4.2 Linear Approximations and Differentials

The linearization of a function at a point and its use for approximating values, the differential as the change along the tangent, the direction of the error decided by concavity and its second-order growth with the step, and relative and percentage error in propagated measurements.

Subject: Calculus I · 70 slides · symbolic lesson

Open the interactive version of this deck

What this lesson covers

The lesson, slide by slide

1. Linear Approximations and Differentials

Title

Calculus I · Section 4.2

Replacing a curve by its tangent, and knowing how far to trust it

2. What this lesson gives you

Objectives

You have spent Chapter 3 computing derivatives. This lesson spends them: the tangent line becomes a machine for estimating values you cannot compute by hand, and the derivative becomes an exchange rate that turns a small change in the input into a change in the output.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, pp. 308-316 — learning objectives 4.2.1 to 4.2.4

Chapter 3 was about computing derivatives. This lesson is the first place you get paid for all that work. A derivative is a slope, and a slope is exactly what you need to estimate how far a function moves when its input moves a little.

There are two faces of the same idea here. The first is the linearization: replace a curve by its tangent line and read values off the line instead of the curve. The second is the differential: the change along that tangent line, used to estimate the change in a quantity, and especially the error in a quantity computed from a measurement.

You will also learn to judge your estimates. Every approximation in this lesson comes with two questions you can answer before touching a calculator: is it too high or too low, and roughly how far off is it? Those two questions are what separate an estimate you can trust from a guess.

3. Before anything new: a tangent line

Warm-up

Discussion prompt

From Chapter 3: write the equation of the tangent line to f(x) = 1/x at the point where x = 2. You need a point and a slope.

Write the line before you reveal the answer. You need two things: a point on the curve and the slope there. The point is two, one half. The slope is the derivative at two, and the derivative of one over x is minus one over x squared, which is minus one quarter at two.

Point-slope form then gives the line immediately. Keep it in that form rather than multiplying it out. Written as one half minus a quarter of x minus 2, it says: start at the known height one half, and move along a slope of minus a quarter for however far you have stepped from 2.

That reading, a known value plus a slope times a step, is the entire lesson in one sentence. Everything that follows is this line, used for new purposes.

4. The linearization

Section

Part 1

5. Zoom in and a curve turns into its tangent

Concept

Figure (svg): Left: the curve y equals one over x and its tangent line at x equals 2, with a small box drawn around the point of contact. Right: the inside of that box magnified, x from 1.9 to 2.2, where the curve and the tangent are two nearly parallel lines; at x equals 2.1 the curve is at 0.47619 and the tangent at 0.475.

Zoom in on the point of contact and the curve straightens out until it is barely distinguishable from its tangent. That is the whole idea of linear approximation.

A differentiable function looks straighter and straighter the closer you look at it, and the straight line it turns into is its tangent. So near the point of contact, the tangent can stand in for the curve.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, pp. 308-309 — Figure 4.7

On the left is the whole picture: the curve one over x and its tangent at 2. They touch at one point and then go their separate ways, the curve flattening toward the axis and the line diving below it.

Now look inside the small yellow box, magnified on the right. At that scale the curve has straightened out, and it runs almost exactly alongside the tangent line. At x equal to 2.1 the two heights are 0.47619 and 0.475, different only in the third decimal place.

This is what differentiable really means geometrically: zoom in far enough on any point and the graph looks like a straight line, and that line is the tangent. If the curve and the line are nearly the same up close, then you can use the line, which is easy to evaluate, in place of the curve, which may not be.

6. The tangent as an estimate: one over x near 2

Concept

Use the warm-up line in place of the curve and read off a value just to the right of the anchor.

\[ y = \frac12 - \frac14(x - 2) \]

\[ x = 2.1: \quad y = \frac12 - \frac14(0.1) = 0.475 \]

\[ f(2.1) = \frac{1}{2.1} \approx 0.47619 \]

The line is off by about one thousandth, and evaluating it needed nothing harder than a quarter of a tenth.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 308 — the tangent to 1/x at a = 2

Here is the idea used for the first time. Substitute 2.1 into the tangent line instead of into the function. The step from the anchor is one tenth, a quarter of that is 0.025, and one half minus 0.025 is 0.475.

The true value, one over 2.1, is 0.47619 to five places. So the line is off by a little over one thousandth. That is a good estimate, and it cost you nothing more than a subtraction.

Of course, you could divide 1 by 2.1 by hand without much trouble. The point of starting here is that you can check everything, so you can see how good the idea is before trusting it with functions like square roots and sines, where there is no easy hand calculation.

7. Near the anchor and far from it

Picture it

Figure (svg): The curve y equals one over x and its tangent at x equals 2, drawn out to x equals 11. At x equals 10 the curve is at 0.1 while the tangent has fallen to minus 1.5, and a dashed red segment marks the gap of 1.6 between them.

One tenth of a unit from the anchor the tangent was right to three decimal places; eight units away it is not even the right sign. A linearization is a local tool.

\[ x = 10: \quad y = \frac12 - \frac14(10 - 2) = -1.5, \qquad f(10) = 0.1 \]

The same line that was right to three decimal places at 2.1 is not even the right sign at 10. Distance from the anchor is what decides the quality.

Now push the same line much further, to x equal to 10. The line has kept falling at a steady quarter per unit, and by 10 it has reached minus one and a half. The curve, meanwhile, is still positive at one tenth. The red dashed segment shows a gap of 1.6, which is sixteen times the true value.

So the tangent line is a local tool. It is excellent close to the anchor and gets steadily worse as you move away, because the curve bends and the line does not.

Keep this picture in mind when you choose an anchor for an estimate. The single most important decision is to anchor as close to the target as you can. Part 4 of this lesson measures exactly how fast the quality falls off with distance.

8. The linearization of f at a

Concept

Write the tangent line as a function of x and give it a name.

\[ f(x) \approx f(a) + f'(a)(x - a) \quad \text{for } x \text{ near } a \]

Linearization (equation 4.1) — The linear function L whose graph is the tangent line to f at a. It is also called the linear approximation or tangent line approximation of f at a.

\[ L(x) = f(a) + f'(a)(x - a) \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 309 — equation 4.1

This slide just gives the tangent line a name and a new job. For any differentiable function and any point a, the function near a is approximately its value at a plus its slope at a times the step from a.

The linear function on the right side is called the linearization of f at a, and the book also calls it the linear approximation or the tangent line approximation. It is equation 4.1, and it is worth memorising in exactly this form, because the form tells you how to use it.

Notice that the linearization is not a new object. You have been writing tangent lines since Chapter 3. What is new is the question you ask of it: not what line touches the curve here, but how well does this line stand in for the curve nearby.

9. Reading equation 4.1 piece by piece

Notation

Annotate

On: \( L(x) = f(a) + f'(a)\,(x - a) \)

  • The anchor: a point near your target where f and f′ are easy to evaluate exactly, like 9 for √x or π/3 for sin x.
  • Where you start: the exact height of the curve at the anchor. The line passes through this point.
  • The slope you travel along: the derivative at the anchor, a single number once a is chosen.
  • The step from the anchor to the target. Small steps give good estimates; the rest of the lesson makes that precise.

Step through the annotations one at a time. The anchor a is your choice, and it is the choice that matters most. You want a point where you know both the function value and the slope exactly, and which is as close to your target as possible.

f of a is the starting height and f prime of a is the slope. Once a is chosen, both are just numbers, so the linearization is an ordinary straight line: a constant plus a constant times x minus a.

The last piece, x minus a, is the step. It is the only thing that changes when you move the target around. That is why a linear approximation is so cheap: all the calculus is done once, at the anchor, and every estimate after that is one multiplication and one addition.

10. Example 4.5: the square root of 9.1

Worked example

Find the linearization of the square root at 9 and use it to estimate the square root of 9.1.

\[ f(x) = \sqrt{x}, \quad a = 9 \]

Evaluate f at the anchor

Why: Nine is a perfect square, which is why it was chosen.

\[ f(9) = \sqrt9 = 3 \]

Differentiate

Why: Power rule with exponent one half.

\[ f'(x) = \frac{1}{2\sqrt{x}} \]

Evaluate the slope at the anchor

Why: Put x equal to 9.

\[ f'(9) = \frac{1}{2\sqrt9} = \frac16 \]

Assemble the linearization

Why: Point plus slope times step.

\[ L(x) = 3 + \frac16(x - 9) \]

Figure (svg): The curve y equals the square root of x from 0 to 17 and its tangent line at the point (9, 3), which has slope one sixth; the tangent runs just above the curve on both sides of the point.

Near 9 the tangent and the square root are almost the same curve, so evaluating the easy line at 9.1 gives almost exactly the square root of 9.1.

Evaluate at 9.1

Why: The step is one tenth.

\[ \sqrt{9.1} \approx L(9.1) = 3 + \frac16(0.1) = 3.01667 \]

Check with a calculator

Why: The true value to five places is 3.01662, so the estimate is high by about 0.00005.

\[ \sqrt{9.1} = 3.01662, \qquad 3.01667 - 3.01662 = 0.00005 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, pp. 309-310 — Example 4.5 and Figure 4.8

This is the model example, and every later one follows the same four moves. Choose the anchor: 9, because it is a perfect square and only one tenth from 9.1. Evaluate the function there: 3. Evaluate the derivative there: one over twice the root of 9, which is one sixth. Assemble the line.

Evaluating the line at 9.1 is then one sixth of a tenth added to 3, which is 3.01667. The picture shows why that is so accurate: at 9.1 you are so close to the point of contact that the line and the curve have barely separated.

The check against a calculator shows the estimate is high by about five hundred-thousandths. That it is high rather than low is not an accident; you can see the orange line riding just above the blue curve. Part 3 turns that observation into a rule you can apply before computing anything.

11. Choose the function and the anchor

Step zero

\[ \sqrt[3]{8.1} \approx \;? \]

Discussion prompt

Checkpoint 4.5 asks for the cube root of 8.1 to five decimal places. Before any calculus: which function f and which anchor a should you use, and why that a?

Answer before revealing. This is the step most people rush, and it decides everything else.

The function is forced by the target: you want a cube root, so f is the cube root function. The anchor is the real decision. It must be close to 8.1, and it must make both the cube root and its derivative easy to evaluate exactly.

Eight does both. It is a perfect cube, so its cube root is exactly 2, and the derivative, which involves 8 to the two thirds, becomes 4. A number like 8.1 itself would be close but useless, since you cannot evaluate the cube root there, which was the whole problem. And 1 or 27 would be easy but far away.

