Fit release height, initial velocity and gravity to a recorded fall, read the residuals, and tell a precise result from an accurate one.
Before you start: Chapters 1 and 2. Fitting and residuals are introduced here; the matrix derivation is an optional section.
After this chapter you can: Derive the constant-acceleration model, fit it to a fall record, and use residuals and repeated drops to judge whether the model and the result can be trusted.
A camera record of one fall is enough to measure the acceleration of gravity. Assume the acceleration is constant and derive height as a quadratic function of time. Then find the curve that comes closest to the record; its quadratic term gives the acceleration of gravity. To judge the result, inspect the deviations left after the fit, repeat the drop, and check the calibration against an independent reference. A small residual alone cannot reveal every error.
Chapters 1 and 2 supply the prerequisites: areas under velocity and acceleration graphs, camera calibration, and the distinction between scatter and offset. By the end, you should be able to fit a drop, interpret its residuals and explain what the reported uncertainty covers.
Picture a ball released from the top of a 30-metre drop tower. A camera at the side takes 20 frames per second, and each frame gives the ball's height with an error of about a centimetre; the camera has been calibrated as in Chapter 2. A photogate is mounted halfway down the tower.
The data in this chapter come from the Free Fall and Gravity Measurement Lab. As in the last chapter, the Lab knows the true motion and can tell us how far an estimate is off; a real experiment has only the checks described here. We treat the acceleration of gravity as a given constant. Why it is the same for every object waits for forces in Chapter 6 and gravitation in Chapter 14.
3.1 Coordinates and what we know
Take upward as positive and the ground as the origin, and write the ball's height as y, in metres. The release height is h and the initial velocity v0, positive upward. The ball is treated as a point, and air resistance is ignored for now.
Under these conditions we assume the ball's acceleration stays constant throughout the fall, equal to −g. The minus sign means the acceleration points down. Near the ground g is about 9.81ms−2, which is also the Lab's default. The task of this chapter is to test that assumption against a record and to measure g.
3.2 From constant acceleration to velocity and position
Chapter 1 showed that the change in velocity is the area under the acceleration graph and the displacement is the area under the velocity graph. With a constant acceleration a, the acceleration graph is a horizontal line, and the area from zero to t is at. So
v(t)=v0+at.
The velocity graph is therefore a sloping line. Its signed area from zero to t is the average of the endpoint velocities multiplied by t. If velocity changes sign, subtract the part below zero; the same expression still gives displacement:
y(t)−h=2v0+v(t)t=v0t+21at2.
For free fall, take a=−g:
y(t)=h+v0t−21gt2,v(t)=v0−gt.
Eliminating t gives one more useful relation. For a=0, solve the velocity equation for t=(v−v0)/a and substitute into the trapezium area to get y−h=(v2−v02)/(2a), that is,
v2=v02+2a(y−h).
Compare these relations with OpenStax University Physics Volume 1, Section 3.4 and Section 3.5. Check the units within each equation: position uses metres, velocity metres per second, and the last relation squared metres per squared second. Setting a=0 in the final relations brings back uniform motion, even though the intermediate division by a is unavailable.
3.3 Initial conditions pick the parabola
For a given g, every height–time curve is a downward-opening parabola of the same shape. h sets the height the curve starts from and v0 its slope at the start. Together they are the initial conditions. Knowing g alone is not enough to predict a particular motion.
For example, throw a ball upward at 14ms−1 from a height of 15 metres. It reaches the top when its velocity is zero:
ttop=gv0≈1.43s,ymax=h+2gv02≈24.99m.
It lands when y=0. Solving 15+14t−4.905t2=0 and taking the positive root gives t≈3.684s. From the velocity–height relation, the landing velocity is −v02+2gh≈−22.14ms−1, the minus sign meaning downward.
The Lab's “Upward launch” preset uses exactly these initial conditions and gives a landing time of 3.68427s. The Lab solves the motion numerically in steps of 0.01s, and its result agrees with the formula to all these digits.
