Math Bridge: Why does a small edge model still depend on sampling physics?

← Back to Edge ML and TinyML Deployment
Math BridgeAnalytics & MLStruggle-friendly runway

Why does a small edge model still depend on sampling physics?

Audit the sensor input before model conversion, quantization, or pruning gets credit.

Data Dora, the guideData Dora guides
The one targetExplain why a 1.00 kHz band needs a 2.56 kHz practical sample rate.
The chapter case1.00 kHz vibration assumption; 12-bit 3.3 V ADC; 100 MB to 200 kB.
What it buys youA release trigger that covers sensor firmware as well as model files.

A field team faces an unresolved physical question: Why does a small edge model still depend on sampling physics? They must answer it before changing firmware rate on the real device. Predict the direction first.

See the relationship before changing it

The figure reads from left to right. The blue card is firmware rate. The middle card applies this page's relationship. The green card is nyquist ceiling. Walk the arrows once: set the input, apply the rule, then read the result with its unit.

The retained audit below checks several chapter fixtures. This added model holds every other chapter fixture fixed, so the numeric fixture does not switch without explanation.

Firmware rate changes nyquist ceiling An input card leads through the page relationship to the nyquist ceiling result. SET INPUT ONE CONTROL APPLY RULE predict calculate check units READ RESULT
Walk the arrows. The firmware rate changes the sensor evidence without changing one model weight. That is why sample rate belongs in the model release ledger.

Derive the baseline in four named moves

  1. 1

    Name the input. The chapter baseline for firmware rate is 2560.

  2. 2

    Name the relationship. fs,min=2(1.00 kHz)=2.00 kHz; fs,practical=2.56(1.00)=2.56 kHz 2.56/(2x1.00)=1.28x Nyquist margin 2¹²=4,096; q=3.3/4,096=0.806 mV; qrms=0.233 mV; SNR=74.0 dB 100 MBx1,024/200 kB=512x

  3. 3

    Substitute the chapter fixture. Set firmware rate to 2560. The page ledger gives nyquist ceiling as 1.3 kHz.

  4. 4

    Read the result. Keep kHz beside the value. Use it only inside the technical boundary on this page.

Predict, then change firmware rate

Try Predict the direction of nyquist ceiling. Move one control, calculate, then check your prediction.

2560
Chapter baseline
Nyquist ceiling

Observe The firmware rate changes the sensor evidence without changing one model weight. That is why sample rate belongs in the model release ledger. Reset the control to 2560 and compare nyquist ceiling.

Explain Only firmware rate moves here. The other chapter fixtures remain fixed.

Check yourself

What should you do before trusting a moved-control result?
Answer: Predict its direction, apply the shown relationship, keep the units, and reset to the worked baseline.
What does this small model leave out?
Answer: Only firmware rate moves. Field effects named in the page's technical boundary stay fixed.

Honesty note. This teaching model changes only firmware rate. It replays the chapter fixture; it is not a field prediction.

1. Begin before the model

The model receives features made from samples. If the ADC folded a fast vibration or rounded away a small change, a smaller model cannot recover the original signal.

Data Dora: Model-weight quantization and sensor-voltage quantization are different operations with different failure modes.

2. Name every algebra move

1

Double the signal bandNyquist minimum=2fmax.

2

Leave filter roomPractical rate=2.56fmax in this stated design convention.

3

Measure marginNyquist margin=fs/(2fmax).

4

Count ADC levelsLevels=2ᴺ; q=Vref/2ᴺ; qrms=q/√12.

5

Compare artifacts100 MB×1,024/200 kB=512× by binary units.

3. Reproduce the teaching case

fs,min=2(1.00 kHz)=2.00 kHz; fs,practical=2.56(1.00)=2.56 kHz
2.56/(2×1.00)=1.28× Nyquist margin
2¹²=4,096; q=3.3/4,096=0.806 mV; qrms=0.233 mV; SNR=74.0 dB
100 MB×1,024/200 kB=512×

The practical factor leaves a transition band for a real anti-alias filter instead of placing useful signal on the exact Nyquist edge.

4. Try the firmware sample rate

TryMove the firmware rate below and above 2.56 kHz and watch the strict and practical margins separate.

Firmware rate
Nyquist ceiling
Minimum rate
Practical target
Nyquist margin
Practical-target share
ADC levels
ADC step
ADC RMS noise
Ideal ADC SNR
Artifact size ratio

ObserveAt 2.56 kHz the Nyquist margin is 1.28× and the practical-target share is 1.00×. At 2.00 kHz the strict rule is only just met, leaving no transition room in this model.

ExplainThe firmware rate changes the sensor evidence without changing one model weight. That is why sample rate belongs in the model release ledger.

Technical boundaries.

This compact engine audits acquisition limits and artifact size, not classifier quality.

Bandwidth
The 1.00 kHz band and 2.56 factor are stated design assumptions, not a measured spectrum
ADC
Ideal resolution omits analogue noise, sensor sensitivity, range, and clipping
Model
File-size ratio says nothing about accuracy, latency, memory arena, power, or operator support

Validate the complete sensor-feature-model chain on the target board.

5. Set the release boundary

Treat sensor firmware, filter, rate, feature recipe, model artifact, runtime, threshold, and fallback as one versioned system. A change to any one can invalidate the test result.

6. Keep the deployment record

Record signal band, sensor range, ADC, filter and rate, feature code, dataset, model hash, operators, tensor arena, latency, power, threshold, fallback, owner, rollback, and retest trigger.

7. Check yourself

Why is 2.00 kHz the strict minimum?
Answer: It is twice the 1.00 kHz retained signal band.
Why use 2.56 kHz in the teaching design?
Answer: It leaves transition room for a realizable anti-alias filter instead of using the exact Nyquist edge.
Does a 512× smaller artifact prove deployment parity?
Answer: No. Size does not prove feature parity, accuracy, timing, memory fit, power, or fallback behavior.
Honesty boundary.

Model sizes come from the chapter; acquisition figures are labelled typical or assumed.

100 MB, 5 MB, 200 kB
The chapter's model-conversion path
1.00 kHz and 2.56×
A stated condition-monitoring band and practical sampling convention
12-bit, 3.3 V
A catalog-typical microcontroller ADC example

Correct, not complete: acquisition arithmetic does not qualify a TinyML release.

Return to Edge ML and TinyML Deployment with the maths unpacked →