12. Checkpoint 4.5: the cube root of 8.1

Worked example

\[ f(x) = x^{1/3}, \quad a = 8, \quad \text{estimate } \sqrt[3]{8.1} \]

Evaluate f at the anchor

Why: Two cubed is eight.

\[ f(8) = 8^{1/3} = 2 \]

Differentiate

Why: Power rule: the exponent drops by one to minus two thirds.

\[ f'(x) = \frac13 x^{-2/3} \]

Evaluate the slope

Why: Eight to the two thirds is 4.

\[ f'(8) = \frac13 \cdot \frac{1}{8^{2/3}} = \frac13\cdot\frac14 = \frac{1}{12} \]

Write the linearization

Why: Point plus slope times step.

\[ L(x) = 2 + \frac{1}{12}(x - 8) \]

Evaluate at 8.1

Why: One twelfth of a tenth is 0.008333.

\[ \sqrt[3]{8.1} \approx 2 + \frac{0.1}{12} = 2.00833 \]

Check by cubing

Why: Cube the estimate: it should land just above 8.1, since the estimate is slightly high; the true cube root is 2.00830.

\[ 2.00833^3 \approx 8.1004, \qquad \sqrt[3]{8.1} = 2.00830 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 310 — Checkpoint 4.5

The same four moves as Example 4.5. The derivative of x to the one third is one third of x to the minus two thirds, and at 8 that is one third times one quarter, so the slope is one twelfth.

The estimate is 2 plus one twelfth of a tenth, which is 2.00833. The checkpoint asks for five decimal places, and here is a good lesson: the true value to five places is 2.00830, so the linear estimate is correct to four places but not quite five. A linear approximation is good, not perfect.

The check by cubing is a nice independent test that needs no cube-root button. Cube 2.00833 and you get about 8.1004, a little over 8.1, which confirms that the estimate is slightly too big. That matches the picture of a concave-down curve with its tangent above.

13. A good anchor is near and easy

Intuition

Every linear estimate starts with the same choice, and the choice is a compromise between two demands.

demandwhytypical anchors
close to the targetthe error grows fast with distance, as the far picture showedthe nearest one available
f(a) exactly knownthe estimate starts from this valueperfect squares and cubes, 0, 1
f′(a) exactly knownthe slope multiplies the steplandmark angles such as π/6, π/4, π/3

The square root of 10 anchored at 9 is excellent; anchored at 4, where the root is just as easy, it is dreadful, because 4 is six units away.

Every linear estimate begins with a trade-off. Nearness keeps the error small, but the anchor also has to be a place where you can compute the function and its derivative exactly. The table lists the kinds of anchors that satisfy the second demand: perfect powers, zero and one, and the landmark angles from trigonometry.

When both demands pull in the same direction, the choice is easy. The square root of 10 anchored at 9 is excellent, because 9 is a perfect square just one unit away. Anchored at 4, also a perfect square, the estimate comes out as 3.5, off by about a third, roughly seventy-seven times worse.

So the rule of thumb is: among the points where everything is easy, take the nearest one.

14. Match each target to its anchor

Matching

Match the pairs

  • a. (2.001)⁶
  • b. sin(0.02)
  • c. (15.99)^(1/4)
  • d. 1/0.98
  • e. cos(0.03)
  • p. f(x) = x⁶ at a = 2
  • q. f(x) = sin x at a = 0
  • r. f(x) = x^(1/4) at a = 16
  • s. f(x) = 1/x at a = 1
  • t. f(x) = cos x at a = 0

Why: In each case the anchor is the nearest input where the function is exactly known: 2 to the sixth is 64, the fourth root of 16 is 2, and sine, cosine and one over x are trivial at 0 and 1. These are Exercises 56 to 60, which ask for each value to within 0.01.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 316 — Exercises 56-60

Try each pair before checking. For every target, ask what function it is a value of and what nearby input makes that function trivial.

The pattern is always the same: split the target into an easy number plus a small step. Two point zero zero one is 2 plus one thousandth, and 2 to the sixth is 64. Fifteen point nine nine is 16 minus a hundredth, and the fourth root of 16 is 2. Point nine eight is 1 minus two hundredths, and one over 1 is 1. The small angles 0.02 and 0.03 radians are tiny steps from 0, where sine and cosine are 0 and 1.

These are Exercises 56 to 60. Carry any one of them through and you will find the linear estimate lands within a hundredth of the true value, which is what those exercises ask for.

15. Example 4.6: the sine of 62 degrees

Worked example

Find the linearization of sine at pi over 3 and use it to estimate the sine of 62 degrees.

Choose the anchor

Why: Pi over 3 radians is 60 degrees, only two degrees from the target, and its sine and cosine are known exactly.

\[ f(x) = \sin x, \quad a = \frac{\pi}{3} \]

Evaluate f and its derivative there

Why: The derivative of sine is cosine.

\[ f\left(\tfrac{\pi}{3}\right) = \frac{\sqrt3}{2}, \qquad f'\left(\tfrac{\pi}{3}\right) = \cos\tfrac{\pi}{3} = \frac12 \]

Write the linearization

Why: Point plus slope times step.

\[ L(x) = \frac{\sqrt3}{2} + \frac12\left(x - \frac{\pi}{3}\right) \]

Figure (svg): The sine curve from minus 1 to 5.5 and its tangent line at the point (pi over 3, root 3 over 2), which has slope one half. The tangent touches the arch near its top and rises away above it on both sides.

Sixty-two degrees is only two degrees past the landmark angle pi over 3, so the point of interest sits right beside the point of contact.

Convert the target to radians

Why: The derivative of sine is cosine only when x is in radians.

\[ 62^\circ = \frac{62\pi}{180} \]

Find the step

Why: Sixty-two degrees minus sixty degrees.

\[ \frac{62\pi}{180} - \frac{\pi}{3} = \frac{62\pi - 60\pi}{180} = \frac{2\pi}{180} = \frac{\pi}{90} \]

Evaluate the linearization

Why: Half of pi over 90 is pi over 180.

\[ \sin 62^\circ \approx \frac{\sqrt3}{2} + \frac12\cdot\frac{\pi}{90} = \frac{\sqrt3}{2} + \frac{\pi}{180} \approx 0.88348 \]

Check with a calculator

Why: The true value is 0.88295, so the estimate is high by about 0.0005: two degrees is a small step.

\[ \sin 62^\circ = 0.88295, \qquad 0.88348 - 0.88295 = 0.00053 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, pp. 310-311 — Example 4.6 and Figure 4.9

The anchor is the landmark angle nearest to 62 degrees, which is 60 degrees, or pi over 3 radians. There the sine is root 3 over 2 and the cosine, which is the slope, is one half.

The step that deserves attention is the conversion. The derivative of sine is cosine only when x is measured in radians, so the step from 60 to 62 degrees must also be in radians: 2 degrees is 2 pi over 180, which is pi over 90. Half of that is pi over 180, and adding it to root 3 over 2 gives 0.88348.

The calculator says 0.88295, so the estimate is high by about five ten-thousandths. The picture explains both facts: 62 degrees is right beside the point of contact, so the error is small, and the arch of the sine curve bends downward there, so the tangent sits above it.

16. Trap: stepping in degrees

Trap

The trap

Measuring the step as two degrees:

\[ \sin 62^\circ \approx \frac{\sqrt3}{2} + \frac12(2) \]

\[ = 1.866 \]

Wrong: a sine above 1.

The fix

The slope one half is per radian, so the step must be in radians too. Two degrees is pi over 90, about 0.0349.

\[ \frac{\sqrt3}{2} + \frac12\cdot 0.0349 = 0.8835 \]

A sine bigger than one should always stop you.

This error is easy to make because the question is posed in degrees and the step, 2, looks so natural. But the slope one half came from differentiating sine, and that derivative is only valid with the angle in radians. Multiplying a per-radian slope by a step in degrees mixes units, like multiplying a speed in kilometres per hour by a time in minutes.

The result gives itself away if you look at it. A sine can never exceed 1, and this answer is 1.866. Whenever an estimate lands outside the range of the function, go back and check the units of the step. The fix is to convert first: two degrees is about 0.0349 radians, and then the estimate is the sensible 0.8835.

17. Checkpoint 4.6: cosine at pi over 2

Worked example

\[ f(x) = \cos x, \quad a = \frac{\pi}{2} \]

Evaluate f at the anchor

Why: Cosine of a right angle.

\[ f\left(\tfrac{\pi}{2}\right) = \cos\tfrac{\pi}{2} = 0 \]

Differentiate and evaluate

Why: The derivative of cosine is minus sine.

\[ f'(x) = -\sin x, \qquad f'\left(\tfrac{\pi}{2}\right) = -\sin\tfrac{\pi}{2} = -1 \]

Write the linearization

Why: The first term is zero.

\[ L(x) = 0 + (-1)\left(x - \frac{\pi}{2}\right) \]

Simplify

Why: Distribute the minus sign.

\[ L(x) = \frac{\pi}{2} - x \]

Figure (svg): The cosine curve from minus 0.5 to 3.6 and its tangent line at (pi over 2, 0), the straight line y equals pi over 2 minus x, which crosses the curve there with slope minus one.

At pi over 2 the cosine crosses the axis at its steepest, so the tangent is a line of slope minus one through that crossing.

Check at a nearby point

Why: At x equal to 1.6, just past pi over 2, the line and the curve agree to four decimal places.

\[ L(1.6) = -0.029204, \qquad \cos 1.6 = -0.029200 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 311 — Checkpoint 4.6

At pi over 2 the cosine is zero and its derivative, minus the sine, is minus one. So the linearization is zero plus minus one times x minus pi over 2, which tidies up to pi over 2 minus x.

The picture shows what this line is: the cosine curve crosses the axis at pi over 2, and it crosses at its steepest, with slope minus one. The tangent is the straight line through that crossing with the same slope.

The check picks a point just past the anchor, x equal to 1.6, and compares. The line gives minus 0.029204 and the cosine gives minus 0.029200, agreeing to four decimal places. That extraordinary agreement is partly luck of the anchor: at pi over 2 the cosine has an inflection point, so it hardly bends near there, and the tangent hugs it especially closely.

18. Example 4.7: roots and powers near 1

Worked example

Find the linearization of one plus x to the n at 0, then use it to estimate 1.01 cubed.

\[ f(x) = (1 + x)^n, \quad a = 0 \]

Evaluate f at the anchor

Why: Any power of 1 is 1.

\[ f(0) = 1^n = 1 \]

Differentiate

Why: Power rule with the chain rule; the inner derivative is 1.

\[ f'(x) = n(1 + x)^{n-1} \]

Evaluate the slope

Why: Again a power of 1.

\[ f'(0) = n\cdot 1^{n-1} = n \]

Write the linearization

Why: Point plus slope times step.

\[ L(x) = 1 + n(x - 0) = 1 + nx \]

Figure (svg): The curve y equals one plus x, cubed, from minus 2.2 to 1.2 and its tangent line y equals 1 plus 3x at the point (0, 1). The curve flattens to the left and climbs steeply to the right; the tangent lies below it near 0 on the right.

