The Gold Standard of Kinematics | Sampled Position Data, Hidden Reversals and Measurement Error

Suppose a camera records a marker’s horizontal location at regular times. Before taking differences, ask what each entry represents. Is it a pixel coordinate, a calibrated distance along a track, a projected horizontal coordinate or a distance along a curved path? These quantities are not interchangeable.

A camera coordinate becomes a physical position only after a suitable calibration. If a straight scale in the plane of motion gives 0.004 m per pixel, a 25-pixel change represents 0.100 m under that calibration. A scale in a different depth plane can produce a different conversion. Perspective can make a single constant conversion inappropriate.

Time deserves the same care. A video described as 30 frames per second may be useful for an ideal exercise with uniform 1/30 s intervals. In actual processing, inspect available time stamps and dropped frames. Frame index alone is not always an exact clock.

Record the coordinate origin, positive direction, units, time origin and any selection rule for the marker. Tracking the front of a cart in one frame and its centre in another creates a false displacement even when both positions appear plausible.

A minimum observation contract

A usable statement might read: “The same marker’s horizontal position, measured in metres along a straight track, was recorded at the listed time stamps. The camera remained fixed and the calibration scale lay in the motion plane.” It tells a reader what was measured and which assumptions support the conversion.

Keep raw observations alongside processed quantities. If later analysis reveals a wrong scale or time interval, the raw record lets you repair the calculation. A graph without the underlying table can hide how the result was obtained.

Contents · Next chapter

For two observations (t₀, x₀) and (t₁, x₁), the interval average velocity is (x₁ − x₀)/(t₁ − t₀). It describes net displacement per elapsed time across that interval. It does not automatically describe the instantaneous velocity at either endpoint.

Consider the original sample record below. Times are exact for the teaching model, and positions are displayed in metres.

The interval velocities are 1, 3 and 5 m/s. On a velocity plot, it is often useful to place them at interval midpoints 0.5, 1.5 and 2.5 s. This labels where each interval sits in time; it does not, by itself, prove that the actual instantaneous velocity equals the plotted value there.

If the underlying model is x = t², differentiation gives v = 2t. For that quadratic model, the midpoint equality happens to be exact. The data table alone does not establish the quadratic. A different trajectory can pass through all four points while having different derivatives between them.

Unequal intervals

Suppose positions 0, 1 and 5 m occur at times 0, 1 and 3 s. The two interval velocities are 1 and 2 m/s. Dividing both displacement changes by 1 s would misread the second interval. Always use the actual elapsed time.

A high-value checking step is to multiply each interval average velocity by its own duration. Summing the resulting signed displacements must recover x_final − x_initial. If it does not, the intervals, signs or units have been mixed.

Contents · Previous chapter · Next chapter

For a sampled one-dimensional record, add the absolute position changes between consecutive observations. This produces the length of the piecewise straight path that joins the observations. If the object may reverse between observations, its true distance travelled can be greater.

Suppose the positions are 0 m at t = 0 s, 2 m at t = 1 s and 0 m at t = 2 s. Net displacement is zero. The sampled path length is |2 − 0| + |0 − 2| = 4 m. If motion is monotonic within each interval, distance is exactly 4 m. If there are hidden excursions, 4 m is only a lower bound.

The inequality follows from the triangle inequality. Any route connecting two endpoint positions must have length at least the straight separation of those endpoints. Summing over the intervals preserves the lower-bound relationship.

Now consider only two observations: x = 0 at both 0 s and 1 s. The sampled path length is zero. One compatible motion is complete rest. Another is x(t) = sin(2πt) metres. That smooth motion travels from 0 to 1, through 0 to −1, and back to 0, covering 4 m during the second.

There is no contradiction. The observations constrain the endpoints but leave the intervening trajectory unresolved. Without a speed, acceleration or other physical bound, endpoint data alone do not supply a useful finite upper bound on all possible travel between samples.

This is why an answer should say “distance under the piecewise linear model” or “minimum distance consistent with the sampled positions” when the assumptions are limited. Using the word measured for a reconstructed total can conceal an inference.

Contents · Previous chapter · Next chapter

Take x(t) = A sin(2πft), where A is an amplitude in metres and f is a frequency in cycles per second. Sample at t_n = n/f, one full cycle apart. Every recorded position is zero, even though the object moves during each cycle.

Sampling at half-cycle intervals can also record zeros for this particular phase. A larger number of entries does not guarantee that the motion becomes visible if the timing aligns with the same uninformative phases.

This illustrates aliasing: different time-dependent motions can produce the same sampled sequence. Rest and a suitably timed oscillation may be indistinguishable from the recorded positions alone. A smooth line through the samples cannot restore information that was never captured.

A useful practical check is to change the sampling interval or phase and repeat the observation, where the setup permits. If an apparently stationary object suddenly displays oscillation under a different sampling pattern, the original record was incomplete.