3.4 Worked example: predicting a drop
A ball is released from rest at the top of a 30-metre tower, with g=9.81ms−2. Find the landing time, the landing velocity, and the ball's velocity as it passes a photogate 15 metres up. The photogate measures speed by timing a 10-centimetre flag carried by the ball; find how long the flag blocks the beam.
Released from rest, v0=0. Landing means y=0, so
timpact=g2h=9.81ms−260m≈2.473s.
The landing velocity is −gt≈−24.26ms−1. An independent check with v2=2g(h−y) gives 2×9.81×30ms−1≈24.26ms−1, in agreement.
At the photogate the ball has fallen 15 metres, so its speed is 2×9.81×15ms−1≈17.16ms−1. In the Lab, timing starts and stops when the ball crosses planes 0.05m above and 0.05m below the gate height. This represents a centred 0.1m travel interval. The exact mean speed over it is the length divided by transit time; it approximates the instantaneous speed at the gate. The flag takes about
17.16ms−10.1m≈5.83ms
to pass. Here the velocity is taken as constant during the passage. In that time it changes by only g×5.83ms≈0.057ms−1, about 0.3% of the speed, so the approximation is reasonable.
The camera takes a frame every 0.05s, so 50 frames are recorded from 0 to 2.45s. The Lab's “Earth tower” preset gives a landing time of 2.4731s, a photogate speed of 17.1552ms−1 and a flag transit time of 5.8292ms, as predicted.
3.5 From a record back to parameters: fitting
Now turn the problem round. We have 50 noisy heights yn with their times tn, and we want h, v0 and g.
For any trial set of values, each frame has a residual, the recorded value minus the model value:
rn=yn−(h+v0tn−21gtn2).
A least-squares fit picks the parameters that make S=∑rn2 as small as possible. Squaring has three advantages: positive and negative residuals cannot cancel; points further off count for more; and because the parameters appear only to the first power, the minimization reduces to three linear equations. The solution is unique when the record contains at least three distinct times; estimating noise from residuals needs more than three points. The method is described in the NIST Engineering Statistics Handbook, Section 4.1.4.1.
The simplest case shows why it is sensible. If the model is just a constant c, then
S(c)=n=1∑N(yn−c)2=Nc2−2c∑yn+∑yn2.
This is a quadratic in c that opens upward. Its slope is zero at the bottom, where c=N1∑yn. Fitting a constant by least squares gives exactly the mean. With three parameters the idea is the same, except that three slope-zero conditions must hold at once; the derivation is in Section 3.9.
For the Earth tower's default record, the fit gives
The number after the ± is the standard error. It is computed from the size of the residuals and the spread of the recording times, and estimates the scatter of repeated fits under the noise and model assumptions of Section 3.9. The interval shown spans one standard error; it covers the modelled random noise, with calibration uncertainty assessed separately. Section 3.7 checks it with repeated drops.
3.6 Residuals: what is the model missing?
A fit always returns some parameters, even when the model is wrong. To judge whether the model is adequate, look at the residuals left after the fit.
If the model is complete and the noise has independent zero-mean errors of equal variance, the residuals scatter above and below zero with no trend, and their RMS is close to the size of the noise. The default record's residual RMS is 8.7mm, comparable with the 1cm noise that was set. The three fitted parameters have absorbed some of the noise, so residuals a little smaller than the noise are normal.
The “Drag biases the fit” preset adds linear air damping at a rate of 0.5s−1, with everything else as in the Earth tower. The drag grows with speed, so the acceleration is no longer constant. Fitting the same quadratic model gives
g^=(4.73±0.08)ms−2,h^=30.61m,v^0=−2.68ms−1.
g^ is less than half the true value, and although the ball was released from rest the fit reports a downward initial velocity. All of these parameters are patching over the drag the model leaves out.
No air resistance: residual (millimetres)−20020Linear air damping: residual (millimetres)−60006000123
Time after release (seconds)
Figure 3.1 Fit residuals for two 30-metre drops, both fitted with the same constant-acceleration model. Top: no air resistance; the residuals scatter randomly and stay within 22mm. Bottom: linear air damping added; the residuals go negative, positive, negative and positive again, reaching half a metre.
Both vertical axes are in millimetres, on different scales. The data come from the Free Fall Lab's model code with the same noise realization.
