Two wearable devices record the same classroom movement, but their timestamps disagree. The disagreement grows during the session. Moving one graph sideways may align the beginning while leaving the end misaligned.
The gold standard of wearable technology in this situation is a comparison whose timing assumptions are visible. Clock offset, clock rate, signal-processing delay and wireless delivery time describe different parts of the record. They need different evidence.
This guide develops an original two-marker clock model for supplied wearable records. It is for learners, parents and teachers exploring measurement, ratios, graphs and engineering. It does not evaluate a particular product or interpret physiological timing as a medical result.
Read The Gold Standard of Wearable Technology for the device system. The earlier wear-time comparison guide addresses unequal observation coverage. This article asks a separate question: whether observations labelled with the same time belong to the same instant.
eduKateSG · Wearable Technology
Find your next learning step
Separate clock offset, rate error, processing delay and packet arrival through original wearable timing cases and calibration checks.
Chapter contents
Understand the evidence · Chapters 1–4
Work the model · Chapters 5–8
Test assumptions · Chapters 9–12
Practise and repair · Chapters 13–16
Report and continue · Chapters 17–20
A sensor can assign a timestamp when it samples a signal. A processor can assign another when it produces an event label. A phone can record when a packet arrives, and a server can record when it receives an upload. These times may all be valid descriptions of different events.
Before comparing numbers, write the timestamp contract. Does “12.4 seconds” mean the acquisition instant, the completion of an algorithm, or receipt by another system? Which clock generated it? Is it elapsed time from session start or a calendar time?
Consider a movement sampled at reference time 100.0 seconds. The device might label a processed event at 100.2 seconds and the phone might receive the packet at 100.8 seconds. Subtracting the phone receipt time from another device’s sample time combines processing and transmission effects with the physical timing difference.
The first repair is therefore descriptive. Keep the original timestamp field names and add a plain-language definition for each. A neatly aligned chart cannot compensate for comparing different stages of the measurement chain.
Clock offset is the difference between two clock readings at a specified instant. A rate difference describes how quickly that difference changes. A clock that starts two seconds ahead and runs at the correct rate keeps the same two-second offset.
For a simple model, let t be reference elapsed time in seconds and d the wearable’s timestamp in its nominal seconds. Write d = a + bt. Here a is the device reading at reference time zero, and b is the ratio of device elapsed time to reference elapsed time.
If b = 1, subtracting a aligns the entire interval. If b differs from one, subtracting a alone leaves an error (b − 1)t that grows with elapsed time. A constant shift and a time-axis rescaling solve different problems.
Everyday discussions often call this accumulated mismatch “clock drift”. In precise time metrology, a constant frequency offset and a changing frequency have different meanings. The main worked model uses a constant rate ratio b. Later chapters show how additional markers can reveal departures from that model.
CHAPTER 3 OF 20 · Understand the evidence
3. Establish markers that really refer to one event
A calibration marker must identify the same reference event in both records. Similar-looking peaks do not prove simultaneity. They could reflect separate movements, different filtering rules or a delay in one device’s event detector.
For teaching, use a supplied record in which two markers are explicitly declared to be simultaneous reference events. A real engineering procedure needs a suitable shared reference and a documented uncertainty. The mathematical exercise does not establish that any convenient gesture is an accurate calibration source.
Record each marker’s reference time, device time, identification method and uncertainty. Keep any failed or ambiguous markers rather than quietly selecting only those that produce a clean line. The selection method can affect the apparent clock behaviour.
When the common event is known only within a window, preserve that window. Reporting 20.000 seconds for a marker located only to the nearest tenth of a second creates false precision. The uncertainty belongs to the calibration evidence, even when the algebra gives many decimal places.
Suppose simultaneous markers occur at reference times 0 and 3,600 seconds. The wearable records them at d = 2.000 and d = 3,602.360 seconds. The device interval is 3,600.360 seconds, compared with 3,600 reference seconds.
The rate ratio is b = (3,602.360 − 2.000)/3,600 = 1.0001. The first marker gives a = 2.000. The complete model is d = 2.000 + 1.0001t.
