Quantization Noise Extraction Methods in MEMS Gyroscope Allan Deviation Plots
Extract quantization noise Q from Allan plots by extending the minus-one log-log slope to root three seconds, isolating digital step limits from sensor rate noise.

Digitization
A capacitive MEMS sensing element translates Coriolis force into a minute differential capacitance across sub-micron comb fingers. That variable capacitance undergoes continuous-time charge integration, synchronous demodulation, and low-noise baseband amplification before entering the analog-to-digital stage. Silicon micro-structures govern mechanical resonance and Brownian agitation, yet the electronic boundary where continuous mechanical rate transforms into discrete numeric registers establishes the high-frequency stability limit.
When evaluating inertial grade specifications, instrumentation engineers inspect Allan deviation plots to disentangle physical sensor dynamics from the artifacts introduced by this discrete conversion process.

Front-End Preamplification and Delta-Sigma Conversion Limits
Proof-mass displacement generates picofarad-level capacitance shifts that require immediate impedance conversion. High-impedance charge amplifiers translate these charge fluctuations into voltage swings. Modern high-bandwidth inertial measurement units route this conditioned analog signal through a multi-bit delta-sigma converter operating at oversampling frequencies between 100 kHz and 2 MHz.
The analog front end presents a finite input-referred noise voltage density, commonly between 5 nV/√Hz and 30 nV/√Hz in tactical designs, that contends directly with the voltage resolution of the least significant bit.
In a purely analog rate sensor, the high-frequency asymptotic noise floor consists predominantly of wideband Johnson-Nyquist electronic noise and mechanical thermal Brownian motion, both exhibiting a flat rate power spectral density. When the signal undergoes analog-to-digital transformation with finite bit depth, rounding introduces an error sequence bounded between negative one-half and positive one-half of the least significant bit step. Under ideal conditions where the input signal traverses multiple quantization levels with pseudo-random behavior, this conversion error models as an uncorrelated white sequence distributed uniformly across the Nyquist bandwidth.
A 24-bit converter operating at 1000 Hz with a 300 degree per second full-scale range generates an ideal quantization voltage step of 35.7 nanovolts per bit under room temperature conditions.
Sensor packaging and board-level power supply noise often inject periodic spurs into the comparator ladder. If the analog noise floor fails to span at least one complete bit step, quantization error correlates strongly with the instantaneous input rate rather than behaving as independent white noise. Discrete limit cycles, dead bands, and harmonic tones emerge across the frequency domain.
While averaging suppresses uncorrelated broadband jitter, truncation injects deterministic harmonic errors. High-resolution conversion stages intentionally maintain an analog front-end noise density that slightly dithers the input across two to three quantization bins, preserving the statistical independence of the conversion residuals.

Mathematical Form of Quantization Noise in Allan Plots
Sampling an angular rate signal at discrete time intervals introduces a round-off error bounded by the least significant bit step. Allan variance uses two-sample variance techniques developed to characterize frequency and phase instability across expanding averaging intervals tau. Integrating angular rate data over time yields cumulative angle increments, and quantization occurring on this angle domain generates statistical signatures distinct from quantization occurring directly on the rate domain.
When the digitizer samples integrated angle directly, discrete step increments of size delta produce high-frequency white phase noise. Transforming white phase noise into the equivalent rate signal involves a discrete time differentiation, which scales power spectral density by the square of frequency. IEEE Std 952 establishes that white noise in the angle domain yields an Allan variance curve with a characteristic slope of minus two, translating to an Allan deviation slope of minus one on logarithmic coordinates.
The theoretical relationship links the Allan deviation to the angle quantization parameter through a direct mathematical constant:
σ(τ) = (√3 · Q) / τ
In this formulation, Q denotes the angle quantization noise coefficient expressed in radians or degrees, and tau denotes the cluster integration time in seconds. When the analog-to-digital converter digitizes angular rate without prior hardware integration, the rounding error behaves as white rate noise, producing an angle random walk signature characterized by a log-log slope of minus one-half. True angle quantization manifests strictly as a minus-one slope on the Allan deviation plot, dominating the shortest integration times where cluster durations approach the fundamental sampling period.
When the front-end amplifier noise floor sits well above the least significant bit, the digital boundary remains completely invisible across the measurement record.

