrepancy; it does not itself assert convergence at every point or uniform convergence. (ocw.mit.edu)
For any orthonormal system, Bessel’s inequality bounds the sum of squared projection coordinates by the squared norm. Here, completeness strengthens that inequality to Parseval’s identity:
Applying orthogonality to the residual gives the exact truncation error
Thus the omitted coefficients contain precisely the omitted mean-square energy. For example, if , its mean square is . Retaining only frequencies leaves squared error , entirely from the third harmonic. (ocw.mit.edu)
Uniqueness extends beyond : if two integrable periodic functions have identical Fourier coefficients, they agree almost everywhere. In particular, an function whose coefficients all vanish is zero almost everywhere; this result does not require ordinary convergence of its partial sums. Fourier coefficients therefore determine an equivalence class, not arbitrary values at individual points. Changing a function at one point—or on any set of Lebesgue measure zero—changes none of its coefficients. If both functions are continuous, almost-everywhere equality implies equality everywhere, so the ambiguity disappears within that class. (noter.math.ku.dk)
Symmetry, Half-Range Expansions, and Representative Examples
For an integrable periodic function, symmetry can force entire families of coefficients to vanish. If , the function is even: its product with each sine is odd, so integration over a symmetric period gives . If , the function is odd, and . These properties depend on the choice of origin. Half-wave antisymmetry, , instead compares successive half-periods. Splitting a coefficient integral into two halves introduces the factor , because shifting the th harmonic by multiplies it by . Thus the mean and all even-numbered harmonics vanish; odd-numbered sine and cosine terms may both remain. Oddness alone does not imply half-wave antisymmetry. (lpsa.swarthmore.edu)
A function prescribed only on admits two standard half-range expansions. Reflecting it evenly, , and then repeating with period produces
Reflecting it oddly, away from zero, produces
The symbol identifies the assigned series without asserting convergence under mere integrability. These are different periodic continuations of the same interior data, not interchangeable descriptions outside . Every sine partial sum vanishes at and ; nonzero one-sided endpoint values therefore create jumps in the odd continuation. Endpoint assignments do not affect the coefficient integrals. (mx.nthu.edu.tw)
The -periodic square wave , equal to on and on , is both odd and half-wave antisymmetric. Direct integration gives
Consequently,
Only odd sine harmonics occur, with amplitudes proportional to . At every jump the series sums to zero, irrespective of the value assigned there. Substitution of yields the numerical identity
For the sawtooth wave obtained by periodically repeating on , oddness again leaves only sines, but now every positive harmonic occurs:
The equality holds for ; at the periodically identified endpoints the sum is zero. Its decay reflects the reset jump from to . (lpsa.swarthmore.edu) By contrast, periodically repeating gives a continuous triangular wave:
Its mean is ; subtracting that mean produces half-wave antisymmetry. Only odd cosine harmonics remain, and their amplitudes decay faster than those of the discontinuous square and sawtooth waves. Here the nonsmooth features are corners rather than jumps in value. (math.mit.edu)
The periodic continuation of on is also continuous, since its endpoint values agree. Evenness eliminates sines, while two integrations by parts give and :
Unlike the triangular example, this cosine expansion contains both even and odd harmonics. Its absolute summability gives uniform convergence, permitting evaluation at the endpoints. Setting , and then , yields respectively
Pointwise, Uniform, and Almost-Everywhere Convergence
For symmetric partial sums , pointwise convergence to means that at each specified point; the required truncation index may depend on . Uniform convergence on a set requires , so one index controls the error throughout . Convergence almost everywhere permits an exceptional set of Lebesgue measure zero. These differ from mean-square convergence, , which controls integrated error rather than values at individual points. In particular, mean-square convergence alone does not establish pointwise convergence of the entire sequence. (people.math.harvard.edu)
The Dirichlet–Jordan theorem states that if a -periodic function has bounded variation over a period, then
at every . Bounded variation means that the supremum of over all ordered partitions of the period is finite; this condition ensures the existence of the one-sided limits. At continuity points the limit is , whereas at a jump it is the midpoint of the limiting values, regardless of the assigned value . Endpoints are interpreted periodically: on , the limit at either endpoint is . Camille Jordan introduced bounded variation in 1881 in connection with this convergence criterion. (noter.math.ku.dk)
