Decimated Signal Processing Moving Average Filter Fundamentals
Decimated moving average filters trade stopband attenuation for low compute density by combining unweighted integration with downsampling.

Comb
A finite impulse response moving average filter calculates the unweighted arithmetic mean over N consecutive samples to act as a digital lowpass filter. Combining this summation with downsampling by an integer factor M yields a decimated structure that cuts the output sample rate while attenuating high-frequency content. When window length matches the downsampling ratio, the impulse response collapses into a single-stage boxcar filter tied directly to a rate compressor.
Because this setup requires no multipliers ~ only adders and sample dropping ~ it remains a standard choice for microcontrollers and FPGAs.
In the z-domain, the discrete-time transfer function of an N-point moving average filter is written as H(z) = frac1N sumk=0N-1 z-k. Factoring the sum into closed recursive form gives H(z) = frac1N frac1 – z-N1 – z-1. In this form, the architecture splits into an undamped integrator running at the full sampling rate fs and a differentiator operating with an N-sample delay.
Cascading P of these integrator-differentiator pairs sharpens stopband attenuation without adding multipliers to the hardware.
Decimated filter structures eliminate multiplier gates by substituting unweighted integration and discrete sample dropping.

Cascaded Integrator Architecture
Accumulators running at the main input clock carry out recursive summation without feedback gain control. Because integration of unsigned or two’s complement integers runs continuously, these registers periodically overflow. So long as the arithmetic uses two’s complement format and register width covers the maximum expected bit growth, wrap-around overflow introduces no error once the downstream differentiator takes sample differences.
The differentiator subtracts the sample delayed by N periods from the current incoming sample. Placing decimation by factor M between the integrator and differentiator shrinks the required differentiator delay from N high-rate samples down to R = N / M low-rate samples. When N = M, that delay becomes just a single clock cycle at the reduced rate fs / M. Moving the delay line into the lower clock domain cuts dynamic switching power across the clock tree.

Boxcar Summation Mechanics
Without decimation, direct averaging needs an N-tap delay line and an adder tree for every output sample. Downsampling before summation means the system only computes every M-th sample, though standard non-recursive filters still evaluate all intermediate states unless configured recursively or via polyphase paths. A boxcar layout copes with high data rates by keeping an accumulator register that updates on every input clock cycle.
Each incoming sample adds directly to the running total. Once M values accumulate, the internal register latches its value to the output port and resets. In high-speed sensor pipelines, resetting registers can create tight timing paths.
Contemporary logic designs avoid explicit resets altogether, letting the integrator roll over naturally and using the differentiation stage to extract exact differences across the decimation boundary.
Data corruption occurs if register bit widths cannot hold the accumulated sum prior to subtraction. An under-provisioned integrator drops most significant bits prematurely, creating large impulse errors in the output stream that pass down the entire processing chain.

Spectrum
The frequency attenuation of a decimated boxcar filter follows a continuous sinc response mapped into the digital domain. Its normalized magnitude spectrum is given by H(ejω) = left| fracsin(ω N / 2)N sin(ω / 2) right|, where ω = 2π f / fs is the digital frequency in radians per sample. Spectral nulls land exactly at integer multiples of fs / N, concentrating attenuation at harmonic intervals determined by filter length.
Filter nulls align with integer multiples of the sampling rate divided by the averaging window length.
When downsampling by M = N, frequencies surrounding the spectral nulls fold straight into the baseband near zero hertz. Out-of-band energy within bands at k · fracfsM ± fb ~ where k is the integer alias index and fb is the signal bandwidth ~ aliases directly into the passband. Because a single-stage sinc profile offers modest suppression near the passband edge, the first sidelobe peak reaches only 13.46 dB of rejection.

Alias Folding Boundaries
Once downsampling folds high-frequency components into the signal band, they cannot be separated from true signal data. Cascading P stages sharpens the mainlobe transition and raises sidelobe attenuation by 13.46 × P dB. However, adding stages increases stopband rejection at the cost of noticeable passband droop across the measurement bandwidth.
Spectral aliasing mechanisms in decimated architectures exhibit structural failure modes when processing wideband noise or out-of-band interference:
- First Sidelobe Intrusion leaks high-frequency noise around 1.43 × fs / N into the low-frequency output, suppressed by only 13.5 dB per stage.
- Harmonic Foldover folds sensor resonance peaks at integer multiples of the output sample rate down into static DC offsets.
- Phase Distortion adds a linear group delay of (N-1)/2 sample periods, altering feedback timing in time-critical control loops.
- Quantization Sidelobe Elevation degrades harmonic distortion when truncated fixed-point registers produce intermodulation products across filter stages.