The tangent at 0 is the line 1 plus 3x. Just to the right of 0 the cube climbs slightly faster than the line, which is why the estimate for 1.01 cubed comes out a hair low.

Apply it with n equal to 3 and x equal to 0.01

Why: 1.01 is 1 plus 0.01.

\[ (1.01)^3 = f(0.01) \approx 1 + 3(0.01) = 1.03 \]

Check by multiplying out

Why: The binomial expansion shows exactly what the line left out: the square and cube of the step.

\[ (1.01)^3 = 1 + 0.03 + 0.0003 + 0.000001 = 1.030301 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, pp. 311-312 — Example 4.7 and Figure 4.10

This example does something new: it keeps the exponent n as a letter, so the result works for every power at once. At zero, one plus x to the n is 1, and its derivative, n times one plus x to the n minus one, is just n. So the linearization is 1 plus n x.

In words: raising a number slightly bigger than 1 to the power n multiplies the small excess by n. So 1.01 cubed is about 1.03, and 1.02 to the fifth is about 1.1.

The check multiplies 1.01 cubed out completely, and it shows you precisely what the linear estimate threw away: the terms 0.0003 and 0.000001, which are the square and the cube of the step, times some constants. Those discarded terms are tiny because the step is tiny, and they are all positive, which is why the estimate is slightly low. Keep this in mind; Part 4 is about exactly those discarded terms.

19. Checkpoint 4.7: one plus x to the fourth

Fill the middle

Find the linearization at 0 directly, without quoting Example 4.7.

Fill in the blanks

f(x) = (1+x)^4, \quad f(0) = 1, \quad f'(0) = 4, \quad L(x) = 1 + 4x

Why: The derivative is 4 times one plus x cubed, which is 4 at x equal to 0, so the linearization is 1 plus 4x: the n equal to 4 case of Example 4.7. It predicts 1.01 to the fourth is about 1.04; the true value is 1.0406.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 312 — Checkpoint 4.7

Fill in both blanks before checking, and do it by differentiating, not by quoting the formula from the previous slide. The checkpoint says so deliberately: the point is to practise the method, not the result.

The derivative of one plus x to the fourth is 4 times one plus x cubed, by the power rule and the chain rule. At zero that is 4. The value at zero is 1. So the linearization is 1 plus 4x, which is exactly what Example 4.7 predicts with n equal to 4.

As a quick use of it, 1.01 to the fourth is about 1.04. Multiplying out gives 1.0406, so again the linear estimate is slightly low, by the discarded higher-power terms.

20. Four linearizations at zero worth knowing

Concept

Anchoring at 0 turns the familiar functions into their simplest lines. Each follows from equation 4.1 with f of 0 and f prime of 0.

\[ \sin x \approx x, \qquad \cos x \approx 1, \qquad e^x \approx 1 + x \]

\[ (1 + x)^n \approx 1 + nx, \qquad \ln(1 + x) \approx x \]

The first is the small-angle approximation of physics; the fourth is the one Example 4.7 derived. All of them are only as good as x is small.

\[ \sin(0.1) = 0.099833 \approx 0.1, \qquad e^{0.1} = 1.10517 \approx 1.1 \]

These five lines are so useful that people carry them around in their heads. Each one is equation 4.1 at a equal to zero, using the value and the slope at zero. Sine is 0 with slope 1, so it is about x. Cosine is 1 with slope 0, so it is about 1. The exponential is 1 with slope 1. The logarithm of one plus x is 0 with slope 1.

You will see them all over physics and engineering, usually without any mention of calculus. The small-angle approximation for a pendulum is the first. Compound interest over a short period uses the exponential one. The power one appears in relativity, in finance and in error analysis.

The numbers at the bottom are the honesty check: at x equal to one tenth, each approximation is right to about two or three decimal places. At x equal to one, they are all poor. Every one of these is only as good as x is small.

21. Exercise 49: when is the linear approximation exact?

Counterexample

Figure (svg): The curve y equals x cubed from minus 2.6 to 2, and its tangent line y equals 3x minus 2 at the point (1, 1). The tangent touches the curve at x equals 1 and crosses it again at the point (minus 2, minus 8).

The linearization is exact wherever the tangent line meets the curve. For x cubed at 1 that happens twice: at the anchor, and again at minus 2, three units away.

Discussion prompt

A classmate says: the linear approximation is exact only at x = a, unless f itself is a line. Use the picture to repair the claim, and say when L equals f everywhere.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 316 — Exercises 46 and 49

Answer the question before revealing. The tempting claim sounds right, and it is half right: for a genuinely curved function, the linearization is certainly exact at the anchor, because the line passes through that point.

But look at the picture. The tangent to x cubed at 1 is the line 3x minus 2, and that line does not stay on one side of the curve. It crosses the cubic again at minus 2, where both equal minus 8. At that second point the linearization gives the exact value, even though it is three units from the anchor. Setting x cubed equal to 3x minus 2 and factoring confirms it: a double root at 1, where the line touches, and a single root at minus 2, where it crosses.

So the correct statement is that L is exact wherever the tangent line meets the curve. It is exact everywhere only when f is itself linear, because the tangent line to a line is that same line. That is Exercise 46.

22. Differentials

Section

Part 2

23. Giving dx and dy a meaning of their own

Concept

Until now dy over dx was one symbol. Now split it: let dx be an independent variable, any nonzero number, and define dy from it.

\[ dy = f'(x)\,dx \qquad (4.2) \]

Then dy depends on two things, the point x and the step dx. Divide both sides by dx and the familiar derivative comes back.

\[ \frac{dy}{dx} = f'(x) \qquad (4.3) \]

Differentials — The quantities dx and dy. Equation 4.2 is the differential form of equation 4.3.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 312 — equations 4.2 and 4.3

Since Chapter 3 you have written dy over dx as a single symbol for the derivative, with a warning that it is not really a fraction. This slide finally gives dx and dy meanings of their own, so that the fraction becomes an honest one.

The definition is simple. Let dx be any nonzero number you like, a step in the input. Then define dy to be the derivative times that step. That is equation 4.2. Now dy is a number that depends on two things: where you are, x, and how far you step, dx.

Divide both sides by dx and you get back dy over dx equals f prime of x, equation 4.3. So the old notation was consistent all along. What you have gained is the ability to write a relationship between small changes in multiplied-out form, dy equals f prime dx, which is exactly the form you need for estimating changes.

24. Reading the differential piece by piece

Notation

Annotate

On: \( dy = f'(x)\,dx \)

  • A step in the input that YOU choose: any nonzero number, like 0.1. It is not an infinitely small thing here.
  • The slope at the starting point x: the exchange rate between input change and output change.
  • The rise along the tangent line over that step. It depends on both x and dx.
  • dy is the change on the tangent, not on the curve. The true change Δy = f(x + dx) − f(x) is a different number.

The first annotation corrects a common misreading. In this section dx is not an infinitely small quantity or a mysterious limit. It is a real step that you choose, like one tenth or two hundredths.

The slope f prime of x converts that step into a change in the output. Think of it as an exchange rate: at x equal to 3, the function x squared plus 2x exchanges each unit of input for 8 units of output, so a step of a tenth buys 0.8.

The last annotation is the one to keep. The differential dy is the rise along the tangent line, not along the curve. The true change in the function is a different number, written delta y, and the whole point of the next few slides is how close the two are.

25. Example 4.8(a): the differential of a polynomial

Worked example

Find dy and evaluate it when x equals 3 and dx equals 0.1.

\[ y = x^2 + 2x \]

Differentiate

Why: The key step: once you have the derivative, dy follows immediately.

\[ f'(x) = 2x + 2 \]

Write the differential

Why: Multiply the derivative by dx.

\[ dy = (2x + 2)\,dx \]

Substitute x equal to 3 and dx equal to 0.1

Why: Evaluate the slope first, then multiply.

\[ dy = (2\cdot 3 + 2)(0.1) = 8(0.1) = 0.8 \]

Check against the actual change

Why: The function really moves from 15 to 15.81 over this step, so 0.8 is within one hundredth of the truth.

\[ f(3.1) - f(3) = 15.81 - 15 = 0.81 \approx 0.8 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, pp. 312-313 — Example 4.8a

As the book says, the key step is the derivative. Once you have it, the differential is just that derivative with dx attached. For x squared plus 2x the derivative is 2x plus 2, so dy is 2x plus 2, times dx.

Evaluating at x equal to 3 and dx equal to a tenth, the slope is 8, so dy is 0.8. Notice the order: evaluate the slope at the point first, then multiply by the step.

The check computes the true change in the function over the same step. It moves from 15 to 15.81, a change of 0.81. So the differential, which needed no function values at all beyond the slope, is within a hundredth of the truth. Example 4.9 comes back to exactly this pair of numbers.

26. Example 4.8(b): the differential of cosine

Worked example

Find dy and evaluate it when x equals 3 and dx equals 0.1.

\[ y = \cos x \]

Differentiate

Why: The derivative of cosine is minus sine.

\[ f'(x) = -\sin x \]

Write the differential

Why: Multiply by dx.

\[ dy = -\sin x\,dx \]

Substitute

Why: Leave the sine of 3 exact, then evaluate.

\[ dy = -\sin(3)(0.1) = -0.1\sin 3 \approx -0.01411 \]

Compare with the actual change

Why: Here the true change is noticeably smaller: at 3 the slope is small, so the curve's bending is comparable with its slope over a step this long.

\[ \cos 3.1 - \cos 3 \approx -0.00914 \]

Check with a smaller step

Why: Shrink dx to 0.001 and the differential and the true change agree to within half a percent, as they must.

\[ dy = -0.000141, \qquad \Delta y = \cos 3.001 - \cos 3 = -0.000141 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 313 — Example 4.8b

The differential is minus sine x times dx, and at x equal to 3 with a step of a tenth that is minus a tenth of sine 3, about minus 0.0141. The book leaves the answer as minus 0.1 sine 3, which is perfectly acceptable.

Now something instructive happens when you compare with the true change. Cosine really moves by only about minus 0.0091 over this step, so dy is off by more than half. Why is it so poor here? At 3 the cosine is near the bottom of its trough, so its slope is small, about 0.14, but its bending is large. Over a step of a tenth, the bending contributes almost as much as the slope does.

