Ringing artifacts, also known as the Gibbs phenomenon, are spurious oscillations or ripples that occur near sharp transitions or discontinuities in signals, arising from the approximation of functions using finite Fourier series or transforms in digital signal processing.^[1][2] These artifacts manifest as wavy patterns or "ghosting" effects, with the overshoot typically reaching about 9% of the jump discontinuity height, and their severity decreases with distance from the transition but persists regardless of the number of terms used in the approximation.^[3] In image processing, ringing appears as alternating bright and dark bands around high-contrast edges, commonly introduced by lossy compression algorithms like JPEG, aggressive sharpening filters, or reconstruction from undersampled data in techniques such as MRI or deconvolution.^[1][4] For instance, in magnetic resonance imaging (MRI), they present as parallel lines adjacent to boundaries like the skull-brain interface, resulting from finite k-space sampling and inverse Fourier transformation.^[2] Mitigation strategies include increasing sampling resolution, applying low-pass filters, or using advanced deringing algorithms based on sparse representations or variational methods.^[1][5] In audio signal processing, ringing manifests as damped oscillations or echo-like artifacts following transients, such as percussive sounds in drums or piano attacks, primarily caused by the impulse response of digital filters like equalizers or anti-aliasing filters in digital-to-analog converters.^[6] Pre-ringing (before the transient) and post-ringing (after) can occur, often in ultrasonic frequencies during resampling, and are reduced by employing minimum-phase filters or higher sampling rates to preserve transient fidelity.^[6] Overall, ringing artifacts degrade perceptual quality across domains—from visual distortions in compressed media to audible "smearing" in sound reproduction—and remain a fundamental challenge in bandlimited signal reconstruction, prompting ongoing research into adaptive suppression techniques.^[7][1]

Fundamentals

Definition

Ringing artifacts refer to unwanted oscillations or ripples that manifest as spurious signals near sharp transitions or discontinuities in processed signals, images, or audio.^[8] These distortions arise from limitations in representation techniques, appearing as repetitive waves or halos adjacent to edges or abrupt changes.^[9] They are primarily associated with the Gibbs phenomenon, which occurs in Fourier-based approximations where a finite number of harmonics fails to accurately reconstruct discontinuities, leading to overshoots and oscillations.^[10] This phenomenon, first described in the context of Fourier series convergence, underpins the oscillatory nature of ringing in various transform-based processing methods.^[11] Unlike random noise, which is stochastic and uniformly distributed across the signal, ringing artifacts are periodic and localized, exhibiting structured, deterministic patterns confined to regions of high contrast.^[12] Historically, these have been termed Gibbs ringing, truncation artifacts, or spectral leakage artifacts, reflecting their origins in frequency truncation and windowing effects.^[9] Such artifacts commonly appear in domains like digital imaging and audio signal processing.^[2]

Contexts of Occurrence

Ringing artifacts manifest across multiple technical domains in digital signal processing (DSP), where they arise from band-limiting or filtering operations that approximate discontinuous signals.^[13] In image compression, such as JPEG or HEVC standards, these artifacts appear as oscillatory halos or ripples around sharp edges due to the truncation of high-frequency components during transform coding and quantization.^[14] Similarly, in magnetic resonance imaging (MRI), Gibbs ringing—also termed truncation artifact—presents as parallel lines adjacent to high-contrast boundaries, like the cerebrospinal fluid-spinal cord interface, stemming from finite k-space sampling in the Fourier reconstruction process.^[2] In audio encoding, particularly perceptual coding schemes like MP3, ringing manifests as pre-echo (sound preceding transients) or post-echo (lingering after transients) in block-based compression, where short attack times smear impulsive signals such as percussion strikes across frame boundaries.^[15] For oscilloscope signal measurements, ringing appears as overshoots and oscillations near voltage transitions in high-speed waveforms, induced by the instrument's band-limited response when capturing fast edges, as seen in time-domain reflectometry (TDR) or eye diagram analysis.^[11] Hardware contexts involving analog-to-digital converters (ADCs) and digital-to-analog converters (DACs) exhibit ringing due to finite sampling and reconstruction filtering; for instance, the sinc interpolation in DAC output stages introduces oscillatory transients around step changes, amplifying with steeper filter roll-offs.^[16] In modern applications, super-resolution techniques in AI-generated images often produce ringing near enhanced edges, as neural networks struggle to reconstruct high-frequency details without introducing Gibbs-like oscillations, particularly in medical imaging upscaling.^[17] Likewise, high-speed printed circuit board (PCB) signal integrity analysis reveals ringing in simulations of transmission lines, where discontinuities like vias cause resonant overshoots in gigahertz-range signals, impacting data rates beyond 10 Gbps.^[18]

