2.6 Kinematics from Experimental Position and Velocity Data

The Mess

[Video: A motion sensor tracks a cart rolling along a metal track. The sensor emits ultrasonic pulses and measures the time for each echo to return. On a computer screen next to the track, position-versus-time data points appear in real time as the cart moves. The cart is given a gentle push, rolls forward, slows, and rolls back. The resulting graph is recognizable as a parabola --- but only barely. The data points scatter above and below the expected curve. Some points jitter visibly. A few look like outright errors. The camera zooms in on a section of the graph where the scatter is most obvious.]

Look at that data. You have been working with clean mathematical functions all chapter --- $x(t) = \frac{1}{2}at^2$, polynomial position functions, piecewise models with sharp boundaries. Those functions are precise. You can differentiate them and get exact velocities. You can differentiate again and get exact accelerations. The calculus is clean because the inputs are clean.

Now here is what actual measurements look like.

[Image: A position-time scatter plot showing approximately 200 data points from a motion sensor. The overall shape is a downward-opening parabola (consistent with an object moving forward and then backward under roughly constant deceleration), but every individual point is displaced slightly from the smooth curve. Some points deviate by a millimeter or two. A few deviate by five or more millimeters. No smooth curve is drawn --- just the raw points.]

The data does not lie on a smooth curve. It clusters near a curve, but every point is a little off. Some are a lot off. There is no equation written on the apparatus telling you what $x(t)$ is. There are only numbers, and they are noisy.

This is not a failure of the experiment. This is what real measurement looks like. Every sensor has finite precision. Every reading includes some random error. The question is not whether the data is noisy --- it always is. The question is: can you still extract velocity and acceleration from this mess?

Before you read on: The motion sensor measures position to within about 1 mm. That sounds pretty good. If your position measurements are accurate to $\pm 1\,\text{mm}$, how accurate do you think your velocity estimates will be? Take a guess --- a number with units --- before continuing.

[Interactive: Predict-Then-Reveal. Student enters a numerical estimate for velocity accuracy (in m/s or mm/s). After submitting, the system stores their answer for comparison with the actual result they will compute later in the section. The reveal is delayed until after the exploration.]

The Guiding Question

What does real motion data add to the ideal mathematical picture --- and what does it take away?

All of Chapter 2 up to this point assumed that you start with a known function: $a(t)$ is given, or $x(t)$ is given, and you compute everything else with calculus. But in a laboratory, nobody hands you a function. You get a table of numbers. The question is how to go from numbers to physics --- and what you can and cannot conclude from imperfect measurements.

From Data to Velocity: Finite Differences

When you have a position function $x(t)$, velocity is the derivative:

$$v(t) = \frac{dx}{dt} = \lim_{\Delta t \to 0} \frac{x(t + \Delta t) - x(t)}{\Delta t}$$

But you cannot take a limit with measured data. You have positions at discrete times --- maybe every 0.05 seconds, maybe every 0.01 seconds --- and nothing in between. The best you can do is approximate the derivative with a finite difference:

$$v(t_i) \approx \frac{x(t_{i+1}) - x(t_i)}{t_{i+1} - t_i}$$

This is the slope between two adjacent data points. It is not the instantaneous velocity --- it is an estimate of the average velocity over one time interval. But if the time interval is small enough, this estimate should be close to the true instantaneous velocity.

Should be. Let's see what actually happens.

The Noise Problem

Here is a concrete example. Suppose the true position of an object is $x(t) = 5t - t^2$ (in meters, with $t$ in seconds), so the true velocity is $v(t) = 5 - 2t$ --- a clean, straight line sloping downward.

Now suppose you measure this position every 0.1 seconds, but each measurement has a random error of up to $\pm 1\,\text{mm}$ ($\pm 0.001\,\text{m}$). Here are the first few data points:

$t$ (s)	True $x$ (m)	Measured $x$ (m)	Error (mm)
0.00	0.0000	0.0000	0.0
0.10	0.4900	0.4908	+0.8
0.20	0.9600	0.9594	$-0.6$
0.30	1.4100	1.4112	+1.2
0.40	1.8400	1.8395	$-0.5$
0.50	2.2500	2.2491	$-0.9$

The errors are tiny --- less than a millimeter in every case. The position measurements look excellent.