If is also continuous throughout the periodic domain, the Dirichlet–Jordan conclusion strengthens to uniform convergence; more locally, convergence is uniform on any closed interval on which is continuous. Continuity must include agreement across the chosen period’s endpoints. For example, applying the theorem to the periodic continuation of on gives convergence to in the interior and to at the seam. Convergence cannot be uniform over the whole period: every partial sum is continuous, and a uniform limit of continuous functions is continuous. Thus midpoint convergence at a jump does not imply uniform approximation across it. (noter.math.ku.dk)
A separate sufficient condition is absolute convergence of the coefficient sequence:
Since , the tail is bounded uniformly by
Consequently, the series converges absolutely at every point and uniformly to a continuous periodic function . If the coefficients were computed from an integrable , uniqueness gives almost everywhere; if is continuous, equality holds everywhere. Sufficient regularity includes on the periodic domain. Continuity alone, however, is insufficient even for pointwise convergence everywhere: Paul du Bois-Reymond constructed continuous periodic functions whose Fourier series diverge at a point. (noter.math.ku.dk)
The Carleson–Hunt theorem states that for every periodic , with , the ordinary symmetric partial sums satisfy
Lennart Carleson proved the case in 1966; Richard Hunt extended it to the stated range of in 1968. This concerns the full sequence of partial sums, not merely a subsequence or an averaging method. Every continuous periodic function belongs to , so its Fourier series converges almost everywhere even though exceptional divergence points may remain. (link.springer.com)
The endpoint does not admit the same general conclusion. In 1923, Andrey Kolmogorov published an example of a Lebesgue-integrable function whose Fourier series diverges almost everywhere; in 1926 he constructed an integrable example with divergence everywhere. These are counterexamples to a universal convergence theorem for , not assertions that every integrable function has a divergent series. Integrability still defines the coefficients, but without additional hypotheses it does not guarantee ordinary pointwise convergence, even outside a null set. (pldml.icm.edu.pl)
Dirichlet Kernels and the Gibbs Phenomenon
For an integrable -periodic function, the symmetric partial sum can be expressed as periodic convolution:
The last identity follows by summing a finite geometric progression; at multiples of , its removable singularities are filled with the value . The Dirichlet kernel therefore implements sharp frequency truncation: harmonics with are retained unchanged, and all higher harmonics are discarded. Its normalized integral is , but it changes sign, so convolution with it is not a positive weighted average of nearby function values. (math.unm.edu)
The kernel has a narrow central peak and alternating positive and negative lobes. Near a jump, these lobes sample the different levels on either side and produce ringing: successive excursions above and below the limiting levels. This Gibbs phenomenon concerns values near the discontinuity, not the convergence value at the discontinuity itself. For a piecewise continuously differentiable periodic function with one-sided limits and , the partial sums at converge to , regardless of the value assigned to . Nearby peaks can nevertheless remain above or below . Josiah Willard Gibbs described the phenomenon in an 1899 letter to Nature. (dlmf.nist.gov)
The limiting shape is resolved by magnifying distances of order . Write for the signed jump and . Under the preceding regularity assumption, for fixed real ,
This profile follows from the local approximation . The sine integral approaches as , but its first positive maximum occurs at , since its derivative is . Consequently, for an upward jump, the first peak approaches , where
The limiting overshoot is thus approximately of the full jump size; the corresponding undershoot approaches . (ocw.mit.edu)
Normalization matters when reporting this percentage. For the square wave taking values and , the jump is , so the first positive peak tends to . Its excess is about of the upper level’s magnitude, but still only of the jump. More explicitly, retaining the first odd harmonics gives
The first positive maximum occurs exactly at , and its height tends to . (ocw.mit.edu)
