The Implicit Function Theorem

Consider the equation x2+y2=1x^2 + y^2 = 1, the unit circle. Near the point (3/5,4/5)(3/5, 4/5), can we solve for yy as a smooth function of xx? Near the point (1,0)(1,0), can we? The equation itself never mentions "solving for yy." It is a single constraint F(x,y)=0F(x,y)=0 treating xx and yy symmetrically. Yet in ordinary life we routinely think of one variable as depending on another. For example, pressure as a function of volume and temperature via an equation of state. Or a budget constraint pinning down one purchase given the others. The Implicit Function Theorem tells us exactly when, and in what sense, an implicit relation F(x,y)=0F(x,y)=0 can be legitimately rewritten as y=g(x)y = g(x) for a smooth function gg. It does this without ever solving the equation explicitly. This question matters because most equations arising in geometry, physics, and optimization are not presented in solved form. Level sets of functions, constraint surfaces in Lagrange Multipliers problems, conservation laws, and equilibrium conditions in economics all come as implicit relations F(x1,…,xn,y)=0F(x_1,\dots,x_n,y)=0. The theorem gives a computable criterion. This criterion is nonvanishing of a partial derivative, or more generally invertibility of a Jacobian block. It guarantees that a local, smooth, unique dependent-variable function exists. It also hands us a formula for that function's derivative, even though we may never write the function down explicitly. The big picture is that the Implicit Function Theorem is really a disguised form of the Inverse Function Theorem. Both theorems answer the same underlying question: when can a map be locally "solved"? They phrase this in two dialects. The Inverse Function Theorem solves y=F(x)y = F(x) for xx globally in all variables. The Implicit Function Theorem solves F(x,y)=0F(x,y)=0 for some of the variables in terms of the rest. We will derive the Implicit Function Theorem directly from the Inverse Function Theorem. This makes the connection between them completely transparent, rather than treating them as two independent black boxes. This article states the precise hypotheses: continuous differentiability, and a nonsingular partial Jacobian. It proves the theorem by reduction to the inverse function case. It derives the implicit-differentiation formula for the derivative of the solved-for variable. It also works through asymmetric examples, including a genuinely vector-valued implicit system. It closes by previewing how the theorem underlies the local parametrization of level sets used in Surface Integrals and Parametrized Surfaces.

Key takeaways
  • The Implicit Function Theorem gives conditions under which F(x,y)=0F(x,y)=0 can be locally solved for yy as a smooth function y=g(x)y=g(x).
  • The key hypothesis is that the partial Jacobian of FF with respect to yy is invertible at the point in question.
  • It is proved by applying the Inverse Function Theorem to an auxiliary map that appends the free variable xx to FF.
  • The derivative of the implicit function is obtained without ever solving for gg explicitly, via Dg=−(DyF)−1DxFDg = -(D_yF)^{-1}D_xF.
  • Failure of the nonsingularity hypothesis (e.g. at (1,0)(1,0) on the unit circle) is exactly where the implicit solution breaks down or becomes non-unique/non-smooth.
  • Key idea: local invertibility of a square sub-block of the Jacobian is precisely the condition that lets an implicit equation be "solved" for the corresponding variables.

Where this lives today

In differential geometry, the Implicit Function Theorem is the tool that certifies a level set {F=0}\{F=0\} of a C1C^1 map with surjective derivative is locally a smooth submanifold, the starting point for defining tangent spaces and, downstream, for Surface Integrals and Parametrized Surfaces, where a surface given implicitly (e.g. a sphere x2+y2+z2=1x^2+y^2+z^2=1) is locally reparametrized as a graph to compute flux or area integrals. In constrained optimization, it justifies the sensitivity analysis behind Lagrange Multipliers: it shows how the optimal point of a constrained problem moves as the constraint or objective is perturbed, by implicitly differentiating the first-order (KKT) conditions with respect to a parameter. In economics, comparative statics — how an equilibrium quantity shifts when a parameter like a tax rate changes — is literally an application of the derivative formula Dg=−(DyF)−1DxFDg = -(D_yF)^{-1}D_xF to equilibrium conditions. In numerical analysis and control theory, implicit ODE solvers and feedback linearization rely on locally solving implicit algebraic equations at each time step, with the theorem certifying well-posedness. In algebraic geometry and singularity theory, points where the Jacobian hypothesis fails are precisely the singular points of a variety, making the theorem a diagnostic for smoothness. In machine learning, implicit differentiation through fixed-point layers (e.g. deep equilibrium models) uses exactly the formula Dg=−(DyF)−1DxFDg=-(D_yF)^{-1}D_xF to backpropagate gradients through an implicitly defined layer without unrolling an iterative solver.

Ask the tutor: How does implicit differentiation of the Lagrange multiplier conditions let us compute how an optimal solution changes when a constraint's constant is perturbed?

DefinitionImplicit Function Theorem — setup and statement

Let U⊆Rn×RmU \subseteq \mathbb{R}^n \times \mathbb{R}^m be open, and let

F:U→RmF : U \to \mathbb{R}^m

be a C1C^1 map. Write points of UU as (x,y)(x,y) with x∈Rnx \in \mathbb{R}^n, y∈Rmy \in \mathbb{R}^m. Let (a,b)∈U(a,b) \in U satisfy

F(a,b)=0.F(a,b) = 0.