Passband Droop Compensation
Signal attenuation near the cutoff frequency affects sensor measurements well before reaching the Nyquist limit of the decimated clock. At half the decimated sample rate, f = fs / (2M), a single-stage moving average filter drops signal level by 3.92 dB. A three-stage cascade increases that loss to 11.76 dB at the same frequency, making inverse-sinc equalization necessary downstream to flatten the response.
| Cascade Order P | First Sidelobe Rejection (dB) | Passband Droop at 0.25 f_s/M (dB) | Passband Droop at 0.5 f_s/M (dB) | Group Delay (Input Samples) |
|---|---|---|---|---|
| 1 | 13.46 | 0.91 | 3.92 | (M – 1) / 2 |
| 2 | 26.92 | 1.82 | 7.84 | M – 1 |
| 3 | 40.38 | 2.73 | 11.76 | 3(M – 1) / 2 |
| 4 | 53.84 | 3.64 | 15.68 | 2(M – 1) |
Equalization relies on FIR filters running at the lowered sample rate fs / M with coefficients calculated from the inverse sinc function. Operating compensation at this lower clock rate keeps dynamic power low while maintaining amplitude accuracy within one percent over eighty percent of the decimated Nyquist zone. However, designing equalization without tracking alias noise rising above signal thresholds destroys the benefit of gain restoration.
Decimated filters maintain consistent phase delay over temperature as long as the master sample clock stays locked to an external reference oscillator.

Register
Internal word lengths in recursive decimation networks expand as samples accumulate. To prevent overflow in two’s complement math from corrupting the final result, internal registers must be wider than the input converter resolution. For a single stage with window size N, maximum gain is N, which requires log2(N) extra bits of headroom.
Cascading P identical stages with a decimation factor of M increases total filter gain to MP. The required bit growth Bgrowth over the input word length Bin is defined by Bgrowth = P · log2(M). For instance, a 16-bit ADC feeding a 4-stage filter with a decimation factor of 16 requires internal registers of 16 + 4 · log2(16) = 32 bits.

Bit Growth Mathematics
Determining register width requires rounding up to the next integer bit. Physical hardware cannot allocate fractional bits, so any non-integer result rounds up to a full bit. If integrator registers fall short of this calculated width, unrecoverable overflow errors corrupt the output.
- Find the input bit width Bin from the ADC specification.
- Identify the decimation factor M and cascade order P needed for target attenuation.
- Calculate total DC gain G = MP.
- Compute bit growth Bgrowth = lceil P · log2(M) rceil using the ceiling function.
- Set accumulator widths to Btotal = Bin + Bgrowth for all high-rate integrator stages.
- Match differentiator register widths to Btotal to retain two’s complement precision.
Register allocation equations determine exact internal bit expansion necessary to prevent unrecoverable integer overflow.

Is Bit Pruning Permissible in Precision Converters?
Truncating intermediate register bits saves silicon area and lowers dynamic power in programmable logic. However, discarding least significant bits in early integrator stages introduces quantization noise that passes through to the output. Selective bit pruning removes low-order bits from early stages while keeping the resulting quantization noise safely below the analog sensor’s thermal noise floor.
| Decimation Factor M | Cascade Order P | Total DC Gain | Bit Growth (Bits) | Required Register Width (Bits) |
|---|---|---|---|---|
| 8 | 2 | 64 | 6 | 22 |
| 8 | 4 | 4096 | 12 | 28 |
| 16 | 3 | 4096 | 12 | 28 |
| 32 | 3 | 32768 | 15 | 31 |
| 64 | 4 | 16777216 | 24 | 40 |
Determining pruning thresholds without degrading the noise floor requires analyzing the stage-by-stage noise transfer function. Quantization noise introduced at stage i undergoes filtering by all remaining P-i stages. Tailored truncation cuts logic gate counts by twenty to thirty-five percent in multi-channel sensor hubs, though maintaining absolute monotonicity across wide temperature swings demands thorough verification.

Polyphase
Decomposing a non-recursive moving average filter into polyphase paths changes how downsampling interacts with memory. Standard decimated filters process data at the high input rate fs, wasting power on samples that the decimation step eventually discards. Polyphase designs split the length-N impulse response into M parallel sub-filters and place downsampling at the input, running all calculations at the lower rate fs / M.
Polyphase expansion splits the transfer function H(z) = sumn=0N-1 h(n) z-n into M branches following H(z) = sumk=0M-1 z-k Ek(zM), where Ek(z) is the k-th polyphase sub-filter. For an unweighted moving average filter with coefficients h(n) = 1/N, each sub-filter Ek(z) simplifies to a shorter FIR filter or delay chain, removing high-speed accumulator updates from the main clock domain.