The check with a smaller step shows that nothing is wrong with the method. Shrink the step to a thousandth and the differential and the true change agree to within half a percent. Differentials are always good for small enough steps; how small is small enough depends on how sharply the function bends.

27. Checkpoint 4.8: the differential of e to the x squared

Fill the middle

Fill in the blanks

y = e^2xe^{x^2} \;\Longrightarrow\; dy = ___\,dx

Why: By the chain rule the derivative of e to the x squared is e to the x squared times the derivative of x squared, which is 2x. The differential is that derivative times dx.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 313 — Checkpoint 4.8

Type your answer before checking. The only work is the derivative, and it needs the chain rule. The outer function is the exponential, whose derivative is itself, and the inner function is x squared, whose derivative is 2x.

So the derivative of e to the x squared is 2x times e to the x squared, and the differential is that times dx. Leaving out the factor 2x is the usual slip, and it is the chain rule, not anything new about differentials, that catches people here.

28. Why dy approximates the true change

Concept

Step from a to a plus dx. The true change is measured on the curve; replace the curve by its linearization and see what is left.

\[ \Delta y = f(a + dx) - f(a) \]

\[ f(a + dx) \approx L(a + dx) = f(a) + f'(a)(a + dx - a) \]

\[ f(a + dx) - f(a) \approx L(a + dx) - f(a) = f'(a)\,dx \]

\[ \Delta y \approx f'(a)\,dx = dy \]

So the differential is the change in the linearization, and it approximates the change in the function exactly as well as L approximates f.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 313 — the derivation before Figure 4.11

This is the short argument that connects the two halves of the lesson so far. Start from the true change, the new value minus the old one. Replace the new value by the linearization evaluated there. The linearization at a plus dx is f of a plus f prime of a times dx, because the step from a is exactly dx.

Subtract f of a from both sides and everything cancels except f prime of a times dx, which is the differential. So the differential is nothing more than the change in the linearization. Wherever the linearization is a good stand-in for the function, the differential is a good stand-in for the true change.

That means everything you learn about the accuracy of linear approximations applies equally to differentials. They are one idea, seen from two angles: the linearization estimates a value, the differential estimates a change.

29. Two climbs from the same point

Picture it

Figure (svg): The curve y equals x squared and its tangent at a equals 1. A dashed horizontal step of length dx equals 0.6 runs from the point (1, 1). At a plus dx equals 1.6, a short vertical segment up to the tangent (height 2.2) is labelled dy equals 1.2, and a longer bracket up to the curve (height 2.56) is labelled delta y equals 1.56. The small leftover piece between the tangent and the curve is the error, 0.36.

Both changes start from the same point. The differential climbs along the tangent; the true change climbs along the curve. The small red piece at the top is what dy misses.

\[ f(x) = x^2,\; a = 1,\; dx = 0.6: \quad dy = 2(1)(0.6) = 1.2, \quad \Delta y = 1.6^2 - 1 = 1.56 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 313 — Figure 4.11

This is Figure 4.11 from the book, drawn with a real function so the lengths are honest. The function is x squared, the anchor is 1, and the step is 0.6, large enough that the pieces are visible.

Both changes start from the same point, one, one, and move right by the same dashed step. The yellow segment climbs up to the tangent line: that is dy, 2 times 0.6, which is 1.2. The green bracket on the right climbs all the way to the curve: that is delta y, 1.6 squared minus 1, which is 1.56.

The short red piece at the top of the yellow segment is the difference, 0.36, and it is exactly 0.6 squared. The curve is concave up here, so it pulls away above its tangent and the differential falls short. Hold on to that red piece; Part 4 is entirely about how big it is.

30. Example 4.9: dy against delta y

Worked example

Compute the actual change and the differential at x equal to 3 with dx equal to 0.1.

\[ y = x^2 + 2x \]

Evaluate f at the new point

Why: x moves from 3 to 3.1.

\[ f(3.1) = (3.1)^2 + 2(3.1) = 9.61 + 6.2 = 15.81 \]

Evaluate f at the old point

Why: Nine plus six.

\[ f(3) = 3^2 + 2(3) = 15 \]

Subtract for the actual change

Why: This is delta y.

\[ \Delta y = 15.81 - 15 = 0.81 \]

Compute the differential

Why: The slope at 3 is 8.

\[ dy = f'(3)\,dx = (2(3) + 2)(0.1) = 0.8 \]

Check the gap against the picture

Why: The difference is exactly dx squared, the corner the tangent misses for a squared term.

\[ \Delta y - dy = 0.81 - 0.8 = 0.01 = (0.1)^2 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 314 — Example 4.9

This example puts the two changes side by side for the same function as Example 4.8(a). The true change needs two function values: 15.81 at 3.1 and 15 at 3, a difference of 0.81. The differential needs only the slope at 3, which is 8, times the step, giving 0.8.

The book's point is that the differential is much simpler to compute and nearly as good. The check adds a second point that the book leaves implicit: the gap between the two is 0.01, which is exactly the step squared.

That is not a coincidence. This function is x squared plus 2x, and the 2x part is a straight line, which the tangent reproduces perfectly. Only the x squared part bends, and for x squared the missing piece is always the step squared, the red corner from the previous picture.

31. Checkpoint 4.9: the same function, twice the step

Worked example

\[ y = x^2 + 2x, \quad x = 3, \quad dx = 0.2 \]

Evaluate f at 3.2

Why: Square and double.

\[ f(3.2) = 10.24 + 6.4 = 16.64 \]

Subtract f of 3

Why: f of 3 is 15.

\[ \Delta y = 16.64 - 15 = 1.64 \]

Compute the differential

Why: Same slope, 8, twice the step.

\[ dy = 8(0.2) = 1.6 \]

Check how the gap grew

Why: Doubling the step doubled dy but quadrupled the gap: 0.04 against 0.01.

\[ \Delta y - dy = 0.04 = (0.2)^2 = 4(0.1)^2 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 314 — Checkpoint 4.9

Work the checkpoint the same way, then compare with Example 4.9. With a step of 0.2, the new value is 16.64, so the true change is 1.64. The differential is the same slope, 8, times 0.2, which is 1.6.

Now look at the two examples together. Doubling the step exactly doubled the differential, from 0.8 to 1.6, because the differential is linear in the step. But the gap between the true change and the differential went from 0.01 to 0.04. It quadrupled.

That is your first measured evidence of the square law: the error of a linear approximation grows like the square of the step. Keep this pair of numbers in mind; the rest of the lesson explains it and uses it.

32. Along the tangent or along the curve?

Sorting

Sort into buckets

Each expression is either the change along the tangent (dy) or the actual change along the curve (Δy). Sort them.

dy: along the tangent
f′(a) dx; L(a + dx) − f(a); 0.8 for x² + 2x at 3, dx = 0.1
Δy: along the curve
f(a + dx) − f(a); 0.81 for x² + 2x at 3, dx = 0.1; the true propagated error in a computed value
dy
f′(a)dx, the change in L, and the 0.8 of Example 4.9 are all computed from the slope at a: they live on the tangent line.
Dy
f(a + dx) − f(a), the 0.81, and the true propagated error all use the function's actual values: they live on the curve.

Sort each expression, then check the rationales. The test for each one is simple: does computing it use the function's actual values, or only the slope at the anchor?

Anything built from f prime of a alone lives on the tangent line. That includes the change in the linearization, because the linearization is the tangent line, and the 0.8 from Example 4.9, which was computed as slope times step.

Anything that uses f at the new point lives on the curve. That includes the 0.81, and also the true propagated error in a measurement, which is the real difference between two function values. The reason the next part of the lesson estimates that error with dy is precisely that the true one is unavailable.

33. Find the error: the differential reported as the change

Error analysis

Annotate

On: \( A = x^2,\; x: 2 \to 3 \;\Longrightarrow\; \Delta A = dA = 2(2)(1) = 4 \)

  • The differential is computed correctly: slope 2x = 4 at x = 2, times a step of 1.
  • This equals sign is the mistake. dA is an approximation to ΔA, not ΔA itself.
  • ΔA = 3² − 2² = 9 − 4 = 5. The missing 1 is the corner square, dx² = 1².
  • With a step as big as 1 the error is a fifth of the answer. Differentials are for SMALL steps.

Read the line and find the problem before stepping through the annotations. The arithmetic is fine: the derivative of x squared is 2x, which is 4 at 2, and times a step of 1 that is 4.

The mistake is the equals sign between the true change and the differential. The true change in area from a side of 2 to a side of 3 is 9 minus 4, which is 5. The differential is 4. The missing 1 is the corner square in the growing-square picture you will see in Part 4: a step of 1, squared.

The deeper problem is the size of the step. A step of 1 on a side of 2 is a fifty percent change, not a small one, and the differential misses a fifth of the answer. Always write approximately equal between delta y and dy, and only trust it when the step is small compared with the scale of the problem.

34. Why settle for dy?

Intuition

If the true change is available, why approximate it? Three reasons.

The third reason is the one that matters most, and Part 5 is built on it.

It is a fair question: if you can compute the true change, why use an approximation? The first two reasons are about convenience. A differential is a single multiplication once the slope is known, and because it is proportional to the step, it is easy to run backwards: you can ask how big a step is allowed before the change exceeds some limit, and solve a linear equation.

The third reason is the important one, and it is not about convenience at all. When the input is a measurement, you do not know the true input. You know only the reading and roughly how accurate the instrument is. So the true change between the reading and the truth cannot be computed, not even in principle.

The differential needs only the slope, and the slope at the measured value is just as good as the slope at the true value. That is why every error estimate in science and engineering is, at heart, a differential.

35. Which way is the error?

Section

Part 3

36. Concavity decides which side the tangent is on

Concept

Figure (svg): Two panels. Left: the square root curve with its tangent at 9; the thin region between them is shaded red and lies below the tangent, so the tangent is above the curve. Right: the curve e to the x with its tangent 1 plus x at 0; the shaded gap lies above the tangent, so the tangent is below the curve.

A curve that bends down falls away beneath its tangent on both sides; a curve that bends up rises above it on both sides. The sign of the second derivative tells you which picture you are in.

\[ f''(a) < 0 \;\Longrightarrow\; L(x) > f(x) \text{ near } a: \text{ an overestimate} \]

\[ f''(a) > 0 \;\Longrightarrow\; L(x) < f(x) \text{ near } a: \text{ an underestimate} \]

You met this picture in Chapter 4's opening: a curve that is concave down lies below its tangents, a curve that is concave up lies above them. The second derivative's sign tells you which way to lean the estimate before computing anything.

In every example so far, the estimate came out either slightly high or slightly low, and you may have noticed that it was never random. This picture shows the reason.

