Multivariable Functions

A multivariable function maps several real inputs to one or more real outputs, $f : R^{n} \to R^{m}$ . Its values depend jointly on all input variables, so it encodes surfaces, fields, and transformations rather than mere curves. Used across calculus, physics, economics, machine learning, and engineering wherever a quantity depends on more than one independent variable. Studied systematically since the 18th–19th centuries (Euler, Cauchy, Jacobi, Weierstrass).

A multivariable function assigns to each point of a subset $D \subseteq R^{n}$ (the domain) one value in $R^{m}$ . When $m = 1$ we call it a scalar field; when $m > 1$ it is a vector-valued function or vector field. These objects matter because almost every real quantity—temperature in a room, profit as a function of price and volume, the loss of a neural network—depends on multiple inputs simultaneously, and single-variable calculus cannot capture how they vary in different directions. A surprising fact: a function of two variables can be separately continuous in each variable along every axis yet fail to be continuous at the origin, revealing that "approaching along lines" is fundamentally insufficient in dimensions greater than one.

The Core Insight

The central conceptual leap from one variable to many is that there is no single "rate of change": variation depends on the direction in which you move. A scalar field over $R^{n}$ has infinitely many directional derivatives at each point, organized by a single linear object—the gradient or, more generally, the derivative (Jacobian) matrix. Local behavior is governed by linear approximation in all directions at once, and limits must respect every path of approach, not just coordinate lines.

Formal Statement

Multivariable Function

Let $D \subseteq R^{n}$ be nonempty. A multivariable function is a rule

f : D \subseteq R^{n} \to R^{m}, x = (x_{1}, \dots, x_{n}) \mapsto f (x) = (f_{1} (x), \dots, f_{m} (x))

⟳a function is a set of ordered pairs

(x, y)

with

x \in D

y \in R^{m}

, in which each

x

appears at most once; the component form follows by projecting onto coordinates.

Here:

$n$ = number of independent variables (dimension of the domain),
$m$ = number of output components (dimension of the codomain),
each $f_{i} : D \to R$ is a scalar component function,
the range is ${f (x) : x \in D} \subseteq R^{m}$ .

When $m = 1$ , $f$ is a scalar field and its graph is the set

Γ_{f} = {(x, f (x)) \in R^{n + 1} : x \in D}

⟳replace the last output coordinate of the codomain

R^{n + 1}

with the function value; this generalizes

y = f (x)

to a hypersurface of dimension

n

Level Sets

For a scalar field $f : D \to R$ and constant $c \in R$ , the level set at value $c$ is

L_{c} = f^{- 1} ({c}) = {x \in D : f (x) = c}

⟳take the preimage of a single point under

f

; in

R^{2}

this is a level curve (contour), in

R^{3}

a level surface.

Limit of a Multivariable Function

Let $a$ be a limit point of $D$ . We write $lim_{x \to a} f (x) = L$ if

\forall ε > 0 \exists δ > 0 : 0 < ∥ x - a ∥ < δ, x \in D ⟹ ∥ f (x) - L ∥ < ε

⟳replace absolute values in the single-variable

ε

–

δ

definition by the Euclidean norm

∥ \cdot ∥

; the punctured ball of radius

δ

must lie entirely inside

D

's relevant neighborhood for every direction simultaneously.

Continuity and Component Decomposition

A function $f : D \to R^{m}$ is continuous at $a \in D$ if and only if every component function $f_{i}$ is continuous at $a$ . Moreover $lim_{x \to a} f (x) = L$ iff $lim_{x \to a} f_{i} (x) = L_{i}$ for each $i$ . The vector limit (when it exists) is unique and must agree along every path approaching $a$ .

Component Reduction of Limits

We show vector convergence is equivalent to coordinatewise convergence.

Forward direction. Suppose $lim_{x \to a} f (x) = L$ . For each $i$ , the projection $π_{i} (y) = y_{i}$ satisfies $∣ y_{i} - L_{i} ∣ \leq ∥ y - L ∥$ . Hence given $ε > 0$ , the same $δ$ from $[4]$ forces $∣ f_{i} (x) - L_{i} ∣ \leq ∥ f (x) - L ∥ < ε$ .