Causes

Gibbs Phenomenon

The Gibbs phenomenon refers to the oscillatory overshoot and ringing that occur in the partial sums of the Fourier series of a piecewise continuously differentiable periodic function near points of jump discontinuity.^[19] This behavior manifests as persistent ripples on either side of the discontinuity, where the approximation exceeds the true function value before settling toward it.^[20] The phenomenon was first described by Henry Wilbraham in 1848 in a paper examining the convergence of Fourier series for discontinuous functions. It was independently rediscovered and popularized by J. Willard Gibbs in 1899 through his analysis of Fourier series convergence, following earlier experimental observations by Albert Michelson in 1898 using a harmonic analyzer. Although the effect bears Gibbs's name, Wilbraham's earlier work highlighted the non-uniform convergence near discontinuities, laying the groundwork for later mathematical scrutiny.^[19] A key characteristic is that the amplitude of the ringing approaches a fixed overshoot of approximately 9% of the jump height, independent of the number of terms in the series, with the oscillations decaying slowly as one moves away from the discontinuity.^[19] For the classic example of a square wave with a jump discontinuity at $x = 0$, the partial sum $S_N(x)$ exhibits oscillations whose maximum height near the discontinuity is given by

$$\frac{1}{2} + \frac{1}{\pi} \int_0^\pi \frac{\sin t}{t} \, dt \approx 1.089$$

times the half-jump value, where the integral evaluates to the sine integral $\mathrm{Si}(\pi) \approx 1.85194$.^[20] This overshoot persists in the limit as $N \to \infty$, illustrating the failure of pointwise convergence at the discontinuity despite uniform convergence elsewhere for smooth functions.^[21] The Gibbs phenomenon underscores a fundamental limitation in Fourier analysis: it is inherent to any band-limited approximation of functions with sharp transitions, such as ideal low-pass filters, where truncating high-frequency components inevitably introduces these oscillations.^[22] This property explains why exact reconstruction of discontinuous signals requires infinite bandwidth, impacting applications in signal processing where finite representations are unavoidable.^[19]

Filter Impulse Responses

The impulse response of an ideal low-pass filter is given by the sinc function, expressed as $ h(t) = \frac{\sin(\pi f_c t)}{\pi t} $, where $ f_c $ is the cutoff frequency.^[23] This continuous-time response exhibits infinite-duration oscillations, or ringing, that decay inversely with time, arising from the abrupt discontinuity in the filter's frequency response at the cutoff.^[24] In discrete-time systems, finite impulse response (FIR) filters approximate this ideal sinc function by truncating it to a finite length, which introduces approximations that manifest as truncated ringing in the time domain.^[25] This truncation preserves the oscillatory nature near the filter's edges but limits the duration of the ripples, often exacerbating localized artifacts around sharp signal transitions. High-pass and band-pass filters exhibit similar ringing behaviors due to their oscillatory impulse responses, particularly at the cutoff or center frequencies, where the frequency response discontinuities lead to prolonged oscillations in the output.^[26] In image compression schemes like JPEG, quantization of discrete cosine transform (DCT) coefficients effectively imposes band-limiting, which triggers Gibbs-like ringing effects along high-contrast edges.^[27] Non-ideal filters with smoother roll-off characteristics in the frequency domain reduce the amplitude and extent of ringing compared to brick-wall designs, though they cannot fully eliminate the phenomenon inherent to band-limiting operations.^[16]

Analysis

Time Domain Characteristics

In the time domain, ringing artifacts manifest as alternating positive and negative peaks that are symmetric around sharp edges or transients in signals and images. These oscillations arise prominently in the response to abrupt changes, such as step inputs or discontinuities, creating ripple-like patterns that extend away from the transition point. The peaks typically exhibit a sinusoidal character, reflecting the underlying frequency content of the filter involved.^[8][28] For signals processed through bandlimited filters, ringing appears as damped sinusoids oscillating at approximately the filter's cutoff frequency immediately following a step input. The decay of these oscillations is characteristically slow, following an inverse-time envelope rather than a rapid exponential falloff, leading to prolonged ripples that can span multiple cycles. This behavior is caused by the sinc-like impulse responses of ideal filters, which introduce these persistent waves.^[8] In contrast to physical signals, where oscillations are naturally attenuated by damping mechanisms like friction or resistance, ringing artifacts in digital or filtered systems persist indefinitely without such inherent decay, potentially distorting the signal for extended durations.^[28] The severity of ringing is quantified through metrics such as peak overshoot percentage, which measures the maximum deviation beyond the steady-state value (often around 9% for the Gibbs phenomenon in square wave approximations), and ringing duration, defined as the time required for the response to settle within 1% of the final value.^[29] For the step response of an ideal low-pass filter, the ringing component can be approximated asymptotically as

$$\frac{1}{\omega_c t} \cos(\omega_c t),$$

where $\omega_c$ is the cutoff angular frequency; this term highlights the 1/t decay modulating the sinusoidal oscillation at the cutoff frequency.^[30] Such characteristics emphasize the artifact's origin in mathematical truncation rather than physical processes, making it a key indicator of insufficient bandwidth or abrupt filtering.^[8]

