How Filter Cutoff Frequency Choice Alters Jump Height

Two athletes step onto the same testing session, both wearing the identical hip-mounted accelerometer, both jumping on a device set to whatever cutoff frequency shipped from the factory. The first, a sprinter with a compact, aggressive takeoff and barely 180 milliseconds of ground contact, reads a countermovement jump of 33 cm. The second, a longer-limbed thrower with a slower, more deliberate loading phase and roughly 320 milliseconds of contact, reads 41 cm on the exact same unit, same settings, same session. Put a force plate under both of them and the numbers tell a different story: 36 cm and 42 cm. The sprinter's true jump got squeezed down by 3 cm. The thrower's barely moved. Nobody adjusted a setting between reps, and neither athlete's actual jump changed between the wearable and the plate. What changed is how much of each athlete's takeoff signal survived a single low-pass filter cutoff that never asked which kind of jump it was about to process.

Every accelerometer feeding a jump-height algorithm records two things at once: the real linear acceleration of the athlete's center of mass, and a layer of higher-frequency content that has nothing to do with jumping, such as mounting vibration against skin or fabric, electrical sensor noise, and sharp transients from footstrike and takeoff that ring through the sensor housing for a few milliseconds longer than the underlying movement itself. A low-pass filter, almost always a second- or fourth-order Butterworth design in commercial jump devices, is applied before the signal gets integrated, and its job is to let content below a chosen cutoff frequency pass through largely untouched while attenuating content above it. That cutoff is not a hard wall. Frequency content sitting close to the cutoff value gets partially reduced rather than cleanly kept or removed, and how aggressively the filter rolls off past that point depends on the filter's order as much as its cutoff value.

What matters for jump height is that this cutoff is usually a single fixed number, chosen once by an engineering team during development and buried in firmware, applied identically to every athlete and every rep a device ever records. It is rarely exposed in a settings menu, and almost never validated against the specific population using it before it ships.

Jump height from an accelerometer comes from double integration: acceleration integrated once gives velocity, integrated again gives displacement across the flight phase. That calculation only works if the algorithm knows, precisely, the instant the athlete's feet left the ground, because that instant sets both the starting velocity the integration begins from and the boundary where ground reaction force effectively drops to zero. In the raw signal, that instant does not look like a smooth transition. It looks like a sharp spike lasting only a handful of milliseconds, as the body goes from pushing hard against the ground to pushing against nothing at all.

A sharp, brief, high-amplitude spike is, by definition, high-frequency content, which places it exactly in the range a low-pass filter is designed to reduce. Bezodis, Salo, and Trewartha (2013, Gait & Posture) demonstrated this mechanism directly, though in a related task: they showed that the choice of low-pass filter cutoff frequency produced systematic, spurious fluctuations in calculated knee joint moments specifically around footstrike, the other high-acceleration transition event common in sprint and jump biomechanics. Their study analyzed sprint stance-phase kinetics rather than vertical jump flight phases, so the specific moment magnitudes they reported do not transfer directly onto a jump-height calculation. What does transfer is the underlying mechanism: any digital low-pass filter distorts a signal most right at the sharp, brief, high-acceleration events a movement analysis most depends on, and a jump's takeoff instant is exactly that kind of event.

That distortion runs in two directions depending on which way the cutoff gets set wrong, and both directions are systematic rather than random.

Set the cutoff too low, and the filter smooths the sharp takeoff spike the way a slow shutter blurs a fast-moving object: its peak amplitude gets reduced and its energy gets smeared across a slightly wider window in time. The algorithm's estimate of takeoff velocity comes in lower than the athlete actually produced, and since displacement scales with the square of velocity in a projectile's vertical rise, even a modest velocity underestimate compounds into a larger height underestimate. The more explosive an athlete's true takeoff, the sharper and higher-frequency that spike is to begin with, and the more of it sits above a conservative cutoff waiting to be discarded, which is why this failure mode hits sprinters, throwers, and other high-rate-of-force-development athletes hardest.

