Quantization Noise Shaping

The Noise Transfer Function defines the spectral filter applied to the quantization error. It is designed to have high attenuation in the signal band and high gain out-of-band. In a sigma-delta modulator, the NTF is typically a high-pass filter, shaping low-frequency noise away from a DC or low-IF signal. The order of the NTF determines the slope of the noise shaping, with higher orders providing more aggressive in-band noise suppression at the cost of potential stability issues.

The Oversampling Ratio is the ratio of the sampling frequency to the Nyquist rate. It is a critical parameter that enables noise shaping. By sampling much faster than required, the total quantization noise power is spread over a wider bandwidth. The noise shaper then pushes the majority of this power outside the narrower signal band. Doubling the OSR, combined with the noise shaping order, yields a significant gain in effective number of bits (ENOB).

The core architecture for noise shaping is the sigma-delta modulator. It consists of a loop filter, a coarse quantizer, and a feedback DAC. The difference between the input and the feedback signal is integrated, creating an error signal that is quantized. This closed-loop structure inherently shapes the quantization noise according to the loop filter's characteristics, forcing the in-band noise to be minimal while the out-of-band noise is high and later removed by a digital decimation filter.

For high-order noise shaping loops, stability is a primary design constraint. Aggressive NTF designs can cause the quantizer to overload, leading to large, unbounded oscillations. Designers must scale the NTF's out-of-band gain to prevent instability. Additionally, idle tones or limit cycles can occur when the input is a DC or very low-level signal, producing periodic patterns in the output that manifest as spurious tones in the spectrum. Dithering is often used to break these cycles.

The internal quantizer can be single-bit (a simple comparator) or multi-bit. Single-bit quantizers are inherently linear, eliminating DAC non-linearity errors in the feedback path. However, they produce a high level of out-of-band noise. Multi-bit quantizers reduce this out-of-band noise and improve loop stability, but require a highly linear feedback DAC. Dynamic element matching (DEM) techniques are often employed to linearize the multi-bit DAC by scrambling the usage of its unit elements.

The output of a noise-shaping modulator is a high-rate, shaped-noise bitstream. A digital decimation filter follows the modulator to remove the out-of-band noise and reduce the sample rate to the Nyquist rate. This is typically a multi-stage process: a coarse CIC filter first decimates the high-speed data, followed by FIR compensation and sharp-cutoff filters to produce the final high-resolution, band-limited digital output.

The core architecture that enables noise shaping. A sigma-delta modulator uses oversampling and a feedback loop containing a quantizer and a loop filter to push quantization error to high frequencies. The loop filter's order determines the noise shaping slope—a first-order modulator provides 9 dB/octave improvement, while a third-order modulator achieves 27 dB/octave. The feedback structure spectrally separates the input signal from the quantization noise, making subsequent digital filtering straightforward.

• Oversampling Ratio (OSR): The ratio of sampling frequency to Nyquist rate, typically 64x–256x in audio converters
• Stability: Higher-order loops (>2nd order) require careful coefficient design to prevent oscillation
• Idle Tones: Limit cycles that produce audible artifacts in low-order modulators without dithering

The mathematical function that describes how quantization noise is spectrally shaped by the modulator. The NTF acts as a high-pass filter on the quantization error while the Signal Transfer Function (STF) passes the input unchanged. Designing the NTF involves placing zeros at DC and optimizing pole locations to maximize in-band noise attenuation while maintaining modulator stability.

• Zero Optimization: Spreading NTF zeros across the signal band rather than clustering at DC improves in-band noise suppression
• Lee's Rule: The peak NTF gain should remain below 1.5 to ensure stable operation
• Complex NTF: Bandpass modulators use complex NTFs to shape noise away from a non-zero intermediate frequency

The digital filter stage following the modulator that removes the shaped out-of-band noise and reduces the sample rate to the Nyquist rate. A decimation filter chain typically consists of a CIC filter for initial high-rate decimation, followed by FIR compensation filters and half-band filters for final downsampling. The filter's stopband attenuation must match the noise floor requirements of the target application.

• Cascaded Integrator-Comb (CIC): Multiplier-less first stage efficient at high OSRs
• Passband Droop: CIC filters introduce magnitude roll-off requiring compensation
• Group Delay: Linear-phase FIR designs preserve signal waveform shape through the decimation chain

The dynamic range metric that quantifies the real-world resolution of a noise-shaped converter after accounting for all noise and distortion sources. ENOB is calculated from the Signal-to-Noise and Distortion Ratio (SINAD) and directly reflects the in-band noise floor achieved by the noise shaping loop. A 1-bit sigma-delta converter with aggressive noise shaping can achieve 20+ ENOB within a narrow bandwidth.

• Formula: ENOB = (SINAD_dB - 1.76) / 6.02
• Bandwidth Trade-off: ENOB decreases as signal bandwidth increases for a fixed sampling rate
• Thermal Noise Limit: At very high OSRs, thermal noise from the input stage dominates over shaped quantization noise

The intentional injection of a small random or pseudo-random signal before the quantizer to break up idle tones and limit cycles in sigma-delta modulators. Dithering linearizes the quantizer's transfer function, making the quantization error behave more like true white noise—a critical assumption for the linear model of noise shaping.

• Subtractive Dither: Dither added before quantization and subtracted digitally afterward, eliminating dither noise from the output
• Non-Subtractive Dither: Simpler implementation that raises the noise floor slightly but eliminates tonal artifacts
• Self-Dithering: High-order modulators with sufficient internal activity may not require explicit dithering

An architecture variant that replaces the single-bit quantizer with a multi-bit ADC (typically 3–5 bits) inside the feedback loop. Multi-bit modulators achieve lower quantization noise power, improved stability, and reduced sensitivity to clock jitter. The trade-off is DAC non-linearity in the feedback path, which directly degrades the signal path unless corrected by Dynamic Element Matching (DEM).

• DEM Algorithms: Scramble DAC element usage to convert static mismatch errors into shaped noise
• Jitter Immunity: Multi-bit feedback reduces the slew rate of the error signal, lowering jitter sensitivity
• Stability Margin: The reduced quantizer gain variation simplifies high-order loop stabilization