Quantization Error

Unlike thermal noise, quantization error is deterministically correlated with the input signal. For low-resolution converters or simple periodic inputs like sine waves, the error creates harmonic distortion rather than a flat noise floor. This correlation is the primary reason quantization error is classified as distortion, not noise. The specific harmonic structure depends on the converter's transfer function non-linearity and can serve as an identifying feature.

• For an ideal ADC, error is bounded by ±0.5 LSB
• Correlation is strongest for low-amplitude, periodic signals
• Dithering is used to decorrelate the error from the input

Under the standard model, quantization error is assumed to have a uniform probability density function over the range [-q/2, q/2], where q is the Least Significant Bit (LSB) size. This assumption holds only when the input signal is complex and spans many quantization levels without saturating the converter. The variance of this error is q²/12, which defines the theoretical signal-to-quantization-noise ratio (SQNR).

• SQNR for a full-scale sine wave: 6.02N + 1.76 dB
• Assumption breaks down for low-amplitude or DC inputs
• Differential Non-Linearity (DNL) distorts this uniform distribution

Without dithering, quantization error power is concentrated at harmonic frequencies of the input. Adding a small amount of non-subtractive wideband noise (dither) before the quantizer randomizes the error, transforming harmonic distortion into a spectrally white noise floor. This is critical for applications requiring high spectral purity. The specific dithering strategy—such as triangular PDF dither—can eliminate first-order statistical moments of the error.

• Triangular PDF dither eliminates noise modulation
• Subtractive dither can be removed digitally for higher SQNR
• Dithering modifies the device's intrinsic fingerprint

The finite resolution of a converter creates a fundamental granularity limit. When Differential Non-Linearity (DNL) exceeds 1 LSB, the converter can skip output codes entirely, a phenomenon known as missing codes. This creates a permanent, non-monotonic gap in the transfer function. Missing codes are a severe, highly distinctive hardware impairment that directly alters the quantization error profile and serves as a robust, unclonable device identifier.

• DNL < -1 LSB guarantees missing codes
• Creates a dead zone in the digital output range
• A strong, static feature for RF fingerprinting

In Sigma-Delta converters, quantization error is not minimized but spectrally shaped. A feedback loop pushes the error power out of the band of interest into higher frequencies, where it is filtered digitally. This technique achieves high in-band resolution with a low-resolution quantizer. The resulting noise transfer function (NTF) is a designed characteristic, but its implementation is affected by analog non-idealities, creating a unique residual signature.

• 1st-order shaping provides 9 dB/octave improvement
• Mismatch shaping in multi-bit DACs further sculpts the error
• Out-of-band error must be removed by a decimation filter

In time-interleaved ADCs, multiple sub-converters sample sequentially. Mismatches in gain, offset, and sampling clock phase between these sub-ADCs create deterministic, periodic errors that appear as fixed spurs in the output spectrum. These spurs are not random quantization noise but a direct product of hardware imperfection, making them a dominant and highly exploitable component of a device's RF fingerprint.

• Gain mismatch spurs appear at f_s/M ± f_in
• Timing skew spurs scale with input frequency
• Spurs are deterministic and repeatable

A severe manifestation of Differential Non-Linearity (DNL) where a data converter completely skips one or more digital output codes. No analog input voltage will ever produce the missing code.

• Occurs when DNL ≤ -1 LSB at a specific code transition.
• Creates a permanent, highly distinctive gap in the transfer function.
• Missing codes are a strong identifying feature for fingerprinting, as they represent a hard, non-recoverable non-ideality that is extremely unlikely to be identical across two devices.