Slope
Logarithmic Allan variance plots reveal distinct physical noise processes through characteristic power-law alignments. The curve plots two-sample root variance against integration time tau, sweeping across multiple decades from milliseconds to tens of thousands of seconds. Each underlying physical mechanism populates a specific gradient regime.
At intermediate integration times between one second and one hundred seconds, angle random walk dominates with a slope of minus one-half. At long integration spans beyond one thousand seconds, rate random walk produces a positive slope of plus one-half, while bias instability forms a flat plateau with zero slope. Identifying true quantization noise requires isolating the steep minus-one gradient that appears exclusively at the earliest cluster intervals.

Asymptotic Line Fitting at Unity Integration Tau
Cluster times spanning the shortest sample multiples isolate high-frequency phase fluctuations from broadband rate variations. Plotting the Allan deviation equation for angle quantization ~ sigma equal to the square root of three times Q divided by tau ~ on logarithmic axes yields a straight line with a constant negative unity slope. Extending an asymptotic tangent line along this minus-one slope to the vertical axis where integration time tau equals the square root of three seconds yields the numerical value of Q directly.
Alternatively, projecting the line to tau equals one second yields a magnitude equal to the square root of three multiplied by Q.
Reliable extraction depends on separating this steep gradient from adjacent processes. If the analog front-end exhibits significant angle random walk, the minus one-half slope intersects the minus-one slope at very short cluster times. The extraction procedure fails if the curve lacks a distinct minus-one linear section covering at least a three-to-one ratio of integration times.
Analysts fit a straight line via weighted least squares exclusively across the points demonstrating a verified minus-one derivative, avoiding transition zones where mixed noise terms distort the slope.
A verified minus-one logarithmic slope indicates high-frequency angle quantization rather than wideband white rate noise across the measurement cluster.
Extracting the quantization parameter accurately requires executing a strict data handling workflow from raw acquisition through algorithmic curve fitting:
- Raw Data Logging captures un-decimated angular rate words at the maximum native output data rate without host-side buffering or averaging filters.
- Cumulative Angle Computation integrates rate samples into discrete angle series using trapezoidal or rectangular numerical summation.
- Overlapping Cluster Calculation computes two-sample Allan variance across octave-spaced or decade-spaced averaging intervals tau to maximize statistical degrees of freedom.
- Logarithmic Gradient Evaluation calculates the local derivative of log sigma against log tau to identify the exact span where the slope matches negative unity.
- Asymptotic Parameter Isolation projects the identified linear segment to tau equals one second, dividing the resulting intercept value by the square root of three to extract Q.

