IMU Sampling Rate Too Low for Plyometrics: How to Diagnose and Fix It

A depth jump session wraps up and the ground contact times come back suspiciously tidy: 80 ms, 90 ms, 90 ms, 100 ms, 80 ms. Five different reps, same athlete, same box height, and every value happens to land on a multiple of 10. That is not a remarkably consistent athlete - it is a sensor confessing, in plain sight, that it cannot resolve anything finer than a 10 millisecond grid. On a contact window that is itself only 80 to 150 milliseconds long, a device reporting in 10 millisecond steps is not measuring the movement so much as sorting it into buckets.

Fast stretch-shortening cycle activity - depth jumps, hurdle bounds, most reactive strength index testing - lives almost entirely under 250 ms of ground contact, and a well-executed rep off a moderate box can sit at 120 to 160 ms. That is the range where sampling rate stops being a spec-sheet number nobody checks and starts directly deciding whether the contact time on your screen is real or an artifact of how coarsely your device sliced time. Get it wrong and you rarely see obvious garbage. You see numbers that look plausible, repeat in a strange stair-step pattern across sessions, and quietly corrupt every reactive strength index calculated downstream.

Every IMU records at a fixed sample period - the inverse of its sampling rate. At 100 Hz that period is 10 ms; at 800 Hz it is 1.25 ms. A contact-time algorithm marks touchdown at the first sample that crosses an acceleration threshold and toe-off at the first sample that crosses back below it. Neither edge can be marked any more precisely than the sample it falls on, so each edge carries an uncertainty of roughly one full sample period, and a contact time depends on two edges, not one.

That uncertainty does not shrink as a fraction of the contact window just because the window is short - it grows. A 20 ms timing error is a rounding error on a 700 ms ground-contact phase in a countermovement jump. On a 180 ms depth-jump contact, the same 20 ms is more than a tenth of the entire measured event.

This is why a device rated at 100 Hz - perfectly fine for step counting or coarse activity classification - can still be the wrong tool for depth jump contact time. The rated frequency was never the limiting spec on its own; the sample period relative to the event duration is.

Low sampling rate rarely announces itself as an error message. It shows up as a data pattern that looks plausible until you know what to look for.

• The quantization staircase. Across 15 to 20 visually different reps - some deeper, some faster off the box - the reported contact times only ever take a handful of distinct values, and every one of them is a multiple of the sample period. Real biological variation is continuous; a device sampling at 100 Hz can only ever report contact time in 10 ms steps, so the underlying continuous variation gets forced onto a coarse grid no matter how different the reps actually were.
• A one-directional bias, not just noise. Threshold-crossing detection with sparse sampling does not average out to zero error. The algorithm confirms touchdown only after the first sample that reads above threshold - a moment that falls after the true touchdown instant - and confirms toe-off only after the first sample that reads below threshold following release, which also falls after the true release moment. Both delays pad the window in the same direction, so low sampling rate tends to systematically inflate reported contact time rather than scatter it randomly around the truth.
• RSI distortion out of proportion to the timing error. Reactive strength index is flight time divided by contact time, and contact time sits in the denominator. A true contact of 130 ms against a flight time of 500 ms gives an RSI of 3.85. Pad that contact time by one 100 Hz sample period to 140 ms and RSI drops to 3.57 - a 7.3% swing produced entirely by a 10 ms sampling artifact, not by anything the athlete did differently between sessions.

None of this is speculative. It connects to a body of biomechanics literature on sampling frequency and short-duration event timing, most of it built on force platforms rather than wearable IMUs specifically, which matters as a limitation worth stating plainly rather than glossing over.

Street, McMillan, Board, Rasmussen and Heneghan (2001, Journal of Applied Biomechanics) examined how sampling frequency affects countermovement jump height calculated from force-platform data using the impulse method, comparing rates from 50 Hz up through several hundred Hz. Error attributable to sampling frequency alone grew sharply, not gradually, as rate dropped, becoming practically meaningful once frequency fell into a range still common on lower-cost wearable IMUs today. The limitation to name directly: their outcome variable was jump height via impulse integration, not ground contact time via accelerometer threshold-crossing, so the specific error magnitudes do not transfer one-to-one. What does transfer is the underlying finding - timing-dependent outputs degrade nonlinearly once sampling frequency approaches the time scale of the event itself, rather than declining smoothly as rate is lowered.

Falbriard, Meyer, Mariani, Millet and Aminian (2018, Frontiers in Physiology) validated foot-worn inertial sensors sampling at 500 Hz for running temporal parameters, including ground contact time, against a reference optical system across a range of running speeds, and reported timing agreement within a few milliseconds - accurate enough to resolve contacts in the 80 to 200 ms range that covers both sprinting and fast stretch-shortening cycle plyometric work. The limitation here: validation ran at a single high sampling rate on running gait, not on depth jumps or bounds, so the paper demonstrates what 500 Hz can achieve rather than directly quantifying how much accuracy is lost stepping down to 100 or 200 Hz on a jump-landing task specifically.

The reason this matters more for plyometrics than for general movement tracking traces back to Schmidtbleicher's (1992) classification of stretch-shortening cycle activity, drawing the line between fast SSC - contact times under roughly 250 ms, covering depth jumps, bounding and most reactive strength testing - and slow SSC, with longer, more countermovement-jump-like ground phases. Fast SSC is exactly the zone where a 10 or 20 ms timing error stops being a rounding error and becomes a double-digit percentage of the entire measured window.

Not every test needs the same timing resolution. A rule of thumb that holds up across the numbers above: aim for a sample period no more than roughly 3-5% of the shortest contact time you intend to measure.

For a 150 ms depth-jump contact, that rule points to a sample period at or under roughly 6-7 ms, which means a minimum of 150-200 Hz and a real margin of safety above 400 Hz. A device that markets itself as jump-capable at 100 Hz is not lying - it can detect that a jump happened. Whether it can time the contact phase precisely enough to trust the RSI it reports is a separate question, and the answer for fast SSC work is usually no.

1. Pull the last 15-20 ground contact time readings from the same jump type and a similar effort level - a batch of depth jumps or repeated bounds works well.
2. List the distinct values in that set and divide the count of unique values by the total rep count. Call this the uniqueness ratio.
3. Compare it to what your device's stated Hz should physically allow. At 100 Hz (a 10 ms sample period), a realistic fast-SSC range of 80-220 ms only contains 14 possible reportable values. If your uniqueness ratio across 15+ visibly varied reps is much lower than that, quantization is doing more of the work than genuine execution differences.
4. Cross-check one rep against smartphone slow-motion video at 240 fps (4.17 ms per frame). Count frames from first visible ground compression to the first visible extension off the ground, convert to milliseconds, and compare against the device reading.
5. If device and video disagree by more than roughly 10-15% of the contact window, and the gap is consistently in the same direction - device reading longer - the sampling rate, not the athlete or the detection algorithm, is the more likely source.

An athlete performs a depth jump off a 30 cm box. A 1000 Hz reference system times the true ground contact at 142.0 ms and flight time at 480 ms, giving a reference RSI of 3.38.

Flight time was captured accurately by both devices in this example - the error lives entirely in the contact-time edges. An 8 ms overestimate on the 100 Hz unit, well within the worst-case edge error the math above predicts, produces an RSI reading 5.3% lower than the true value. That is large enough to read as a meaningful week-to-week decline in reactive strength when nothing about the athlete's neuromuscular output changed at all. The 800 Hz unit's 0.5 ms error moves RSI by less than 1%, comfortably inside normal biological session-to-session variability.