The shape of the residuals gives the problem away. In the bottom panel the residual RMS is 232mm, more than twenty times the noise, and the residuals form a wave: about −618mm at the start, positive at frames 10 and 20, negative from frame 40 to 50, and about +499mm at the last frame. A pattern this large compared with the imposed noise is evidence against the constant-acceleration fit. In this simulation we know its cause is drag. In a real experiment, camera distortion or correlated measurement errors could also produce structure; the pattern alone does not identify the cause.
The standard error here is only 0.08ms−2, which looks very precise, yet the result is wrong by half. The standard error describes only the scatter of the data and does not check whether the model is right. Looking at the residuals is an essential model check; see the NIST Handbook, Section 4.4.4, for how to read them.
3.7 Precision and accuracy
Metrology separates two words that are often used interchangeably. Precision is how closely repeated measurements agree with one another; accuracy is how closely a measured value agrees with the true value. See the International Vocabulary of Metrology, VIM 2.15 and VIM 2.13. A result can be very precise and still inaccurate.
In the Lab, “Capture a new drop” keeps the physical parameters and swaps in a new set of noise. We repeated the Earth tower 500 times, with noise realizations numbered 431 to 930:
Table 3.1 Fit results for 500 repeated Earth tower drops.
Quantity
Value (ms−2)
Mean of the 500 values of g^
9.80988
Standard deviation of the 500 values of g^
0.00621
Mean of the reported standard errors
0.00606
A single drop reports a standard error of about 0.0061, and the actual standard deviation of 500 results is about 0.0062. They agree. When the model is right and the noise independent, the standard error of a single fit reliably predicts the scatter of repeated experiments, that is, the precision.
The mean of the 500 results differs from the true 9.81 by 0.00012. The uncertainty of the mean itself is about 0.00621/500≈0.00028, larger than the difference, so no systematic bias shows. Under these conditions the result is both precise and accurate.
With drag, the fitted value is biased by −5.08ms−2, more than sixty times the reported standard error of 0.08ms−2. That standard error is not a valid precision estimate here: the residuals contain model error, not just measurement noise. The “Imperfect camera” preset is a third case: g^=9.29±0.43, off by about 1.2 standard errors. Its offset is not surprising on the scale of that reported error, but clock rounding violates the exact-time assumption of the fit, so the quoted standard error is only a rough guide.
3.8 Photogates and clock resolution
A photogate takes no pictures. It records how long a flag blocks its light beam, Δtgate, and the speed estimate is the flag length divided by that time. It measures the velocity around a single moment, which makes it an independent check on the camera fit: using all three fitted parameters, predict vgate2=v^02+2g^(h^−ygate). The simpler v2=2g(h−y) applies only to release from rest. A real gate shares any flag-length or clock calibration errors with other measurements that use those same standards.
A timer can read time only to some smallest unit, its clock resolution. In this Lab the complete transit interval is rounded once to the nearest clock unit, so its error is at most half a unit. A device that rounds two endpoint timestamps separately can have an interval error of up to one unit. The shorter the interval being timed, the larger the share of that error.
The “Clock misses the gate” preset sets the clock resolution to 0.05s. The flag blocks the beam for only 5.83ms, which rounds to zero, so the photogate cannot give a speed. The camera fit is untouched: the camera takes a frame every 0.05s anyway, so the frame times are already whole multiples of the clock unit and rounding leaves them unchanged.
The “Imperfect camera” preset has a clock resolution of 0.02s, a release height of 8 metres and the gate at 4 metres. The true speed at the gate is about 8.86ms−1 and the transit time about 11.3ms, which rounds to 20ms, so the photogate reports 5.00ms−1, 44% low. The same clock also rounds each camera frame time by up to 0.01s, and that enters the fit error too.
The rule of thumb is that the clock resolution must be much smaller than the interval being timed. For the Lab’s once-rounded interval τ and clock unit q, the time error is at most q/2. If q<2τ, the resulting worst-case relative speed error is bounded by
v∣v^−v∣≤τ−q/2q/2.
Thus q<0.0198τ keeps this speed error below 1%. The familiar “resolution below 2% of the interval” is a first-order estimate, not a strict bound.