Let DyF(a,b)D_yF(a,b) denote the m×mm\times m matrix of partial derivatives of FF with respect to the yy-variables only, evaluated at (a,b)(a,b):

(DyF(a,b))ij=∂Fi∂yj(a,b),1≤i,j≤m.\big(D_yF(a,b)\big)_{ij} = \frac{\partial F_i}{\partial y_j}(a,b), \qquad 1 \le i,j \le m.

The theorem asserts: if DyF(a,b)D_yF(a,b) is invertible, then there exist open sets V∋aV \ni a in Rn\mathbb{R}^n and W∋bW \ni b in Rm\mathbb{R}^m, with V×W⊆UV \times W \subseteq U, and a unique C1C^1 function

g:V→Wg : V \to W

such that g(a)=bg(a) = b and, for every (x,y)∈V×W(x,y)\in V\times W,

F(x,y)=0  ⟺  y=g(x).F(x,y) = 0 \iff y = g(x).

Moreover the derivative of gg is given at every x∈Vx\in V by

Dg(x)=−[DyF(x,g(x))]−1DxF(x,g(x)),Dg(x) = -\big[D_yF(x,g(x))\big]^{-1} D_xF(x,g(x)),

where DxFD_xF is the m×nm\times n matrix of partials of FF with respect to the xx-variables.

Intuition

Think of F(x,y)=0F(x,y)=0 as a system of mm equations in n+mn+m unknowns — generically this cuts out an nn-dimensional solution set, matching the count of free variables xx. The role of yy is to be pinned down by xx; the condition that makes this pinning-down work smoothly and uniquely is that the "yy-part" of the linearization of FF is nonsingular, exactly as in solving a linear system Ay=cAy = c for yy requires AA invertible. The Implicit Function Theorem is the nonlinear, local version of "invert the coefficient matrix of the variables you want to solve for." Geometrically, for the classic case n=m=1n=m=1: F(x,y)=0F(x,y)=0 is a curve in the plane, and DyF(a,b)≠0D_yF(a,b) \ne 0 means the curve is not tangent to the yy-direction at (a,b)(a,b) — equivalently its tangent line is not vertical, so it locally looks like the graph of a function y=g(x)y=g(x). Where the curve does have a vertical tangent, such as at (1,0)(1,0) on the unit circle where ∂F/∂y=2y=0\partial F/\partial y = 2y = 0, no such local function can exist: two values of yy (or none) correspond to nearby xx. The proof strategy is to build a bigger map that keeps xx as a passenger. Define Φ(x,y)=(x,F(x,y))\Phi(x,y) = (x, F(x,y)), a map from U⊆Rn+mU\subseteq\mathbb{R}^{n+m} to Rn+m\mathbb{R}^{n+m}. Its Jacobian at (a,b)(a,b) is block triangular with the identity in the top-left and DyF(a,b)D_yF(a,b) in the bottom-right, so Φ\Phi is a local diffeomorphism there by the Inverse Function Theorem. Inverting Φ\Phi and reading off the yy-component of Φ−1(x,0)\Phi^{-1}(x,0) produces exactly the implicit function g(x)g(x).

Figure
The unit circle: where it is, and is not, a graph over the xx-axis
x y (3/5, 4/5) D_yF ≠ 0: graph y = g(x) (1, 0) D_yF = 0: vertical tangent, fails
TheoremImplicit Function Theorem

Let U⊆Rn+mU \subseteq \mathbb{R}^{n+m} be open, F:U→RmF:U\to\mathbb{R}^m a C1C^1 map, (a,b)∈U(a,b)\in U with F(a,b)=0F(a,b)=0, and suppose the m×mm\times m matrix DyF(a,b)D_yF(a,b) is invertible. Then there exist open sets V⊆RnV\subseteq \mathbb{R}^n containing aa and W⊆RmW\subseteq\mathbb{R}^m containing bb with V×W⊆UV\times W \subseteq U, and a unique function g:V→Wg:V\to W of class C1C^1, such that

g(a)=b,and∀(x,y)∈V×W:  F(x,y)=0  ⟺  y=g(x).g(a) = b, \qquad \text{and} \qquad \forall (x,y) \in V\times W:\; F(x,y)=0 \iff y=g(x).

Furthermore for all x∈Vx \in V,

Dg(x)=−[DyF(x,g(x))]−1DxF(x,g(x)).Dg(x) = -\left[D_yF(x,g(x))\right]^{-1}D_xF(x,g(x)).
Proof

Technique: reduction to the Inverse Function Theorem via an auxiliary map. Step 1 — Construct the auxiliary map. Define Φ:U→Rn×Rm\Phi : U \to \mathbb{R}^n\times\mathbb{R}^m by

Φ(x,y)=(x, F(x,y)).\Phi(x,y) = (x,\, F(x,y)).

Since FF is C1C^1 and the identity map on the xx-slot is C∞C^\infty, Φ\Phi is C1C^1 on UU. Step 2 — Compute and invert the Jacobian at (a,b)(a,b). In block form, writing coordinates (x,y)∈Rn×Rm(x,y)\in\mathbb{R}^n\times\mathbb{R}^m, the derivative of Φ\Phi is

DΦ(x,y)=(In0DxF(x,y)DyF(x,y)),D\Phi(x,y) = \begin{pmatrix} I_n & 0 \\ D_xF(x,y) & D_yF(x,y) \end{pmatrix},

because the first block-row is the identity map's derivative (In, 0)(I_n,\,0), and the second block-row is DF(x,y)=(DxF(x,y)  DyF(x,y))DF(x,y) = (D_xF(x,y)\ \ D_yF(x,y)) by definition of the partial Jacobians. This is a block lower-triangular matrix, so its determinant is

det⁡DΦ(x,y)=det⁡(In)⋅det⁡(DyF(x,y))=det⁡DyF(x,y).\det D\Phi(x,y) = \det(I_n)\cdot \det\big(D_yF(x,y)\big) = \det D_yF(x,y).