On the left is the square root, which bends downward. Its tangent at 9 lies above the curve on both sides, and the red sliver between them is the error. Every estimate made with that tangent is too high. On the right is the exponential, which bends upward. Its tangent at 0 lies below the curve on both sides, the green sliver, and every estimate is too low.

Bending downward is concave down, which means a negative second derivative, and bending upward is a positive second derivative. So one extra derivative at the anchor tells you, before any calculation, which way your estimate is wrong. Section 4.5 develops concavity fully; here you need only its sign.

37. Every estimate so far, sorted by concavity

Concept

Go back through the examples and compare the sign of the second derivative with what the calculator said.

estimatef″ at the anchortangent isestimate vs truth
√9.1 ≈ 3.01667−1/108 < 0abovehigh: truth 3.01662
∛8.1 ≈ 2.00833−1/144 < 0abovehigh: truth 2.00830
sin 62° ≈ 0.88348−√3/2 < 0abovehigh: truth 0.88295
1.01³ ≈ 1.036 > 0belowlow: truth 1.030301
x² + 2x: dy = 0.82 > 0belowlow: Δy = 0.81

Five predictions, five hits. The direction of the error was knowable in advance every time.

This table puts the rule to the test on everything you have done. For each estimate, the second column is the sign of the second derivative at the anchor, the third is where that says the tangent lies, and the last says what the calculator actually found.

The three concave-down cases, the square root, the cube root and sine near 60 degrees, all gave estimates that were too high. The two concave-up cases, the cube near 1 and the quadratic from Example 4.9, gave estimates that were too low. Five predictions, five hits.

This matters in practice. Knowing that your estimate of the square root of 9.1 is high means you can state that the true value is below 3.01667, which is a guaranteed one-sided bound from a calculation you did in your head.

38. Overshoot or undershoot?

Sorting

Sort into buckets

Without a calculator, decide whether each tangent-line estimate is too high or too low.

Too high: f″ < 0
√10 from a = 9; sin(0.1) from a = 0; √8.9 from a = 9
Too low: f″ > 0
(1.02)⁵ from a = 1; 1/0.98 from a = 1; e^0.1 from a = 0
hi
The square root is concave down everywhere, and sine is concave down just to the right of 0 (f″ = −sin x < 0 there). √10 ≈ 3.1667 vs 3.1623; sin 0.1 ≈ 0.1 vs 0.09983; √8.9 ≈ 2.98333 vs 2.98329.
lo
x⁵, 1/x (for positive x) and eˣ are all concave up. 1.02⁵ ≈ 1.1 vs 1.10408; 1/0.98 ≈ 1.02 vs 1.02041; e^0.1 ≈ 1.1 vs 1.10517.

Sort each estimate before checking. For each one, you need only the sign of the second derivative near the anchor, not the estimate itself.

The square root is concave down everywhere, so both of its estimates are high, whichever side of 9 you step to. Sine near 0 on the right is concave down too, because its second derivative is minus sine, which is negative just past zero. So sine of 0.1 is overestimated by 0.1.

The fifth power, one over x for positive x, and the exponential all have positive second derivatives, so their tangent-line estimates are low. The rationales show the actual numbers: in every case the estimate falls on the side the concavity predicted.

39. Trap: a step to the left flips the direction

Trap

The trap

Estimating the square root of 8.9 from 9:

\[ L(8.9) = 3 - \frac{0.1}{6} = 2.98333 \]

The step is negative, so the estimate must be too low.

The fix

No. The square root is concave down on both sides of 9, and the tangent lies above it on both sides. The estimate is high whichever way you step.

\[ \sqrt{8.9} = 2.98329 < 2.98333 \]

Only the sign of f″ decides the direction.

This mistake comes from a reasonable-sounding intuition: stepping backwards should reverse everything. But look at the concavity picture again. A concave-down curve falls away beneath its tangent on the left of the anchor just as it does on the right. The tangent sits above the curve on both sides.

So the direction of the error depends only on the sign of the second derivative, not on the sign of the step. The estimate of the square root of 8.9 from 9 is 2.98333, and the true value is 2.98329, so it is high, exactly like the estimate of the square root of 9.1.

The sign of the step does affect the sign of dy, of course: stepping left makes dy negative. It just has no effect on which side of the truth the estimate lands.

40. Slide the anchor and watch the gap

Tweak it

Parameter explorer

The curve is the square root minus its own linearization at a, plotted against x. Slide the anchor a. Where is the gap zero, is it ever positive, and when is it widest?

\[ \sqrt{x} - L_a(x), \quad a = {a} \]

  • a — from 1 to 16: anchor a

The curve plotted here is the square root minus its linearization at a, so it measures the error directly. Start with a at 9 and look: the curve touches zero at x equal to 9 and is negative everywhere else. Negative means the linearization is bigger than the root, so the tangent is above the curve, as concavity promised.

Now slide a down to 1 or 2. The gap opens up very quickly on both sides, because the square root bends sharply near zero, so its second derivative is large there. Slide a up to 16 and the gap stays shallow over a wide range, because the root is nearly straight far out.

Two lessons from one slider. The gap is never positive, which is the direction rule. And how fast the gap grows depends on how strongly the curve bends at the anchor, which is exactly the second derivative. That is the bridge to the next part.

41. How big is the error?

Section

Part 4

42. For x squared the error is exactly the step squared

Concept

Take the simplest curved function and compute the true change and the differential in general.

\[ \Delta y = (a + dx)^2 - a^2 \]

\[ \Delta y = a^2 + 2a\,dx + (dx)^2 - a^2 = 2a\,dx + (dx)^2 \]

\[ dy = f'(a)\,dx = 2a\,dx \]

\[ \Delta y - dy = (dx)^2 \]

The error does not depend on a at all, and it is the square of the step. That is exactly what Example 4.9 and Checkpoint 4.9 found: 0.01, then 0.04.

Here is the cleanest possible case, worked in general. Expand a plus dx, squared, and subtract a squared. What is left is 2a dx plus dx squared. The differential is the first part, 2a dx, because the slope of x squared at a is 2a.

So the error, the true change minus the differential, is exactly dx squared. Not approximately, exactly. And notice what it does not depend on: the anchor a has disappeared completely. For x squared, a step of a tenth always leaves an error of one hundredth, wherever you start.

This explains the two numbers you found earlier. In Example 4.9 the step was 0.1 and the gap was 0.01. In Checkpoint 4.9 the step was 0.2 and the gap was 0.04. Both are the step squared.

43. The corner the differential misses

Picture it

Figure (svg): A square of side a with area a squared, enlarged to side a plus dx. The added area is two thin strips of area a times dx, one along the right and one along the top, plus a small corner square of area dx squared.

The differential counts the two long strips and ignores the corner. The corner is the error, and its area is the step squared: halve the step and the corner shrinks to a quarter.

Grow a square of side a by dx. The differential counts the two long strips; the true change also includes the little corner, whose area is the step squared.

This is the same algebra drawn as area. The blue square has side a and area a squared. Grow each side by dx and the area grows by three pieces: two long yellow strips, each a by dx, and a small red corner, dx by dx.

The differential of the area, 2a dx, is exactly the two yellow strips. It ignores the corner. The true change includes it. So the error is the red square, whose area is dx squared.

The picture makes the square law obvious. If you halve dx, each yellow strip gets half as thin, so the differential halves. But the red corner shrinks in both directions at once, so its area becomes a quarter. When the step is small, the corner becomes negligible compared with the strips, and the differential becomes an excellent approximation.

44. In general: about half of f double prime times the step squared

Concept

Try x cubed at a, and expand the true change the same way.

\[ (a + dx)^3 - a^3 = 3a^2\,dx + 3a(dx)^2 + (dx)^3 \]

\[ \Delta y - dy = 3a(dx)^2 + (dx)^3 \approx 3a(dx)^2 \]

And 3a is exactly half of the second derivative 6a. The same holds for any function with a continuous second derivative:

\[ \Delta y - dy \approx \tfrac12 f''(a)(dx)^2 \]

\[ |\text{error}| \le \tfrac12 M(dx)^2, \quad M = \max|f''| \text{ between } a \text{ and } x \]

The bound is the first term of the Taylor series you will meet in Calculus II. For now, read it as: the error is second order in the step.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 316 — Exercise 48: why accuracy falls with distance

Try the same expansion with x cubed. The true change is 3a squared dx, which is the differential, plus 3a dx squared plus dx cubed. When the step is small, the cube of it is far smaller still, so the error is essentially 3a times dx squared.

Now compare with the second derivative of x cubed, which is 6a. Half of it is 3a. And for x squared, half of the second derivative 2 is 1, which matches the error of exactly one times dx squared. The pattern holds for every smooth function: the error is about half the second derivative times the step squared.

The practical version is the bound on the last line. If M is the largest size of the second derivative between the anchor and the target, the error is at most half of M times the step squared. You will prove this properly in Calculus II with Taylor polynomials. For now, take away the shape of it: error is second order in the step, and it is scaled by how sharply the function bends.

45. Measuring the square law on the square root

Worked example

Anchor the square root at 9 and measure the error of L at four steps.

\[ L(x) = 3 + \frac{x - 9}{6}, \qquad E(dx) = L(9 + dx) - \sqrt{9 + dx} \]

Step 1

Why: L gives 3.16667; the root of 10 is 3.16228.

\[ E(1) = 3.166667 - 3.162278 = 0.004389 \]

Step one half

Why: L gives 3.08333; the root of 9.5 is 3.08221.

\[ E(0.5) = 3.083333 - 3.082207 = 0.001126 \]

Step one quarter

Why: Halve again.

\[ E(0.25) = 3.041667 - 3.041381 = 0.000285 \]

Take the ratios

Why: Each halving of the step divides the error by very nearly four.

\[ \frac{E(1)}{E(0.5)} = 3.90, \qquad \frac{E(0.5)}{E(0.25)} = 3.95 \]

Figure (svg): The error of the tangent-line estimate for the square root, anchored at 9, plotted against x from 5 to 13. It is zero at 9 and rises on both sides in a parabola-like cup; a dashed parabola, the step squared over 216, follows it closely. Dots mark the errors at x equals 9.25, 9.5 and 10.

The error is flat at the anchor and then curves up like a parabola. Doubling the distance from 9 roughly quadruples the error.

Check against the formula

Why: Half of f double prime at 9 is one over 216, so the prediction for a step of one tenth is 0.0000463; the measured error is 0.0000460.

\[ \frac{(0.1)^2}{216} = 0.0000463, \qquad E(0.1) = 0.0000460 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 316 — Exercise 48

Here the square law is tested on a function that is not a polynomial, so nothing is exact. The error at each step is the linearization minus the true root, each computed to six places.