3.9 Optional: least squares with matrices, and the standard error
This section needs matrix multiplication, transposes and inverses, plus the idea of variance from Section 2.4. Skipping it does not affect later chapters.
Write the N recorded heights as a column vector y and the parameters as β=(h,v0,g)T. The design matrix X has N rows; row n is (1,tn,−tn2/2). The model and the sum of squared residuals are
y=Xβ+e,S(β)=(y−Xβ)T(y−Xβ).
Changing only the jth parameter while holding the others fixed, the slope of S is called a partial derivative. At the minimum all three partial derivatives vanish. The one for parameter j is −2∑nXnj(yn−(Xβ)n)=0, and together the three give
XTXβ^=XTy,β^=(XTX)−1XTy.
The inverse exists when the three columns are linearly independent, which here requires at least three distinct recording times. These are the normal equations. They say that the residuals of the best fit are "orthogonal" to every column of X: the sum of their products is zero.
The standard error comes from how the noise passes into the parameters. Substituting y=Xβ+e gives β^−β=(XTX)−1XTe, so each parameter error is a weighted sum of the frame noises. If the frame noises are unrelated, average zero and share the same standard deviation σ, the add-the-squares rule of Section 2.4 gives
Var(β^)=σ2(XTX)−1.
The true σ is unknown, so it is estimated from the residuals as s2=Smin/(N−3). We divide by N−3 rather than N because the three parameters have already fitted away some of the noise, leaving the residuals systematically small. The standard error of g^ is then s times the square root of the diagonal element of (XTX)−1 that belongs to g.
An independent check: the Earth tower record has 50 frames at tn=n/(20Hz) with σ=0.01m. Computing the matrix directly gives an expected standard deviation of g^ of 0.00608ms−2, in line with the 0.00621 from 500 repeats in Table 3.1.
This result needs correct model shape, known recording times, and zero-mean, uncorrelated frame errors of equal variance. With omitted drag, deterministic model mismatch also enters s2, so the reported standard error need not predict repeated-fit scatter. Correlated frame errors or rounded times require another uncertainty model.
A scale error can be harder to spot than drag. Divide every recorded height by 1.02: the fitted height, initial velocity and gravity are all divided by 1.02, yet the record is still a perfect parabola without noise. Residuals can look excellent while gravity is wrong by about 2%. Repeats test scatter; an independently calibrated ruler tests this shared scale error.
3.10 Checking in the Free Fall Lab
Open the Free Fall and Gravity Measurement Lab. A complete record is ready when the page opens, and the data row below shows the camera record and the “Fit residuals · bias leaves a pattern” plot.
Predict, then compare. Choose “Earth tower”. Use Section 3.4 to predict the landing time, the photogate speed and the flag transit time, then compare with the “Complete record duration”, “Photogate speed” and “Flag transit time” readouts.
Repeat the drop. Press “Capture a new drop” five times and note “Gravity fit and standard error” each time. Work out the mean and standard deviation of your five values of g^ and compare them with the single-drop standard error. The standard deviation of five values itself fluctuates a lot, which is exactly why more repeats are needed.
Make the model fail. Choose “Drag biases the fit”. Before looking, predict whether g^ will come out high or low. Look at the shape of the residual plot, compare the “Photogate speed” readout with the drag-free prediction of Section 3.4, and say which assumption has broken down.
Check the clock. Choose “Clock misses the gate” and explain why the photogate gives no speed while the camera fit is unchanged. In the advanced controls, reduce “Clock resolution” step by step until the photogate gives a speed, then see how far off it is at 1ms.
Transfer to the Moon. Choose Moon tower, where g=1.62ms−2 and the position noise is 2mm. Predict the landing time and the number of camera frames first, then explain why the standard error is so much smaller than on Earth.
Table 3.2 gives each preset's results with the default noise realization, so a reader without the Lab can follow the argument.
Table 3.2 Representative results for each Free Fall Lab preset (noise realization 431).