Frequency Domain Characteristics

Ringing artifacts in the frequency domain stem from the presence of discontinuities in the time-domain signal, which generate theoretically infinite high-frequency content in the spectrum. Band-limiting this content through truncation, as occurs in practical filtering or sampling processes, results in ripples within the passband and stopband of the frequency response, distorting the ideal rectangular shape of a low-pass filter. This truncation effectively multiplies the ideal infinite sinc impulse response by a rectangular window in the time domain, leading to a convolution in the frequency domain that spreads spectral energy.^[31] The Gibbs phenomenon manifests in the frequency domain as this convolution of the ideal frequency response with the sinc function derived from the truncation window, producing prominent side lobes that leak energy across frequency bands. For a rectangular window, the first side lobe level is approximately -13 dB relative to the main lobe peak, with subsequent lobes rolling off at about 6 dB per octave; this fixed side-lobe structure arises from the abrupt discontinuities at the window edges. The discrete-time Fourier transform (DTFT) of the rectangular window approximates a sinc function, whose envelope decays as $1/\omega$, contributing to the persistent leakage even as the main lobe narrows with increased window length.^[32][33] The frequency response of a truncated sinc filter, given by the convolution of the ideal rectangular spectrum with the window's transform, exhibits ripples whose amplitude remains largely independent of the truncation length $N$ for a rectangular window, though the spacing between ripple peaks scales inversely with $N$. In mathematical terms, for an ideal low-pass filter with cutoff $\omega_c$, the truncated response is:

$$H(\omega) = \frac{1}{2\pi} \int_{-\pi}^{\pi} \operatorname{rect}\left(\frac{\theta}{2\omega_c}\right) \cdot \frac{\sin\left(N(\omega - \theta)/2\right)}{N \sin\left((\omega - \theta)/2\right)} \, d\theta$$

where the Dirichlet kernel approximates the sinc for large $N$, leading to oscillatory deviations from the ideal flat response near the band edges.^[31] Spectral leakage occurs because finite observation windows in the discrete Fourier transform (DFT) cause energy from a true frequency component to spill into adjacent bins, particularly pronounced for signals with components near bin boundaries. This leakage is a direct consequence of the window's non-ideal frequency response, manifesting as ringing during inverse transformation back to the time domain.^[34][35] In DFT and FFT implementations, zero-padding the signal—appending zeros to increase the transform length—interpolates the underlying DTFT, thereby reducing discretization artifacts like scalloping loss but failing to eliminate the core spectral leakage or associated ringing, as the window's side lobes persist.^[36] These frequency-domain imperfections correspond to the temporal ripples observed in the time domain.

Mitigation

Filter Design Improvements

To minimize ringing artifacts inherent in filter designs, particularly those arising from sharp frequency cutoffs, engineers often employ filters with smoother roll-off characteristics instead of ideal brick-wall responses. Ideal low-pass filters, which abruptly transition from passband to stopband, exhibit pronounced ringing due to the Gibbs phenomenon in their time-domain impulse responses. In contrast, Butterworth filters provide a maximally flat passband response with a gradual roll-off, reducing the amplitude of these oscillations by distributing energy more evenly across frequencies. For instance, a higher-order Butterworth filter sharpens the transition while limiting ringing duration, though it may amplify oscillation amplitude if not balanced properly. Chebyshev filters further optimize this by allowing controlled ripple in the passband or stopband to achieve steeper roll-offs with less overall ringing compared to the ideal case, trading minimal distortion for reduced sidelobe effects.^[37][38] A primary linear method to suppress ringing in finite impulse response (FIR) filters involves applying windowing functions to the ideal sinc impulse response, which otherwise produces infinite sidelobes leading to persistent oscillations. Windowing tapers the filter coefficients to zero at the edges, lowering sidelobe levels in the frequency domain and thereby attenuating the ripples associated with ringing. Common windows include the Hamming, Blackman, and Kaiser types; for example, the Hamming window reduces sidelobes to approximately -43 dB, significantly mitigating Gibbs ringing at the cost of a wider transition band. The windowed sinc filter is defined as

