Calculus › Section N · Functions of several variables › Topic N.3
Partial derivatives
A quantity with several inputs changes differently depending on which input you nudge. Partial derivatives measure one direction of change at a time — and they turn out to be ordinary derivatives in disguise.
Hold everything else still
The temperature at a point on a metal rod depends on where you stand and on the time; a factory's output depends on labour and machinery. When several inputs influence one output, a single number for "the rate of change" is not enough — you must say which input is moving. The fix is to freeze all but one input. The partial derivative of \(f\) with respect to \(x\) measures how \(f\) responds when \(x\) changes and every other input holds still. It answers questions like: how much extra output does one more worker add with the machinery fixed? Economists call that partial the marginal product of labour.
The definition
Let \(z = f(x, y)\). The partial derivative with respect to \(x\) is the ordinary derivative taken while \(y\) is pinned down:
Only \(x\) moves, from \(x\) to \(x + h\), so the point \((x, y)\) travels horizontally. The partial with respect to \(y\) is defined the same way with the roles swapped:
Each definition is a one-variable derivative in disguise — the new feature is only that the other variables are locked.
Computing them: freeze and differentiate
In practice you never use the limit. To find \(f_x\), treat \(y\) as a constant and differentiate with respect to \(x\) exactly as in one-variable calculus; to find \(f_y\), treat \(x\) as the constant. A term that involves only \(y\), such as \(5y\), counts as a constant when differentiating with respect to \(x\), so its \(x\)-partial is zero. Several notations mean the same thing: \(f_x\), \(\frac{\partial f}{\partial x}\), \(z_x\) and \(\frac{\partial z}{\partial x}\) all name the partial with respect to \(x\). The rounded \(\partial\) signals that other variables are being held fixed, distinguishing it from the ordinary \(\frac{d}{dx}\).
Worked example: first partials
Find both first partials of \(f(x, y) = 2x^3y^2 + 5xy - 3y\) and evaluate them at \((1, 2)\). For \(f_x\), freeze \(y\): the derivative of \(2x^3y^2\) is \(6x^2y^2\), the derivative of \(5xy\) is \(5y\), and \(-3y\) is a constant. So \(f_x = 6x^2y^2 + 5y\). For \(f_y\), freeze \(x\): \(2x^3y^2\) gives \(4x^3y\), \(5xy\) gives \(5x\), and \(-3y\) gives \(-3\), so \(f_y = 4x^3y + 5x - 3\). At \((1, 2)\): \(f_x = 6(1)(4) + 10 = 34\) and \(f_y = 8 + 5 - 3 = 10\). The surface climbs 34 units per unit step in the \(x\)-direction there, but only 10 per unit step in the \(y\)-direction.
Geometry: slopes of a sliced surface
The graph of \(f\) is a surface, and each partial derivative is the slope of a slice. Cut the surface with the vertical plane \(y = b\) and you get a curve; \(f_x(a, b)\) is its slope at \(x = a\). Cut with the plane \(x = a\) instead and the slope of that curve is \(f_y(a, b)\). For \(f(x, y) = 12 - x^2 - 3y^2\), we get \(f_x = -2x\) and \(f_y = -6y\), so at the point \((2, 1)\), where \(f = 12 - 4 - 3 = 5\), the surface falls 4 units per unit step in the \(x\)-direction and 6 per unit step in the \(y\)-direction. One point on a surface can have different slopes in different directions — that is the whole point.
Worked example: a cooling rod
A metal rod has one end heated, and \(t\) hours after the heater is switched off the temperature at distance \(x\) metres from that end is modelled as \(T(x, t) = 20 + 80e^{-0.2x - 0.5t}\) degrees Celsius; the model assumes the room stays at 20 degrees. The two partials answer different physical questions. Freezing time, \(T_x = -16e^{-0.2x - 0.5t}\): at \(x = 0, t = 0\) the temperature drops about 16 degrees per metre as you walk away from the warm end. Freezing position, \(T_t = -40e^{-0.2x - 0.5t}\): at that same point the rod is cooling at 40 degrees per hour. One function, two separate rates of change.
Three or more variables
Nothing changes when more inputs arrive: to differentiate with respect to one variable, freeze all the others. For \(w = x^2y + yz^3 + 2xz\), the partial with respect to \(z\) treats \(x\) and \(y\) as constants, so \(w_z = 3yz^2 + 2x\). Likewise \(w_x = 2xy + 2z\) and \(w_y = x^2 + z^3\). Each partial singles out one input, and the rest become spectators.