The relative rate error is b − 1 = 0.0001, or 100 parts per million. This is an invented value for arithmetic clarity, not a performance claim about wearable devices. Over one reference hour, that rate difference contributes 0.360 seconds of additional offset.
To recover reference time, invert the relation: t = (d − 2.000)/1.0001. The order matters. Subtract the offset in device-time coordinates before dividing by the rate ratio. Writing d/1.0001 − 2.000 is a slightly different transformation and does not reproduce the first calibration marker exactly.
At reference time 1,800 seconds, the model predicts d = 1,802.180 seconds. Subtracting only the initial two-second offset gives 1,800.180 seconds, which remains 0.180 seconds late relative to reference time.
Using the full inverse gives (1,802.180 − 2.000)/1.0001 = 1,800 seconds. The scale correction removes the modelled accumulation. At the second marker, a constant shift would leave 0.360 seconds of error.
| Reference elapsed time | Predicted device time | Device time minus 2 seconds | Remaining error after shift |
|---|---|---|---|
| 0 s | 2.000 s | 0.000 s | 0.000 s |
| 900 s | 902.090 s | 900.090 s | 0.090 s |
| 1,800 s | 1,802.180 s | 1,800.180 s | 0.180 s |
| 3,600 s | 3,602.360 s | 3,600.360 s | 0.360 s |
The table also explains why a short visual check can be misleading. Near the first marker, the uncorrected rate error is small. Agreement over a few seconds does not establish agreement over the full session. Choose a verification interval that reflects the intended comparison.
CHAPTER 6 OF 20 · Work the model
6. Convert a later event without inventing its cause
Suppose a wearable event is timestamped d = 2,702.270 seconds. Under the fitted model, its reference time is (2,702.270 − 2.000)/1.0001 = 2,700 seconds. This is a coordinate conversion of the event timestamp.
It does not establish what produced the event, whether the sensor detected the movement promptly, or whether the event labels on two devices represent the same physical feature. Those questions require the signal definition and processing chain.
If another device reports the corresponding event at reference time 2,700.120 seconds, there is a residual difference of 0.120 seconds after clock conversion. Possible explanations include marker uncertainty, event-detection rules, filtering, mismatched events and physical timing differences.
Do not label that residual “human reaction time” merely because it has units of seconds. A valid conversion narrows the problem; it does not identify every remaining mechanism. Keep clock calibration and interpretation as separate steps in the written analysis.
Two distinct exact marker pairs determine a straight-line relation. Consequently, perfect agreement at those two points does not independently validate the straight-line model. The model has used both observations to choose its two parameters.
Introduce a third simultaneous marker at reference time 1,800 seconds. The model predicts device time 1,802.180 seconds. If the observed marker is 1,802.480 seconds, the residual in device-time units is +0.300 seconds.
That residual could reflect a marker identification error, a clock adjustment, varying clock rate or another unmodelled effect. It is evidence to investigate rather than an instruction to bend the line until every point fits. A more complex model also needs enough independent evidence to assess its behaviour.
A useful exercise asks the learner to explain why the first two markers fit exactly while the third is informative. The answer connects elementary simultaneous equations with model checking. Fitting and checking are different jobs even when the same graph displays both.
Small irregular residuals may arise from timestamp resolution, marker selection or varying delays. A sudden persistent change in residuals may suggest a clock step, restart or change in timestamp convention. A gradual curve may suggest a changing rate or another time-dependent effect.
These shapes are clues, not diagnoses. A repeated movement chosen as the wrong marker can imitate a sudden timing change. A change in processing delay can resemble a clock change if the timestamp is assigned after processing.
Inspect the device log and field definitions before dividing a session into separate calibration intervals. If a documented restart occurs, preserve the segment boundary and estimate the necessary mapping for each segment using appropriate evidence.
Do not force all records into one continuous affine relation merely to simplify a spreadsheet. Conversely, do not fit a separate line to every short interval without justification. Excessive flexibility can hide errors and make later timing claims look more certain than the observations support.