Reverse direction. Suppose each $f_{i} \to L_{i}$ . Given $ε > 0$ , choose $δ_{i}$ so $∣ f_{i} (x) - L_{i} ∣ < ε / m$ on the punctured $δ_{i}$ -ball. Let $δ = min_{i} δ_{i}$ .

Then on the punctured $δ$ -ball,

∥ f (x) - L ∥ = (i = 1 \sum m ∣ f_{i} (x) - L_{i} ∣^{2})^{1/2} < (i = 1 \sum m \frac{ε ^{2}}{m})^{1/2} = ε

⟳square and sum the componentwise bounds, then take the square root; the bound

ε / m

is chosen precisely so the sum of

m

terms collapses to

ε^{2}

Uniqueness follows from the Hausdorff property of $R^{m}$ : two distinct limits $L \neq = L^{'}$ would force $∥ f (x) - L ∥$ and $∥ f (x) - L^{'} ∥$ both below $∥ L - L^{'} ∥/2$ , contradicting the triangle inequality. $■$

Directional Derivative and Gradient

For scalar $f$ , the directional derivative at $a$ along a unit vector $u$ is

D_{u} f (a) = t \to 0 lim \frac{f ( a + t u ) - f ( a )}{t}

⟳restrict

f

to the line

t \mapsto a + t u

, producing a one-variable function whose ordinary derivative at

t = 0

D_{u} f (a)

The gradient collects all first partials, and for differentiable $f$ ,

\nabla f (a) = (\frac{\partial f}{\partial x _{1}}, \dots, \frac{\partial f}{\partial x _{n}})_{a}, D_{u} f (a) = \nabla f (a) \cdot u

⟳apply the multivariable chain rule to

g (t) = f (a + t u)

;

g^{'} (0) = \sum_{i} \partial_{i} f u_{i}

, which is the dot product.

Best Linear Approximation / Differentiability

f : D \to R^{m}

is differentiable at an interior point

a

if there exists a linear map

D f (a) \in R^{m \times n}

(the Jacobian) with

h \to 0 lim \frac{f ( a + h ) - f ( a ) - D f ( a ) h}{∥ h ∥} = 0

⟳this demands the error of the affine approximation

f (a) + D f (a) h

vanish faster than

∥ h ∥

; it is the exact multivariable analogue of

f (a + h) = f (a) + f^{'} (a) h + o (h)

The Jacobian entries are the partials $[D f (a)]_{ij} = \partial f_{i} / \partial x_{j}$ . Existence of all partials does not imply differentiability, but if the partials exist and are continuous near $a$ , then $f$ is differentiable there ( $C^{1} \Rightarrow$ differentiable).

Augustin-Louis Cauchy | 1789–1857

Cauchy gave the modern rigorous footing for limits, continuity, and convergence that multivariable analysis rests upon. His insistence on the $ε$ – $δ$ spirit (later formalized by Weierstrass) turned vague talk of "infinitely small" multivariable increments into precise statements. Cauchy also contributed foundational results on systems of functions and the determinant theory that underlies the Jacobian, the key object measuring how a multivariable map distorts volume locally.

Building Intuition

Picture a scalar field $f (x, y)$ as a landscape: the height above each point $(x, y)$ is the function value. The graph $[2]$ is the terrain surface. Level curves $[3]$ are the contour lines on a topographic map—walking along one keeps your altitude fixed. The gradient $[7]$ points in the direction of steepest ascent, and its magnitude tells you how steep. Crucially, the gradient is always perpendicular to the level curve through your point: along a contour the height doesn't change, so the "uphill" arrow can have no component tangent to it.

For a vector field $f : R^{2} \to R^{2}$ , drop the landscape picture and imagine an arrow attached to every point—like wind velocity over a map. The Jacobian $[8]$ then tells you how a tiny disk of points near $a$ gets stretched, rotated, and sheared into a tiny ellipse near $f (a)$ .