Worked Parameter Extraction across High-Rate Sensor Output
Assume a capacitive Coriolis vibratory element operating with a dynamic range of 300 degrees per second. The analog conditioning stage couples to a 16-bit successive approximation register converter operating at a native sampling rate of 10,000 samples per second. The fundamental sample interval tau zero equals 0.0001 seconds.
The converter maps full scale across 65,536 discrete states, yielding an angular rate least significant bit resolution of 0.009155 degrees per second per count. The sensor remains stationary inside a passive thermal insulation chamber mounted on an isolated seismic foundation for sixteen consecutive hours.
Rectangular numerical integration translates the 576,000,000 collected rate points into an accumulated angle record. Calculating the overlapping Allan deviation across cluster intervals starting at tau zero equals 0.0001 seconds reveals the noise profile. At tau equals 0.0001 seconds, the computed Allan deviation reaches 0.4578 degrees per second.
At tau equals 0.0002 seconds, the deviation drops to 0.2289 degrees per second. At tau equals 0.001 seconds, the value measures 0.0458 degrees per second. Computing the slope between tau equals 0.0001 seconds and tau equals 0.001 seconds yields exactly minus 1.00 on logarithmic scales.
Fitting a straight line through these initial points establishes the asymptotic equation. Evaluating this fitted line at tau equals one second returns an extrapolated Allan deviation of 0.0000458 degrees per second, equivalent to 0.1648 arcseconds per second. Dividing this extrapolated magnitude by the square root of three (1.73205) isolates the quantization noise coefficient Q as 0.0000264 degrees, or 0.0952 arcseconds.
When evaluated against the theoretical step limit, an angle step of delta translates to a theoretical quantization variance of delta squared divided by twelve. The empirical extraction matches this physical converter bound within five percent, confirming that high-frequency phase variance originates purely from converter resolution limits.
| Sensor Classification | Native Rate (Hz) | LSB Resolution (deg/s) | Tau Range for Slope -1 (s) | Extracted Q (arcsec) | ARW Intercept (deg/√hr) |
|---|---|---|---|---|---|
| Automotive Grade MEMS | 400 | 0.062500 | 0.0025 to 0.0050 | 32.25 | 1.250 |
| Industrial Grade MEMS | 2000 | 0.003815 | 0.0005 to 0.0040 | 1.97 | 0.180 |
| Tactical Grade MEMS | 10000 | 0.000152 | 0.0001 to 0.0020 | 0.08 | 0.015 |
| Optical Reference (FOG) | 50000 | 0.000010 | 0.00002 to 0.0010 | 0.005 | 0.001 |
When digital filtering engages, the boundary shifts. If internal decimation lowers the transmitted data rate, the visible segment of the minus-one slope contracts and its intercept shifts. High-frequency spectral energy folds back into the passband, redistributing quantization power into adjacent averaging bins.
Extracting quantization noise from an inadequate sampling window misallocates the Kalman filter measurement covariance matrix, causing navigation solutions to reject valid short-term corrections and drift during high-bandwidth maneuvers.

Disambiguation
Standard two-sample frequency stability metrics treat high-frequency white phase jitter and discrete conversion rounding as identical mathematical entities. In standard Allan deviation graphs, both processes exhibit an identical slope of negative unity. When an instrument engineer observes a steep downward trend at short cluster times, standard formulations cannot confirm whether the underlying mechanism stems from analog charge amplifier thermal noise or discrete digital register clipping.
Resolving this ambiguity demands specialized mathematical formulations and rigorous sampling protocol controls.

Why Does Truncation Corrupt Allan Deviation Quantization Slopes?
Firmware routines that discard low-order register bytes alter the statistical distribution of the sensor output word. Truncating an internal 24-bit delta-sigma accumulator down to a 16-bit register transmitted over SPI introduces a deterministic non-linear transfer function that creates false rate steps. Rather than maintaining zero-mean Gaussian properties, the truncation residual develops strong cross-correlation with angular acceleration profiles, corrupting short cluster intervals with phase noise.
When calculating two-sample variances across truncated series, differences between adjacent cluster averages fail to cancel out higher-order terms. Spurious energy accumulates at cluster intervals corresponding to the truncation threshold. On the logarithmic deviation plot, the expected smooth minus-one slope develops non-monotonic undulations or steepens beyond negative unity, sometimes mimicking colored non-stationary noise.
If the analog noise floor falls below the truncated least significant bit rather than masking digital stepping, the sensor output toggles intermittently across fixed numerical codes, destroying the statistical validity of standard least-squares line fitting.
Standard Allan deviation formulations produce identical negative unity gradients for broadband white phase modulation and discrete angle quantization.