CHAPTER 9 OF 20 · Test assumptions
9. Keep wireless latency out of acquisition time
Assume a sample has a trustworthy device acquisition timestamp. It may wait in a buffer before wireless transmission. Its arrival at the phone is later, but the original sample time can still be converted through the clock mapping.
If a sample acquired at reference time 500 seconds arrives at 501.2 seconds, its delivery latency is 1.2 seconds. A second sample acquired at 501 seconds might arrive in the same batch at 501.2 seconds, giving latency 0.2 seconds. Equal arrival times do not imply simultaneous acquisition.
If only arrival times are retained, the acquisition times are generally not recoverable without additional information about buffering and delay. A smooth interpolation through arrival timestamps does not create the missing acquisition record.
The existing Sensor Fusion Engineering guide explains why evidence from moving systems must refer to compatible times. Here the original batch example shows specifically why receipt time cannot simply replace observation time.
Suppose the device timestamp for each calibration marker is uncertain by at most 0.010 seconds, while reference times are treated as exact in this simplified exercise. The device interval could then differ from its stated value by as much as 0.020 seconds.
Across a 3,600-second interval, that gives a worst-case rate-ratio uncertainty of 0.020/3,600, approximately 5.56 parts per million. Across only 60 seconds, the same endpoint uncertainty permits approximately 333.33 parts per million.
A longer calibration separation can reduce the relative effect of a fixed endpoint uncertainty. It does not guarantee a better model if the clock rate changes over the longer interval. The choice balances measurement resolution with the stability of the relation being estimated.
These are deterministic bounds under the stated assumptions, not confidence intervals. If reference markers are also uncertain or errors are correlated, the analysis must include those facts. Do not attach a statistical probability to a worst-case interval unless a probability model has actually been supplied.
Clock conversion assigns reference-time coordinates to existing samples. Resampling constructs values on a chosen time grid. They are different operations. A record may be correctly timed while its samples do not occur at exactly the same instants as another record.
Suppose device A has corrected samples at 10.00 and 10.02 seconds, while device B has one at 10.01 seconds. Comparing at 10.01 may require an interpolation assumption for A. That assumption concerns how the signal varies between observed samples.
A straight-line interpolation may be adequate for a supplied slowly varying toy signal and unsuitable for a sharp event. Increasing the number of interpolated points does not increase the number of physical observations. It changes the representation of the existing evidence.
Record the original sampling interval, clock mapping, resampling grid and interpolation rule separately. When a result changes, this record helps identify whether the change came from timing calibration or from the assumed signal shape between samples.
CHAPTER 12 OF 20 · Test assumptions
12. Use cross-correlation as evidence with limits
Comparing waveform patterns can suggest a relative lag. However, a repeating signal can align well at several lags separated by its period. A walking-like cycle repeated every second may offer multiple plausible matches if the available record lacks a distinctive marker.
Differences in filtering can also shift or reshape peaks. A lag estimated from one feature may not apply uniformly across frequencies or event types. A high correlation is evidence of a particular pattern match under the chosen analysis, rather than automatic proof of shared acquisition time.
For a teaching comparison, state whether the lag is estimated from known simultaneous markers or from signal similarity. Do not describe the second method using the certainty of the first. Preserve competing alignments when the signal is ambiguous.
An effective transfer question is: “If I repeat the same pattern every two seconds, what additional evidence would distinguish a zero-second lag from a two-second lag?” The learner should propose a unique event or independent timing reference rather than merely asking for a smoother plot.
A common error divides a millisecond difference by a time interval expressed in seconds without converting units. A 360-millisecond gain over 3,600 seconds is 0.360/3,600 = 0.0001, not 360/3,600 = 0.1.
Another error uses the reciprocal rate in the wrong direction. If the device gains time, b is greater than one in d = a + bt. Recovering reference elapsed time requires division by b after removing a. Multiplying by b would increase the discrepancy.