Set the cutoff too high, and the opposite problem shows up. More of the broadband noise floor around the true signal passes through untouched, and double integration does not treat that noise gently: it accumulates. Integrating a signal that carries a small persistent noise component does not average that noise away, it sums it, and the resulting error grows across the flight-time window rather than canceling out. Casartelli, Müller, and Maffiuletti (2010, Journal of Strength and Conditioning Research) validated a commercial accelerometer, the Myotest, against force-plate criterion measures of countermovement jump height and found a systematic overestimation on the device, alongside a good but imperfect correlation with the force plate. Their paper evaluated one commercial device's fixed processing pipeline rather than running a controlled sweep across multiple cutoff values, so it cannot isolate cutoff frequency as the single cause the way a dedicated filtering study would. What it does confirm is the direction and reality of the artifact: a fixed accelerometer-based algorithm can produce a consistent, non-trivial bias against a criterion measure, in exactly the direction this filtering mechanism predicts when a cutoff sits on the high side.

Because the size and direction of this bias depends on how sharp an individual athlete's true takeoff transient is, one fixed cutoff applied across a whole roster does not add uniform error, it adds error correlated with jumping style. An explosive athlete tested on a conservatively low cutoff gets pulled toward underestimation more than a longer-loading teammate tested on the identical device and setting; a device tuned toward a higher cutoff to protect that explosive athlete's peak signal will, on the same setting, start inflating the longer-loading teammate's reading as more noise slips through relative to their smaller true spike.

The practical consequence shows up in two places coaches actually look at. Across a roster, it can flatten or invert a ranking that would hold up under a force plate, making a genuinely more explosive athlete look shorter on paper than a teammate with a slower, more compliant style. Within a single athlete across a session, it can manufacture a false trend: fatigue commonly lengthens ground contact time and softens the sharpness of the takeoff transient as a session wears on, which under a fixed cutoff tuned for an early-session, sharper jump can nudge later reps toward the device's noise-driven overestimate side, reading as a smaller true decline, or even a false improvement, right when the athlete is actually fading.

Run this whenever a new accelerometer-based jump device joins your testing setup, or whenever you start relying on its output to compare athletes with visibly different jump styles.

1. Confirm whether the device's cutoff frequency is fixed in firmware or adjustable in a settings menu or companion software, and note the default value if one is published.
2. Recruit at least two athlete profiles with clearly different takeoff styles for the diagnostic: one with a short, sharp ground-contact phase (roughly 180-200 ms or less) and one with a longer, more compliant loading phase (300 ms or more), each performing 5 countermovement jumps.
3. Record every rep simultaneously on the device and on a force plate or a validated flight-time method, such as 240 fps video read with flight time squared times gravity, divided by eight.
4. If the cutoff is adjustable, reprocess the same raw trials at a low setting (6-8 Hz), a mid setting (12-18 Hz), and a high setting (25-30 Hz), and calculate percentage bias against the criterion for each athlete profile at each setting.
5. Choose the cutoff that minimizes bias magnitude across both profiles at once, not the value that happens to flatter one athlete's number.
6. If the cutoff is fixed and cannot be changed, characterize the bias direction and rough size separately for explosive-style and compliant-style athletes using the same criterion cross-check, and apply that as a documented correction rather than trusting the raw figure equally for every athlete on the roster.

The same 10 raw accelerometer recordings from two athletes, one explosive and one compliant-loading, were reprocessed offline through three fixed low-pass cutoffs, then compared against a force plate criterion recorded on the same reps.

The compliant-loading athlete's reading barely moved across the three cutoffs, drifting from a 2% underestimate at 8 Hz to a 5% overestimate at 30 Hz, a 7-percentage-point swing. The explosive athlete's reading swung far more, from a 9% underestimate at 8 Hz to an 8% overestimate at 30 Hz, a 17-percentage-point swing on the identical filter settings and the identical processing pipeline. The mid-range 15 Hz cutoff landed closest to criterion for both athletes in this test, but that is not a universal recommendation; it is the result for these two takeoff profiles on this device's noise floor. The actionable finding is narrower and more durable: an explosive athlete's jump-height reading is far more sensitive to cutoff choice than a compliant-loading athlete's, so any comparison across athletes with visibly different takeoff styles needs its own bias check rather than an assumption that one factory cutoff serves everybody equally.