Navigation

Approximation · Topic 19 of 23

Linear Approximation and Differentials

Near a point where ff is differentiable, replace the curve by its tangent: f(x)L(x)=f(a)+f(a)(xa)f(x)\approx L(x)=f(a)+f'(a)(x-a). The differential dy=f(a)dxdy=f'(a)\,dx estimates the change in ff, the sign of ff'' says whether the estimate is too big or too small, and f(x)L(x)M2(xa)2\left|f(x)-L(x)\right|\le\frac{M}{2}(x-a)^{2} bounds the error.

5 min readFrequent on exams7 formulas6 quick checks
01

Key ideas

5 things to remember
  1. 1

    The tangent line is the estimate

    L(x)=f(a)+f(a)(xa)L(x)=f(a)+f'(a)(x-a) is exactly the tangent at (a,f(a))(a,f(a)), and f(x)L(x)f(x)\approx L(x) for xx near aa. Two hypotheses carry everything: f(a)f'(a) exists, and xx is close to aa.

  2. 2

    Choose the nearest exact base point

    Pick the closest aa at which f(a)f(a) and f(a)f'(a) are computable exactly — a perfect square or cube, 00, 11, π/6\pi/6. For 263\sqrt[3]{26} take a=27a=27, not a=8a=8.

  3. 3

    Differentials measure the tangent rise

    dxdx is an independent variable you may set to anything; dy=f(x)dxdy=f'(x)\,dx is then defined. With dx=Δxdx=\Delta x, dydy is the rise along the tangent and Δy=f(x+Δx)f(x)\Delta y=f(x+\Delta x)-f(x) the rise along the curve.

  4. 4

    Concavity decides over or under

    If f0f''\ge0 between aa and xx the tangent lies below the curve, so LL under-estimates; if f0f''\le0 it over-estimates. Which side of aa the point sits on is irrelevant.

  5. 5

    Errors propagate by the exponent

    For y=Cxny=Cx^{n}, dyy=ndxx\frac{dy}{y}=n\frac{dx}{x}: a 1%1\% error in a radius gives about 2%2\% in an area and 3%3\% in a volume, but only 0.5%0.5\% in a square root.

246810121234xyL(x) = 2 + (x − 4)/4gap 0.25point of tangency (4, 2)y = √x
The tangent at x=4x=4 is L(x)=2+x44L(x)=2+\frac{x-4}{4}. Close in it is excellent — L(4.4)=2.1L(4.4)=2.1 against 4.4=2.0976\sqrt{4.4}=2.0976 — but x\sqrt{x} is concave down, so LL sits above the curve everywhere and the gap grows: L(9)=3.25L(9)=3.25 against 33.
1.01.52.0123xytangent at x = adx = Δxdy = f′(a)ΔxΔy − dy(a, f(a))f(a + Δx)y = f(x)
The whole climb from f(a)f(a) to f(a+Δx)f(a+\Delta x) is Δy\Delta y; the amber piece is the differential dy=f(a)Δxdy=f'(a)\,\Delta x and the dark piece on top is what the tangent misses. Here f(x)=x2f(x)=x^{2}, a=1a=1, Δx=0.8\Delta x=0.8: dy=1.6dy=1.6, Δy=2.24\Delta y=2.24, gap 0.640.64.
02

Formulas

What to have memorised
  • Linearisation at x=ax=a

    L(x)=f(a)+f(a)(xa)L(x)=f(a)+f'(a)(x-a)

    The derivative is frozen at aa, which is what makes LL linear.

  • Tangent-line estimate

    f(a+Δx)f(a)+f(a)Δxf(a+\Delta x)\approx f(a)+f'(a)\,\Delta x

    Needs f(a)f'(a) to exist and Δx\Delta x to be small — nothing is promised far from aa.

  • Differential

    dy=f(x)dx,Δy=f(x+Δx)f(x)dydy=f'(x)\,dx,\qquad \Delta y=f(x+\Delta x)-f(x)\approx dy

    dydy rides the tangent line, Δy\Delta y rides the curve.

  • Standard approximations near 00

    (1+x)k1+kx,ex1+x,sinxx,ln(1+x)x(1+x)^{k}\approx 1+kx,\quad e^{x}\approx 1+x,\quad \sin x\approx x,\quad \ln(1+x)\approx x

    Also tanxx\tan x\approx x and cosx1\cos x\approx 1; the trig ones need radians.