A third error treats calendar dates as simple decimal numbers without documenting the software’s date representation. For the learning exercise, elapsed seconds from a clearly defined session origin keep the relation visible. Real records also need timezone and clock-scale information when those affect interpretation.
Repair each error with a round-trip check. Convert a reference marker to device time, then invert it. Recovering the original marker is necessary, although it does not by itself validate the physical calibration.
In the repair pathway, begin with units and ratios. Ask the learner to calculate how much a clock gains in ten minutes if it gains 0.360 seconds per hour. The answer is 0.060 seconds under a constant-rate assumption.
In the stabilisation pathway, use two supplied markers and require the learner to find a, b and one converted event. Then introduce a third marker and ask whether it supports the model. Keep the arithmetic small enough that attention can remain on interpretation.
In the extension pathway, add bounded marker uncertainty, a documented restart or a repeating signal. Ask which observations can estimate a parameter and which can check it. The learner should recognise when the available evidence is insufficient.
Parents can ask, “What time does this number describe?” Teachers can ask, “Which observation would test the rate you fitted?” These questions make the reasoning visible without suggesting that all wearable results are unreliable or that a single calibration makes every result trustworthy.
CHAPTER 15 OF 20 · Practise and repair
15. Complete independent practice with answers
Task A: simultaneous reference markers are at 0 and 2,000 seconds; device markers are at 5.000 and 2,005.200 seconds. Find the relation. Answer: b = 1.0001 and a = 5.000, so d = 5.000 + 1.0001t.
Task B: under that relation, convert d = 1,005.100 seconds. Answer: t = (1,005.100 − 5.000)/1.0001 = 1,000 seconds. A constant five-second shift would leave 0.100 seconds of rate-related error.
Task C: two samples arrive together but retain different acquisition timestamps. Were they necessarily acquired together? Answer: no. Buffering can group their delivery. Use the appropriate clock mapping for acquisition timestamps.
Task D: two calibration markers fit exactly and a third differs by 0.4 seconds. Has the model passed an independent check? Answer: the third marker exposes a discrepancy needing investigation. The first two fits are consequences of fitting the line.
Task E: an event remains 0.2 seconds later after calibration. Is that a proven physiological delay? Answer: no. The event definition, processing effects and calibration uncertainty still need examination.
Does the correct timezone synchronise two devices? No. A timezone convention describes calendar representation. It does not remove a device’s offset, rate difference or processing delay.
Does a nominal sampling rate establish exact elapsed time? No. The actual timing depends on the relevant clock and acquisition system. A label such as 100 samples per second is not independent evidence that every nominal second matches the reference exactly.
Can one marker correct the whole session? One marker can establish an offset at that instant under suitable assumptions. It cannot independently estimate a rate difference. Additional separated timing evidence is needed for the two-parameter model.
Should all records be stretched to fit? Only when the timestamp contract and calibration evidence justify the transformation. A mismatch caused by different event definitions should not be disguised as a clock-rate correction.
Does this guide determine whether a wearable is suitable for a health decision? No. Its original examples concern measurement coordinates and model checking. Intended-use validation requires evidence beyond this timing exercise.
CHAPTER 17 OF 20 · Report and continue
17. Map two device clocks through a shared reference
Suppose device A follows dA = aA + bA t and device B follows dB = aB + bB t over a common validated interval. Eliminating t gives dB = aB + (bB/bA)(dA − aA). This maps timestamps directly between the two device coordinates while retaining the shared reference model.
For an original exact case, let dA = 2 + 1.0001t and dB = −1 + 0.9999t. At reference time 1,000 seconds, A reads 1,002.1 and B reads 998.9. Their displayed difference is 3.2 seconds, even though both readings refer to the same stipulated instant.
Subtracting the initial difference of three seconds leaves a 0.2-second mismatch at that time. The mismatch grows because the devices’ rates differ. The direct mapping uses the ratio 0.9999/1.0001, rather than simply subtracting a constant.
This is a useful algebraic extension because it exposes the direction of conversion. If the task requests B-time coordinates, the output must reproduce B’s calibration markers. If it requests reference time, use the inverse relation for the relevant device. A formula should be tested against the coordinate system it claims to produce.