Preset
Camera frames
g^ (ms−2)
Bias against truth
Residual RMS (mm)
Photogate speed (ms−1)
Earth tower
50
9.8182±0.0055
+0.08%
8.7
17.16
Moon tower
122
1.61978±0.00011
−0.01%
1.7
6.97
Upward launch
74
9.8121±0.0020
+0.02%
8.8
18.52
Imperfect camera
11
9.29±0.43
−5.3%
84
5.00
Drag biases the fit
63
4.73±0.08
−51.8%
232
12.56
Clock misses the gate
50
9.8182±0.0055
+0.08%
8.7
unresolved
3.11 Exercises
In the Free Fall Lab, which quantities are recorded directly and which are estimated? Sort the camera frame times, heights, flag transit time, g^, h^ and photogate speed.
A ball is thrown straight up at 8ms−1 from a height of 2 metres, with g=9.81ms−2. Find the time to the top, the maximum height, the landing time and the landing velocity.
A photogate is 1.2 metres below the release point, the ball is released from rest, and the flag is 2 centimetres long. Find the transit time. With a clock resolution of 1ms, give a conservative bound on the relative speed error from interval rounding alone. Also find the actual rounded reading. Treat the gate speed as constant during the short passage.
Someone fits a straight line (a uniform-motion model) to the height record of a fall. What shape will the residuals have? Which term does that shape say is missing, and with what sign?
Two groups of students measure g. Both quote the standard error of the mean under independent random errors. Group A repeats five times and gets 9.62±0.02ms−2; group B gets 9.9±0.3ms−2. Which group is more precise? Which agrees with 9.81? What might cause group A's difference?
Continuing: if group A assumed a camera scale 2% larger than the true one, how far off would the fitted g be? Compare with group A's result.
A single drop has a standard error of 0.0055ms−2. Roughly how many independent drops are needed to bring the uncertainty of the mean down to 0.001ms−2? What does this estimate assume?
Transfer. A cart is released from rest on a slope and its position is recorded every second: 0.00, 0.31, 1.18, 2.72 and 4.80 metres. Use the differences of Chapter 2 to test whether it is roughly in constant acceleration, and estimate the acceleration. If the second differences kept getting smaller, what would you suspect?
Optional, needs Section 3.9. Use the matrix formula to show that when a constant is fitted to N data points with noise standard deviation σ, the mean has a standard deviation of σ/N.
3.12 Answers and explanations
Exercise 1Show the answer
The Lab supplies calibrated heights and rounded frame times as its camera observations, and a rounded transit interval as its gate observation. Chapter 2’s more detailed camera model starts one step earlier, with pixels; a height has already undergone calibration here. g^ and h^ are estimates from fitting a model. The photogate speed is an estimate too, the flag length divided by the transit time, giving the mean speed over the passage and approximating the instantaneous gate speed when the speed changes little. Note that the frame times have been rounded by the clock and the heights depend on the camera calibration, so even the recorded quantities carry errors.
Exercise 2Show the answer
The top comes at t=v0/g≈0.815s, and the maximum height is 2m+v02/(2g)≈5.26m.
For the landing, solve 2+8t−4.905t2=0; the positive root is t≈1.85s. The landing speed is v02+2gh=64+39.24ms−1≈10.16ms−1, downward. Checking with v=v0−gt gives 8−9.81×1.85≈−10.15, the same within rounding.
Exercise 3Show the answer
The speed at the gate is 2×9.81×1.2ms−1≈4.85ms−1, so the transit time is about 0.02/4.85s≈4.12ms.
For a once-rounded interval the time error is at most 0.5ms. Since speed is inversely proportional to time, a conservative speed-error bound is 0.5/(4.12−0.5)≈14%, rather than the first-order 12% estimate. For this particular interval, 4.12ms rounds to 4ms: the reported speed is 5.00ms−1, about 3% high. A longer flag reduces the relative clock error but increases the change in speed during passage; a finer clock avoids that trade-off.
Exercise 4Show the answer
The height curve of a fall bends downward, and a straight line can only cut through its middle. The best line lies above the curve at both ends of the record and below it in the middle, so the residuals are negative at the ends and positive in the middle, forming an arch.