  • Relative error propagation

    y=Cxn  dyy=ndxxy=Cx^{n}\ \Longrightarrow\ \frac{dy}{y}=n\,\frac{dx}{x}

    Percentage error in yy is about n\left|n\right| times the percentage error in xx.

  • Direction from concavity

    f0 on If(x)L(x);f0f(x)L(x)f''\ge 0\ \text{on}\ I\Rightarrow f(x)\ge L(x);\qquad f''\le 0\Rightarrow f(x)\le L(x)

    Concave up: LL under-estimates. Concave down: LL over-estimates.

  • Error bound (first-order Taylor)

    f(x)L(x)M2(xa)2\left|f(x)-L(x)\right|\le\frac{M}{2}(x-a)^{2}

    MM must bound f\left|f''\right| on the whole interval between aa and xx.

03

Estimate a value by linear approximation

The steps, in order
  1. 1

    Name the function ff and pick the nearest base point aa where f(a)f(a) and f(a)f'(a) are exact.

  2. 2

    Compute f(a)f(a) and f(a)f'(a), then write L(x)=f(a)+f(a)(xa)L(x)=f(a)+f'(a)(x-a).

  3. 3

    Set dx=xadx=x-a — converting degrees to radians first — and evaluate f(x)f(a)+f(a)dxf(x)\approx f(a)+f'(a)\,dx.

  4. 4

    Check the sign of ff'' between aa and xx: concave up means under-estimate, concave down means over-estimate.

  5. 5

    If a bound is asked, find MM with fM\left|f''\right|\le M on that interval and quote M2(xa)2\frac{M}{2}(x-a)^{2}.

  6. 6

    For measurement error instead, differentiate the model, then divide by it: dyy=ndxx\frac{dy}{y}=n\frac{dx}{x}.

04

Watch out

The mistakes that cost marks
  • ✗ Wrong

    L(x)=f(a)+f(x)(xa)L(x)=f(a)+f'(x)(x-a)

    ✓ Right

    L(x)=f(a)+f(a)(xa)L(x)=f(a)+f'(a)(x-a) — the slope is the number f(a)f'(a).

    Why: With f(x)f'(x) left in it, LL is not a linear function at all.

  • ✗ Wrong

    Estimating sin31\sin 31^{\circ} with dx=1dx=1

    ✓ Right

    Convert first: a=π/6a=\pi/6 and dx=π1800.01745dx=\frac{\pi}{180}\approx0.01745.

    Why: The derivative formulas for sine and cosine assume radians.

  • ✗ Wrong

    "x>ax>a, so the estimate is too small."

    ✓ Right

    The side of aa is irrelevant — only the sign of ff'' decides the direction.

  • ✗ Wrong

    dy=f(x)dy=f'(x)

    ✓ Right

    dy=f(x)dxdy=f'(x)\,dx, and Δydy\Delta y\approx dy is an approximation, never an equality.

    Why: Without the dx the equation is dimensionally meaningless.

  • ✗ Wrong

    Taking M=f(a)M=\left|f''(a)\right| in the error bound

    ✓ Right

    MM must dominate f\left|f''\right| on the entire interval from aa to xx.

    Why: Any valid overestimate of M is fine; a value only at a is not.

  • ✗ Wrong

    Reporting a relative error of 0.001670.00167 as the percentage error

    ✓ Right

    Percentage error is 100×100\times the relative error: 0.167%0.167\%.

05

Quick check

Commit to an answer before you reveal one
  1. Q1easy

    Find the linearisation of f(x)=xf(x)=\sqrt{x} at a=9a=9 and use it to approximate 9.1\sqrt{9.1}. Is the estimate too large or too small? Justify the direction.

    Hint

    99 is a perfect square, which is why it was chosen; the sign of ff'' settles the direction.

    Show answer

    Answer

    L(x)=3+x96L(x)=3+\frac{x-9}{6}, so 9.1181603.01667\sqrt{9.1}\approx\frac{181}{60}\approx3.01667 — an over-estimate.

    Steps

    f(9)=3f(9)=3 and f(x)=12xf'(x)=\frac{1}{2\sqrt{x}}, so f(9)=16f'(9)=\frac16 and

    L(x)=3+16(x9),L(9.1)=3+0.16=181603.016667.L(x)=3+\tfrac16(x-9),\qquad L(9.1)=3+\tfrac{0.1}{6}=\tfrac{181}{60}\approx3.016667.

    f(x)=14x3/2<0f''(x)=-\frac14 x^{-3/2}<0 on (0,)(0,\infty), so ff is concave down and the tangent lies above the graph: the estimate is too large.

