Calculus › Section N · Functions of several variables › Topic N.4
Differentials in several variables
Change one input and a partial derivative tracks the effect. Change several inputs at once and their effects stack — the total differential adds them up, and it powers practical error estimates.
The one-variable idea, recalled
For a one-variable function, the differential \(dy = f'(x)\,dx\) approximates the true change \(\Delta y = f(x + \Delta x) - f(x)\) when the step \(\Delta x\) is small: the curve behaves like its tangent line over a short stretch. Now let \(z = f(x, y)\). Both inputs can shift at once, and the true change is the increment
This is harder to compute, because \(f\) must be re-evaluated at the new point. The question of this topic is what one-variable-style quantity replaces \(dy\) — and how good that replacement is.
The total differential
If \(x\) and \(y\) both move a little, the effects add, one direction at a time. Define \(dx = \Delta x\) and \(dy = \Delta y\), and combine the two partial rates:
This is the total differential of \(z\). Each term is a one-variable change: slope in the \(x\)-direction times the step in \(x\), plus slope in the \(y\)-direction times the step in \(y\). The pattern extends to any number of inputs; for \(w = f(x, y, z)\), \(dw = w_x\,dx + w_y\,dy + w_z\,dz\).
Worked example: building a total differential
Find the total differential of \(z = x^2\cos y + 4xy^2\). First the partials: with \(y\) frozen, \(z_x = 2x\cos y + 4y^2\); with \(x\) frozen, \(z_y = -x^2\sin y + 8xy\). Then
For three variables, the same recipe applies to \(w = x^2 + y^2 + z^2\): freeze two variables at a time to get \(w_x = 2x\), \(w_y = 2y\) and \(w_z = 2z\), so \(dw = 2x\,dx + 2y\,dy + 2z\,dz\). No limits are needed — a total differential costs two (or three) ordinary differentiations.
Approximating a change
For small steps, the increment and the differential agree closely: \(\Delta z \approx dz\). The reason is that each partial rate barely changes over a short step, so the two one-variable tangent-line estimates, one per direction, already account for almost everything. Geometrically, \(dz\) is exactly the change in height of the plane tangent to the surface at the starting point, so the approximation says: near the point, the surface and its tangent plane rise and fall together. The tangent-plane view returns in the topic on tangent planes; here the practical form is what matters: substitute the starting point into the partials, multiply by the steps and add.
Worked example: approximating a move
A drone's altitude is modelled as \(z = \sqrt{x^2 + y^2}\) metres above flat ground, \(x\) and \(y\) being grid positions. Estimate the change in altitude when it moves from \((3, 4)\) to \((3.02, 3.97)\). The partials are \(z_x = \frac{x}{\sqrt{x^2 + y^2}}\) and \(z_y = \frac{y}{\sqrt{x^2 + y^2}}\); at \((3, 4)\), where \(z = 5\), they are \(\frac{3}{5}\) and \(\frac{4}{5}\). The steps are \(dx = 0.02\) and \(dy = -0.03\), so
The exact change is \(\sqrt{3.02^2 + 3.97^2} - 5 \approx -0.0119\) — the differential is spot on.
What differentiability really means
In one variable, \(f\) is differentiable at a point when \(f'(x)\) exists, and that is all. In two variables the honest definition is stronger: \(f\) is differentiable at \((x_0, y_0)\) if the increment can be written as
where \(\varepsilon_1\) and \(\varepsilon_2\) both shrink to 0 as \((\Delta x, \Delta y)\) approaches \((0, 0)\). In words: the tangent-plane-style change accounts for \(\Delta z\) up to leftover errors that vanish faster than the steps themselves. Mere existence of both partials does not achieve this — a function can have both partials at a point and still fail to be continuous there, as the path examples of the previous topic show.
When differentiability is guaranteed
Two theorems sort the situation out. First, the sufficient condition: if \(f_x\) and \(f_y\) are continuous on an open region, then \(f\) is differentiable there. Since every function built from polynomials, roots, exponentials and trigonometric functions by sums, products, quotients and compositions has continuous partials wherever it is defined, differentiability comes free in practice. Second, the consequence: if \(f\) is differentiable at a point, then \(f\) is continuous there. So the hierarchy is strict — continuous partials beat differentiability, and differentiability beats continuity — but at the top sits the everyday class of functions, which enjoy all three.
Worked example: propagated error in a volume
A workshop measures a storage box as 40 by 25 by 20 centimetres, each dimension correct to within 0.1 centimetres. How much could the computed volume be off? With \(V = xyz\), the total differential is \(dV = yz\,dx + xz\,dy + xy\,dz\). Each measurement error is at most 0.1, so the worst case is \(dV = (25)(20)(0.1) + (40)(20)(0.1) + (40)(25)(0.1) = 50 + 80 + 100 = 230\) cubic centimetres. The measured volume is \(V = 20{,}000\) cubic centimetres, so the relative error is about \(\frac{230}{20000} = 0.0115\), roughly 1.2 percent. A full 1.2 percent swing from millimetre slips — the products \(yz\), \(xz\) and \(xy\) magnify small inputs.
Common mistakes to avoid
The most frequent slip is reporting only the \(f_x\,dx\) term; the total differential always has one term per input, so count them. The second is treating \(dz\) as the exact change: it is an estimate whose quality degrades as the steps grow, as the drone example shows in miniature. The third is assuming that existing partials make a function differentiable — the definition demands more, and continuous partials are the standard guarantee. The fourth is a sign slip: a negative \(dy\) must be carried through, since moving down counts as negative change. Finally, when tolerances are given as plus-or-minus, take absolute values of every term to bound the worst case.
Key takeaways
The increment is \(\Delta z = f(x + \Delta x, y + \Delta y) - f(x, y)\), and the total differential \(dz = f_x\,dx + f_y\,dy\) (with \(dx = \Delta x\), \(dy = \Delta y\)) approximates it for small steps, one term per input. Differentiability means \(\Delta z\) equals that linear change up to errors vanishing faster than the steps; it is guaranteed when the partials are continuous, and it implies continuity. In practice \(\Delta z \approx dz\) turns re-evaluation into multiplication and addition, and it converts measurement tolerances into propagated error bounds, as in the volume estimate \(dV = yz\,dx + xz\,dy + xy\,dz\).
Flashcards
Flashcard
f(x + Δx, y + Δy) − f(x, y)
Test your understanding
Question 1 of 5
What is the total differential of z = 5x²y?
Question 2 of 5
For f(x, y) = √(x² + y²), use dz to estimate the change when (x, y) moves from (3, 4) with dx = 0.02 and dy = −0.03.
Question 3 of 5
Which condition guarantees that f is differentiable on an open region?
Question 4 of 5
A box measures 40 × 25 × 20 cm, each dimension correct to within 0.1 cm. Estimate the worst-case propagated error in its computed volume.
Question 5 of 5
If f is differentiable at (x₀, y₀), what follows for certain?
Standards: Beyond AP