At (a,b)(a,b), by hypothesis DyF(a,b)D_yF(a,b) is invertible, so det⁡DΦ(a,b)≠0\det D\Phi(a,b) \ne 0, i.e. DΦ(a,b)D\Phi(a,b) is an invertible linear map. Step 3 — Apply the Inverse Function Theorem. Since Φ\Phi is C1C^1 near (a,b)(a,b) and DΦ(a,b)D\Phi(a,b) is invertible, the Inverse Function Theorem gives open sets Ω∋(a,b)\Omega \ni (a,b) in UU and Ω′∋Φ(a,b)=(a,0)\Omega' \ni \Phi(a,b) = (a,0) in Rn+m\mathbb{R}^{n+m}, such that Φ:Ω→Ω′\Phi:\Omega\to\Omega' is a bijection with C1C^1 inverse Ψ=Φ−1:Ω′→Ω\Psi = \Phi^{-1}:\Omega'\to\Omega. Step 4 — Identify the structure of Ψ\Psi. Since Φ(x,y)=(x,F(x,y))\Phi(x,y) = (x,F(x,y)) has identity first coordinate, for any (x,z)∈Ω′(x,z)\in\Omega' writing Ψ(x,z)=(α(x,z),β(x,z))\Psi(x,z) = (\alpha(x,z), \beta(x,z)) we have

Φ(Ψ(x,z))=(x,z)  ⟹  (α(x,z),F(α(x,z),β(x,z)))=(x,z).\Phi(\Psi(x,z)) = (x,z) \implies (\alpha(x,z), F(\alpha(x,z),\beta(x,z))) = (x,z).

The first coordinate forces α(x,z)=x\alpha(x,z) = x for all (x,z)∈Ω′(x,z) \in \Omega'. So Ψ\Psi has the form

Ψ(x,z)=(x, β(x,z))\Psi(x,z) = (x,\ \beta(x,z))

for some C1C^1 function β:Ω′→Rm\beta:\Omega' \to \mathbb{R}^m, and the second coordinate equation reads

F(x,β(x,z))=zfor all (x,z)∈Ω′.(∗)F(x,\beta(x,z)) = z \qquad \text{for all } (x,z)\in\Omega'. \tag{$\ast$}

Step 5 — Define gg and the neighborhoods. Set g(x):=β(x,0)g(x) := \beta(x,0). Since Ω′\Omega' is open and contains (a,0)(a,0), the set

V0:={x∈Rn:(x,0)∈Ω′}V_0 := \{x \in \mathbb{R}^n : (x,0) \in \Omega'\}

is open (it is the preimage of the open set Ω′\Omega' under the continuous, indeed C1C^1, map x↦(x,0)x \mapsto (x,0)) and contains aa. On V0V_0, gg is C1C^1 as a composition/restriction of the C1C^1 map β\beta. By (∗)(\ast) with z=0z=0:

F(x,g(x))=0for all x∈V0.(∗∗)F(x, g(x)) = 0 \quad \text{for all } x \in V_0. \tag{$\ast\ast$}

Also g(a)=β(a,0)=g(a) = \beta(a,0) = the yy-component of Ψ(a,0)=Ψ(Φ(a,b))=(a,b)\Psi(a,0) = \Psi(\Phi(a,b)) = (a,b), so g(a)=bg(a) = b. Since Ω\Omega is open in U⊆Rn×RmU\subseteq \mathbb{R}^n\times\mathbb{R}^m and contains (a,b)(a,b), choose open boxes V⊆V0V \subseteq V_0 containing aa and W⊆RmW\subseteq \mathbb{R}^m containing bb small enough that V×W⊆ΩV \times W \subseteq \Omega (possible since Ω\Omega is open, using that products of small enough open balls/boxes around (a,b)(a,b) lie in any open set containing (a,b)(a,b)) and additionally so that g(V)⊆Wg(V) \subseteq W (possible by continuity of gg at aa with g(a)=b∈Wg(a)=b\in W, shrinking VV further if necessary). Restrict gg to this VV; by (∗∗)(\ast\ast), g:V→Wg:V\to W satisfies F(x,g(x))=0F(x,g(x))=0 for all x∈Vx\in V. Step 6 — Prove the equivalence F(x,y)=0  ⟺  y=g(x)F(x,y)=0 \iff y=g(x) on V×WV\times W. The direction y=g(x)  ⟹  F(x,y)=0y=g(x) \implies F(x,y)=0 is (∗∗)(\ast\ast) restricted to V×WV\times W. For the converse, suppose (x,y)∈V×W⊆Ω(x,y)\in V\times W \subseteq \Omega and F(x,y)=0F(x,y)=0. Then Φ(x,y)=(x,0)\Phi(x,y) = (x,0). Also Φ(x,g(x))=(x,F(x,g(x)))=(x,0)\Phi(x,g(x)) = (x, F(x,g(x))) = (x,0) by (∗∗)(\ast\ast), and (x,g(x))∈V×W⊆Ω(x,g(x))\in V\times W\subseteq\Omega since g(x)∈Wg(x)\in W. Both (x,y)(x,y) and (x,g(x))(x,g(x)) lie in Ω\Omega and map under the injective map Φ∣Ω\Phi|_\Omega (injective since Φ:Ω→Ω′\Phi:\Omega\to\Omega' is a bijection by Step 3) to the same point (x,0)(x,0). Injectivity forces (x,y)=(x,g(x))(x,y) = (x,g(x)), i.e. y=g(x)y = g(x). Step 7 — Uniqueness of gg. Suppose h:V→Wh:V\to W is any function (not assumed C1C^1 a priori) with h(a)=bh(a)=b and F(x,h(x))=0F(x,h(x))=0 for all x∈Vx \in V. Fix x∈Vx\in V; then (x,h(x))∈V×W(x,h(x))\in V\times W and F(x,h(x))=0F(x,h(x))=0, so by Step 6 (the equivalence just proved) h(x)=g(x)h(x) = g(x). Hence h=gh=g, proving uniqueness among functions V→WV\to W satisfying these implicit relations, in particular uniqueness of a C1C^1 solution. Step 8 — Differentiate to get the derivative formula. Since F(x,g(x))=0F(x,g(x)) = 0 identically for x∈Vx\in V and both FF and gg are C1C^1, the chain rule applies to the composite map x↦F(x,g(x))x \mapsto F(x,g(x)). Differentiating the identity F(x,g(x))≡0F(x,g(x)) \equiv 0 with respect to xx gives, for each x∈Vx \in V,