With a step of 1 the error is 0.00439; with one half it is 0.00113; with a quarter it is 0.00029. The ratios of successive errors are 3.90 and 3.95, closing in on 4 as the step shrinks. The graph shows the whole error curve: flat at the anchor, then climbing like a parabola, with the dashed parabola from the formula tracking it closely.

The check compares the formula's prediction at a step of a tenth with the real error. Half of the second derivative at 9 is one over 216, so the prediction is 0.0000463. The measured error is 0.0000460. That is agreement to within about one percent, from a formula that needed nothing but a second derivative.

46. Shrink the step tenfold

Estimation

\[ \sqrt{10} \approx 3.16667, \quad \text{error } 0.0044 \quad (dx = 1) \]

Predict first

Anchored at 9, the estimate of √10 is off by about 0.0044. Estimate √9.1 the same way. Roughly how big is that error?

  • About 0.00044
  • About 0.000044
  • About 0.0000044
  • Exactly zero

Correct: About 0.000044

Why: The step shrinks by a factor of 10, so the error shrinks by a factor of 10 squared, 100: from 0.0044 to about 0.000044. The measured error is 0.000046. Guessing 0.00044 treats the error as proportional to the step, which is the most common mistake here.

Commit to an answer before revealing. The question is really about which law the error follows.

If the error were proportional to the step, shrinking the step tenfold would shrink the error tenfold, to about 0.00044. But the error is proportional to the square of the step, so a tenfold smaller step gives a hundredfold smaller error: about 0.000044. The measured value is 0.000046.

If you chose 0.00044, you are in good company; it is the most common answer. The square law is the thing to remember from this part of the lesson: small steps are not just a little better, they are dramatically better.

47. What second order means in practice

Intuition

An error proportional to the square of the step behaves very differently from one proportional to the step.

change the step bydy changes bythe error changes by
× 2× 2about × 4
× 1/2× 1/2about × 1/4
× 1/10× 1/10about × 1/100

So as the step shrinks, the error becomes a smaller and smaller fraction of dy itself. That is the precise sense in which the tangent is the best line through the point: any other line would leave an error proportional to the step.

The table compares how the differential and its error respond when you change the step. The differential scales exactly with the step, because it is slope times step. The error scales with the step squared.

The consequence is in the last paragraph, and it is subtle but important. As the step gets smaller, the error becomes a smaller and smaller fraction of the differential itself. At a step of a tenth, the error in Example 4.9 was one eightieth of dy; at a hundredth it would be one eight-hundredth.

That is what makes the tangent line special among all lines through the point. Any other line, with a slightly wrong slope, would leave an error proportional to the step, which does not shrink relative to the change. The tangent is the only line whose error vanishes faster than the step. That is, in fact, one way to define the derivative.

48. Bounding the error before you know the answer

Worked example

Bound the error in the estimate of the square root of 10 from 9 without using a calculator for the root.

\[ f(x) = x^{1/2}, \quad a = 9, \quad dx = 1 \]

Find the second derivative

Why: Differentiate twice.

\[ f'(x) = \tfrac12 x^{-1/2}, \qquad f''(x) = -\tfrac14 x^{-3/2} \]

Find its largest size between 9 and 10

Why: The size of x to the minus three halves shrinks as x grows, so the maximum is at 9.

\[ M = \tfrac14 \cdot 9^{-3/2} = \tfrac14\cdot\tfrac1{27} = \tfrac{1}{108} \]

Apply the bound

Why: Half of M times the step squared.

\[ |\text{error}| \le \tfrac12\cdot\tfrac{1}{108}\cdot 1^2 = \tfrac{1}{216} \approx 0.00463 \]

State the conclusion with its direction

Why: f double prime is negative, so the estimate is high, by at most 0.00463.

\[ 3.16204 \le \sqrt{10} \le 3.16667 \]

Check with a calculator

Why: The true root is 3.16228: inside the interval, and the actual error 0.00439 is just under the bound.

\[ 3.16667 - 3.16228 = 0.00439 \le 0.00463 \]

This example puts the direction rule and the size rule together to produce a guaranteed interval, without knowing the true answer. The second derivative of the square root is minus one quarter of x to the minus three halves.

Its size is largest where x is smallest, so on the stretch from 9 to 10 the biggest value is at 9: one quarter times one over 27, which is one over 108. Half of that times a step of 1 squared is one over 216, about 0.00463. And since the second derivative is negative, the estimate is high. So the root of 10 lies between 3.16667 minus 0.00463 and 3.16667 itself.

The calculator confirms it: the true value 3.16228 is inside the interval, and the actual error 0.00439 is just under the bound. You have trapped an irrational number to within half a hundredth using only derivatives, which is the same idea calculators use, with higher-degree polynomials.

49. Errors in measurement

Section

Part 5

50. Propagated error

Concept

A quantity is computed from a measurement. The true input is a, but you read a plus dx, where dx is the measurement error. The computed value is off too.

\[ \Delta y = f(a + dx) - f(a) \qquad \text{(the propagated error)} \]

You cannot compute this: you never know a. But you do know the size of dx from the instrument, and the differential needs only a slope.

\[ \Delta y \approx dy = f'(a)\,dx \approx f'(a + dx)\,dx \]

The last step swaps the unknown true value for the measured one; the slope barely changes over so small a distance.

Propagated error — The error in a computed quantity that results from an error in the measurement it was computed from.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 314 — calculating the amount of error

Now the differential gets its most important job. Many quantities are not measured directly but computed from something that is: an area from a radius, a volume from a side, a height from an angle. Any error in the measurement travels through the formula into the computed quantity. That travelling error is the propagated error.

If the true input is a and you read a plus dx, the propagated error is the difference of the function values. You can never compute it, because you never know a. What you do know is the accuracy of the instrument, which bounds dx. The differential needs only that and a slope.

The last line is a small but necessary swap. The slope should really be taken at the unknown true value, but you only have the measured one. Since the two are close, the slopes are close too, and the difference is a second-order effect that you can safely ignore.

51. Example 4.10: the volume of a cube

Worked example

A cube's side is measured as 5 cm, accurate to within 0.1 cm. (a) Estimate the error in the computed volume. (b) Compare with the volumes at 4.9 and 5.1 cm.

Record the measurement error

Why: Accurate to within a tenth means the step lies between minus 0.1 and 0.1.

\[ -0.1 \le dx \le 0.1 \]

Write the differential of the volume

Why: V equals x cubed.

\[ V = x^3 \;\Longrightarrow\; dV = 3x^2\,dx \]

Use the measured side

Why: At 5 cm the slope is 75.

\[ -3(5)^2(0.1) \le dV \le 3(5)^2(0.1) \]

Simplify

Why: Seventy-five times a tenth.

\[ -7.5 \le dV \le 7.5 \]

Figure (svg): The curve V equals x cubed for x from 4 to 6, with its tangent at x equals 5. A vertical strip marks the measured side 5 plus or minus 0.1 centimetres; a horizontal strip marks the resulting volume band 125 plus or minus 7.5 cubic centimetres. The two strips cross at the point (5, 125).

The slope at 5 is 75, so a wobble of a tenth in the side becomes a wobble of 7.5 in the volume. The derivative is the exchange rate between the two errors.

Compute the extreme volumes for (b)

Why: Cube the two possible true sides.

\[ V(4.9) = 117.649, \qquad V(5.1) = 132.651 \]

Subtract the computed volume

Why: The computed volume is 5 cubed, 125.

\[ -7.351 \le \Delta V \le 7.651 \]

Check the estimate against the exact range

Why: The differential's 7.5 sits between the two actual extremes, 7.351 and 7.651, off by about 2 percent: the cube is concave up, so the true change is larger on the right.

\[ |dV| = 7.5 \quad \text{vs} \quad |\Delta V| \in [7.351,\, 7.651] \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, pp. 314-315 — Example 4.10

Part (a) is the standard method. The side is known to within a tenth, so dx lies between minus a tenth and a tenth. The differential of x cubed is 3x squared dx, and at the measured side of 5 the slope is 75, so the volume is uncertain by about 7.5 cubic centimetres either way.

The picture shows what that means. The yellow vertical strip is the band of possible sides; the tangent line at 5, with slope 75, carries it to the green horizontal band of possible volumes, 125 plus or minus 7.5.

Part (b) computes the extreme volumes exactly, and the check compares. The true errors are 7.351 on the low side and 7.651 on the high side, so the differential's single number, 7.5, sits right between them and is off by about 2 percent in each direction. The asymmetry is concavity again: the cube bends upward, so stepping right gains more than stepping left loses.

52. Checkpoint 4.10: a bigger cube, a coarser ruler

Prediction

\[ x = 6 \text{ cm}, \quad \text{accurate to within } 0.2 \text{ cm} \]

Predict first

Use differentials to estimate the error in the computed volume.

  • ±7.2 cm³
  • ±21.6 cm³
  • ±43.2 cm³
  • ±0.6 cm³

Correct: ±21.6 cm³

Why: The slope 3x squared is 108 at 6, and 108 times 0.2 is 21.6. The exact changes are 22.3 upward and 20.9 downward, so the estimate is again in the middle. Seven point two forgets to square the side; 43.2 doubles the step; 0.6 multiplies the step by 3 but forgets x squared.

\[ dV = 3(6)^2(0.2) = 108(0.2) = 21.6 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 315 — Checkpoint 4.10

Pick an answer before revealing. The method is exactly Example 4.10 with new numbers, so the question is really whether you can carry it out without slipping.

The slope of x cubed at 6 is 3 times 36, which is 108. Times a step of 0.2 that is 21.6 cubic centimetres. The exact changes, if you want to confirm, are about 22.3 upward and 20.9 downward, so again the differential lands in the middle.

Each wrong option is a specific slip. Seven point two forgets to square the side. Forty-three point two doubles the error bound for no reason. Zero point six is 3 times the step, having lost the x squared entirely.

53. Relative and percentage error

Concept

An absolute error of one unit can be tiny or enormous; it depends on what it is an error in. So divide by the size of the quantity.

\[ \text{relative error} = \frac{\Delta q}{q}, \qquad \text{percentage error} = 100\,\frac{\Delta q}{q}\,\% \]

Figure (svg): Two bars drawn the same length. The top bar is the 62-inch ladder, with its 1-inch error shaded red at the end, a sliver of 1.6 percent. The bottom bar is the 8-inch cardboard, with its quarter-inch error shaded red, about twice as wide a sliver at 3.1 percent.