The arch says a term in t2 is missing, and since it bends the curve downward its coefficient is negative: the −21gt2 term. With that term added, the residuals should lose their pattern.
Exercise 5Show the answer
Group A's standard error is much smaller, so group A is more precise. Group B's result differs from 9.81 by 0.09, less than one standard error, so it agrees. Group A differs by 0.19, more than nine standard errors, which would be very unlikely under the quoted independent random-error model.
Group A probably has a systematic error, such as a wrong calibration scale or air resistance that cannot be ignored. Repeating the same experiment cannot reveal such errors. Look at the residuals, measure another way, or check the calibration.
Exercise 6Show the answer
With the scale assumed 2% too large, every height is divided by 1.02. The model is linear in the heights, so the fitted parameters are divided by 1.02 as well: g^≈9.81/1.02ms−2≈9.62ms−2, exactly group A's result. That makes it a reasonable suspect, but an independent calibration measurement is needed to check it. The parabola can fit perfectly even with this wrong scale.
Exercise 7Show the answer
The uncertainty of the mean falls like 1/N, so N≥(0.0055/0.001)2=30.25; this estimate requires at least 31 drops.
This assumes the errors of different drops are unrelated and the model has no systematic bias. Averaging does not shrink a systematic bias: with drag present, more repeats only give a wrong number more precisely.
Exercise 8Show the answer
The first differences are 0.31, 0.87, 1.54 and 2.08 metres, and the second differences 0.56, 0.67 and 0.54 metres. For constant acceleration every second difference should equal a(Δt)2. Their mean is 0.59m, suggesting a≈0.59ms−2. With only three second differences, compatibility depends on how uncertain the positions are; absence of an obvious trend is not proof of constant acceleration.
The second differences scatter by about ±0.06m. If the position error is about 2cm, the rule from Chapter 2 gives second-difference noise of about 6×0.02m≈0.05m, the same size. If the second differences kept shrinking, the acceleration would be falling with time or speed. Suspect friction or air resistance growing with speed; the constant-acceleration model would then not be enough.
Exercise 9Show the answer
When only a constant is fitted, X is a column of N ones and XTX=N. The normal equations give c^=N1∑yn, the mean. The variance formula gives Var(c^)=σ2/N, a standard deviation of σ/N. That is the 1/N rule used in Exercise 7.
Next: a different observer
This ends the first unit. We have learned to describe a motion, estimate it from a camera record and test a model against the record. Every description so far came from one camera standing still on the ground. Chapter 4 (in preparation) asks: for an observer who is moving, which of position, velocity and acceleration change, and which stay the same?
Figure and data notes
The constant-acceleration and free-fall formulas follow OpenStax University Physics Volume 1, Sections 3.4 and 3.5. Least squares and residual checks follow the NIST Engineering Statistics Handbook, Sections 4.1.4.1 and 4.4.4. Precision and accuracy use the definitions of VIM 2.15 and 2.13.
Tables 3.1 and 3.2 and Figure 3.1 were computed with the Free Fall Lab's model code: a point mass, upward positive, fourth-order Runge–Kutta integration with a 0.01s step, Gaussian camera noise, default noise realization 431. “Drag biases the fit” changes only the linear air damping, to 0.5s−1, from “Earth tower”; “Clock misses the gate” changes only the clock resolution, to 0.05s.
The drag fit was checked independently against the exact solution for linear drag: without noise, the fit gives g≈4.72ms−2 and a residual RMS of about 231mm, consistent with the Lab's noisy result. The independent check of the standard error is in Section 3.9.
In Table 3.2 the photogate speeds for Earth, Moon and Upward launch match the true speed at the gate to the digits shown. The 5.00 for Imperfect camera is the reading after clock rounding; the true value is about 8.86.
Test it in the Lab
Change one setting at a time, then compare what you see with the tables and answers in this chapter.
Least squares chooses the parameters that minimize the sum of squared deviations between data and model. The chapter’s standard-error formula and its assumptions are stated in the chapter.
Definitions of accuracy (closeness to a true value) and precision (closeness of replicate measurements), including the note that precision should not be used to mean accuracy. Entry 2.15 is at https://jcgm.bipm.org/vim/en/2.15.html.