However, “sample faster” is not an unlimited cure. Exposure time can blur the marker, location selection can vary, and the measurement instrument has its own limits. The question is whether the observation process resolves the shortest motion feature relevant to the task.

An original transfer example

A rotating marker completes exactly one revolution between photographs. Its image can return to the same location in every photograph. That record does not prove zero angular speed. Adding a distinctive second marker may help resolve orientation, but its success depends on geometry and sampling.

The central lesson is modest and powerful: an unchanging sampled coordinate does not establish an unchanging continuous coordinate. State which variations the recording process was designed to detect, and avoid claims about movements faster than that process can resolve.

Contents · Previous chapter · Next chapter

For equally spaced times with spacing h, a centred difference at time t is [x(t + h) − x(t − h)]/(2h). It uses observations on both sides of the target time. It is an estimate of the derivative, with accuracy depending on the smoothness and time scale of the underlying motion.

A Taylor expansion explains its behaviour. Expanding the two positions around t and subtracting cancels the even-order terms. For a sufficiently smooth trajectory, the centred estimate is x′(t) + h²x‴(t)/6 plus higher-order terms. The leading approximation error is proportional to h² when the needed derivatives remain controlled.

Consider x(t) = t³ metres, with time measured in seconds. At t = 2 s, the exact velocity is 12 m/s. With h = 1 s, the centred estimate is (27 − 1)/2 = 13 m/s. With h = 0.5 s, it becomes (15.625 − 3.375)/1 = 12.25 m/s. Reducing h improves the approximation in this noise-free example.

This does not mean the smaller interval must be better with real observations. A difference divides position errors by time. Reducing the time interval can magnify the influence of measurement noise even while reducing the mathematical approximation error.

At the first or last sample, a centred estimate needs data that may not exist. A forward or backward difference can be used, but label its different construction. Do not silently invent observations beyond the record.

A responsible result includes method and scale: “Velocity estimated by a centred difference with h = 0.5 s.” That statement allows another reader to reproduce the value and judge whether its time resolution suits the question.

Contents · Previous chapter · Next chapter

For equally spaced positions x₋, x₀ and x₊ at times t − h, t and t + h, the familiar second-difference acceleration estimate is (x₊ − 2x₀ + x₋)/h². It is exact for a quadratic position function and an approximation for more general smooth motion.

The denominator explains a difficulty: position errors are divided by h². If each observed position has an unknown error bounded in magnitude by ε, the acceleration estimate’s error contribution from those observations can be as large as 4ε/h². This is a worst-case bound from adding the magnitudes of coefficients 1, −2 and 1.

Let ε = 0.001 m and h = 0.1 s. The bound is 0.4 m/s². If h is reduced to 0.01 s without improving position accuracy, the bound becomes 40 m/s². More frequent sampling has not made each position more accurate.

Random uncertainty is a different calculation

If the three position errors are independent, each has zero mean and standard deviation σ, the second-difference noise standard deviation is √6 σ/h². This combines variances rather than adding worst-case magnitudes. It requires the independence and equal-variance assumptions; a worst-case bound and a standard deviation answer different questions.

The distinction matters in lab reports. Writing “±0.4” without explaining whether it is a bound, a standard uncertainty or a confidence interval leaves the reader unable to interpret it.

For a checkable example, use x(t) = 0.5at². At times −h, 0 and h, the noiseless positions are 0.5ah², 0 and 0.5ah². Substituting gives a exactly. Now add a positive error to the middle observation: the estimated acceleration falls because that observation carries coefficient −2. The direction of the effect follows from the formula, not from intuition about a noisy-looking graph.

Contents · Previous chapter · Next chapter

Smoothing can reduce the influence of noisy observations by pooling information across time. A fitted line estimates a local trend; a fitted quadratic can provide a local acceleration estimate. Both introduce assumptions about how motion behaves within the chosen window.

A wider window usually averages over more observations, but it can also mix genuinely different stages. A trolley that accelerates, then coasts, should not automatically receive one constant-acceleration fit over its whole journey. A fitted coefficient can be numerically stable while describing none of the stages well.

Inspect residuals: observed positions minus model-predicted positions. If residuals show a systematic curve or a clear change around a known event, the model may be missing structure. Random-looking residuals do not prove the model is physically correct, but structured residuals are a useful warning.

Record the window length and polynomial order. “Acceleration from a quadratic fit over 0.8–1.2 s” is reproducible. “The software gave acceleration” is not enough to judge time resolution or possible smoothing bias.

Do not confuse precision with validation

A computer can display six decimal places from measurements accurate only to millimetres. The number of displayed digits reflects calculation settings, not necessarily information in the experiment.

Likewise, a smooth graph is not automatically a better account. Abrupt but real changes can be blurred; noise can be made attractive. Compare the processed result with the original points and with a physically plausible description of the experiment.