Increasing the cutoff therefore moves the principal ringing toward the jump and contracts its spatial scale in proportion to , without eliminating its limiting relative height. There is no contradiction with pointwise convergence: a fixed point away from the jump eventually lies outside this shrinking transition region, whereas the peak’s location changes with . On an interval spanning the jump, uniform convergence fails. Even assigning the midpoint value at the discontinuity cannot remove that failure, because each partial sum is continuous while the limiting function retains different one-sided limits. (ocw.mit.edu)
Cesàro and Abel Summation
For a -periodic integrable function , let denote its complex Fourier coefficients and its symmetric partial sums. The Fejér means apply Cesàro summation to these partial sums:
The triangular coefficient weights arise because the term with frequency occurs in precisely of the averaged sums. Thus the constant coefficient remains unchanged, while higher frequencies are progressively attenuated rather than retained at full strength up to a sharp cutoff. This is Cesàro summation of order one, conventionally denoted . (ocw.mit.edu)
Averaging the Dirichlet kernels gives the nonnegative Fejér kernel:
With normalized periodic convolution,
Its normalized integral is one, and, for every fixed , its mass outside tends to zero. Indeed, there . These properties make the kernels an approximate identity: they average nearby values with increasing concentration around the evaluation point. Fejér’s theorem gives uniform convergence for every continuous periodic . For f\inL^p, , it gives . The latter follows from concentration of the kernels together with continuity of translations in the norm. (ocw.mit.edu)
Abel means instead damp every nonzero frequency exponentially:
For integrable , its coefficients are bounded, so this series converges absolutely and uniformly for each fixed . Summing the geometric series produces the Poisson kernel and its convolution representation:
Unlike a Fejér mean, an Abel mean generally contains infinitely many frequencies. The Poisson kernel is positive, has normalized integral one, and concentrates at zero as . Consequently, uniformly for continuous periodic , and in norm for . Its interpretation as a harmonic extension into the unit disk connects this summation method with boundary-value problems. (math.uchicago.edu)
Both methods are regular: whenever the ordinary symmetric partial sums converge at a point, their Cesàro and Abel means converge to the same value. The converse need not hold. In particular, every continuous periodic function has uniformly convergent Fejér and Abel means, although continuity alone does not guarantee ordinary Fourier-series convergence at every point. Summability therefore asserts convergence of a transformed family of approximations, not convergence of the original partial sums. (math.uchicago.edu)
At a jump where finite one-sided limits exist, both means tend to , independently of the assigned value . Their positivity also explains their different behavior from sharp truncation: for real-valued data satisfying almost everywhere, both approximations remain between and , since each is a weighted average with total weight one. Thus, for the square wave taking values and , neither approximation exceeds those levels. This range-preservation argument explains why positive averaging avoids the persistent jump overshoot of the Gibbs phenomenon; it does not remove the transition region or permit uniform convergence across a discontinuity. (addi.ehu.es)
Coefficient Decay, Calculus, and Algebraic Operations
For a -periodic function, write and . The Riemann–Lebesgue lemma states that implies as . Oscillatory cancellation therefore suppresses high-frequency coefficients even without continuity. This is only a necessary decay property: it neither implies absolute summability nor guarantees pointwise convergence of the Fourier series. Convergence requires additional information about the function or its coefficients. (dlmf.nist.gov)
Integration by parts makes the dependence on regularity explicit. If is absolutely continuous, , and the endpoint values match, for , then, for ,
The lemma strengthens this bound to . Endpoint matching removes the boundary terms; smoothness on the interval alone is insufficient. Piecewise smooth functions with jumps generally have coefficients, while a continuous function with jumps in its first derivative generally has coefficients. For example, the periodic continuation of on has , whereas that of has , for . (math.ucdavis.edu)
A periodic function consequently has coefficients decaying faster than every inverse power, although this need not be exponential decay. For an analytic function, complex extension gives a stronger estimate: if its periodic extension is holomorphic and bounded by on the closed strip , contour shifting yields
The width of the available strip thus controls the exponential rate. As a concrete -periodic example, has nonzero-frequency coefficients ; its complex extension has poles at . (math.umd.edu)