Now compute the velocity by finite differences. Between $t = 0.10$ and $t = 0.20$:

$$v \approx \frac{0.9594 - 0.4908}{0.20 - 0.10} = \frac{0.4686}{0.10} = 4.686 \, \text{m/s}$$

The true velocity at $t = 0.15\,\text{s}$ is $v = 5 - 2(0.15) = 4.70\,\text{m/s}$.

The estimate is 4.686 instead of 4.700 --- an error of 0.014 m/s, or 14 mm/s.

Wait. The position error was less than 1 mm. But the velocity error is 14 mm/s. Where did the extra error come from?

Pause and think: The position errors were about 1 mm. The velocity errors turned out to be about 14 mm/s. The time step was 0.1 s. Do you see a relationship between these three numbers?

Here is what happened. The velocity estimate divides the difference between two noisy measurements by the time interval $\Delta t$. Each position has an error of about $\pm 1\,\text{mm}$. The difference between two positions can have an error of up to $\pm 2\,\text{mm}$ (if one is high and the other is low). And then you divide by $\Delta t = 0.1\,\text{s}$:

$$\text{velocity error} \approx \frac{2 \times 1\,\text{mm}}{0.1\,\text{s}} = \frac{0.002\,\text{m}}{0.1\,\text{s}} = 0.02\,\text{m/s} = 20\,\text{mm/s}$$

A 1 mm position error has been amplified into a 20 mm/s velocity error. The act of computing a finite difference --- which approximates differentiation --- amplified the noise by a factor of $2/\Delta t$.

This is a general and important result: differentiation amplifies noise.

[Interactive: Noise Amplification Explorer. The screen shows two plots stacked vertically. The top plot displays position data: a smooth underlying curve (e.g., $x(t) = 5t - t^2$) with noisy data points scattered around it. The bottom plot shows velocity estimates computed by finite differences from those noisy positions. Two sliders control the display: (1) "Position noise" adjusts the amplitude of the random scatter in the position data, from 0 mm up to 5 mm. (2) "Time step $\Delta t$" adjusts the spacing between data points, from 0.01 s to 0.5 s. As the student increases the noise, the velocity plot becomes dramatically noisier --- far more than the position plot. As the student decreases $\Delta t$, the velocity plot gets worse, not better, because the same position error is divided by a smaller time interval. A readout shows: "Position uncertainty: $\pm$ ___ mm. Velocity uncertainty: $\pm$ ___ mm/s. Amplification factor: ___."]

Before you read on: You just saw that smaller time steps make the velocity noisier. But larger time steps mean you are computing an average over a longer interval, so you lose information about how the velocity changes during that interval. This is a genuine tradeoff. Can you think of a strategy that reduces noise without sacrificing too much time resolution?

[Interactive: Open-response prompt. Student types their strategy before continuing. No right/wrong judgment --- the response is stored and displayed alongside the formal discussion below.]

The Smoothing Tradeoff

There are two opposing pressures when you estimate velocity from data:

Small $\Delta t$ gives better time resolution --- you can detect rapid changes in velocity --- but it amplifies noise because you divide small, noisy differences by a small number.
Large $\Delta t$ reduces noise because the signal (the actual change in position) grows relative to the random errors --- but it smears out the details. If the velocity is changing quickly, a large $\Delta t$ gives you only a blurred average.

This is not a problem you solve once and forget. It is a tradeoff you manage every time you work with experimental data. There is no single "correct" choice of $\Delta t$. The right choice depends on how noisy your data is and how rapidly the quantity you are measuring is changing.

One practical approach: instead of using just two adjacent points, use a multi-point average. For example, a 5-point centered difference estimates the velocity at time $t_i$ using data from $t_{i-2}$ through $t_{i+2}$:

$$v(t_i) \approx \frac{x(t_{i+2}) - x(t_{i-2})}{t_{i+2} - t_{i-2}}$$

This is equivalent to computing the finite difference over a wider interval (4 time steps instead of 1), which reduces the amplification factor. The cost: you lose the ability to detect velocity changes that happen faster than 4 time steps.

[Interactive: Smoothing Window Explorer. A single plot shows noisy velocity estimates overlaid on the true velocity curve. A slider labeled "Averaging window" adjusts from 2 points (raw finite difference) to 3, 5, 7, 11, and 21 points. As the window increases, the noisy velocity estimates smooth out and approach the true curve --- but they also begin to round off sharp features. A second dataset option shows a motion with a sudden velocity change (e.g., a bounce). With the wide averaging window, the sharp change is smeared into a gradual transition. Guided prompts appear: "Set the window to 2 points. Describe what you see." "Now set it to 5 points. What changed?" "Now set it to 21 points. Where does the smoothed curve fail to capture the real motion?"]

