Variable latency is harder to adapt to than constant latency

What it is

Variable, or jittery, latency is harder to adapt to than constant latency: even at the same mean delay, operators perform worse and struggle more to build a stable adaptation strategy under jittery conditions.

Why it happens

This counterintuitive result comes down to how human motor control builds an internal timing model. Under constant delay, an operator can learn through repeated practice to fold that fixed time gap into their predictive model — effectively adding a fixed offset to an internal clock, after which a stable timing expectation re-forms with practice. Variable delay means that offset itself keeps changing: the predictive model an operator just tuned to one delay value is wrong again on the next interaction because the delay has shifted, so prediction stays off and the operator can never converge, through practice, on a stable internal timing model. This is harder to compensate for through experience than a simply larger delay would be — a constant large delay is at least a learnable target, whereas a jittery delay offers no stable target to learn in the first place. This is also why two links with identical mean delay can produce completely different operating experiences: what drives performance is often not the mean but the tail of the delay distribution — a small number of high-delay events far from the mean can disrupt a predictive rhythm the operator just established, and their disruptive cost is out of proportion to how often they occur.

Studying it

A common design contrasts constant delay against variable delay (same mean, different variance), measuring differences in task performance and learning curves — if the hypothesis holds, the constant-delay group should improve and stabilize with practice while the variable-delay group improves more slowly or plateaus below the same level. Other work looks at whether the specific pattern of variation matters — periodic variation, which follows a predictable rule, is in principle more learnable in part than fully random jitter. Beyond mean and variance, this line of work typically reports a tail statistic of the delay distribution separately (the 95th-percentile delay, for instance), because a network that looks fine "on average" can still hide occasional long-tail delays that do real damage to operation. Study designs also need to watch for carryover across practice order — if the same participants experience both delay conditions in sequence, the predictive model built under the first condition can interfere with adaptation to the second, which usually calls for a between-subjects design or strict order counterbalancing.

Where it stops holding

If the jitter magnitude is small relative to the mean, its extra cost may not be noticeable; this result mainly holds when jitter is substantial (a swing of half the mean delay or more, say). Network-induced delay in practice usually carries both a higher mean and higher jitter together, so the two factors tend to co-occur in real systems and are hard to fully separate — which is also why this result is harder to verify in isolation in a real deployment than in a controlled lab setting.

Applying it

System design should prioritize reducing latency variance — using a buffer to smooth jitter, prioritizing timing stability over chasing a lower mean delay alone — and, where necessary, can even deliberately add a fixed extra buffer delay to flatten jitter into a constant delay, trading predictability for adaptability. Buffer size is itself a trade-off: a larger buffer gets delay closer to constant and easier to adapt to, but it also raises the effective floor of that delay, so buffer size cannot simply be pushed up indefinitely in pursuit of "perfect" stability. Validate by monitoring both the variance and the tail percentile of delay on the target network link; if variance is large, first check whether performance with a smoothing buffer beats the raw jittery delay before treating mean delay as the only metric that matters.

Related

• Same group: X7.02.1 Delay turns closed-loop control into predictive control · X7.02.2 Compensating delay requires a predictive display or adjustable autonomy
• Nearby: X7.03 Force and haptic feedback · X4.04 Shared control
• Search terms: latency jitter · teleoperation · motor adaptation · latency compensation