Drawn to the same length, the cardboard's red sliver is about twice the ladder's, even though its absolute error is a quarter of the size.

\[ \text{ladder: } \frac{1}{62} \approx 0.016 = 1.6\%, \qquad \text{cardboard: } \frac{0.25}{8} = \frac{1}{32} \approx 3.1\% \]

The cardboard measurement is the worse one, although a quarter of an inch is smaller than an inch.

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 315 — relative error and percentage error

An absolute error means nothing on its own. An error of one inch is negligible in the height of a ladder and ruinous in the width of a microchip. So scientists report errors relative to the size of the thing measured: the absolute error divided by the value, or that times one hundred as a percentage.

The book's two measurements make the point. The ladder is off by an inch out of 62, which is 1.6 percent. The cardboard is off by a quarter inch out of 8, which is 3.1 percent. The figure draws both bars the same length so you can compare shares directly: the cardboard's red sliver is about twice the ladder's.

So the cardboard measurement is the worse one, even though its absolute error is a quarter of the ladder's. When someone reports an error, the first question to ask is: out of how much?

54. Example 4.11: the volume of Earth

Worked example

An astronaut measures Earth's radius as 4000 miles with an error of up to 80 miles. Estimate the relative and percentage error in the computed volume, treating Earth as a sphere.

Record the measurement error

Why: Accurate to within 80 miles.

\[ -80 \le dr \le 80 \]

Write the differential of the volume

Why: V is four thirds pi r cubed.

\[ V = \tfrac43\pi r^3 \;\Longrightarrow\; dV = 4\pi r^2\,dr \]

Form the relative error

Why: Divide by the volume, estimated from the measured radius.

\[ \frac{dV}{V} = \frac{4\pi r^2\,dr}{\tfrac43\pi r^3} \]

Cancel

Why: Four pi r squared over four thirds pi r cubed is 3 over r.

\[ \frac{dV}{V} = \frac{3\,dr}{r} \]

Substitute

Why: Three times 80 over 4000.

\[ -\frac{3(80)}{4000} \le \frac{dV}{V} \le \frac{3(80)}{4000} \;\Longrightarrow\; -0.06 \le \frac{dV}{V} \le 0.06 \]

Check against the exact ratios

Why: Cube the ratio of the extreme radii to the measured one: the true relative changes are 6.1 percent and 5.9 percent, straddling the estimate of 6 percent.

\[ \left(\tfrac{4080}{4000}\right)^3 - 1 = 0.0612, \qquad 1 - \left(\tfrac{3920}{4000}\right)^3 = 0.0588 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, pp. 315-316 — Example 4.11

Follow the steps in order. The radius is uncertain by up to 80 miles. The differential of the sphere's volume is 4 pi r squared dr, the surface area times the step, which is a nice fact in itself: a thin shell of thickness dr has volume about its area times dr.

To get the relative error, divide by the volume. Most of it cancels: four pi r squared over four thirds pi r cubed is simply 3 over r. So the relative error in the volume is 3 times dr over r. With 80 over 4000, that is 3 times 0.02, or 0.06, which is 6 percent.

The check cubes the ratios of the extreme radii to the measured one. A radius 80 miles too long gives a volume 6.1 percent too big; 80 miles too short gives one 5.9 percent too small. The estimate of 6 percent sits between them, as it did for the cube.

55. A power multiplies the relative error by its exponent

Concept

The cancellation in Example 4.11 is not special to spheres. Take any power law.

\[ q = k\,x^n \;\Longrightarrow\; dq = k n x^{n-1}\,dx \]

\[ \frac{dq}{q} = \frac{k n x^{n-1}\,dx}{k\,x^n} = n\,\frac{dx}{x} \]

The constant k cancels, and so does the size of x. A 2 percent error in a radius becomes a 4 percent error in an area and a 6 percent error in a volume.

The cancellation in the Earth example was not a coincidence of spheres. Take any quantity that is a constant times a power of the measured variable. Its differential is the constant times n times x to the n minus one, times dx. Divide by the quantity itself and the constant cancels, and so does most of the power, leaving n times dx over x.

In words: the relative error in the output is the exponent times the relative error in the input. Areas double relative errors, volumes triple them, and square roots halve them.

This is one of the most useful rules in experimental science, because it lets you judge an error in your head. It also tells you which measurement to worry about: in a formula with a cube in it, the cubed quantity deserves your most careful ruler.

56. Reading the relative-error rule

Notation

Annotate

On: \( \frac{dV}{V} = 3\,\frac{dr}{r} \)

  • The relative error in the computed volume: what fraction of V you might be off by.
  • The relative error in the measured radius: 80/4000 = 0.02 in Example 4.11.
  • The exponent of r in V = (4/3)πr³. It magnifies the input's relative error threefold.
  • Both cancelled. The answer needs only the exponent and the measurement's own relative error.

Step through the annotations. The left side is what you want, the fraction of the volume you might be off by. The fraction on the right is what the instrument gives you, the fraction of the radius you might be off by.

The 3 is the exponent of r in the volume formula, coming straight out of the derivative. And the last annotation is the surprising part: pi, the four thirds and the actual size of the radius have all cancelled. You do not even need to compute the volume to know its relative error.

That is why Checkpoint 4.11 on the next slide takes only two lines.

57. Checkpoint 4.11: a different astronaut

Worked example

\[ r = 3950 \text{ mi}, \quad |dr| \le 100 \text{ mi} \]

Use the rule for a cube of the radius

Why: The volume is a constant times r cubed.

\[ \frac{dV}{V} = 3\,\frac{dr}{r} \]

Substitute the bounds

Why: Three hundred over 3950.

\[ \left|\frac{dV}{V}\right| \le \frac{3(100)}{3950} = \frac{300}{3950} \]

Divide and convert to a percentage

Why: Multiply by 100.

\[ \frac{300}{3950} \approx 0.0759 = 7.59\% \]

Check with the exact ratios

Why: The true relative changes at 4050 and 3850 miles are 7.8 and 7.4 percent, either side of 7.59.

\[ \left(\tfrac{4050}{3950}\right)^3 - 1 = 0.0779, \qquad 1 - \left(\tfrac{3850}{3950}\right)^3 = 0.0740 \]

OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials §4.2, p. 316 — Checkpoint 4.11

With the rule in hand this checkpoint is short. The volume is still a constant times r cubed, so the relative error is 3 times dr over r, and here that is 3 times 100 over 3950.

That division gives about 0.0759, or 7.59 percent. Notice that the answer is bigger than in Example 4.11 for two reasons: a larger error, 100 miles instead of 80, and a slightly smaller radius to divide by.

The check computes the true relative changes at the extreme radii, 4050 and 3850 miles. They are 7.8 percent and 7.4 percent, on either side of the differential's 7.59, the same pattern as before.

58. Trap: the volume inherits the radius's percentage

Trap

The trap

Radius 4000 miles, error 80 miles:

\[ \frac{dr}{r} = \frac{80}{4000} = 2\% \]

\[ \Longrightarrow \text{ volume error } 2\% \]

Wrong by a factor of three.

The fix

The volume goes like the cube of the radius, so the relative error triples.

\[ \frac{dV}{V} = 3\,\frac{dr}{r} = 6\% \]

Always differentiate before dividing; the exponent comes out of the derivative.

The temptation is to think that an error of 2 percent in the measurement must mean an error of 2 percent in the answer. That is only true when the answer is proportional to the measurement, a power of exactly one.

The volume depends on the cube of the radius, and cubing magnifies relative errors threefold. You can see this without calculus: a radius 2 percent too large gives a volume 1.02 cubed, about 1.061 times too large, which is 6 percent. The safeguard is procedural: always write the differential first and divide afterwards, so the exponent comes out of the derivative instead of being forgotten.

59. Working backwards from a tolerance

Worked example

A steel ball's radius is measured as 10 cm. (a) If the radius is good to 0.1 cm, estimate the error in the volume. (b) How accurately must the radius be measured for the volume to be good to 1 percent?

Differential of the volume at 10 cm

Why: Four pi r squared times the step.

\[ dV = 4\pi(10)^2(0.1) = 40\pi \approx 125.7 \text{ cm}^3 \]

Relative error for (a)

Why: Three times the radius's relative error.

\[ \frac{dV}{V} = 3\cdot\frac{0.1}{10} = 0.03 = 3\% \]

Set up (b) as an inequality

Why: Require the volume's relative error to be at most one hundredth.

\[ 3\,\frac{dr}{r} \le 0.01 \]

Solve for the radius's relative error

Why: Divide by 3.

\[ \frac{dr}{r} \le \frac{1}{300} \approx 0.33\% \]

Convert to centimetres

Why: Multiply by the radius.

\[ dr \le \frac{10}{300} = \frac{1}{30} \approx 0.033 \text{ cm} \]

Check with the exact volumes

Why: A radius one thirtieth too long gives a volume 1.003 percent too large: right at the 1 percent target.

\[ \left(\frac{10 + 1/30}{10}\right)^3 - 1 = 0.01003 \]

Part (a) runs the method forwards, as in Examples 4.10 and 4.11. The volume's differential at a radius of 10 is 40 pi, about 126 cubic centimetres, and the relative error is 3 times one percent, which is 3 percent.

Part (b) is the question an engineer actually asks: how good must my measurement be? Run the relative-error rule backwards. If the volume must be good to one percent, then 3 times the radius's relative error must be at most one percent, so the radius's relative error must be at most one three-hundredth, a third of a percent. On a radius of 10 centimetres that is one thirtieth of a centimetre, about a third of a millimetre.

The check confirms it with the exact formula. A radius one thirtieth too long gives a volume 1.003 percent too large, right at the target. This backwards use is where the linearity of the differential pays off: an inequality in dy becomes a simple inequality in dx.

60. How much each formula magnifies a measurement error

Comparison

Comparison matrix

computed quantityformularelative error in the result
area of a circleA = πr²2 × dr/r
volume of a cubeV = x³3 × dx/x
volume of a sphereV = (4/3)πr³3 × dr/r
side of a square from its areas = √A1/2 × dA/A
kinetic energy from speedK = (1/2)mv²2 × dv/v

Fill in the blanks before checking. Each row applies the rule from two slides ago: find the exponent on the measured quantity, and that is the multiplier on its relative error.

The cube and the sphere both have exponent 3; the constant four thirds pi makes no difference. The side of a square computed from its area is the square root of the area, exponent one half, so it halves the area's relative error. That is good news: computing a length from an area makes the answer more reliable than the measurement.

Kinetic energy depends on the square of the speed, so a speedometer that is 5 percent off makes the computed energy about 10 percent off. The mass enters to the first power and would pass its relative error through unchanged.