The coefficient rule for differentiation is whenever is absolutely continuous and periodic; in particular, . This coefficient identity does not itself justify pointwise termwise differentiation. A sufficient condition is : the differentiated series then converges absolutely and uniformly and differentiates the continuous representative of . Integration is less restrictive. For , the Cauchy–Schwarz inequality gives , so its primitive has the uniformly convergent representation
The zero-frequency term must be treated separately: a periodic primitive exists exactly when . (math.ucdavis.edu)
Linearity gives , while translation gives , changing phase but not magnitude. If and have absolutely summable coefficient sequences, multiplication corresponds to discrete convolution:
Absolute convergence justifies regrouping the double series. Conversely, define normalized periodic convolution by
For , interchange of the integrals yields . Omitting the normalization introduces a factor on the right-hand side. (math.ucdavis.edu)
Differential Equations and Periodic Physical Systems
For a homogeneous rod of length , with its ends held at zero temperature, the one-dimensional heat equation is , where is the thermal diffusivity. Separation of variables, , gives and . The boundary conditions select the eigenvalues and sine modes. Consequently, the initial temperature determines the solution through
Each spatial harmonic evolves independently: its decay time is , so finer-scale variations disappear faster. For example, in dimensionless units with , , and initial temperature in the interior, only odd modes occur, with ; the long-time profile is dominated by . The initial endpoint mismatch does not prevent the series from satisfying the zero-temperature boundary conditions for . (math.mit.edu)
For an ideal stretched string, the wave equation produces oscillating rather than decaying mode amplitudes. With fixed ends, initial displacement , and initial velocity , its standing-wave expansion is
where and are the sine coefficients of and . Thus both initial shape and initial motion determine the vibration. Free ends instead impose , selecting cosine modes, including a spatially constant mode whose displacement can vary as . One fixed end and one free end select , . These choices arise from the physical boundary conditions, not from an arbitrary preference for sine or cosine notation. (people.math.harvard.edu)
In the unit disk, Laplace’s equation becomes . Expanding prescribed boundary data into angular harmonics gives
Regularity at the center excludes the singular radial solutions and, for the zero mode, . The resulting harmonic extension can model steady temperature in a circular plate with insulated faces and prescribed edge temperature. For continuous , it approaches the boundary values uniformly as . The factors suppress high angular frequencies inside the disk; this is the same extension represented by the Poisson kernel in Abel summation. (math.mit.edu)
In acoustics, a periodic pressure signal can be written as . The amplitudes specify harmonic strengths, while the phases specify their relative timing and help determine waveform shape. For an ideal fixed string, the frequencies are integer multiples of ; different excitation patterns produce different harmonic mixtures. These mixtures contribute to musical timbre, although an actual instrument’s attack, decay, and resonating body are not fully described by a single stationary periodic waveform. (openstax.org)
A stable linear time-invariant system responds to each input harmonic without changing its frequency: in periodic steady state, input coefficients become . Thus the th amplitude is multiplied by , and its phase is shifted by . In an alternating-current series RLC circuit, the voltage and current coefficients satisfy , for , with impedance
At , the reactive contributions cancel and that harmonic’s current is in phase with its voltage. Harmonic-by-harmonic calculation therefore determines the circuit’s response to nonsinusoidal periodic forcing, after transients have decayed. (ocw.mit.edu)
Numerical Computation and Sampled Data
Continuous Fourier coefficients are defined by integration, whereas sampled coefficients depend on a finite set of function values. For a -periodic function, , with , may be evaluated analytically or approximated by numerical quadrature. Equally spaced quadrature uses , , excluding the duplicated endpoint, and gives
These numbers are normalized discrete Fourier transform (DFT) coefficients. Continuous Fourier projection minimizes an integral mean-square error; trigonometric interpolation instead matches the sample values exactly. For odd , the interpolant passes through all samples, but generally differs from the orthogonal projection using the exact coefficients . (dlmf.nist.gov)
