# Implicit function theorem and its multivariate generalization

Originally published at 狗和留美者不得入内. You can comment here or there.

The implicit function theorem for a single output variable can be stated as follows:

Single equation implicit function theorem. Let $F(\mathbf{x}, y)$ be a function of class $C^1$ on some neighborhood of a point $(\mathbf{a}, b) \in \mathbb{R}^{n+1}$. Suppose that $F(\mathbf{a}, b) = 0$ and $\partial_y F(\mathbf{a}, b) \neq 0$. Then there exist positive numbers $r_0, r_1$ such that the following conclusions are valid.

a. For each $\mathbf{x}$ in the ball $|\mathbf{x} - \mathbf{a}| < r_0$ there is a unique $y$ such that $|y - b| < r_1$ and $F(\mathbf{x}, y) = 0$. We denote this $y$ by $f(\mathbf{x})$; in particular, $f(\mathbf{a}) = b$.

b. The function $f$ thus defined for $|\mathbf{x} - \mathbf{a}| < r_0$ is of class $C^1$, and its partial derivatives are given by $\partial_j f(\mathbf{x}) = -\frac{\partial_j F(\mathbf{x}, f(\mathbf{x}))}{\partial_y F(\mathbf{x}, f(\mathbf{x}))}$.

Proof. For part (a), assume without loss of generality positive $\partial_y F(\mathbf{a}, b)$. By continuity of that partial derivative, we have that in some neighborhood of $(\mathbf{a}, b)$ it is positive and thus for some $r_1 > 0, r_0 > 0$ there exists $f$ such that $|\mathbf{x} - \mathbf{a}| < r_0$ implies that there exists a unique $y$ (by intermediate value theorem along with positivity of $\partial_y F$) such that $|y - b| < r_1$ with $F(\mathbf{x}, y) = 0$, which defines some function $y = f(\mathbf{x})$.To show that $f$ has partial derivatives, we must first show that it is continuous. To do so, we can let $r_1$ be our $\epsilon$ and use the same process to arrive at our $\delta$, which corresponds to $r_0$.

For part (b), to show that its partial derivatives exist and are equal to what we desire, we perturb $\mathbf{x}$ with an $\mathbf{h}$ that we let WLOG be $\mathbf{h} = (h, 0, \ldots, 0)$.

Then with $k = f(\mathbf{x}+\mathbf{h}) - f(\mathbf{x})$, we have $F(\mathbf{x} + \mathbf{h}, y+k) = F(\mathbf{x}, y) = 0$. From the mean value theorem, we can arrive at $0 = h\partial_1F(\mathbf{x}+t\mathbf{h}, y + tk) + k\partial_y F(\mathbf{x}+t\mathbf{h}, y+tk)$

for some $t \in (0,1)$. Rearranging and taking $h \to 0$ gives us $\partial_j f(\mathbf{x}) = -\frac{\partial_j F(\mathbf{x}, y)}{\partial_y F(\mathbf{x}, y)}$.

The following can be generalized to multiple variables, with $k$ implicit functions and $k$ constraints.     ▢

Implicit function theorem for systems of equations. Let $\mathbf{F}(\mathbf{x}, \mathbf{y})$ be an $\mathbb{R}^k$ valued functions of class $C^1$ on some neighborhood of a point $\mathbf{F}(\mathbf{a}, \mathbf{b}) \in \mathbb{R}^{n+k}$ and let $B_{ij} = (\partial F_i / \partial y_j)(\mathbf{a}, \mathbf{b})$. Suppose that $\mathbf{F}(\mathbf{x}, \mathbf{y}) = \mathbf{0}$ and $\det B \neq 0$. Then there exist positive numbers $r_0, r_1$ such that the following conclusions are valid.

a. For each $\mathbf{x}$ in the ball $|\mathbf{x} - \mathbf{a}| < r_0$ there is a unique $\mathbf{y}$ such that $|\mathbf{y} - \mathbf{b}| < r_1$ and $\mathbf{F}(\mathbf{x}, \mathbf{y}) = 0$. We denote this $\mathbf{y}$ by $\mathbf{f}(\mathbf{x})$; in particular, $\mathbf{f}(\mathbf{a}) = \mathbf{b}$.