DxF(x,g(x))+DyF(x,g(x)) Dg(x)=0,D_xF(x,g(x)) + D_yF(x,g(x))\, Dg(x) = 0,

where the left side is the chain rule expansion of D[F(x,g(x))]D[F(x,g(x))] into its dependence through the first slot (DxFD_xF term) and through gg in the second slot (DyF⋅DgD_yF \cdot Dg term). We must justify that DyF(x,g(x))D_yF(x,g(x)) is invertible for xx near aa to solve for Dg(x)Dg(x): the determinant map det⁡∘DyF(x,g(x))\det \circ D_yF(x,g(x)) is continuous in xx (as a composition of the continuous maps x↦(x,g(x))x\mapsto (x,g(x)), DyFD_yF, and det⁡\det), and it is nonzero at x=ax=a by hypothesis; hence by continuity it is nonzero on a neighborhood of aa. Shrinking VV to this neighborhood (intersected with the earlier VV; this does not disturb Steps 1–7 since neighborhoods can always be shrunk while preserving those conclusions), DyF(x,g(x))D_yF(x,g(x)) is invertible for all x∈Vx \in V, and we may solve:

Dg(x)=−[DyF(x,g(x))]−1DxF(x,g(x)).Dg(x) = -\left[D_yF(x,g(x))\right]^{-1}D_xF(x,g(x)).

This completes the proof. ■\blacksquare

∎
CorollaryCkC^k and analytic refinements — stated, proof beyond scope

If in addition FF is CkC^k (k≥1k\ge 1, possibly k=∞k=\infty) or real-analytic on UU, then the implicit function gg produced above is CkC^k, respectively real-analytic, on VV. The CkC^k case follows by bootstrapping the derivative formula Dg=−(DyF)−1DxFDg = -(D_yF)^{-1}D_xF: the right side is a composition of Ck−1C^{k-1} maps (matrix inversion is smooth where defined) once gg is known to be Ck−1C^{k-1}, inductively raising the regularity of gg by one order at each step. The analytic case requires convergent power series majorization arguments and is beyond the present scope.

Worked examples

Example 1 (scalar case, asymmetric constants): the curve x3+y3−3xy=2x^3 + y^3 - 3xy = 2. This is a variant of the folium of Descartes. Let F(x,y)=x3+y3−3xy−2F(x,y) = x^3+y^3-3xy-2. Look for an implicit solution near the point (a,b)=(2,−1)(a,b) = (2,-1): check

F(2,−1)=8+(−1)−3(2)(−1)−2=8−1+6−2=11≠0,F(2,-1) = 8 + (-1) - 3(2)(-1) - 2 = 8 - 1 + 6 - 2 = 11 \ne 0,

so (2,−1)(2,-1) is not on the curve; instead try (a,b)=(1,1)(a,b)=(1,1):

F(1,1)=1+1−3−2=−3≠0.F(1,1) = 1+1-3-2 = -3 \ne 0.

Try (a,b)=(2,0)(a,b) = (2,0): F(2,0)=8+0−0−2=6≠0F(2,0) = 8+0-0-2 = 6\ne 0. We want a genuine root; solve directly: try y=1,x=?y=1,x=? satisfying x3+1−3x−2=0⇒x3−3x−1=0x^3+1-3x-2=0 \Rightarrow x^3-3x-1=0. Instead pick a clean root: set x=−1x=-1: F(−1,y)=−1+y3+3y−2=y3+3y−3F(-1,y) = -1+y^3+3y-2 = y^3+3y-3; at y=1y=1: 1+3−3=1≠01+3-3=1\ne0. Try the cleanest asymmetric root by inspection: x=0x=0: F(0,y)=y3−2=0⇒y=21/3F(0,y) = y^3-2=0 \Rightarrow y = 2^{1/3}. Take (a,b)=(0,21/3)(a,b) = (0, 2^{1/3}). Check the hypothesis: ∂F/∂y=3y2−3x\partial F/\partial y = 3y^2 - 3x, so at (0,21/3)(0,2^{1/3}),

DyF(a,b)=3(21/3)2−0=3⋅22/3≠0.D_yF(a,b) = 3(2^{1/3})^2 - 0 = 3\cdot 2^{2/3} \ne 0.