The DFT represents the sampled sequence exactly, not necessarily the continuous function between samples. Frequencies whose indices differ by a multiple of have identical values on the sampling grid. Consequently, if the Fourier expansion is absolutely convergent, substitution into the sampling formula yields
This is aliasing: unresolved harmonics enter coefficients assigned to lower frequencies. For example, with , and produce identical samples. Exact recovery therefore requires restrictions on the permitted frequencies; a symmetric band strictly within avoids ambiguity. For even , the two indices and coincide on the grid, requiring a convention for the Nyquist component. (math.mit.edu)
Evaluating all DFT coefficients directly requires arithmetic operations. A fast Fourier transform (FFT) computes the same transform through factorizations and reuse of intermediate sums. In a radix-2 algorithm, splitting the input into even- and odd-indexed samples produces two transforms of half the size, followed by combinations; recursion gives work for power-of-two lengths. Other FFT algorithms accommodate composite and prime lengths. Speed does not remove errors from floating-point arithmetic: rounding and inaccurate phase factors perturb the coefficients, and small coefficients can have large relative errors. Different implementations may consequently return slightly different numerical results. (fftw.org)
Truncation error arises from omitting harmonics, while sampling error includes the aliasing contributions above. For smooth periodic functions, rapid coefficient decay can make both errors small; smoothness must include compatibility across the period boundary. In measured records, an acquisition interval containing a noninteger number of oscillations generally imposes an artificial boundary mismatch. The resulting spectral leakage distributes a sinusoid’s energy among several frequency bins rather than confining it to a single bin. This is a consequence of the finite record and its implied periodic extension, not a failure of the FFT algorithm. (math.mit.edu)
Windowing multiplies the record by a taper, such as a Hann window, to reduce boundary discontinuities and suppress spectral sidelobes. It changes the analyzed signal: reduced sidelobes usually accompany a wider main lobe, and amplitude measurements require attention to the window’s scaling. Leakage, aliasing, and roundoff are distinct from the intrinsic Gibbs phenomenon. Gibbs ringing occurs when a sharply truncated Fourier expansion approximates a jump, even with exact coefficients and exact arithmetic; a boundary mismatch can introduce such a jump, but numerical imprecision is not its cause. (ni.com)
Fourier spectral methods approximate periodic partial differential equations using finitely many harmonics: differentiation multiplies coefficient by , and second differentiation by . Galerkin methods project the equation onto retained modes; pseudospectral methods evaluate nonlinear terms on a grid and transform them back. Multiplication generates higher frequencies, which can alias into retained modes. For quadratic products, the dealiasing rule pads the coefficient arrays with zeros to approximately their original length, evaluates the product on the enlarged grid, transforms back, and discards excess modes. With appropriate cutoff conventions, this recovers the retained coefficients of the projected product without aliasing. (math.cmu.edu)
Fourier Transforms and Generalized Harmonic Expansions
A periodic function has frequencies confined to the discrete set , where , whereas the Fourier transform uses a continuous frequency variable. With the angular-frequency convention
a nonzero, locally integrable periodic function is not integrable over the whole real line, so this integral does not define its transform in the ordinary sense. It does, however, define a tempered distribution, whose transform is
Here is the Dirac distribution concentrated at . The resulting line spectrum encodes the same amplitudes and phases as the Fourier coefficients; it is not an ordinary function with narrow peaks. For example, transforms into . (oca.eu)
Periodization provides a direct bridge between the two representations. If is a Schwartz function—smooth, with all derivatives decreasing faster than every inverse power—then
is a smooth -periodic function. Splitting the coefficient integral into translated intervals gives . Evaluating its absolutely convergent Fourier expansion at zero yields the Poisson summation formula
Thus a sum over a spatial lattice equals a sum over its reciprocal frequency lattice. The Schwartz hypothesis is sufficient, not necessary; weaker versions require explicit convergence conditions. For instance, continuity and integrability, together with summability of the translated suprema and sampled transform values, suffice. Neither side should be assumed convergent for an arbitrary integrable function. (ocw.mit.edu)
