So far dy/dx has been a single symbol: the derivative, not a quotient. Yet scientists and geometers constantly write things like dA=π(bda+adb) or xdx+ydy=0 and manipulate the pieces separately. §160 shows how to give the isolated symbols dx, dy, dz a precise meaning, so that the approximate equation δz≈fxδx+fyδy of the last lesson becomes an exact equation between differentials. Its great advantage, which Hardy singles out, is that the formula does not care whether x and y are independent variables or functions of something else.
The idea
Take the area z=xy of a rectangle and increase its sides by δx and δy. The new area exceeds the old by two strips, yδx and xδy, plus a small corner δxδy. The strips are linear in the increments; the corner is of second order. The differential dz is, by definition, the linear part: dz=ydx+xdy.
ExploreStrips and a corner
Try this.Press “Halve both increments” several times and watch the two strips and the corner, and the numbers in the read-out.
z=xy: δz=yδx+xδy+δxδy
xy
original area xy
the differential dz = y dx + x dy (two strips)
the error δz − dz = δx δy (corner)
δz=2.2, dz=ydx+xdy=1.96, error δz−dz=δxδy=0.24. Error compared with the larger step: max(δx,δy)δxδy=min(δx,δy)=0.4, which tends to 0 with the step. Relative change: zdz=xdx+ydy=0.49.
What to notice
Each halving halves the strips but quarters the corner. The differential dz (the strips) captures the change to first order; the error δxδy becomes negligible compared with the step. The relative form dz/z=dx/x+dy/y says: percentage errors in the sides add up to the percentage error in the area.
Precise statements
For one variable, y=f(x), we know that δy={f′(x)+ε}δx with ε→0, so δy≈f′(x)δx.
DefinitionDifferentials of a function of one variableHardy: §160 (3)–(6)
Let y=f(x) be differentiable. For the independent variable, the differential dx is an arbitrary increment, dx=δx. The differential of y is defined by
dy=f′(x)δx.
Taking y=x gives dx=δx, consistently, so dy=f′(x)dx, and dividing, dy/dx=f′(x): the derivative is the quotient of the differentials.
So dy/dx now has two meanings, the derivative and the quotient of two differentials, and they agree. Note that dy is not δy: it is the change along the tangent line, while δy is the change along the curve.
DefinitionThe total differentialHardy: §160 (7)–(9)
Let z=f(x,y), with x and y independent and fx, fy continuous. Its differential is
dz=fxδx+fyδy.
Since dx=δx and dy=δy (apply the definition to z=x and z=y), this reads
dz=∂x∂zdx+∂y∂zdy,
an exact equation corresponding to the approximate equation δz≈fxδx+fyδy of §159.
TheoremInvariance of the differentialHardy: §160
The formula dz=zxdx+zydy holds also when x and y are not independent but are functions x=ϕ(t), y=ψ(t) of one variable t (with continuous derivatives), the differentials then being dx=ϕ′(t)dt, dy=ψ′(t)dt and dz=dtdzdt.
TheoremReading off partial derivativesHardy: §160
If x and y are independent and dz=λdx+μdy for all dx, dy, then λ=∂z/∂x and μ=∂z/∂y.
The same definitions and results hold for any number of variables: du=uxdx+uydy+uzdz, and so on.
Proofs
ProofInvariance of the differential
Strategy
Compute dz in the only way available when z is a function of the single variable t, namely dz=dtdzdt, and expand dz/dt with the chain rule of §157.
z as a function of t
Now z=f{ϕ(t),ψ(t)} is a function of t alone, so by the one-variable definition dz=dtdzdt.
Apply the chain rule
Since fx and fy are continuous, dtdz=fxϕ′(t)+fyψ′(t).
Multiply by dt
dz=fxϕ′(t)dt+fyψ′(t)dt=fxdx+fydy, because dx=ϕ′(t)dt and dy=ψ′(t)dt by the one-variable definition. This is the same formula as for independent x and y.
∎
ProofReading off partial derivatives
Strategy
Since dx and dy are arbitrary for independent variables, we may choose one of them to be zero.
Compare with the definition
We have fxdx+fydy=λdx+μdy for all dx, dy.
Choose dy=0, then dx=0
With dy=0 and dx=1: λ=fx. With dx=0 and dy=1: μ=fy. (Hardy derives it from §159 instead; the content is the same uniqueness.)
∎
The invariance is what makes the notation so efficient. To differentiate f(x,y)=0 implicitly, just write fxdx+fydy=0 (valid although y now depends on x) and divide: dy/dx=−fx/fy whenever fy=0. The tangent to the curve f(x,y)=0 at (x0,y0) is therefore (x−x0)fx+(y−y0)fy=0 (Examples LXII, 9–10). And at a maximum of f(x,y) on a curve g(x,y)=0 we have df=0 for every allowed displacement, which forces fx/gx=fy/gy (Examples LXII, 13), the germ of the method of Lagrange multipliers.