    Check by squaring: (3+160)2=9.100278>9.1\left(3+\frac{1}{60}\right)^{2}=9.100278>9.1. (True value 3.01662063.0166206.)

  2. Q2easy

    Let y=x32xy=x^{3}-2x.

    (a) Find the differential dydy. (b) Evaluate dydy and the actual change Δy\Delta y as xx goes from 22 to 2.12.1. (c) What happens to the gap Δydy\Delta y-dy if the step is 0.010.01 instead?

    Hint

    dy=f(x)dxdy=f'(x)\,dx uses the tangent; Δy\Delta y must be computed from ff itself.

    Show answer

    Answer

    (a) dy=(3x22)dxdy=(3x^{2}-2)\,dx. (b) dy=1dy=1 and Δy=1.061\Delta y=1.061. (c) The gap falls from 0.0610.061 to 0.0006010.000601.

    Steps

    (a) f(x)=3x22f'(x)=3x^{2}-2, so dy=(3x22)dxdy=(3x^{2}-2)\,dx.

    (b) At x=2x=2 with dx=0.1dx=0.1: dy=(122)(0.1)=1dy=(12-2)(0.1)=1. Along the curve, f(2.1)=9.2614.2=5.061f(2.1)=9.261-4.2=5.061 and f(2)=4f(2)=4, so Δy=1.061\Delta y=1.061.

    (c) For a general step hh, Δy=10h+6h2+h3\Delta y=10h+6h^{2}+h^{3} while dy=10hdy=10h, so

    Δydy=6h2+h3.\Delta y-dy=6h^{2}+h^{3}.

    At h=0.1h=0.1 that is 0.0610.061; at h=0.01h=0.01 it is 0.0006010.000601. A tenth of the step gives about a hundredth of the gap — the error is quadratic.

  3. Q3medium

    Use a linear approximation to estimate cos29\cos 29^{\circ}. Give an exact expression and a decimal to six places, state whether it over- or under-estimates, and bound the error.

    Hint

    Base at 30=π/630^{\circ}=\pi/6 and work in radians, so dx=π/180dx=-\pi/180.

    Show answer

    Answer

    cos2932+π3600.874752\cos 29^{\circ}\approx\frac{\sqrt3}{2}+\frac{\pi}{360}\approx0.874752: an over-estimate, with error at most 12(π180)21.5×104\frac12\left(\frac{\pi}{180}\right)^{2}\approx1.5\times10^{-4}.

    Steps

    Take f(x)=cosxf(x)=\cos x, a=π6a=\frac{\pi}{6}, dx=29π180π6=π180dx=\frac{29\pi}{180}-\frac{\pi}{6}=-\frac{\pi}{180}. Then f(a)=32f(a)=\frac{\sqrt3}{2} and f(a)=sinπ6=12f'(a)=-\sin\frac{\pi}{6}=-\frac12, so

    cos293212(π180)=32+π3600.874752.\cos 29^{\circ}\approx\frac{\sqrt3}{2}-\frac12\left(-\frac{\pi}{180}\right)=\frac{\sqrt3}{2}+\frac{\pi}{360}\approx0.874752.

    f(x)=cosx<0f''(x)=-\cos x<0 across [29,30][29^{\circ},30^{\circ}], so ff is concave down and this over-estimates.

    Since f(t)=cost1\left|f''(t)\right|=\left|\cos t\right|\le1, take M=1M=1: error12(π180)21.5×104\left|\text{error}\right|\le\frac12\left(\frac{\pi}{180}\right)^{2}\approx1.5\times10^{-4}. (True value 0.8746200.874620.)

  4. Q4medium

    The radius of a sphere is measured as 2121 cm with a possible error of at most 0.050.05 cm. Use differentials to estimate the maximum error in the calculated volume, and give the relative and percentage errors.

    Hint

    Divide dVdV by VV symbolically before substituting any numbers.

    Show answer

    Answer

    dV88.2π277\left|dV\right|\le88.2\pi\approx277 cm3^{3}; relative error 11400.00714\frac{1}{140}\approx0.00714; percentage error about 0.71%0.71\%.