Graph, level curves, and gradient of $f(x,y)=x^2+y^2$

The Field Picture

Physicists meet multivariable functions as fields. A scalar field like temperature $T (x, y, z)$ or electric potential $V (r)$ assigns one number to each point of space; its gradient $- \nabla V = E$ gives the electric field, a vector field. The recurring theme is that physical laws are statements about how fields vary in space and time: $\nabla \cdot E$ , $\nabla \times B$ , and $\nabla^{2} ϕ$ all package directional rates of change of multivariable functions. The Jacobian's determinant appears when changing coordinates in integrals (e.g., to spherical coordinates) and measures local volume scaling.

Worked Examples

Computing a Gradient and Steepest-Ascent Rate

Given:

f (x, y) = x^{2} y + 3 y

(a dimensionless potential), evaluate at

a = (2, 1)

, and find the maximum rate of increase and its direction. Steps:

Partials: $\frac{\partial f}{\partial x} = 2 x y$ , $\frac{\partial f}{\partial y} = x^{2} + 3$ .
At $(2, 1)$ : $\partial_{x} f = 2 (2) (1) = 4$ , $\partial_{y} f = 4 + 3 = 7$ .
Gradient: $\nabla f (2, 1) = (4, 7)$ via $[7]$ .
Maximum directional derivative equals $∥\nabla f ∥$ (achieved when $u ∥ \nabla f$ ):

∥ u ∥ = 1 max D_{u} f = ∥\nabla f (2, 1) ∥ = 4^{2} + 7^{2} = 65

⟳

D_{u} f = \nabla f \cdot u = ∥\nabla f ∥ cos θ

is maximized at

θ = 0

, giving

∥\nabla f ∥

Unit direction of steepest ascent: $u^{*} = \frac{1}{65} (4, 7)$ .

Answer:

\nabla f (2, 1) = (4, 7), max rate = 65 \approx 8.06 (units of f per unit length), u^{*} = \frac{1}{65} (4, 7)

A Limit That Fails (Path Dependence)

Given:

f (x, y) = \frac{x y}{x ^{2} + y ^{2}}

for

(x, y) \neq = (0, 0)

. Does

lim_{(x, y) \to (0, 0)} f

exist? Steps:

Along $y = 0$ : $f (x, 0) = 0/ x^{2} = 0$ , so the limit along the $x$ -axis is $0$ .
Along $x = 0$ : $f (0, y) = 0$ likewise gives $0$ .
Along $y = x$ : $f (x, x) = \frac{x ^{2}}{2 x ^{2}} = \frac{1}{2}$ for all $x \neq = 0$ .
Compute the candidate values:

x \to 0 lim f (x, 0) = 0 \neq = \frac{1}{2} = x \to 0 lim f (x, x)

⟳substitute the path equation into

f

to reduce it to a single-variable limit; differing path limits violate uniqueness from the Continuity and Component Decomposition theorem.

Because two paths give different values, the joint limit cannot exist.

Answer:

The limit does not exist; f is discontinuous at (0, 0) regardless of how it is defined there.

Common Misconceptions

Existence of all partial derivatives does NOT imply continuity or differentiability. Partials probe only the coordinate-axis directions. A function can have

\partial_{x} f

and

\partial_{y} f

defined everywhere yet not be continuous, let alone differentiable, at a point. The correct local linearization criterion is the limit

[8]

, not the mere existence of partials. The reliable shortcut is the

C^{1}

theorem: continuous partials guarantee differentiability.

Partials Exist, Function Is Discontinuous

Let

f (x, y) = ⎩ ⎨ ⎧ \frac{x y}{x ^{2} + y ^{2}}, 0, (x, y) \neq = (0, 0) (x, y) = (0, 0)

⟳this is the function of

[10]

with the value

0

assigned at the origin.