The core insight: There is no free lunch. Smoothing reduces noise but also reduces your ability to see rapid changes. Every choice of smoothing window is a compromise between noise and resolution. The physicist's job is to understand the tradeoff and make a defensible choice --- not to pretend the noise isn't there.

Differentiation Amplifies; Integration Smooths

You have now seen that computing velocity from position data (an approximation of differentiation) makes noise worse. But there is a counterpart that goes the other direction.

Suppose instead of differentiating, you integrate. Suppose you have noisy velocity data and you want to compute position by numerical integration:

$$x(t_n) \approx x_0 + \sum_{i=0}^{n-1} v(t_i) \, \Delta t$$

Each term $v(t_i) \, \Delta t$ has some random error. But when you add up many small terms, the errors partially cancel. Some velocity measurements are too high, some are too low, and the sum tends to average them out. The more terms in the sum, the more cancellation you get.

This is not a coincidence. It is a fundamental mathematical fact:

Differentiation (or its finite-difference approximation) takes differences between nearby values. If those values have random noise, the differences amplify it.
Integration (or its finite-sum approximation) takes sums of many values. If those values have random noise, the sums tend to cancel it.

Think of it this way. If you measure 100 noisy positions and take the difference between the first and the second, a single bad measurement throws off your answer. But if you add up all 100 values, a single bad measurement barely changes the total. Sums are robust against individual errors. Differences are fragile.

Before you read on: This asymmetry --- differentiation amplifies noise, integration smooths it --- has a practical consequence for experimental physics. If you could choose between measuring position and computing velocity (by differentiating), or measuring velocity and computing position (by integrating), which would give you more reliable results? Why?

[Interactive: Predict-Then-Reveal. Student selects one of two options before the explanation unlocks. After submitting: "Measuring position and computing velocity requires differentiation, which amplifies noise. Measuring velocity and computing position requires integration, which smooths noise. The second approach generally gives more reliable derived quantities. This is one reason why many experimental setups measure velocity directly (using Doppler sensors, for example) when they can."]

What Can You Conclude from Real Data?

Here is a complete dataset from a cart on a track. The cart was given a push and then left to roll freely.

[Image: A position-time plot with approximately 100 noisy data points. The overall trend is a downward-opening parabola: the cart moves forward, slows, and comes back. The data spans about 2 seconds. The scatter is visible but moderate --- the general shape is clear even without smoothing.]

$t$ (s)	$x$ (cm)
0.00	0.0
0.10	4.8
0.20	9.4
0.30	13.6
0.40	17.9
0.50	21.5
0.60	24.8
0.70	27.6
0.80	30.3
0.90	32.2
1.00	33.9
1.10	35.0
1.20	35.7
1.30	35.8
1.40	35.3
1.50	34.4
1.60	33.0
1.70	31.0
1.80	28.6
1.90	25.5
2.00	22.2

Looking at this data, several conclusions seem reasonable:

The cart moved forward, reached a maximum position near $t = 1.3\,\text{s}$, and then began moving backward.
The overall shape is roughly parabolic, suggesting roughly constant deceleration.
The initial velocity was positive, and the velocity decreased over time, passing through zero near $t = 1.3\,\text{s}$.

But be careful. How much of this can you actually justify from the data, and how much is you fitting a story to noisy points?

Before you read on: Look at the data table. Can you tell whether the acceleration was exactly constant, or just approximately constant? What evidence would you need to distinguish between the two?

The honest answer: from this data alone, with this level of noise, you probably cannot distinguish constant acceleration from an acceleration that varies slowly. Both would produce a position curve that looks "roughly parabolic" at this resolution. The noise hides the fine structure.

This is an important lesson. Data tells you what the motion is consistent with, not what it definitely is. The statement "the acceleration appears to be approximately constant" is scientifically honest. The statement "the acceleration is exactly $-10.2\,\text{m/s}^2$" is overconfident given the noise.

Models and Data: Two Directions

This section connects directly back to the modeling ideas from the very beginning of the course. In Section 1.1, you learned that physics is about building simplified models. In Sections 2.1 through 2.5, you used those models --- clean mathematical functions --- to predict motion exactly.