Worked examples and exercises
ExerciseArea of an ellipse●○○Hardy: Examples LXII, 1
The area of an ellipse with semi-axes a and b is A=πab. Prove that AdA=ada+bdb, and estimate the change of area when a increases by 1% and b decreases by 2%.
Hint
dA=π(bda+adb). Divide by A.
Solution
dA=Aada+Abdb=πbda+πadb, and dividing by A=πab gives dA/A=da/a+db/b. With da/a=0.01 and db/b=−0.02, dA/A=−0.01: the area decreases by about 1%. (Exactly: 1.01×0.98=0.9898, a decrease of 1.02%; the difference is the product of the two small changes, as in the widget’s corner.)
ExerciseThe chain rule with differentials●○○Hardy: Examples LXII, 5
Let z be a function of u and v, which are functions of x and y. Show that ∂x∂z=∂u∂z∂x∂u+∂v∂z∂x∂v, and similarly for y.
Hint
Write dz=zudu+zvdv (valid although u,v are not independent), then substitute du=uxdx+uydy and dv=vxdx+vydy.
Solution
By invariance, dz=zudu+zvdv=zu(uxdx+uydy)+zv(vxdx+vydy)=(zuux+zvvx)dx+(zuuy+zvvy)dy. Since x and y are independent, the coefficients of dx and dy are zx and zy, by the theorem on reading off partial derivatives.
ExerciseImplicit differentiation and tangents●○○Hardy: Examples LXII, 9–10
Find dy/dx on the circle x2+y2=25 and on the folium x3+y3=3xy, and the tangent to the circle at (3,4).
Hint
Take the differential of each side. For the folium, collect the terms in dy.
Solution
Circle: 2xdx+2ydy=0, so dy/dx=−x/y; at (3,4) the slope is −43 and the tangent is 3(x−3)+4(y−4)=0, that is 3x+4y=25.
Folium: 3x2dx+3y2dy=3ydx+3xdy, so (y2−x)dy=(y−x2)dx and dy/dx=(y−x2)/(y2−x), wherever y2=x.
ExerciseTriangles of constant area●●●Hardy: Examples LXII, 3
The sides a,b,c of a triangle vary so that its area Δ stays constant; regard a as a function of b and c. Prove that ∂b∂a=−cosAcosB.
Hint
Use Δ=21bcsinA and a2=b2+c2−2bccosA, with c fixed and dΔ=0. At the end use the projection formula c=acosB+bcosA.
Solution
Keep c fixed. From dΔ=0: 21c(sinAdb+bcosAdA)=0, so dA=−bcosAsinAdb. From the cosine rule:
Find the largest value of f=xy on the line g=x+y−1=0, first by substitution, then by the condition fx/gx=fy/gy.
Hint
At the maximum, df=ydx+xdy=0 for every displacement along the line, where dx+dy=0.
Solution
Substitution: f=x(1−x), maximal at x=21, value 41.
Differentials: on the line dy=−dx, so df=(y−x)dx, which vanishes for all dx only if x=y. That is fx/gx=y/1=x/1=fy/gy. With x+y=1 we again get x=y=21. The method needs no solving for y, which is its point when the constraint is complicated.
Check yourself
Check yourselfEnter a number
A rectangle has sides x=3, y=2. They increase by δx=0.1 and δy=0.05. What is δz−dz for the area z=xy?
Why
dz=ydx+xdy=0.2+0.15=0.35, while δz=3.1×2.05−6=0.355. The difference is δxδy=0.005, the corner of the picture.
Check yourselfTrue or false?
If z=f(x,y) and x, y are both functions of t, the formula dz=fxdx+fydy is still correct (with dx=x′(t)dt, dy=y′(t)dt).
Why
True; this is the invariance Hardy stresses. Dividing by dt gives exactly the chain rule of §157. What changes is only that dx and dy are no longer arbitrary: they are tied to dt.
Check yourselfChoose one
For y=x2 at x=1 with δx=0.1, which is correct?
Why
dy=f′(1)δx=2×0.1=0.2 is the change along the tangent; δy=1.12−1=0.21 is the change along the curve. In Hardy’s treatment a differential is an ordinary finite number, not an “infinitely small quantity”; what is small is the difference δy−dy=(δx)2 compared with δx.
Check yourselfEnter a number
On the circle x2+y2=25, what is dy/dx at the point (3,4)?
Why
2xdx+2ydy=0 gives dy/dx=−x/y=−3/4.
Hardy’s language
the formula which expresses dz in terms of dx and dy is the same whether the variables x and y are independent or not.
— G. H. Hardy, §160
Summary
Summary
For independent variables, dx=δx is an arbitrary increment; for y=f(x), dy=f′(x)dx, so dy/dx is a genuine quotient equal to the derivative.
For z=f(x,y): dz=zxdx+zydy, the linear part of δz; the error δz−dz is small compared with the step.
The formula holds whether or not x and y are independent; that is what makes differentials convenient for implicit differentiation, changes of variable and constrained maxima.
If dz=λdx+μdy for independent x,y, then λ=zx and μ=zy.