Commutator Structural Decomposition
An input commutator sequences incoming data sequentially across M parallel register paths: sample x(nM) goes to the first path, x(nM – 1) to the second, and x(nM – k) to path k. Each branch operates at the decimated rate fs / M, executing additions simultaneously on a single lower-frequency clock edge.
A multi-branch polyphase commutator layout provides structural advantages when integrating moving average filters into high-density ASIC processes:
- Clock Domain Consolidation runs logic exclusively at the lower output rate fs / M, cutting dynamic power by up to 1 – 1/M.
- Sub-filter Simplification reduces sub-filters to basic delays when window length N equals decimation factor M.
- Bus-Width Expansion converts fast serial data into wide, low-speed parallel buses suited for low-power DSP cores.
- Area Optimization removes the high-speed wrap-around logic needed in recursive integrators.

Clock Domain Decoupling
Crossing clock boundaries safely requires careful handshaking. Dual-port circular buffers or ping-pong registers store fast input bursts before polyphase taps read them out at fs / M. In multi-channel sensor layouts, parallel polyphase paths can process incoming streams through shared coefficient networks without adding pipeline latency.
Executing polyphase operations requires careful consideration of structural tradeoffs before committing target designs to silicon layout or programmable logic fabric:
- Check that the sample clock jitter is low enough to avoid phase alignment errors during commutator routing.
- Confirm that register widths across parallel branches accommodate sub-filter bit growth without bit truncation.
- Balance register footprint against available FPGA block RAM to keep routing congestion down.
- Verify that pipeline latency remains within control loop timing budgets when using multi-stage downsampling.
Dynamic simulation models indicate zero power penalty during high-ratio downsampling, yet physical current measurements show substantial switching losses in input commutator routing trees at maximum input rates.

Yield
Overall measurement accuracy relies on matching filter topology to sensor bandwidth and converter noise characteristics. High-resolution delta-sigma ADCs use large decimation factors M to convert fast, low-resolution bitstreams into high-precision baseband data. Unweighted moving average filters form the first decimation stage in these front-ends, where keeping gate counts low directly improves die yield and silicon margins.
Noise shaping in the analog front-end determines the minimum filter order P. A first-order delta-sigma modulator requires at least a second-order filter (P=2) so stopband rejection stays ahead of rising high-frequency noise. A second-order modulator calls for P=3 to keep high-frequency noise from aliasing back into the signal band and reducing effective resolution.

Dynamic Range Quantification
SNR improvements from decimation follow standard processing gain equations. With uncorrelated white noise, averaging N samples reduces noise variance by a factor of N, giving a theoretical dynamic range gain of 10 · log10(N) dB ~ or 0.5 · log2(N) bits. Downsampling by M=64 with ideal anti-aliasing yields an 18.06 dB SNR boost, adding 3 bits of resolution.
Processing gain increases peak signal-to-noise ratio by three decibels for every doubling of the filter window length under white noise conditions.
Practical sensor circuits face flicker noise (1/f noise), supply ripple, and clock jitter that break white-noise assumptions. If 1/f noise dominates below the corner frequency, widening the decimation window N does not increase resolution. Averaging over longer time windows under 1/f noise simply integrates low-frequency drift, degrading measurement repeatability.

Silicon Overhead Calculations
Implementing decimated moving average filters requires balancing logic resources against processing speed. Recursive CIC filters minimize gate count by eliminating hardware multipliers, though they need wider adders to support bit growth. Non-recursive polyphase structures use more storage registers but run at lower clock rates, giving designers a clear tradeoff between silicon area and dynamic power.
| Architecture Type | Multiplier Units Required | Adder Bit-Width Operations | Memory / Register Elements | Relative Power Density |
|---|---|---|---|---|
| Direct Form FIR | N | N – 1 | N Registers | High (1.00) |
| Direct Decimated FIR | N / M | N – 1 | N Registers | Medium (0.45) |
| Recursive Integrator (CIC) | 0 | 2 P | 2 P Registers | Low (0.12) |
| Polyphase FIR Decomposed | 0 (Unweighted) | N – 1 | N Registers (Low Speed) | Very Low (0.08) |
Silicon area maps directly to wafer costs during ASIC synthesis. Selecting a recursive structure cuts gate count from several thousand flip-flops down to a few hundred logic cells, improving yield per wafer. For high-frequency vibration or acoustic sensing, a smaller die footprint also reduces thermal stress on the package, helping preserve long-term calibration in harsh industrial settings.
Subsystem verification involves testing filter operations across temperature and supply voltage corners. Test equipment feeds deterministic sweep vectors into input ports while automated stations check bit-exact output streams against software reference models. This step confirms that register wrap-around works cleanly without pipeline stalls across all decimation settings and stage counts.