Now the arrow runs in the other direction. You have data, and you want to find a model that fits it.

This is a genuinely different activity. When you start with a model, the math tells you exactly what to expect. When you start with data, you have to propose a model and then check whether the data is consistent with it.

For the cart data above, a reasonable proposal is:

$$x(t) = x_0 + v_0 t + \frac{1}{2}a t^2$$

with three unknown parameters: $x_0$, $v_0$, and $a$. You can estimate these by fitting the parabola to the data (a topic you will explore more fully in later courses). A rough fit by eye gives something like $x_0 \approx 0$, $v_0 \approx 48\,\text{cm/s}$, and $a \approx -36\,\text{cm/s}^2$.

But notice what you have done. You assumed the constant-acceleration model and then estimated its parameters. You did not prove that acceleration is constant. You showed that the data is consistent with that assumption. A different model --- say, $x(t) = A\sin(\omega t + \phi)$ --- might also fit the data over this short interval. The data alone does not tell you which model is "right." Physics judgment does.

The connection to idealization: The clean mathematical models from Sections 2.1--2.5 are not just convenient simplifications. They are hypotheses that you can test against data. The idealization is valuable precisely because it gives you something specific to check.

What About Acceleration from Data?

If differentiating position to get velocity amplifies noise, what happens when you differentiate again to get acceleration?

It gets worse. Much worse.

To estimate acceleration from position data, you need to take finite differences twice:

$$a(t_i) \approx \frac{v(t_{i+1}) - v(t_i)}{\Delta t} \approx \frac{\frac{x(t_{i+2}) - x(t_{i+1})}{\Delta t} - \frac{x(t_{i+1}) - x(t_i)}{\Delta t}}{\Delta t} = \frac{x(t_{i+2}) - 2x(t_{i+1}) + x(t_i)}{(\Delta t)^2}$$

Now the position error is divided by $(\Delta t)^2$. If $\Delta t = 0.1\,\text{s}$ and the position error is 1 mm:

$$\text{acceleration error} \approx \frac{3 \times 1\,\text{mm}}{(0.1\,\text{s})^2} = \frac{0.003\,\text{m}}{0.01\,\text{s}^2} = 0.3\,\text{m/s}^2$$

A 1 mm position error has become a 0.3 m/s$^2$ acceleration error. For context, the acceleration due to gravity is about 9.8 m/s$^2$. So a 1 mm position error can produce an acceleration uncertainty that is about 3% of $g$ --- from a single pair of data points. And that is the best case; with smaller time steps, the situation deteriorates further.

[Interactive: Second Derivative Noise. Three plots stacked vertically: position, velocity, and acceleration, all computed from the same noisy dataset. The position plot looks clean. The velocity plot is noticeably noisy. The acceleration plot is wildly noisy --- large spikes and dips that bear little resemblance to the expected constant value. A smoothing slider lets students apply progressively wider averaging windows. Guided prompts: "Without smoothing, can you read the acceleration from the bottom plot?" "Apply enough smoothing to see a trend. What acceleration value does the data suggest?" "How confident are you in that value?"]

This is why experimentalists treat acceleration estimates from position data with great caution. When possible, they measure acceleration directly (using accelerometers) rather than computing it from position through double differentiation.

Sampling and Resolution

There is one more limit that real data imposes, beyond noise.

Your motion sensor records a measurement every $\Delta t$ seconds. That means it takes a "snapshot" of the position at regular intervals, and you see nothing in between. If the object does something interesting between snapshots --- a brief vibration, a momentary pause, a quick reversal --- your data will miss it entirely.

This is the sampling limit: you cannot detect motion features that happen faster than your sampling interval.

Consider a concrete example. Suppose a cart bounces off a bumper at the end of the track. The actual collision takes about 5 milliseconds (0.005 s). If your sensor records every 50 ms (0.05 s), the entire bounce happens between two data points. Your data shows the cart approaching the bumper and then moving away --- but the collision itself is invisible. You would see no difference between a sharp elastic bounce and a gentle reversal.

Pause and think: If you wanted to study the collision in detail, what would you need to change about your measurement setup? Is the issue noise, or something else?

The issue is not noise --- it is sampling rate. You need a faster sensor, one that records positions every millisecond or faster. More precision per measurement would not help; more measurements per second would.