This is nonzero, so the Implicit Function Theorem applies: there is an open interval VV around x=0x=0 and a unique C1C^1 (in fact C∞C^\infty, since FF is a polynomial) function gg with g(0)=21/3g(0) = 2^{1/3} and F(x,g(x))=0F(x,g(x)) = 0 near x=0x=0. Compute Dg(0)Dg(0) using the formula, with DxF=∂F/∂x=3x2−3yD_xF = \partial F/\partial x = 3x^2 - 3y:

DxF(0,21/3)=0−3⋅21/3=−3⋅21/3.D_xF(0,2^{1/3}) = 0 - 3\cdot 2^{1/3} = -3\cdot 2^{1/3}.

Then

g′(0)=−DxF(0,21/3)DyF(0,21/3)=−−3⋅21/33⋅22/3=21/322/3=21/3−2/3=2−1/3.g'(0) = -\frac{D_xF(0,2^{1/3})}{D_yF(0,2^{1/3})} = -\frac{-3\cdot 2^{1/3}}{3\cdot 2^{2/3}} = \frac{2^{1/3}}{2^{2/3}} = 2^{1/3-2/3} = 2^{-1/3}.

So near x=0x=0, y≈21/3+2−1/3xy \approx 2^{1/3} + 2^{-1/3}x, a genuine local linear approximation to the curve obtained without solving the cubic x3+y3−3xy−2=0x^3+y^3-3xy-2=0 for yy in closed form.

Example 2 (vector-valued implicit system, n=1n=1, m=2m=2): solving two equations for two unknowns in terms of one parameter. Let x∈Rx\in\mathbb{R} be the free variable and y=(y1,y2)∈R2y=(y_1,y_2)\in\mathbb{R}^2 the dependent pair. Define F:R3→R2F:\mathbb{R}^3\to\mathbb{R}^2 by

F(x,y1,y2)=(y12+y22−x−5,  y1y2−x2+2x−3).F(x,y_1,y_2) = \big(y_1^2 + y_2^2 - x - 5,\ \ y_1 y_2 - x^2 + 2x - 3\big).

We seek a base point. Try x=1x=1: the system becomes

y12+y22=6,y1y2=12−2(1)+3=2.y_1^2+y_2^2 = 6, \qquad y_1y_2 = 1^2 - 2(1) + 3 = 2.

(Note the sign: y1y2−x2+2x−3=0⇒y1y2=x2−2x+3y_1y_2 - x^2+2x-3=0 \Rightarrow y_1y_2 = x^2-2x+3; at x=1x=1 this is 1−2+3=21-2+3=2.) Look for y1,y2y_1,y_2 with sum of squares 66 and product 22: these are roots of t2−st+2=0t^2 - st + 2=0 paired with y12+y22=s2−4=6⇒s2=10y_1^2+y_2^2 = s^2-4=6 \Rightarrow s^2=10, not integral. Choose instead y1=2, y2=1y_1=2,\ y_2=1: then y12+y22=5y_1^2+y_2^2=5, y1y2=2y_1y_2=2. Solve for xx: need y12+y22−x−5=0⇒5−x−5=0⇒x=0y_1^2+y_2^2-x-5=0 \Rightarrow 5-x-5=0\Rightarrow x=0, and check the second equation at x=0x=0: y1y2−x2+2x−3=2−0+0−3=−1≠0y_1y_2-x^2+2x-3 = 2-0+0-3=-1\ne0. Not consistent. Instead, fix (y1,y2)=(2,1)(y_1,y_2)=(2,1) and choose FF's constant so it is consistent — but to keep FF fixed as stated, find an actual common root. Solve directly: from equation 1, x=y12+y22−5x = y_1^2+y_2^2-5. Substitute into equation 2:

y1y2−(y12+y22−5)2+2(y12+y22−5)−3=0.y_1y_2 - (y_1^2+y_2^2-5)^2 + 2(y_1^2+y_2^2-5) - 3 = 0.

Try y1=2,y2=0y_1=2, y_2=0: then x=4+0−5=−1x = 4+0-5=-1. Check eq. 2: y1y2−x2+2x−3=0−1−2−3=−6≠0y_1y_2 - x^2+2x-3 = 0 - 1 -2-3 = -6 \ne 0. Try y1=1,y2=2y_1=1,y_2=2 (asymmetric swap): x=1+4−5=0x = 1+4-5=0. Eq. 2 at x=0x=0: y1y2−0+0−3=2−3=−1≠0y_1y_2 - 0+0-3 = 2-3=-1\ne 0. Try y1=3,y2=−1y_1=3,y_2=-1: x=9+1−5=5x=9+1-5=5. Eq 2: y1y2−x2+2x−3=−3−25+10−3=−21≠0y_1y_2-x^2+2x-3 = -3-25+10-3=-21\ne0. Rather than hunt further, use a cleaner but still asymmetric FF built to have an exact, checkable root. Redefine:

F(x,y1,y2)=(y12−y2−x,  y1+y22−2x−1).F(x,y_1,y_2) = \big(y_1^2 - y_2 - x,\ \ y_1 + y_2^2 - 2x - 1\big).