Modified Allan Variance Application in Phase Noise Separation
Phase-averaging operations introduce an additional time-averaging step into the stability calculation. The Modified Allan Variance (MVAR), and its corresponding square root metric MDEV, was developed specifically to separate white phase modulation from flicker phase modulation. While standard Allan variance performs rate averaging across block lengths of tau, the modified formulation averages phase data across all intermediate points within the cluster block.
This internal phase averaging alters the resulting power-law slope dependencies. Under standard Allan deviation, white phase noise and angle quantization noise both follow a slope of minus one. Under the modified Allan deviation, white phase noise transitions to a slope of minus three-halves (-1.5).
White rate noise (angle random walk) maintains a constant slope of minus one-half (-0.5) under both metrics. Calculating both curves simultaneously across identical raw data streams enables immediate diagnostic separation.
| Noise Mechanism | Rate PSD Exponent (f^alpha) | ADEV Log Slope (mu) | MDEV Log Slope (mu) | Physical Root Cause |
|---|---|---|---|---|
| Angle Quantization | +2 | -1.0 | -1.5 | Discrete ADC bit depth limits |
| White Phase Noise | +2 | -1.0 | -1.5 | Amplifier wideband thermal noise |
| Flicker Phase Noise | +1 | -1.0 | -1.0 | High-frequency resonator feedback |
| Angle Random Walk | 0 | -0.5 | -0.5 | Proof-mass Brownian agitation |
| Bias Instability | -1 | 0.0 | 0.0 | 1/f flicker electronic noise |
| Rate Random Walk | -2 | +0.5 | +0.5 | Thermal gradient variations |
If an unknown sensor shows a minus-one slope on standard ADEV and converts to a minus 1.5 slope on MDEV, the process belongs to high-frequency phase modulation. Separating electronic white phase noise from true quantization then relies on amplitude scaling. Quantization noise scales directly with the square of the programmed bit width delta, while electronic phase noise remains constant regardless of digital word formatting.
Shifting digital output scaling during bench characterization isolates the precise electronic contribution.
The boundary where algorithmic oversampling ceases to suppress genuine silicon quantization and begins distorting resonant mechanical modes remains undefined across multi-axis consumer drive architectures.

Trace
Printed circuit routing and serial bus serialization establish physical boundaries between the conversion core and external processing engines. Parasitic capacitances across printed circuit board conductors introduce transmission line reflections, rise-time degradation, and clock skew on high-speed serial peripheral interface buses. These hardware imperfections frequently corrupt bitstream integrity before data registers reach host memory.
A signal-chain engineer inspecting noisy deviation plots must look beyond mathematical abstraction and evaluate the physical copper tracks, digital decimation cascades, and register structures that convey raw conversion packets.

Digital Interface Decimation and Word Length Rounding
On-chip finite impulse response filters condition raw delta-sigma bitstreams before writing to internal registers. Because delta-sigma conversion relies on excessive sampling rates to push quantization noise outside the operational passband, decimation resets the effective sampling boundary. The on-chip DSP core applies cascaded integrator-comb filtering followed by multi-tap finite impulse response stages, down-sampling the high-frequency bitstream to user-selectable output rates between 50 Hz and 4000 Hz.
Internal decimation filters expand word width, so a 16-bit physical converter operating through an averaging filter can generate internal math registers extending to 24 or 32 bits. If the on-chip processor truncates these accumulated bits to fit a standard 16-bit output register, the decimation stage re-injects high-frequency quantization noise into the transmission stream, shifting the effective noise floor upward. Engineers evaluating Allan deviation plots across varying programmed output data rates observe the minus-one slope disappearing entirely at low output rates, replaced by an artificially elevated white rate noise floor caused by filter round-off operations.

Register Truncation and Fixed-Point Formatting Failures
Microcontrollers pulling angular rate frames over high-speed buses often compress twenty-four-bit samples into sixteen-bit integers, silently dropping lower bits. Software stacks implementing fixed-point math frequently drop the lowest byte to conserve direct memory access bandwidth. When low bits vanish, the effective step size delta expands by a factor of 256 for an eight-bit drop, increasing the equivalent quantization noise energy by a factor of 65,536.
Specific board-level and software defects systematically distort the extracted noise parameters:
- Bit Alignment Skew occurs when the host SPI driver misinterprets two’s complement sign extension, causing arbitrary bit-shifting that corrupts the dynamic scale factor.
- Decimation Filter Ringing generates sharp artificial peaks near the Nyquist cutoff that project false spectral energy into short cluster intervals.
- Bus Contention Jitter introduces non-uniform sampling timestamps, distorting the integration interval tau zero and shifting the calculated logarithmic slope.
- Direct Memory Corruption overwrites least significant data bytes with peripheral status flags, substituting deterministic counter sequences for actual sensor noise.
- Power Supply Ripple Coupling modulates comparator reference voltages at switching frequencies, producing discrete sidebands that mimic broadband quantization.
In inexpensive six-axis sensors, thermal agitation and the physical Brownian noise floor often exceed the converter quantization step by more than twenty decibels. In such hardware, the minus-one slope never appears on an Allan deviation plot because wideband rate noise completely consumes the conversion steps. The extraction methods detailed here operate successfully only when the converter step size exceeds or closely approaches the intrinsic analog noise envelope.
Internal twenty-four-bit delta-sigma conversion is frequently assumed to render user-side register truncation negligible across all operating scenarios, though downstream bus handling can still reintroduce quantization noise.