This is a different kind of limitation from noise, and it requires a different response. Noise is managed by smoothing and averaging. Insufficient sampling is managed by collecting data faster.

Practice

Layer 1: Concrete --- Compute Finite-Difference Velocities

Here is position data from a motion sensor recording every 0.2 seconds:

$t$ (s)	$x$ (m)
0.0	1.000
0.2	1.196
0.4	1.384
0.6	1.556
0.8	1.720
1.0	1.860
1.2	1.984
1.4	2.084

(a) Estimate the velocity at $t = 0.1\,\text{s}$, $t = 0.3\,\text{s}$, $t = 0.5\,\text{s}$, and $t = 0.7\,\text{s}$ using two-point forward differences.

(b) Estimate the velocity at $t = 0.4\,\text{s}$ using a centered difference: $v(0.4) \approx \frac{x(0.6) - x(0.2)}{0.6 - 0.2}$.

(c) Compare your answer for (b) with the forward difference estimate at $t = 0.4\,\text{s}$: $v(0.4) \approx \frac{x(0.6) - x(0.4)}{0.2}$. Are they the same? Which do you think is a better estimate, and why?

Check your answer

(a) Using forward differences: - $v(0.1) \approx \frac{1.196 - 1.000}{0.2} = \frac{0.196}{0.2} = 0.980 \, \text{m/s}$ - $v(0.3) \approx \frac{1.384 - 1.196}{0.2} = \frac{0.188}{0.2} = 0.940 \, \text{m/s}$ - $v(0.5) \approx \frac{1.556 - 1.384}{0.2} = \frac{0.172}{0.2} = 0.860 \, \text{m/s}$ - $v(0.7) \approx \frac{1.720 - 1.556}{0.2} = \frac{0.164}{0.2} = 0.820 \, \text{m/s}$ (b) Centered difference at $t = 0.4\,\text{s}$: $$v(0.4) \approx \frac{1.556 - 1.196}{0.4} = \frac{0.360}{0.4} = 0.900 \, \text{m/s}$$ (c) Forward difference at $t = 0.4\,\text{s}$: $$v(0.4) \approx \frac{1.556 - 1.384}{0.2} = \frac{0.172}{0.2} = 0.860 \, \text{m/s}$$ These are different. The centered difference (0.900 m/s) is generally a better estimate of the velocity *at* $t = 0.4\,\text{s}$ because it uses data symmetrically on both sides of that time, rather than only looking forward. The forward difference actually estimates the average velocity over $[0.4, 0.6]$, which is centered at $t = 0.5$, not $t = 0.4$. The centered difference averages over $[0.2, 0.6]$, which is genuinely centered at $t = 0.4$.

Layer 2: Pattern --- Compare Step Sizes

Using the same dataset from Layer 1, compute velocity estimates at $t = 0.6\,\text{s}$ using three different approaches:

(a) A 2-point difference using $t = 0.6$ and $t = 0.8$ (step of 0.2 s).

(b) A centered difference using $t = 0.4$ and $t = 0.8$ (step of 0.4 s).

(d) These three estimates are not identical. Explain why. Which would be most affected by noise? Which would be least sensitive to rapid changes in velocity?

Check your answer

(a) $v(0.6) \approx \frac{1.720 - 1.556}{0.2} = \frac{0.164}{0.2} = 0.820 \, \text{m/s}$ (b) $v(0.6) \approx \frac{1.720 - 1.384}{0.4} = \frac{0.336}{0.4} = 0.840 \, \text{m/s}$ (c) $v(0.6) \approx \frac{1.860 - 1.196}{0.8} = \frac{0.664}{0.8} = 0.830 \, \text{m/s}$ The three estimates differ because they average over different time intervals. The 2-point difference (a) has the best time resolution --- it most closely approximates the instantaneous velocity at $t = 0.6\,\text{s}$ --- but it is most vulnerable to noise because it divides by the smallest $\Delta t$. The wide centered difference (c) is most robust against noise but least sensitive to rapid changes, because it averages over 0.8 seconds of motion. Approach (b) is a compromise between the two. This is the fundamental tradeoff: narrow windows give resolution but amplify noise; wide windows reduce noise but blur the signal.

Layer 3: Structure --- Why Differentiation Amplifies and Integration Smooths

Explain, in your own words, why computing a derivative from data amplifies noise while computing an integral from data smooths it. Use specific reasoning, not just a restatement of the rule.