    Steps

    V=43πr3V=\frac43\pi r^{3}, so dV=4πr2drdV=4\pi r^{2}\,dr. With r=21r=21 and dr0.05\left|dr\right|\le0.05:

    dV4π(441)(0.05)=88.2π277.1 cm3.\left|dV\right|\le4\pi(441)(0.05)=88.2\pi\approx277.1\ \text{cm}^{3}.

    For the relative error divide first: dVV=4πr2dr43πr3=3drr\frac{dV}{V}=\frac{4\pi r^{2}\,dr}{\frac43\pi r^{3}}=3\,\frac{dr}{r}, hence

    dVV30.0521=11400.00714,\frac{\left|dV\right|}{V}\le3\cdot\frac{0.05}{21}=\frac{1}{140}\approx0.00714,

    i.e. about 0.71%0.71\% — three times the 0.238%0.238\% error in rr, as the exponent 33 predicts.

  5. Q5medium

    The area of a disk is computed from its radius, nominally 2424 cm, and must be correct to within a percentage error of 0.5%0.5\%. Using differentials, how accurately must the radius be measured — as a percentage, and in centimetres?

    Hint

    Run the propagation rule backwards: dAA=2drr\frac{dA}{A}=2\frac{dr}{r}, then solve for dr\left|dr\right|.

    Show answer

    Answer

    drr0.0025\left|\frac{dr}{r}\right|\le0.0025, i.e. 0.25%0.25\%; with r=24r=24 that is dr0.06\left|dr\right|\le0.06 cm.

    Steps

    A=πr2A=\pi r^{2} gives dA=2πrdrdA=2\pi r\,dr, so dAA=2drr\frac{dA}{A}=2\,\frac{dr}{r} — the power rule with n=2n=2.

    Impose the tolerance dAA0.005\left|\frac{dA}{A}\right|\le0.005:

    2drr0.005  drr0.0025.2\left|\frac{dr}{r}\right|\le0.005\ \Longrightarrow\ \left|\frac{dr}{r}\right|\le0.0025.

    So the radius needs 0.25%0.25\% accuracy — half the tolerance on the area, because the exponent is 22. In centimetres, dr0.0025(24)=0.06\left|dr\right|\le0.0025(24)=0.06 cm.

  6. Q6hard

    ff is twice differentiable on R\mathbb{R} with f(2)=5f(2)=5, f(2)=3f'(2)=-3 and f(x)>0f''(x)>0 for every xx.

    (a) Use the linearisation at a=2a=2 to estimate f(1.9)f(1.9) and f(2.2)f(2.2). (b) Is each estimate too large or too small, and why is the answer the same for both? (c) If also f4\left|f''\right|\le4 on [1.9,2.2][1.9,2.2], give an interval that certainly contains f(2.2)f(2.2).

    Hint

    The over/under question is settled by the sign of ff'' alone — not by which side of aa the point lies on.

    Show answer

    Answer

    (a) f(1.9)5.3f(1.9)\approx5.3, f(2.2)4.4f(2.2)\approx4.4. (b) Both too small. (c) 4.4<f(2.2)4.484.4<f(2.2)\le4.48.

    Steps

    (a) L(x)=53(x2)L(x)=5-3(x-2), so L(1.9)=5+0.3=5.3L(1.9)=5+0.3=5.3 and L(2.2)=50.6=4.4L(2.2)=5-0.6=4.4.

    (b) f>0f''>0 everywhere means ff is concave up, so its graph lies above every tangent line on both sides of 22: f(x)>L(x)f(x)>L(x) for x2x\ne2. Both are under-estimates. That ff is decreasing at 22 changes nothing.

    (c) Taylor's inequality with M=4M=4 gives

    f(2.2)4.442(0.2)2=0.08,\left|f(2.2)-4.4\right|\le\frac{4}{2}(0.2)^{2}=0.08,

    so 4.32f(2.2)4.484.32\le f(2.2)\le4.48. Combining with (b): 4.4<f(2.2)4.484.4<f(2.2)\le4.48.

06

On the exam

How this topic is marked
  • Nearly every question carries three separate marks: the linearisation, the number, and a sentence naming the direction with the sign of ff'' as the reason. Writing only the number loses two of them.

  • For measurement error, divide symbolically first — dVV=3drr\frac{dV}{V}=3\frac{dr}{r} — then substitute. It removes the arithmetic and hands you the percentage directly.

  • Work in radians, quote the final decimal with units, and remember the bound is quadratic: halving xa\left|x-a\right| quarters it.

Keep going