61. The small-angle approximation in physics

Real world

Figure (svg): The line y equals theta and the curve y equals sine theta for theta from 0 to 90 degrees, measured in radians. They are indistinguishable up to about 10 degrees and have visibly separated by 45 degrees; dashed vertical lines mark 10, 20 and 45 degrees.

The small-angle approximation is the linearization of sine at 0. It is excellent near 0 and fades as the angle grows, faster than you might expect.

Discussion prompt

The pendulum formula every physics course uses replaces sine of theta by theta. That is the linearization of sine at 0. How large is the relative error at 10 degrees, and at 20 degrees?

Answer the question before revealing it. The small-angle approximation says that sine of theta is about theta, for theta in radians. It is exactly the linearization of sine at zero, one of the four from earlier in the lesson.

The picture shows how good it is. The line and the curve are indistinguishable up to about 10 degrees and have visibly parted by 45. At 10 degrees the relative error is about half a percent; at 20 degrees it is about 2 percent.

Notice that doubling the angle roughly quadrupled the relative error. That is the square law once more, with a twist: sine is an odd function, so its second derivative at zero is zero, and the absolute error actually grows like the cube of the angle, making the relative error grow like the square. Physicists who use this approximation for pendulums keep the swings small for exactly this reason.

62. Putting it together

Section

Part 6

63. Pattern: estimating with a tangent line

Pattern

Figure (svg): A flow diagram with four boxes left to right: choose an anchor a near the target where f is easy; compute f of a and f prime of a; write L of x and evaluate; then two branches, the sign of f double prime giving the direction of the error and half of M times dx squared giving its size.

Anchor, derivative, line; then two questions about the error, which way and how big.
  1. Choose the anchor a: as close to the target as possible, where f(a) and f′(a) are exact.
  2. Compute f(a) and f′(a), in radians for any trigonometric function.
  3. Write L(x) = f(a) + f′(a)(x − a) and evaluate at the target; for a change, dy = f′(a)dx is the same line.
  4. Direction: f″ < 0 means the estimate is high, f″ > 0 means low.
  5. Size: at most ½M(dx)²; halve the step and the error quarters.
  6. Measurements: for a power law, the relative error is multiplied by the exponent.

This is the whole lesson as a procedure. The diagram runs left to right through the three moves that build the estimate, anchor, derivative, line, and then branches into the two questions you should always ask about the result.

The list adds the details that most often go wrong. Choose the anchor for nearness and ease together. Use radians for any trigonometric function. Remember that a differential is the same calculation measured from f of a, so estimating a change and estimating a value are one skill.

The last three lines are what make you a careful estimator rather than a mechanical one: say which way the estimate is off, from the sign of the second derivative; say roughly how far, from the square law; and for measurements, multiply the relative error by the exponent.

64. Put the estimate in order

Ranking

Put in order

Order the steps for estimating ∛8.1 with a linearization and judging the answer.

  1. Choose f(x) = ∛x and the anchor a = 8
  2. Compute f(8) = 2 and f′(8) = 1/12
  3. Write L(x) = 2 + (x − 8)/12
  4. Evaluate L(8.1) = 2.00833
  5. Use f″ < 0 to say the estimate is slightly high

Why: Anchor first, because every later number depends on it. The line is assembled from f(a) and f′(a) and only then evaluated. Judging the direction comes last and needs one more derivative, not a calculator.

Drag the steps into order before checking. Most of the order is forced, because each step uses the result of the one before: you cannot compute f of a before you have chosen a, and you cannot evaluate L before you have written it.

The step that people tend to misplace is the judgement of direction. It comes last, after the number, and it needs one more derivative rather than a calculator. It is also the step most often skipped entirely, so make a habit of including it.

65. Check: a logarithm near 1

Check

Check your understanding

Using the linearization of ln x at a = 1, what is the estimate of ln 1.1, and is it too high or too low?

  • A. 0.1, too high (correct)
  • B. 0.1, too low
  • C. 1.1, too high
  • D. 0.0953, exact

Answer: A

Why: At 1 the logarithm is 0 and its slope is 1, so L(x) = x − 1 and L(1.1) = 0.1. The second derivative, minus one over x squared, is negative, so the tangent lies above the curve and the estimate is high; the true value is 0.0953.

Why B tempts people
The value is right, but ln x is concave down, so its tangent lies above it and the estimate is high, not low.
Why C tempts people
That is 1 plus the step: it forgets that ln 1 is 0, not 1.
Why D tempts people
0.0953 is the true value, not the linear estimate; a tangent-line estimate of a curved function is almost never exact.

Work it through: the logarithm of 1 is 0 and its derivative, one over x, is 1 at x equal to 1. So the linearization is x minus 1, and at 1.1 it gives one tenth.

For the direction, differentiate again: the second derivative is minus one over x squared, which is negative. The logarithm is concave down, its tangent lies above it, and so the estimate is too high. The true value, 0.0953, confirms it. This is also the fifth of the linearizations at zero from earlier, the logarithm of one plus x being about x, in disguise.

66. Check: dy or delta y

Check

Check your understanding

For y = x³ at x = 2 with dx = 0.05, what is the differential dy?

  • A. 0.6 (correct)
  • B. 0.615125
  • C. 0.015125
  • D. 12

Answer: A

Why: The derivative 3x² is 12 at x = 2, and dy = 12 × 0.05 = 0.6. The true change 2.05³ − 8 = 0.615125 is delta y, and the gap 0.015125 is close to 3a(dx)² = 0.015, as the square law predicts.

Why B tempts people
That is the actual change delta y, computed on the curve; dy is the change along the tangent.
Why C tempts people
That is the difference between delta y and dy: the error of the approximation, not dy.
Why D tempts people
That is the slope f′(2); it still has to be multiplied by the step dx.

The question asks for the differential, so all you need is the slope times the step. The derivative of x cubed is 3x squared, which is 12 at 2, and 12 times 0.05 is 0.6.

The other options are all real numbers from the same situation, which is what makes them tempting. The true change, 0.615125, is delta y. Their difference, 0.015125, is the error, and it matches the square-law estimate: 3 times 2 times the step squared is 0.015. The 12 is only the slope, before multiplying by the step.

67. Check: propagating a percentage

Check

Check your understanding

A circle's radius is measured with a relative error of 2 percent. About what relative error does that cause in the computed area?

  • A. 2 percent
  • B. 4 percent (correct)
  • C. 8 percent
  • D. 1 percent

Answer: B

Why: The area is πr², so dA/A = 2 dr/r: the exponent 2 doubles the relative error, giving 4 percent.

Why A tempts people
That assumes the area inherits the radius's percentage unchanged; the exponent 2 doubles it.
Why C tempts people
That would be the effect of a cube, or of doubling twice; the area has exponent 2.
Why D tempts people
Halving belongs to a square root, as when a side is computed from an area, not to r squared.

The area of a circle is pi r squared, so its differential is 2 pi r dr, and dividing by the area leaves 2 times dr over r. The exponent 2 doubles the relative error: 2 percent in the radius becomes 4 percent in the area.

Each wrong option corresponds to a different exponent. Two percent is what you would get for exponent 1, and it is the most common error. Eight percent would need exponent 4. One percent would come from exponent one half, the case of computing a length from an area rather than an area from a length.

68. Explain the square law

Explain it to yourself

Discussion prompt

In two or three sentences, explain why halving the step roughly quarters the error of a linear approximation, using the growing-square picture or the expansion of x squared.

Write your explanation before revealing. Use whichever picture made it click for you, the growing square or the algebra. The expansion is on the reveal as a reminder, but the explanation should be in words.

The key is that the differential already captures every part of the change that is proportional to the step. In the square picture those parts are the two long strips. Whatever is left must be made from the step multiplied by itself, the corner, and a piece built from two small factors shrinks twice as fast. Halving the step halves each side of the corner and so quarters its area.

If you can explain this, you can also predict it for any smooth function: the leftover is always built from the step squared, scaled by half the second derivative.

69. Exit ticket

Exit ticket

\[ \sqrt{26} \approx \;? \]

Discussion prompt

Estimate the square root of 26 with a linearization, say whether your estimate is high or low, and bound its error.

This problem uses the whole lesson in one go. Choose the anchor, 25, because it is the nearest perfect square. The slope of the square root at 25 is one tenth, so the linearization gives 5 plus one tenth, 5.1.

For the direction, the square root is concave down, so 5.1 is too high. For the size, the second derivative's largest value between 25 and 26 is at 25: one quarter times one over 125, which is one five-hundredth. Half of that times the step squared gives a bound of one thousandth.

So the square root of 26 is between 5.099 and 5.1, and the true value, 5.09902, sits inside that interval, high by just under a thousandth. If you got all three parts, the estimate, the direction and the bound, you have the lesson.

70. Recap

Recap

ideaformulawhat to remember
linearizationL(x) = f(a) + f′(a)(x − a)the tangent line, used as an estimate near a
differentialdy = f′(x)dxthe change along the tangent; Δy is the change along the curve
directionsign of f″(a)concave down: estimate high; concave up: estimate low
size|error| ≤ ½M(dx)²second order: halve the step, quarter the error
measurementdq/q = n·dx/x for q = kxⁿrelative error, multiplied by the exponent

\[ \Delta y \approx dy = f'(a)\,dx \]

Next, Section 4.3 uses the derivative to find maxima and minima. Calculus II returns to this lesson when it replaces the tangent line by higher-degree Taylor polynomials.

Stewart, Calculus: Early Transcendentals 8e, §3.10 Linear Approximations and Differentials §3.10, pp. 251-258 — the same material in Stewart

Five ideas, one table. The linearization is the tangent line used as an estimate near the anchor. The differential is the same line used to estimate a change, and it is always the change along the tangent, never the change along the curve.

The error has a direction and a size. The direction comes from the sign of the second derivative: concave down means the estimate is high, concave up means low. The size is second order: at most half of the largest second derivative times the step squared, so halving the step quarters the error. For measurements, the relative error in a power law is the exponent times the relative error in the measurement.

Section 4.3 turns to maxima and minima. This lesson comes back in Calculus II, where the tangent line becomes the first of a sequence of Taylor polynomials, each one adding a term that captures the error the previous one left behind.

Sources

  1. OpenStax Calculus Volume 1, §4.2 Linear Approximations and Differentials — Strang & Herman, OpenStax / Rice University, CC BY-NC-SA 4.0, pp. 308-316
  2. Stewart, Calculus: Early Transcendentals 8e, §3.10 Linear Approximations and Differentials — James Stewart, Cengage Learning, 2016, pp. 251-258

Want this taught 1-on-1? Alexander tutors Calculus I — $55/session, free consultation.

Book on Wyzant · Text (657) 465-8108