For a classroom repair, start with a noise-free synthetic trajectory whose derivative is known. Sample it, apply the proposed processing and compare the estimates with the true values. Then add controlled errors. This separates approximation behaviour from data quality before the method is applied to an unknown real trajectory.

Contents · Previous chapter · Next chapter

A single unit mistake can overwhelm careful numerical differentiation. A position difference in centimetres divided by seconds gives cm/s. Converting to m/s requires dividing the numerical value by 100. If the time is recorded in milliseconds, convert that scale as well.

A clock-scale error affects velocity and acceleration differently. If reported times are all c times the true times, calculated velocity is the true velocity divided by c, and calculated acceleration is the true acceleration divided by c², for an otherwise correct trajectory and scale.

For example, if an analysis treats every actual 0.02 s interval as 0.04 s, the reported speeds are half the correct values and the reported accelerations are one quarter. This is not random scatter; it is a systematic scaling error.

A position-scale error by factor k multiplies both velocity and acceleration by k. A constant position offset, in contrast, cancels in differences. Changing the coordinate origin does not change the calculated derivatives. These facts supply useful audit questions.

Time origins also cancel from intervals, but synchronisation between instruments remains important. Two accurate clocks can have a relative offset. Combining one instrument’s position with another’s force or velocity at supposedly matching times may then create a misleading relationship.

The audit should end with one independent physical scale check. Does a slow classroom trolley appear to cross metres in milliseconds? Does the estimated speed conflict with the total observed displacement and elapsed time? Simple plausibility checks often expose mistakes that a smooth plotted curve leaves hidden.

Contents · Previous chapter · Next chapter

Complete worked record

Suppose x(t) = 0.5t² m for 0 ≤ t ≤ 2 s in an ideal model. At times 0, 0.5, 1, 1.5 and 2 s, positions are 0, 0.125, 0.5, 1.125 and 2 m.

The four interval velocities are 0.25, 0.75, 1.25 and 1.75 m/s. Their intervals are equal. Multiplying each by 0.5 s and summing gives 2 m, consistent with the net displacement.

The centred velocity estimate at 1 s is (1.125 − 0.125)/1 = 1 m/s. The second-difference estimate there is (1.125 − 1.0 + 0.125)/0.25 = 1 m/s². These match the derivatives of the known quadratic model.

Now change only the observed middle position at 1 s from 0.500 to 0.510 m. The centred velocity estimate at 1 s is unchanged because it uses the two neighbouring observations. The acceleration estimate becomes (1.125 − 1.020 + 0.125)/0.25 = 0.92 m/s². A small central error affects the two estimators differently.

Practice 1: endpoints do not determine distance

A marker is at 1 m at both 0 s and 2 s. What displacement is established? Zero. What distance is established? Only a lower bound of zero without additional motion information. Rest is possible, but so is a return journey.

Practice 2: unequal intervals

Positions 0, 3 and 7 m are recorded at 0, 2 and 3 s. Interval velocities are 1.5 and 4 m/s. Their simple average is 2.75 m/s, while the whole-record average velocity is 7/3 m/s. The intervals have unequal durations, so duration weighting is necessary.

Practice 3: error bound

With bounded position error ε = 0.002 m and centred spacing h = 0.2 s, the second-difference error contribution is bounded by 4ε/h² = 0.2 m/s². That result is a worst-case bound under the stated error assumption, not a confidence interval.

Contents · Previous chapter · Next chapter

Frequently asked questions

Does a finer sampling interval always produce a more accurate acceleration? No. It can reduce approximation error while magnifying position noise. The measurement and processing choices must be considered together.

Is the sum of sample-to-sample distances the true distance? It is exact under a suitable straight, monotonic-between-samples model. More generally it is a lower bound on travel in the recorded coordinate space.

Can identical recorded positions prove rest? No. They show no resolved displacement at the sampled times. They do not exclude motion between samples.

Is a central difference the instantaneous velocity? It is an estimate. Exact agreement occurs for particular underlying functions; real data need an explicit approximation and uncertainty discussion.

Connected reading directory

Use Kinematics to connect the observations with displacement, velocity and acceleration. Calculus develops the derivative, and Physics connects measurements with models and explanations.

For the broader learning aim, see Learning That Transfers into Life. Transfer here means carrying the measurement conditions into a new experiment, rather than repeating an estimator without inspecting its inputs.

Sources and scope

Numerical differentiation and its step-size trade-off were checked against McMaster University numerical methods notes. The continuous kinematics foundations and further study were checked through MIT OpenCourseWare: Kinematics. The finite-difference expansions, synthetic records, error bounds and exercises are derived explicitly in this guide. Checked 22 October 2026. The examples assume a one-dimensional calibrated coordinate unless stated otherwise; they do not certify an instrument, clinical measurement or engineering tracking system.

Contents · Previous chapter · Continue to Kinematics