Try the point x=1, y1=2, y2=3x=1,\ y_1=2,\ y_2=3:

F1=22−3−1=4−3−1=0,F2=2+32−2(1)−1=2+9−2−1=8≠0.F_1 = 2^2 - 3 - 1 = 4-3-1=0, \qquad F_2 = 2 + 3^2 - 2(1) - 1 = 2+9-2-1 = 8 \ne 0.

Try y1=2,y2=2,x=2y_1=2,y_2=2,x=2: F1=4−2−2=0F_1 = 4-2-2=0; F2=2+4−4−1=1≠0F_2 = 2+4-4-1=1\ne0. Try y1=1,y2=0,x=1y_1=1,y_2=0,x=1: F1=1−0−1=0F_1=1-0-1=0; F2=1+0−2−1=−2≠0F_2=1+0-2-1=-2\ne0. Try y1=1,y2=1,x=0y_1=1,y_2=1,x=0: F1=1−1−0=0F_1=1-1-0=0; F2=1+1−0−1=1≠0F_2=1+1-0-1=1\ne0. Try y1=2,y2=1,x=3y_1=2,y_2=1,x=3: F1=4−1−3=0F_1=4-1-3=0; F2=2+1−6−1=−4≠0F_2=2+1-6-1=-4\ne0. Try y1=2,y2=4,x=0y_1=2,y_2=4,x=0: F1=4−4−0=0F_1=4-4-0=0; F2=2+16−0−1=17≠0F_2=2+16-0-1=17\ne0. Solve systematically instead: from F1=0F_1=0, x=y12−y2x = y_1^2-y_2. Substitute into F2=0F_2=0:

y1+y22−2(y12−y2)−1=0  ⟺  y1+y22−2y12+2y2−1=0.y_1+y_2^2-2(y_1^2-y_2)-1=0 \iff y_1+y_2^2-2y_1^2+2y_2-1=0.

Pick y1=1y_1=1: 1+y22−2+2y2−1=0  ⟺  y22+2y2−2=0  ⟺  y2=−1±31+y_2^2-2+2y_2-1=0 \iff y_2^2+2y_2-2=0 \iff y_2 = -1\pm\sqrt{3}. Take y2=−1+3y_2 = -1+\sqrt3 (irrational but exact and checkable). Then x=y12−y2=1−(−1+3)=2−3x = y_1^2-y_2 = 1-(-1+\sqrt3) = 2-\sqrt3. So (a,b)=(a,(b1,b2))=(2−3, (1, −1+3))(a,b) = (a,(b_1,b_2)) = \big(2-\sqrt3,\ (1,\ -1+\sqrt3)\big) is an exact root: F1=1−(−1+3)−(2−3)=1+1−3−2+3=0F_1 = 1-(-1+\sqrt3)-(2-\sqrt3) = 1+1-\sqrt3-2+\sqrt3=0 ✓, and F2=1+(−1+3)2−2(2−3)−1F_2 = 1+(-1+\sqrt3)^2-2(2-\sqrt3)-1. Compute (−1+3)2=1−23+3=4−23(-1+\sqrt3)^2 = 1-2\sqrt3+3=4-2\sqrt3. So F2=1+4−23−4+23−1=0F_2 = 1+4-2\sqrt3-4+2\sqrt3-1 = 0 ✓. Now compute the Jacobian blocks. We have

DF=(∂F1/∂x∂F1/∂y1∂F1/∂y2∂F2/∂x∂F2/∂y1∂F2/∂y2)=(−12y1−1−212y2).DF = \begin{pmatrix} \partial F_1/\partial x & \partial F_1/\partial y_1 & \partial F_1/\partial y_2 \\ \partial F_2/\partial x & \partial F_2/\partial y_1 & \partial F_2/\partial y_2 \end{pmatrix} = \begin{pmatrix} -1 & 2y_1 & -1 \\ -2 & 1 & 2y_2 \end{pmatrix}.

At (y1,y2)=(1, −1+3)(y_1,y_2) = (1,\,-1+\sqrt3):

DyF=(2y1−112y2)=(2−112(−1+3))=(2−11−2+23).D_yF = \begin{pmatrix} 2y_1 & -1 \\ 1 & 2y_2 \end{pmatrix} = \begin{pmatrix} 2 & -1 \\ 1 & 2(-1+\sqrt3) \end{pmatrix} = \begin{pmatrix} 2 & -1 \\ 1 & -2+2\sqrt3 \end{pmatrix}.

Its determinant is

det⁡DyF=2(−2+23)−(−1)(1)=−4+43+1=−3+43≈−3+6.928=3.928≠0.\det D_yF = 2(-2+2\sqrt3) - (-1)(1) = -4+4\sqrt3+1 = -3+4\sqrt3 \approx -3+6.928 = 3.928 \ne 0.

So DyFD_yF is invertible: the Implicit Function Theorem applies, giving a C1C^1 (indeed C∞C^\infty, since FF is polynomial) function g(x)=(g1(x),g2(x))g(x) = (g_1(x),g_2(x)) near x=2−3x=2-\sqrt3 with g(2−3)=(1, −1+3)g(2-\sqrt3) = (1,\,-1+\sqrt3) and F(x,g(x))=0F(x,g(x))=0. Compute DgDg at this point using Dg=−(DyF)−1DxFDg = -(D_yF)^{-1}D_xF, where DxF=(−1−2)D_xF = \begin{pmatrix}-1\\-2\end{pmatrix}. The inverse of DyFD_yF is

(DyF)−1=1−3+43(−2+231−12).(D_yF)^{-1} = \frac{1}{-3+4\sqrt3}\begin{pmatrix} -2+2\sqrt3 & 1 \\ -1 & 2 \end{pmatrix}.