$$h[n] = \mathrm{sinc}\left(\frac{n - M/2}{M}\right) \cdot w[n],$$

where $ M $ is the filter length and the Hamming window is given by

$$w[n] = 0.54 - 0.46 \cos\left(\frac{2\pi n}{M}\right), \quad 0 \leq n \leq M.$$

This approach trades filter sharpness for lower ripple amplitude, with unwindowed sinc filters exhibiting more severe ringing near discontinuities.^[25][39] Infinite impulse response (IIR) filter designs can reduce the effective length of ringing by placing poles farther from the unit circle to enable faster exponential decay in the impulse response, though this may require higher orders or compromise sharpness compared to FIR equivalents. This pole placement allows for efficient approximation of desired frequency responses with shorter effective lengths, minimizing transient oscillations. However, IIR filters risk instability if poles lie outside the unit circle, necessitating careful design constraints such as bilinear transformation from stable analog prototypes to ensure bounded outputs.^[40][41] Optimal filter design requires balancing multiple tradeoffs to curb ringing without excessive performance loss: passband flatness to preserve signal integrity, stopband attenuation for effective noise rejection, and transition width to control sharpness versus ripple extent. Narrower transitions exacerbate ringing, while wider ones reduce it but may allow unwanted frequencies to leak through; thus, the design selects parameters like order and window type to meet specifications while keeping oscillations below perceptible thresholds.^[42]

Post-Processing Techniques

Post-processing techniques for ringing artifacts focus on suppressing oscillations in signals or images after initial acquisition or compression, targeting remnants of phenomena like the Gibbs effect without altering upstream processes. De-ringing filters, such as those using adaptive median operations, dampen periodic ripples by selectively smoothing oscillatory regions while preserving sharp transitions. These filters adapt to local signal characteristics, applying median-based smoothing only where oscillations exceed thresholds derived from neighborhood statistics, thereby minimizing edge degradation. Morphological operations, including adaptive directional morphological filters (ADMF), further enhance this by employing erosion and dilation in edge-aligned directions to erode ringing halos around discontinuities in compressed images like those from JPEG-2000, achieving perceptual quality improvements without introducing blur.^[43][44] Wavelet-based denoising represents another cornerstone of post-processing, where the signal is decomposed into wavelet coefficients, and thresholding is applied to suppress high-frequency components responsible for ripple-like artifacts. This approach exploits the sparsity of wavelet representations to attenuate Gibbs-induced ringing by setting small coefficients near edges to zero or shrinking them adaptively, often using soft-thresholding functions tuned to noise levels. A notable method reverses the standard wavelet denoising pipeline to explicitly target and reconstruct ringing-free subbands, reducing artifact visibility in denoised images while maintaining overall fidelity. Such techniques are particularly effective in transform-coded signals, where they mitigate ripples originating from truncated basis functions.^[45][46] Advancements in machine learning have introduced convolutional neural networks (CNNs) trained on datasets of artifact-affected and clean images for targeted de-ringing via inpainting and restoration. Post-2018 developments, especially in super-resolution and video compression, leverage deep architectures like sub-layered deeper CNNs (SDCNN) to predict and correct ringing in block-based codecs such as HEVC, using strided convolutions for efficient feature extraction and mean squared error optimization during training. These models outperform traditional filters by learning contextual patterns, yielding bit-rate reductions of up to 4.1% and PSNR gains of 2-3 dB on benchmark sequences. Recent advancements as of 2023 include attention-based CNNs for Gibbs-ringing suppression in MRI.^[47][48] In MRI applications, parallel imaging accelerates k-space sampling to undersample high frequencies less severely, reducing truncation ringing, while direct k-space filtering applies low-pass attenuation to suppress Gibbs artifacts, though it risks over-smoothing. CNN variants trained on simulated ringing further refine MRI reconstructions, eliminating artifacts with minimal structural loss.^[49][50] A key trade-off in these post-processing methods is the balance between artifact suppression and preservation of fine details, where aggressive filtering can induce blurring, as quantified by metrics like peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). For instance, wavelet thresholding may reduce ringing by 20-30% in PSNR terms but requires careful parameter tuning to avoid low-frequency attenuation, while CNN-based approaches demonstrate superior SSIM scores (e.g., 0.95+ on compressed videos) yet demand computational resources for inference. Effectiveness is evaluated through such metrics on standardized datasets, ensuring ringing reduction enhances perceptual quality without compromising diagnostic or visual utility.^[47][48][45]

Examples

In Image Processing

In JPEG compression, ringing artifacts manifest as halo-like oscillations around sharp edges and high-contrast boundaries, resulting from the quantization of Discrete Cosine Transform (DCT) coefficients in 8x8 pixel blocks.