Hint: Think about what each operation does to random errors. Does it combine them in a way that makes errors cancel, or in a way that makes errors grow?

Check your answer

**Differentiation** computes *differences* between nearby values and divides by a small number ($\Delta t$). If two adjacent measurements each have a small random error, the difference between them can have an error as large as the sum of the two individual errors. Then dividing by $\Delta t$ magnifies this combined error. The smaller $\Delta t$ is, the worse the magnification. In the extreme, if $\Delta t$ is very small, you are dividing a tiny (and noisy) difference by a tiny number --- the noise dominates. **Integration** computes *sums* of many values multiplied by $\Delta t$. Random errors in individual measurements are equally likely to be positive or negative. When you add many such errors together, they partially cancel: some push the sum up, others push it down. The more terms you add, the more cancellation occurs. (Statistically, the total error grows only as $\sqrt{N}$ while the signal grows as $N$, so the signal-to-noise ratio *improves* with more data.) The key difference is that differentiation isolates individual measurement errors and magnifies them, while integration pools many measurements together and lets errors average out.

Layer 4: Creation --- Propose a Model

Here is position data from a different experiment:

$t$ (s)	$x$ (cm)
0.0	0.0
0.5	12.1
1.0	23.5
1.5	34.2
2.0	43.8
2.5	52.0
3.0	58.9
3.5	64.1
4.0	67.5
4.5	69.2
5.0	69.0

(a) Compute finite-difference velocity estimates for this data. Does the velocity appear to be constant, linearly changing, or something else?

(b) Propose the simplest mathematical model $x(t)$ that captures the essential features of this motion. State what type of function you chose and why.

(c) What aspects of the data does your model capture well? What aspects does it miss?

(d) Is there enough information in this data to determine whether the object will continue past its apparent turning point, or stop and return? What additional measurement would help you decide?

Check your answer

(a) Finite-difference velocities (using 0.5 s intervals): | Interval | $v$ (cm/s) | |:---:|:---:| | 0.0--0.5 | 24.2 | | 0.5--1.0 | 22.8 | | 1.0--1.5 | 21.4 | | 1.5--2.0 | 19.2 | | 2.0--2.5 | 16.4 | | 2.5--3.0 | 13.8 | | 3.0--3.5 | 10.4 | | 3.5--4.0 | 6.8 | | 4.0--4.5 | 3.4 | | 4.5--5.0 | $-0.4$ | The velocity is decreasing, but not linearly --- it decreases more rapidly at later times. This suggests the acceleration is not constant but is itself changing (becoming more negative over time), or that a nonlinear model is more appropriate. (b) A reasonable simple model is a quadratic: $x(t) = v_0 t + \frac{1}{2}at^2$. With $v_0 \approx 25\,\text{cm/s}$ and $a \approx -5\,\text{cm/s}^2$, this gives $x(t) \approx 25t - 2.5t^2$. However, the velocity data suggests the deceleration is not perfectly constant (the velocity drops faster at the end), so a quadratic may not be perfect. A quadratic $x(t) = 25t - 2.5t^2$ is the simplest model. If you want to capture the apparent increase in deceleration, you might try $x(t) = At - Bt^2 - Ct^3$ with a small cubic term, but this adds complexity and you should only do so if the quadratic clearly fails. (c) The quadratic captures the overall shape --- initial forward motion, deceleration, and approach to a turning point. It may miss the subtle increase in deceleration at later times, but with only 11 data points and some noise, the discrepancy may not be statistically significant. (d) The data shows the velocity passing through zero near $t = 5\,\text{s}$, but you cannot determine from this data alone whether the object will reverse direction or simply stop. If $a \neq 0$ at the moment $v = 0$, the object will reverse. If $a = 0$ at that moment, it might stay at rest. You would need either more data points beyond $t = 5\,\text{s}$ or a direct measurement of acceleration at that time.

The Gap Between Models and Measurements

If working with messy data feels uncomfortable after the clean mathematics of the earlier sections, that is entirely normal. The first five sections of this chapter gave you precise tools: exact equations, exact derivatives, exact integrals. This section deliberately muddied the water --- because real physics always involves muddy water.

