Dithering

The primary purpose of dithering is to break the statistical correlation between the quantization error and the input signal. Without dither, quantization error is signal-dependent, creating harmonic distortion and idle tones that are highly audible or spectrally damaging. By adding a noise signal with a specific probability density function (PDF), such as triangular or Gaussian, the error becomes a random, uncorrelated process. This transforms deterministic, non-linear distortion into a benign, broadband noise floor, effectively linearizing the converter's average transfer function and eliminating missing codes.

Two primary architectures exist for dither injection:

• Subtractive Dithering: The dither signal is added before the quantizer and then digitally subtracted from the output. This completely removes the dither noise from the final signal, leaving only the linearization benefit. It requires precise synchronization and is common in high-precision audio mastering.
• Non-Subtractive Dithering: The dither is added but not removed. This is simpler to implement but raises the system's noise floor by a small, controlled amount. It is the standard method in most real-time analog-to-digital conversion systems where a slight SNR penalty is acceptable for the elimination of harmonic distortion.

Dithering fundamentally alters the intrinsic hardware fingerprint of a data converter. By randomizing the quantization process, it masks the static Differential Non-Linearity (DNL) and Integral Non-Linearity (INL) patterns that would otherwise create a unique, deterministic signature. For RF fingerprinting systems, a dithered converter presents a more generic, noise-like output, making device identification more challenging. However, the specific characteristics of the dither source itself—such as its PDF accuracy, bandwidth, and amplitude stability—can become a new, secondary component of the device's unique signature.

The statistical distribution of the dither signal is critical to its effectiveness:

• Rectangular (Uniform) PDF: The simplest to generate but only partially decorrelates the first moment of the quantization error. It can still leave some signal-dependent modulation.
• Triangular PDF: Generated by summing two independent uniform random sources. This is the minimum requirement to fully decorrelate both the first and second moments of the error, making it the standard for high-fidelity audio applications.
• Gaussian PDF: Provides robust decorrelation and is naturally generated by thermal noise sources, making it common in analog dither circuits. It theoretically requires infinite amplitude but is practically clipped.

In Time-Interleaved ADC architectures, dithering serves a dual purpose. Beyond linearizing each individual sub-ADC, a large-amplitude dither can be injected to randomize and measure the interleaving mismatch spurs (gain, offset, and timing skew errors between channels). By correlating the known dither with the digital output, a background calibration engine can continuously estimate and correct these mismatches. This technique, known as correlation-based dithering, dynamically suppresses the periodic, architecture-specific spurs that would otherwise dominate the device's RF fingerprint.

A specialized application involves injecting a dither signal with an amplitude spanning several Least Significant Bits (LSBs). This is specifically used to eliminate missing codes caused by severe DNL errors. The large dither sweeps the input across the dead zone, ensuring that the converter toggles through all codes over time. While this dramatically increases the in-band noise, it guarantees monotonicity and averages out the non-linearity. For fingerprinting, this process actively destroys the most distinctive static identifier—a permanent gap in the transfer function—replacing it with a high-power noise signature.

The inherent difference between an analog input value and its discrete digital representation. This fundamental, signal-dependent noise source is the primary target of dithering. By adding noise, dithering decorrelates this error from the input signal, linearizing the converter's average transfer function. The residual statistical properties of the quantization noise floor, shaped by the converter's architecture, remain a persistent component of the device's fingerprint.

A measure of a data converter's static linearity, defined as the maximum deviation of the actual transfer function from an ideal straight line. While dithering can smooth out the step-like errors of quantization, it does not eliminate the smooth, low-frequency curvature caused by INL. This process-dependent, unique signature remains in the output waveform and is a primary target for high-precision device identification.

The deviation between an actual step width and the ideal 1 Least Significant Bit (LSB) step. Large DNL errors can lead to missing codes—permanent gaps in the transfer function. Dithering effectively 'fills in' these missing codes on average by smearing the signal across adjacent quantization levels, but the underlying DNL pattern still modulates the noise statistics, leaving a distinct, device-specific fingerprint.

The sample-to-sample variation in the precise instant a sample-and-hold circuit captures a signal. This timing uncertainty introduces a phase modulation that is fundamentally different from amplitude noise. Dithering, typically added as amplitude noise, does not mask aperture jitter. Instead, the jitter creates a unique, clock-related phase noise signature that remains a highly robust identifying feature, especially at higher input frequencies.

The ratio of the fundamental signal's RMS amplitude to the highest spurious component in the output spectrum. Dithering is often intentionally applied to break up high-amplitude harmonic spurs caused by static non-linearity, spreading their energy into a noise-like floor. This improves SFDR but replaces a distinct, easily identifiable spur with a shaped noise pedestal that still contains the unique spectral signature of the converter's non-linear transfer function.

A technique used in high-resolution DACs to dynamically scramble the usage of mismatched unit elements. Like dithering, DEM converts static mismatch errors into shaped, high-frequency noise. However, the specific noise-shaping transfer function and the residual in-band noise power are implementation-dependent. The interaction between an applied dither signal and the DEM algorithm creates a complex, highly distinctive noise-shaped residual that is extremely difficult to clone.