Calculus › Section D · Applications of derivatives › Topic D.9
Linear approximation and differentials
Near any point where a function is smooth, the curve behaves like its tangent line. That single observation gives you fast estimates and honest error margins without a calculator.
The tangent as a local stand-in
Newton's method just exploited it: near a point of tangency, a smooth curve and its tangent line are nearly indistinguishable. Turn that around and the tangent line becomes a computational tool. If you know a function and its derivative at one point \(a\), you know an excellent approximation of the function at nearby points, computed with one multiplication and one addition. This was a standard hand technique before calculators, and it is still how engineers get quick first answers on the back of an envelope. The formal name for the tangent-line approximation is the linearisation of \(f\) at \(a\).
The linearisation formula
Write the tangent line at \(x = a\) in point-slope form and it reads
The approximation is \(f(x) \approx L(x)\) for \(x\) near \(a\). It is a straight-line formula with two matching guarantees at the centre: \(L(a) = f(a)\) and \(L'(a) = f'(a)\), so the line shares both the value and the slope of \(f\) at \(a\). Everything in this topic flows from that one formula: choose \(a\) so that \(f(a)\) and \(f'(a)\) are exactly computable, keep \(x\) close to \(a\), and read off an estimate.
Worked example: a square root without a calculator
Estimate \(\sqrt{4.02}\). Use \(f(x) = \sqrt{x}\) centred at \(a = 4\), chosen because \(\sqrt{4} = 2\) exactly and \(f'(x) = \frac{1}{2\sqrt{x}}\) gives \(f'(4) = \frac{1}{4}\). The linearisation says
The true value is \(2.0049938\ldots\), so the error is about \(6 \times 10^{-6}\) from one mental step. The choice of anchor was the whole art: \(4\) is the nearest perfect square, so its root is known exactly, and \(4.02\) sits close to it.
Differentials: dy versus the real change
The same idea wears a compact notation. Let \(dx\) be a small change in the input from \(x\), and define the differential \(dy = f'(x)\,dx\): the rise along the tangent line over that run. The true change in the function is \(\Delta y = f(x + dx) - f(x)\), the rise along the curve. For small \(dx\) they are nearly equal, and the difference \(\Delta y - dy\) measures how far the curve has peeled away from its tangent. Because \(dy = f'(x)\,dx\) rearranges to \(\frac{dy}{dx} = f'(x)\), this notation is also why Leibniz's fraction-like writing behaves so well; in Section E the differential officially reappears inside every integral.
Worked example: how close is close?
For \(y = x^2\), take \(x = 3\) and \(dx = 0.01\). The differential estimate is \(dy = f'(3)\,dx = 6 \cdot 0.01 = 0.06\), while the true change is \(\Delta y = 3.01^2 - 9 = 0.0601\). The two differ by \(0.0001\), which is exactly \((dx)^2\), because \((3 + dx)^2 - 9 = 6\,dx + (dx)^2\). So halving \(dx\) divides the gap between \(\Delta y\) and \(dy\) by four. For other smooth functions the gap behaves roughly like a constant times \((dx)^2\). That is what "good near \(a\)" means in numbers, and it is why these estimates are trusted over short distances.
Propagating measurement error
Differentials answer the question engineers actually ask: I measured this with some error; how wrong is my computed answer? If the input has uncertainty \(\Delta x\), the output's uncertainty is about \(|f'(x)|\,\Delta x\). Relative error is the fairer measure: divide by the function value to get a percentage. For the sphere volume \(V = \frac{4}{3}\pi r^3\), we have \(\frac{dV}{V} = 3\frac{dr}{r}\): a 1% radius uncertainty means about a 3% volume uncertainty. The exponent 3 turned into a multiplier.
Worked example: the measured sphere
A workshop measures the radius of a metal sphere as \(10\) cm, with up to \(0.1\) cm of error, and asks how wrong the computed volume could be. With \(V = \frac{4}{3}\pi r^3\), the differential is \(dV = 4\pi r^2\,dr\). Substituting \(r = 10\) and \(dr = 0.1\):
Against the true volume \(\frac{4}{3}\pi(1000) \approx 4188.8\) cm³, that is about \(3\%\), matching the relative-error prediction \(3 \cdot \frac{0.1}{10}\) exactly. Assuming the measuring error really is at most 0.1 cm and the object is truly spherical, one differential line turns a measurement tolerance into a volume tolerance.
Over or under? Let concavity answer
The sign of the error is free information, courtesy of D.4. Where the graph is concave up, it lies above its tangent lines, so the linear approximation lands below the truth: an underestimate. Concave down flips it to an overestimate. The square root curves downward since its second derivative \(-\frac{1}{4}x^{-3/2}\) is negative, and indeed \(2.005\) exceeded the true \(2.0049938\ldots\): an overestimate, exactly as the bend predicts. Two free conclusions per estimate, then: the error's rough size from the \((dx)^2\) scaling, and its direction from the concavity.
Common mistakes to avoid
The classics: anchoring at an \(a\) that is far from the target, so the quadratic error is no longer small; anchoring \(\sqrt{4.02}\) at \(a = 3\) just because 3 is a whole number, when \(\sqrt{3}\) is not known exactly (the anchor must be both near and exact); confusing \(\Delta y\) with \(dy\), reporting the tangent's rise as the curve's; forgetting absolute values for errors, since an uncertainty is a size; believing the estimate is exact, when a linearisation always carries some error; and misreading the over-versus-under rule by skipping the concavity check.
Key takeaways
Near a differentiable point, the function equals its tangent line to first order: \(f(x) \approx f(a) + f'(a)(x-a)\), the linearisation. Choose an anchor close by whose value and derivative you know exactly. The differential \(dy = f'(x)\,dx\) is the tangent's rise and the estimate of the true change \(\Delta y\); their gap scales like the gap squared. Differentials propagate measurement error in one line, and relative error turns tolerances into percentages, as with the sphere's \(3\frac{dr}{r}\). Concavity signs the error: cup means under, cap means over. This finishes Section D; the course now turns to reversing differentiation itself, the integral.
Flashcards
Flashcard
L(x) = f(a) + f′(a)(x − a)
Test your understanding
Question 1 of 5
What is the linearisation of f at x = a?
Question 2 of 5
Using linear approximation anchored at a = 4, what is √4.02 ≈ ?
Question 3 of 5
For y = x³ at x = 2 with input change dx = 0.1, what is dy?
Question 4 of 5
A radius is measured with up to 0.5% error. About how much error can the computed sphere volume have?
Question 5 of 5
If f is concave up near the anchor, its linear approximation gives…
Standards: 4