But notice something. The clean mathematical models from earlier are not made irrelevant by noisy data. They are made more valuable. When you proposed a constant-acceleration model for the cart data, you used the idealization from Section 2.2 as a hypothesis. When you computed finite-difference velocities, you used the definition of the derivative from the very beginning of the course, adapted for discrete data. When you discussed the smoothing tradeoff, you were reasoning about what kind of model the data can support.

The idealized mathematics gives you a framework for thinking about imperfect data. Without the clean theory, you would have no way to interpret the noisy measurements. Without the noisy measurements, the clean theory would have no connection to the physical world.

Real physics lives in the gap between the two.

Chapter 2 Summary

This chapter developed the mathematics of one-dimensional motion from the ground up. Here is what we built:

Sign conventions (2.1): Positive and negative encode direction, not "good" and "bad." A consistent sign convention is the foundation for everything else.
Constant acceleration (2.2): When $a$ is constant, integration gives exact formulas: $v(t) = v_0 + at$ and $x(t) = x_0 + v_0 t + \frac{1}{2}at^2$. These are not arbitrary equations --- they are the inevitable result of integrating a constant.
Nonuniform acceleration (2.3): When $a(t)$ varies, the general integral relationships $v(t) = v_0 + \int_0^t a(t')\,dt'$ and $x(t) = x_0 + \int_0^t v(t')\,dt'$ always apply. Constant acceleration is the special case where the integrals are trivial.
Piecewise motion (2.4): Real motion often involves multiple stages with different acceleration rules. Piecewise models handle this by enforcing continuity of position and velocity at stage boundaries.
Qualitative analysis (2.5): Turning points, extrema, and the signs of derivatives let you read the story of motion from graphs and equations without solving anything numerically.
Experimental data (2.6): Real measurements are noisy. Differentiating data amplifies noise; integrating smooths it. Data tells you what the motion is consistent with --- not what it definitely is.

The single most important idea in this chapter:

$$x(t) \xleftrightarrow{\text{differentiate}} v(t) \xleftrightarrow{\text{differentiate}} a(t)$$

$$x(t) \xleftrightarrow{\text{integrate}} v(t) \xleftrightarrow{\text{integrate}} a(t)$$

This chain --- running upward through derivatives and downward through integrals --- is the complete machinery of one-dimensional kinematics. Every section in this chapter was an application of moving up or down this chain under different conditions.

Chapter-End Retrieval

Close your notes. Scroll back to the top if you need to, but try not to. Answer these from memory.

1. What are the three fundamental kinematic quantities? How are they related through calculus?

2. Write the constant-acceleration kinematic equations. What single assumption makes them valid?

3. What changes when acceleration is not constant? What mathematical tool replaces the simple formulas?

4. In a piecewise motion problem, what conditions must hold at the boundary between two stages? Why?

5. Why does differentiating experimental data amplify noise? What is the practical consequence for estimating acceleration from position measurements?

6. In one sentence, what is the big idea of Chapter 2?

After you have attempted all six, check your answers against the chapter summary above.

Reflection

Chapter reflection: You started Chapter 2 knowing that position, velocity, and acceleration are connected by calculus. You now know how to use those connections --- under constant acceleration, variable acceleration, piecewise conditions, and with real experimental data.

What is the most important thing you learned in this chapter about describing motion mathematically? Write one sentence --- not a sentence about what was covered, but a sentence about what you understand now that you did not understand before.

Looking Ahead

Chapter 2 kept every object on a straight line. Position was a single number. Velocity was a single number with a sign. The mathematics was powerful but confined to one dimension.

In Chapter 3, the line becomes a plane --- and eventually, three-dimensional space. A drone traces a curve through the sky. A projectile arcs over a field. A satellite orbits the Earth. These motions cannot be captured by a single coordinate. You will need vectors: quantities with both magnitude and direction. Position becomes $\vec{r}(t)$, velocity becomes $\vec{v}(t)$, and the derivative chain from this chapter carries over unchanged --- only now the objects it connects are richer.

Here is the surprising part. When motion escapes the line and enters the plane, some things change dramatically --- velocity now has a direction that can rotate even when speed is constant, and acceleration can point sideways instead of along the path. But the deepest structure does not change at all. The relationship $\vec{v} = \frac{d\vec{r}}{dt}$ is the same derivative you have been using all chapter. The integral relationships work the same way. The kinematic chain survives the jump from one dimension to two and three.

Everything you built in Chapter 2 is about to become more powerful --- not by being replaced, but by being extended.