The Gibbs phenomenon is the oscillatory overshoot and undershoot that occur near a jump discontinuity when a function is approximated by a finite Fourier series. As more terms are retained, the oscillations become concentrated in a narrower region around the jump, but the largest overshoot approaches a nonzero fraction of the jump size—approximately 8.949% for ordinary Fourier partial sums. The phenomenon illustrates how an approximation can converge at every fixed point while retaining a persistent maximum error near a discontinuity. Related effects occur in truncated Fourier transforms and other spectral approximations.
Mathematical setting
Let be a -periodic, piecewise smooth function. Its symmetric Fourier partial sum is
Suppose has a jump at , with one-sided limits
and signed jump size
Under standard piecewise smoothness assumptions, the Fourier series converges at the jump to the midpoint:
This limit is independent of the value assigned to , since changing a function at a single point does not change its Fourier coefficients.
The Gibbs phenomenon concerns the behavior near the jump, not a failure of convergence at the jump itself. For an upward jump, nearby partial sums exceed the upper limiting value and fall below the lower limiting value . Increasing moves these excursions closer to without eliminating their limiting relative height.
The universal overshoot constant
Define the sine integral by
where the integrand is defined to be at .
For a piecewise smooth function with an isolated jump, the local shape of its Fourier partial sums has the scaling limit
for each fixed real . This formula describes the transition layer on the spatial scale .
For an upward jump, the first maximum of the limiting profile occurs at , because
The limiting amount by which that maximum exceeds is
The Gibbs–Wilbraham constant is therefore
Consequently:
- the overshoot above the upper side approaches ;
- the undershoot below the lower side approaches ;
- the span between the first upper and lower extrema approaches , approximately .
The percentage is measured relative to the jump size, not necessarily to either plateau value. For a jump from to , the jump size is , so the upper peak approaches approximately : its excess above is about of the upper plateau value but of the jump.
These constants apply to ordinary, sharply truncated Fourier partial sums. Other summation rules or spectral filters can produce different overshoot behavior.
Example: the square wave
Consider the -periodic square wave
Its Fourier expansion is
Retaining the first nonzero terms gives
Its derivative is
with the quotient interpreted by continuity where appropriate. The first positive maximum occurs at
As ,
Meanwhile, . Thus the peak retains a nonzero excess while moving toward the discontinuity.
Why sharp truncation causes ringing
Fourier partial sums can be written as a convolution:
where the Dirichlet kernel is
The kernel has a narrow central peak surrounded by alternating positive and negative lobes. It is therefore not a positive local average. Near a discontinuity, its signed contributions sample both sides of the jump and create oscillations that extend beyond the local range of the function.
In frequency terms, forming preserves all modes with and discards the rest abruptly. This sharp frequency cutoff produces an oscillatory spatial kernel. Integrating the locally rescaled kernel across a step yields the sine-integral transition profile.
An abrupt jump also produces Fourier coefficients whose leading decay is proportional to . This relatively slow decay means that many modes are required to represent the discontinuity. Smoothness permits faster coefficient decay and generally improves approximation accuracy, although discontinuities in derivatives can still produce weaker oscillatory effects.
Relation to convergence
Pointwise convergence
The Gibbs phenomenon is compatible with pointwise convergence. At a fixed point away from a jump, the partial sums converge to under the usual piecewise smoothness assumptions. The largest nearby extrema occur at points that move with , rather than remaining fixed.
The distinction is between taking at a fixed and examining on the shrinking scale
The latter retains the nontrivial transition profile.
Failure of uniform convergence
Every Fourier partial sum is a continuous function. A uniform limit of continuous functions is continuous, so ordinary partial sums cannot converge uniformly to a discontinuous function across its jump. The persistent Gibbs overshoot provides a quantitative manifestation of this failure of uniform convergence.
For sufficiently regular piecewise smooth functions, uniform convergence can nevertheless hold on closed intervals that stay a positive distance from every jump.
Mean-square convergence
For a square-integrable periodic function, Fourier partial sums converge in the sense:
This remains true when has jump discontinuities. Persistent pointwise peak errors need not prevent integrated error from vanishing, because the regions containing the largest errors become narrower.
From the viewpoint of Hilbert space theory, is the orthogonal projection of onto the finite-dimensional space spanned by the retained Fourier modes. Its optimality in mean-square error does not imply small maximum error near a jump.
Suppression and alternative approximations
Fejér averaging
Instead of using a single partial sum, one can average the first partial sums:
These are the Fejér means. Their convolution kernel is nonnegative and has normalized total mass, so they act as weighted averages rather than signed averages.
For a bounded real-valued function, Fejér means remain within its global range. In particular, they eliminate the classical range overshoot of a two-level square wave. They do not make the transition discontinuous: smoothing and finite transition width remain. For continuous periodic functions, Fejér means converge uniformly.
Spectral filtering
A filtered approximation has the form
where the filter reduces high-frequency coefficients. A gradual taper generally reduces ringing compared with a sharp cutoff, but the result depends on the filter. Suppressing oscillations may broaden sharp transitions or reduce resolution, and not every filter guarantees an overshoot-free reconstruction.
Piecewise and edge-aware methods
When jump locations are known or can be estimated, a reconstruction can treat smooth regions separately. Piecewise approximation, domain decomposition, and specialized reconstruction methods can avoid requiring one global smooth expansion to represent a jump directly.
These approaches introduce additional assumptions or computational steps. Errors in locating discontinuities, limited data, and noise can affect their performance.
Applications and related effects
The Gibbs phenomenon is important in numerical analysis, signal reconstruction, and image reconstruction.
- Spectral computation. Global Fourier approximations used to solve partial differential equations can ring near shocks or other discontinuities. Additional stabilization may be needed, especially when subsequent calculations amplify oscillations.
- Signal processing. Abruptly limiting a signal's frequency content creates ringing near sudden changes. This is a trade-off between frequency selectivity and spatial or temporal localization.
- Image reconstruction. Truncated Fourier data can produce bright and dark bands near sharp edges. Such truncation artifacts occur, for example, in magnetic resonance imaging.
- Other global expansions. Related effects occur in approximations based on orthogonal polynomials, although the detailed profile and constants depend on the basis, location, and truncation scheme.
Gibbs ringing is distinct from aliasing: the former arises from truncating an expansion, whereas the latter concerns the inability of sampled data to distinguish certain frequencies. It is also distinct from the Runge phenomenon in high-degree polynomial interpolation. The fast Fourier transform may efficiently compute data used in a reconstruction, but the Gibbs phenomenon is not a numerical defect of that algorithm.
History
Henry Wilbraham analyzed the persistent overshoot of a trigonometric approximation to a discontinuous function in 1848. His work preceded the better-known discussion associated with the American mathematical physicist Josiah Willard Gibbs.
In 1898, Albert A. Michelson's work on harmonic analysis prompted correspondence about the behavior of Fourier approximations near discontinuities. Gibbs discussed the issue in 1898 and clarified the limiting overshoot in 1899. The name Gibbs phenomenon became standard, while Gibbs–Wilbraham phenomenon acknowledges Wilbraham's earlier analysis.