Second and mixed partials
Because \(f_x\) and \(f_y\) are themselves functions of \(x\) and \(y\), you can differentiate again. There are four second partials: \(f_{xx}\) and \(f_{yy}\) (differentiate the same variable twice) and the mixed partials \(f_{xy}\) (first \(x\), then \(y\)) and \(f_{yx}\) (first \(y\), then \(x\)). Subscripts record the order left to right; in the other notation, \(f_{xy} = \frac{\partial^2 f}{\partial y\,\partial x}\), it reads right to left. For the function above, \(f_x = 6x^2y^2 + 5y\), so \(f_{xy} = 12x^2y + 5\); going the other way, \(f_y = 4x^3y + 5x - 3\), so \(f_{yx} = 12x^2y + 5\). The two mixed partials came out equal, and that is no accident.
Clairaut's theorem
If both mixed partials \(f_{xy}\) and \(f_{yx}\) are continuous on an open region containing the point, then they are equal there: \(f_{xy} = f_{yx}\). The continuity condition matters — it is what lets you compute the mixed partial in whichever order is less painful. Almost every function you meet in applications satisfies it, so in practice the order is a free choice. One caution: the theorem does not say \(f_{xx} = f_{yy}\). Differentiating twice with respect to \(x\) and twice with respect to \(y\) are genuinely different operations, and \(f = 2x^3y^2 + 5xy - 3y\), with \(f_{xx} = 12xy^2\) but \(f_{yy} = 4x^3\), shows the results usually differ.
Worked example: partials that need the chain rule
When a one-variable rule like the chain rule appears inside a partial, it applies as usual to the moving variable. Take \(f(x, y) = \ln(x^2 + y^2)\). For \(f_x\), the inner function is \(x^2 + y^2\) and \(y^2\) is constant, so the inner derivative with respect to \(x\) is \(2x\):
By symmetry, \(f_y = \frac{2y}{x^2 + y^2}\). At \((3, 4)\), \(f_x = \frac{6}{25}\) and \(f_y = \frac{8}{25}\). Likewise, for \(e^{xy}\) the inner derivative with respect to \(x\) is \(y\), so \(f_x = ye^{xy}\) and \(f_y = xe^{xy}\). The same combination of a frozen variable and a one-variable technique handles exponentials, roots and trigonometric functions of \(x\) and \(y\).
Common mistakes to avoid
The classic slip is differentiating with respect to the wrong variable, so write the variable symbol next to each step. The second is mishandling the frozen variable: treating \(y\) as constant means \(y\) stays in place as a factor — the \(x\)-partial of \(5xy\) is \(5y\), not 5. The third is inventing products of partials: \(f_{xy}\) means differentiate twice, never multiply \(f_x\) by \(f_y\). The fourth is quoting Clairaut's theorem without its continuity hypothesis, or stretching it to claim \(f_{xx} = f_{yy}\). Finally, do not forget that terms free of the variable you are differentiating have partial derivative zero.
Key takeaways
The partial derivative \(f_x = \lim_{h \to 0} \frac{f(x+h, y) - f(x, y)}{h}\) differentiates with respect to \(x\) while every other input is frozen — computed by treating the others as constants. Notations \(f_x\), \(\frac{\partial f}{\partial x}\) and \(z_x\) all mean the same thing. Geometrically, \(f_x(a, b)\) is the slope of the slice by the plane \(y = b\). Partials apply to any number of variables and support second partials; when the mixed partials are continuous, Clairaut's theorem guarantees \(f_{xy} = f_{yx}\). Applications read each partial as a rate of change in one input, from cooling rods to marginal output.
Flashcards
Flashcard
Rate of change of f when x moves and y is held fixed
Test your understanding
Question 1 of 5
For f(x, y) = 3x²y + y³, what is the partial derivative with respect to x?
Question 2 of 5
For f(x, y) = 2x³y² + 5xy − 3y, what is f_y at the point (1, 2)?
Question 3 of 5
For f(x, y) = ln(x² + y²), what is f_x?
Question 4 of 5
For f(x, y) = 12 − x² − 3y², what is the slope of the surface at (2, 1) in the y-direction?
Question 5 of 5
If f_xy and f_yx are both continuous at a point, what does Clairaut's theorem let you conclude?
Standards: Beyond AP