The shared reference also carries uncertainty. Two separately fitted device mappings do not make their errors independent. If the same imperfect marker source affects both, the comparison needs to consider that shared dependence. Reporting two calibration lines does not automatically double the available timing evidence.
For learners, begin with the exact construction and round-trip checks. Only then introduce uncertain markers or estimated rates. That order keeps a conceptual coordinate-conversion error distinct from the harder question of how precisely the parameters are known.
CHAPTER 18 OF 20 · Report and continue
18. Build a timing budget before interpreting a residual
A timing budget records the sources that limit a comparison. In a simplified bound exercise, suppose converted marker alignment is uncertain by at most 0.020 seconds, locating event A adds at most 0.010 seconds and locating event B adds at most 0.015 seconds.
Without assuming independence or a probability distribution, a conservative bound on the combined timing discrepancy is the sum: 0.020 + 0.010 + 0.015 = 0.045 seconds. An observed difference of 0.030 seconds cannot establish a non-zero physical lag from these bounds alone.
An observed difference of 0.100 seconds exceeds that particular bound, but its interpretation still depends on whether the budget includes all relevant mechanisms. An omitted filtering delay or a mismatched event definition can invalidate the claim that the remaining difference is physical.
Do not use a root-sum-of-squares calculation merely because it gives a smaller number. Such a calculation needs an appropriate uncertainty model and treatment of dependence. The simple sum here is a worst-case construction based on explicit bounds, not a universal measurement standard.
Keep units and sign conventions visible. A positive residual might mean device A’s event occurs later, or it might mean A’s timestamp is larger before conversion. Those statements are not equivalent. Write the subtraction order next to the reported result.
A useful final question is, “What size of lag could this record reliably distinguish under the stated budget?” This turns calibration into a decision about the evidence’s resolution. It also prevents reporting a delay to three decimal places when the underlying event identification is uncertain by several hundredths of a second.
CHAPTER 19 OF 20 · Report and continue
19. Check time intervals as well as individual timestamps
The affine mapping changes durations in a particularly simple way. For two device timestamps d1 and d2 within the same validated segment, the corresponding reference interval is Δt = (d2 − d1)/b. The offset a cancels because it affects both endpoints equally.
In the main construction, a device interval of 600.060 seconds corresponds to 600 reference seconds when b = 1.0001. Subtracting the initial offset from each timestamp would leave the interval unchanged at 600.060 seconds. This is another way to see why a shift cannot repair a rate error.
Conversely, a single event time needs both the offset and rate. A learner who correctly converts a duration may still misplace the event on the reference timeline if a is omitted. Ask separately for “how long between the events?” and “when did the second event occur?”
The cancellation works only when the same affine relation applies to both endpoints. If a clock step or restart occurs between them, subtracting timestamps across that boundary can combine incompatible coordinate segments. Use the documented piecewise mapping and reference-time endpoints instead.
For a final independent check, let a = 7 seconds and b = 0.9998. A reference interval of 500 seconds produces a device interval of 499.9 seconds. The device is accumulating time more slowly in this model. Dividing by 0.9998 recovers 500 seconds; the seven-second offset has no role in this duration calculation.
A useful report retains the original files, timestamp definitions, marker pairs, fitted mapping, uncertainty assumptions and residual checks. Store corrected times in an additional field so the original evidence remains available.
For the main example, the concise report is: “Two stipulated simultaneous markers give d = 2.000 + 1.0001t over the stated hour. The device timestamp 2,702.270 seconds maps to reference time 2,700 seconds. Independent marker checks are required before extending the model to other periods.”
Read the NIST study of computer-clock statistics and synchronisation for the distinction between clock differences, frequency behaviour and synchronisation evidence. The affine datasets, uncertainty bounds and teaching tasks here are original constructions, rather than reported NIST device tests.
Continue through Wearable Technology, Clock Synchronisation and the Gold Standard series. The next learning step is to identify exactly which timing question a new record can answer before comparing its curves.