Both partials at the origin exist:

f_{x} (0, 0) = t \to 0 lim \frac{f ( t , 0 ) - 0}{t} = t \to 0 lim \frac{0}{t} = 0, f_{y} (0, 0) = 0

⟳along each axis

f \equiv 0

, so the difference quotient defining each partial is identically zero.

Yet $f$ is not continuous at $(0, 0)$ , since the limit along $y = x$ is $\frac{1}{2} \neq = f (0, 0) = 0$ . Existence of partials therefore fails to control behavior off the axes—the essential warning of multivariable calculus.

Connections

Multivariable functions sit at the crossroads of analysis and geometry. The single-variable theory of Differentiation generalizes to the Jacobian Matrix and the differentiability criterion $[8]$ . Integration extends to Multiple Integrals via Fubini's theorem, where the Change of Variables Theorem uses the Jacobian determinant. The directional structure $[7]$ launches the study of Gradient Vector Fields, Divergence and Curl, and ultimately Vector Calculus. Level sets $[3]$ connect to Implicit Function Theorem and the geometry of Manifolds. In optimization, $\nabla f = 0$ characterizes critical points, linking to Lagrange Multipliers and the Hessian Matrix second-derivative test.

Real-World Applications

Machine learning (gradient descent): A loss $L (θ)$ is a scalar field over millions of parameters; training follows $- \nabla L$ to minimize error. Backpropagation is repeated Jacobian application via the chain rule $[8]$ .
Thermodynamics: State functions like Gibbs free energy $G (T, P, N)$ are multivariable; partial derivatives give entropy, volume, and chemical potential.
Economics: A production function $Q (K, L)$ (capital, labor) uses partials for marginal products and the gradient for optimal input mixes under constraints.
Computer graphics: Surfaces $z = f (x, y)$ and parametric maps $R^{2} \to R^{3}$ render terrain; normals come from $\nabla f$ .
Meteorology: Pressure $P (x, y)$ fields produce wind via the pressure gradient force $- \nabla P$ ; isobars are level curves.
Electromagnetism: Potential fields $V (r)$ yield $E = - \nabla V$ , and Maxwell's equations are differential statements on vector fields.
Medical imaging: CT/MRI reconstruct a 3D density function $ρ (x, y, z)$ ; level surfaces segment organs.

Historical Development

1667–1684: Newton and Leibniz develop single-variable calculus, laying the conceptual groundwork for rates of change later extended to many variables.
1734: Euler systematically studies functions of several variables and introduces notation for partial differentiation in mechanics and fluid problems.
1748: Euler's Introductio in analysin infinitorum treats functions abstractly, broadening the function concept beyond single inputs.
1797–1813: Lagrange develops the calculus of variations and the multiplier method for constrained multivariable optimization.
1821: Cauchy's Cours d'analyse gives rigorous limit and continuity definitions adaptable to multivariable settings.
1841: Jacobi publishes on functional determinants—the Jacobian—formalizing how multivariable maps transform volumes.
1854–1868: Riemann's work on manifolds and multivariable integration reframes functions on curved domains.
1870s: Weierstrass cements the $ε$ – $δ$ formalism, and pathological examples (nowhere-differentiable, path-dependent limits) sharpen the distinction between partial and total differentiability.
1900s: Fréchet and others generalize differentiability to abstract normed spaces, with $[8]$ as the template (the Fréchet derivative).

Subtleties & Limits

While the theory of multivariable functions is classical and complete, practical and computational frontiers remain open. High-dimensional optimization (where $n$ is in the millions) is dominated by the geometry of saddle points and non-convex loss landscapes, whose structure is not fully understood. The behavior of generic smooth functions in very high dimensions—where almost all critical points are saddles and the curse of dimensionality makes sampling intractable—drives active research in optimization theory and statistics. Even classifying singularities of smooth maps $R^{n} \to R^{m}$ (catastrophe and singularity theory, following Whitney and Thom) yields rich, still-expanding taxonomies, and questions about the regularity of solutions to nonlinear PDEs involving multivariable fields (e.g., Navier–Stokes smoothness) remain among the deepest unsolved problems in mathematics.