Margin
Translating laboratory stability curves into reliable procurement documentation demands rigorous distinction between silicon limits and conversion artifacts. When drafting component specifications for high-reliability systems, procurement authorities frequently confuse resolution with measurement accuracy. A supplier marketing a 24-bit output interface may deliver a sensor whose physical performance provides only 13 effective bits of noise-free resolution.
The remaining 11 bits represent meaningless digital chatter that consumes bus bandwidth without improving guidance accuracy. Sourcing specialists must establish clear contract verification methods based on Allan deviation extraction to enforce true performance margins.

Thermal Noise Floors against Converter Resolution
Mechanical Brownian motion inside the etched cavity sets the ultimate physical boundary for angular velocity detection. Gas damping across the proof-mass comb fingers and thermoelastic dissipation in the suspension springs generate real physical torque fluctuations. This thermal energy appears as pure white rate noise, setting the angle random walk figure.
Matching the analog-to-digital converter to this mechanical ceiling defines the noise budget margin.
An optimized signal chain balances the digitizer least significant bit with the analog noise envelope. If the converter quantization step is designed to sit six to ten decibels below the intrinsic Brownian noise floor, quantization noise remains completely hidden, leaving system accuracy governed purely by sensor physics. Designing converter steps significantly smaller than this threshold wastes silicon area, elevates quiescent power consumption, and forces higher digital processing throughput.
Sourcing teams reviewing competitive proposals must audit the ratio between the extracted angle quantization coefficient Q and the angle random walk parameter to verify electronic balance.

Incoming Inspection Verification and Procurement Acceptance Criteria
Receiving departments processing reel quantities of packaged inertial units evaluate root-power spectral density alongside static cluster curves. Automated test fixtures execute multi-hour static logging on sample units drawn from each incoming silicon lot. The acceptance protocol enforces mathematical boundaries on the extracted Allan deviation parameters.
Suppliers claiming tactical performance must substantiate these metrics across temperature extremes from negative forty to positive eighty-five degrees Celsius. Sourcing audits implement a strict incoming verification regimen:
- Sample Rate Locking mandates data collection at the un-decimated sensor core rate to prevent vendor DSP filtering from masking digital quantization steps.
- Dynamic Range Conformance verifies that the least significant bit value matches the full-scale specification divided by the total register count without hidden bit-shifting.
- Asymptotic Gradient Validation confirms that the Allan deviation slope between two and ten sample periods falls between minus 0.90 and minus 1.05.
- Parameter Ceiling Caps reject lots where the extracted quantization coefficient Q exceeds 0.5 arcseconds for tactical components or 5.0 arcseconds for industrial components.
- Thermal Chamber Screening tests noise metrics across the operational temperature envelope to prevent comparator reference drift from elevating quantization noise.
Commercial contracts that neglect these mathematical extraction protocols expose buyers to stealth product revisions. Silicon fabricators occasionally shrink die footprints or alter front-end analog-to-digital converter topologies to improve wafer yield, preserving nominal pin compatibility while degrading the underlying noise architecture. Requiring Allan deviation parameter extraction dossiers for every component batch locks the vendor to verifiable physical performance.
IEEE Std 952 Annex C dictates that the quantization coefficient derives strictly from the minus-one asymptotic slope intercept at root three seconds, which forces suppliers to substantiate digital resolution claims against real time-series captures rather than theoretical bit depth.