So

Dg=−1−3+43(−2+231−12)(−1−2)=−1−3+43((2−23)−21−4)=−1−3+43(−23−3).Dg = -\frac{1}{-3+4\sqrt3}\begin{pmatrix} -2+2\sqrt3 & 1 \\ -1 & 2 \end{pmatrix}\begin{pmatrix}-1\\-2\end{pmatrix} = -\frac{1}{-3+4\sqrt3}\begin{pmatrix} (2-2\sqrt3) -2 \\ 1-4 \end{pmatrix} = -\frac{1}{-3+4\sqrt3}\begin{pmatrix} -2\sqrt3 \\ -3 \end{pmatrix}.

Numerically, −3+43≈3.928-3+4\sqrt3 \approx 3.928, so

g1′(x)≈233.928≈3.4643.928≈0.882,g2′(x)≈33.928≈0.764,g_1'(x) \approx \frac{2\sqrt3}{3.928} \approx \frac{3.464}{3.928} \approx 0.882, \qquad g_2'(x) \approx \frac{3}{3.928} \approx 0.764,

both evaluated at x=2−3≈0.268x = 2-\sqrt3 \approx 0.268. This gives the exact local linear behavior of (y1,y2)(y_1,y_2) as functions of xx near this point, obtained entirely from Jacobian data without ever inverting the nonlinear system symbolically.

CounterexampleDropping invertibility of DyFD_yF: the unit circle at (1,0)(1,0)

Let F(x,y)=x2+y2−1F(x,y) = x^2+y^2-1, so m=n=1m=n=1, and consider the point (a,b)=(1,0)(a,b) = (1,0), which satisfies F(1,0)=1+0−1=0F(1,0) = 1+0-1=0. Here

DyF(1,0)=∂F∂y(1,0)=2y∣(1,0)=0,D_yF(1,0) = \frac{\partial F}{\partial y}(1,0) = 2y\Big|_{(1,0)} = 0,

so the key hypothesis of the theorem — invertibility (here, nonvanishing) of DyF(a,b)D_yF(a,b) — fails. Consequently no local C1C^1 function gg can exist with F(x,g(x))=0F(x,g(x))=0 near x=1x=1 and g(1)=0g(1)=0: solving x2+y2=1x^2+y^2=1 for yy near x=1x=1 gives y=±1−x2y = \pm\sqrt{1-x^2}, which is only defined for x≤1x \le 1 (undefined for x>1x>1, since 1−x2<01-x^2<0 there), so no function is even defined on a full open interval around x=1x=1. Even restricting to x≤1x\le 1, y=1−x2y=\sqrt{1-x^2} has lim⁡x→1−g′(x)=lim⁡x→1−−x1−x2=−∞\lim_{x\to 1^-} g'(x) = \lim_{x\to1^-}\frac{-x}{\sqrt{1-x^2}} = -\infty, so gg fails to be differentiable at x=1x=1 from the relevant side, let alone C1C^1 on an open neighborhood. This shows the nonvanishing hypothesis is not a technical convenience but is exactly what fails at points where the level curve has a vertical tangent — precisely where solving for yy as a function of xx breaks down. Contrast: DxF(1,0)=2x∣(1,0)=2≠0D_xF(1,0) = 2x|_{(1,0)} = 2 \ne 0, so instead xx can be solved smoothly as a function of yy near (1,0)(1,0) (x=1−y2x = \sqrt{1-y^2}), illustrating that the theorem's hypothesis singles out which variable may be solved for, not whether the level set is smooth at all.

Scope, edge cases & common traps

SituationWhat happens
DyF(a,b)D_yF(a,b) invertibleTheorem applies fully: unique local C1C^1 solution y=g(x)y=g(x), with derivative formula.
DyF(a,b)D_yF(a,b) singular but DxF(a,b)D_xF(a,b) (or some other m×mm\times m block) invertibleSolve for a different set of mm variables instead; the roles of "independent" and "dependent" variables are not intrinsic to FF, only to the chosen partition.
FF only continuous, not C1C^1Theorem can fail entirely: continuity of FF alone does not guarantee even a continuous implicit function exists locally in general non-degenerate settings; C1C^1 (or at least differentiability with continuity of the relevant partials) is essential to the proof via the Inverse Function Theorem.
Whole Jacobian DF(a,b)DF(a,b) has rank <m<mNo m×mm\times m sub-block is invertible; the level set F=0F=0 may fail to be a manifold at all near (a,b)(a,b) (e.g. a self-intersection or cusp).
n=0n=0 (F:Rm→RmF:\mathbb{R}^m\to\mathbb{R}^m)Degenerates exactly to the Inverse Function Theorem: there is no xx, and "solving F(y)=0F(y)=0" becomes finding the local inverse of FF near a point where F(b)=0F(b)=0 under DF(b)DF(b) invertible.

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered
AC
Augustin-Louis Cauchy
1789-1857 · Foundational rigor in 19th-century analysis

Cauchy is the mathematician most closely associated with turning implicit-function reasoning, long used informally, into a rigorous existence theorem grounded in careful estimates, as part of his broader program to place calculus on a foundation of precise limits and inequalities rather than infinitesimal intuition. The modern Banach-space and Jacobian-based formulation used today, tying the Implicit Function Theorem directly to the Inverse Function Theorem, was consolidated later, but the core existence argument — controlling a solution branch by dominating series or fixed-point estimates — traces to Cauchy's analytic methods and was extended by his contemporaries and successors working on differential equations and complex analysis.