For several periodic variables, the natural domain is the torus . Frequencies become integer vectors , and the expansion and coefficients are
These exponentials form an orthonormal basis of . Unlike the one-dimensional symmetric cutoff, multidimensional truncation involves additional choices: rectangles , cubes, or balls . Every increasing exhaustion by finite frequency sets converges in , by completeness and square summability, but this does not establish pointwise convergence or make different cutoff shapes interchangeable pointwise. Absolute summability removes the ordering issue and gives uniform convergence. (math.mit.edu)
Within harmonic analysis, the functions are characters: continuous homomorphisms from the torus into the unit circle. On a general compact abelian group , characters form an orthonormal basis of with respect to normalized Haar measure, the translation-invariant measure satisfying . Coefficients become , indexed by the discrete dual group of characters. A related, but distinct, generalization replaces characters with orthogonal eigenfunctions of an operator. For example, Hermite functions form a basis on the real line adapted to the harmonic-oscillator operator , rather than to translation and periodicity. Such expansions preserve the projection mechanism while changing the domain and meaning of a “mode.” (maths.qmul.ac.uk)
Fourier coefficients also describe objects rougher than functions. Periodic distributions correspond precisely to coefficient sequences of at most polynomial growth, with their expansions interpreted through testing against smooth periodic functions. Smooth functions, conversely, have coefficients decreasing faster than every inverse power. Between these extremes, periodic Sobolev spaces quantify regularity through
For nonnegative integer , this agrees with square integrability of weak derivatives through order ; real orders describe fractional regularity, and negative orders admit distributions. In this formulation, differentiation multiplies coefficients by , while the negative Laplacian multiplies them by . (users.jyu.fi)
References
- DLMF: §1.8 Fourier Seriesdlmf.nist.gov
- Fourier Series for Periodic Functionsmath.mit.edu
- Mathematicians and Music 3mathshistory.st-andrews.ac.uk
- The Beginning of Fourier Seriesbibliotheque.imo.universite-paris-saclay.fr
- Fourier, Josephmathshistory.st-andrews.ac.uk
- Théorie analytique de la chaleur, page 542fr.wikisource.org
- Dirichlet, Peter Gustav Lejeunemathshistory.st-andrews.ac.uk
- Riemann, Georg Friedrich Bernhardmathshistory.st-andrews.ac.uk
- The integralarchimede.dimai.unifi.it
- For Lectures 11-14math.mit.edu
- Chapter 4: Differential and Integral Operatorsmath.mit.edu
- 102 S2021 Complete Lecture Notesocw.mit.edu
- Fourier Series Exampleslpsa.swarthmore.edu
- PowerPoint Presentationmx.nthu.edu.tw
- Partial Differential Equationspeople.uncw.edu
- Real Analysispeople.math.harvard.edu
- An Analogy of the Carleson–Hunt Theorem with Respect to Vilenkin Systemslink.springer.com
- Harmonic Analysis: from Fourier to Haarmath.unm.edu
- LECTURE 8: CONVOLUTIONbpb-us-w2.wpmucdn.com
- ES.1803 S24: Reading: Topic Enrichment: Gibbs phenomenonocw.mit.edu
- DLMF: §6.16 Mathematical Applicationsdlmf.nist.gov
- 102 S2021 Lecture 16. Fejer's Theorem and Convergence of Fourier Seriesocw.mit.edu
- Lecture Notes in Harmonic Analysisbpb-us-w2.wpmucdn.com
- Convergence of Fourier Seriesmath.uchicago.edu
- Lecture Notes on Measure Theorymath.cmu.edu
- Fourier Series, Chapter 7math.ucdavis.edu
- The (Pseudo)Spectral Fourier Expansions — Exponential Accuracymath.umd.edu
- Introduction to Numerical Analysis, Lecture 6ocw.mit.edu
- 085: More on Fourier Seriesmath.mit.edu
- Fourier Seriespeople.math.harvard.edu
- Lecture 7: Wave Equation and Standing Wavesocw.mit.edu
- Problem Set Number 04, 18.306math.mit.edu
- 5 Sources of Musical Soundopenstax.org
- Lecture 9: Frequency Responseocw.mit.edu
- Sinusoidal Steady State Response of Linear Circuitsocw.mit.edu
- DLMF: §3.11 Approximation Techniquesdlmf.nist.gov
- 1. Interpolationmath.mit.edu
- Notes on FFT-based differentiationmath.mit.edu
- The Design and Implementation of FFTW3fftw.org
- Introduction (FFTW 3.3.11)fftw.org
- FFT Accuracy Benchmark Commentsfftw.org
- FFTW FAQ - Section 3fftw.org
- Understanding FFTs and Windowing - NIni.com
- Archived: LabVIEW Analysis Concepts - National Instrumentsdownload.ni.com
- Notes on Mathematical and Computational Fluid Dynamics: Fourier Spectral Methodsmath.cmu.edu
- Fourier analysis courseoca.eu
- RES.18-015 S24 Full Lecture Notes: Topics in Fourier Analysisocw.mit.edu
- Notes for 18.117math.mit.edu
- Volume I: Ergodic Theory with a View Towards Number Theorymaths.qmul.ac.uk
- Fourier analysisusers.jyu.fi
- Periodic Distributionsusers.jyu.fi
- The Sobolev inequality on the torus revisitedpure.ed.ac.uk