History
- 1821 — Cauchy's Cours d'Analyse establishes the rigorous limit-based framework for continuity and differentiability that implicit-function arguments require.
- 1831 — Cauchy develops majorant-series techniques while studying implicit solutions of analytic equations, precursors to the analytic Implicit Function Theorem.
- 1877 — Ulisse Dini gives a systematic real-variable treatment of implicit functions of several variables, the version most textbooks now call "Dini's theorem" in this context.
- 1908 — Édouard Goursat's Cours d'Analyse Mathématique presents the theorem via successive approximations in the now-standard multivariable form.
- 1950s — Formulation via Banach space fixed-point/contraction arguments and explicit reduction to the Inverse Function Theorem becomes the standard modern presentation in analysis textbooks.
Why is the theorem called "implicit" rather than just an existence theorem for functions?

The name reflects that the function gg is never constructed by an explicit formula — it is defined only implicitly, through the relation F(x,g(x))=0F(x,g(x))=0. Unlike explicit function definitions such as g(x)=1−x2g(x) = \sqrt{1-x^2}, the theorem provides no closed form in general; for a degree-5 polynomial equation in two variables, for instance, no algebraic closed form for gg may even exist by Galois-theoretic obstructions (unsolvability of the general quintic), yet the theorem still guarantees gg exists locally, is C1C^1, and has a computable derivative. This is the theorem's real power: existence and differentiability of gg are established without needing an explicit formula for gg itself.

How does this theorem relate to the Inverse Function Theorem exactly?

They are logically equivalent in the sense that each can be derived from the other. This article derives the Implicit Function Theorem from the Inverse Function Theorem by considering the augmented map Φ(x,y)=(x,F(x,y))\Phi(x,y)=(x,F(x,y)) and inverting it. Conversely, the Inverse Function Theorem is the special case n=0n=0 of the Implicit Function Theorem: given F(y)−z=0F(y) - z = 0 with F:Rm→RmF:\mathbb{R}^m\to\mathbb{R}^m and parameter zz playing the role of xx, solving for y=g(z)y=g(z) as a function of zz is precisely constructing the local inverse of FF.

Does the theorem tell us anything about second derivatives of the implicit function?

Yes, by further implicit differentiation. If FF is C2C^2, differentiate the identity DxF(x,g(x))+DyF(x,g(x))Dg(x)=0D_xF(x,g(x)) + D_yF(x,g(x))Dg(x) = 0 once more with respect to xx, using the product and chain rules, to obtain a linear equation for the second derivative (Hessian) of gg in terms of the second partials of FF and the already-known Dg(x)Dg(x). This is exactly how implicit differentiation is extended to compute concavity/convexity of implicitly defined curves, e.g. in constrained optimization sensitivity analysis tied to Lagrange Multipliers.

Can the theorem fail even when FF is a polynomial and looks perfectly smooth?

Yes — smoothness of FF is necessary but not sufficient; the Jacobian condition can still fail at special points. For example F(x,y)=y2−x3F(x,y) = y^2 - x^3 (a cuspidal cubic) is a polynomial, hence C∞C^\infty, but at (0,0)(0,0), DyF(0,0)=2y∣(0,0)=0D_yF(0,0) = 2y|_{(0,0)} = 0 and also DxF(0,0)=−3x2∣(0,0)=0D_xF(0,0) = -3x^2|_{(0,0)}=0, so neither variable can be locally solved for smoothly there, matching the fact that the curve has an actual cusp singularity at the origin.

Why must we shrink the neighborhood VV in the proof, rather than using the full domain UU?

The Inverse Function Theorem itself only guarantees a local diffeomorphism, valid on some possibly small open set around (a,b)(a,b) — invertibility of DΦ(a,b)D\Phi(a,b) says nothing about behavior far away, where Φ\Phi could fail to be injective or its Jacobian could become singular again. Since the whole proof is built by pulling back through this local inverse, the resulting implicit function gg inherits the same local-only validity. This is why the theorem's conclusion is always phrased with "there exist open sets V,WV,W" rather than any global domain.

What goes wrong if DyFD_yF is invertible but FF is only continuous, not differentiable?

The entire proof strategy collapses because it relies on differentiating Φ\Phi and invoking the Inverse Function Theorem, both of which require at least C1C^1 regularity. Without differentiability there is no linearization to test for invertibility, so the hypothesis "DyF(a,b)D_yF(a,b) invertible" cannot even be stated. Purely continuous implicit relations can behave pathologically — failing to have any continuous local branch, or having infinitely many — so C1C^1 (or comparable) smoothness is not a technical nicety but load-bearing throughout the argument.

Is the implicit function gg ever unique globally, not just on the small set V×WV\times W?

Not in general. Uniqueness in the theorem is only asserted within the specific box V×WV \times W constructed in the proof; outside that box, or even inside a larger neighborhood of aa, there can be other branches of solutions to F(x,y)=0F(x,y)=0. The unit circle again illustrates this: near (3/5,4/5)(3/5,4/5) the unique local branch in a small box is y=1−x2y=\sqrt{1-x^2}, but globally there is also the branch y=−1−x2y=-\sqrt{1-x^2} solving the same equation, just not passing through (3/5,4/5)(3/5,4/5) or lying in the same small WW.

Before this