b. The function $\mathbf{f}$ thus defined for $|\mathbf{x} - \mathbf{a}| < r_0$ is of class $C^1$, and its partial derivatives $\partial_j \mathbf{f}$ can be computed by differentiating the equations $\mathbf{F}(\mathbf{x}, \mathbf{f}(\mathbf{x})) = \mathbf{0}$ with respect to $x_j$ and solving the resulting linear system of equations for $\partial_j f_1, \ldots, \partial_j f_k$.

Proof: For this we will be using Cramer’s rule, which is that one can solve a linear system $Ax = y$ (provided of course that $A$ is non-singular) by taking matrix obtained from substituting the $k$th column of $A$ with $y$ and letting $x_k$ be the determinant of that matrix divided by the determinant of $A$.

From this, we are somewhat hinted that induction is in order. If $B$ is invertible, then one of its $k-1 \times k-1$ submatrices is invertible. Assume WLOG that such applies to the one determined by $B^{kk}$. With this in mind, we can via our inductive hypothesis have $F_1(\mathbf{x}, \mathbf{y}) = F_2(\mathbf{x}, \mathbf{y}) = \cdots = F_{k-1}(\mathbf{x}, \mathbf{y}) = 0$

determine $y_j = g_j(\mathbf{x}, y_k)$ for $j = 1,2,\ldots,k-1$. Here we are making $y_k$ an independent variable and we can totally do that because we are inducting on the number of outputs (and also constraints). Substituting this into the $F_k$ constraint, this reduces to the single variable case, with $G(\mathbf{x}, y_k) = F_k(\mathbf{x}, \mathbf{g}(\mathbf{x}, y_k), y_k) = 0$.

It suffices now to show via our $\det B \neq 0$ hypothesis that $\frac{\partial G}{\partial y_k} \neq 0$. Routine application of the chain rule gives $\frac{\partial G}{\partial y_k} = \displaystyle\sum_{j=1}^{k-1} \frac{\partial F_k}{\partial y_j} \frac{\partial g_j}{\partial y_k} + \frac{\partial F_k}{\partial y_k} = \displaystyle\sum_{j=1}^{k-1} B^{kj} \frac{\partial g_j}{\partial y_k} + B^{kk}. \ \ \ \ (1)$

The $\frac{\partial g_j}{\partial y_k}$s are the solution to the following linear system: $\begin{pmatrix} \frac{\partial F_1}{\partial y_1} & \dots & \frac{\partial F_1}{\partial y_{k-1}} \\ \; & \ddots \; \\ \frac{\partial F_{k-1}}{\partial y_1} & \dots & \frac{\partial F_{k-1}}{\partial y_{k-1}} \end{pmatrix} \begin{pmatrix} \frac{\partial g_1}{\partial y_k} \\ \vdots \\ \frac{\partial g_{k-1}}{\partial y_k} \end{pmatrix} = \begin{pmatrix} \frac{-\partial F_1}{\partial y_k} \\ \vdots \\ \frac{-\partial F_{k-1}}{\partial y_k} \end{pmatrix}$.

Let $M^{ij}$ denote the $k-1 \times k-1$ submatrix induced by $B_{ij}$. We see then that in the replacement for Cramer’s rule, we arrive at what is $M^{kj}$ but with the last column swapped to the left $k-j-1$ times such that it lands in the $j$th column and also with a negative sign, which means $\frac{\partial g_j}{\partial y_k}(\mathbf{a}, b_k) = (-1)^{k-j} \frac{\det M^{jk}}{\det M^{kk}}$.

Now, we substitute this into $(1)$ to get \begin{aligned}\frac{\partial G}{\partial y_k}(\mathbf{a}, b_k) &= \displaystyle_{j=1}^{k-1} (-1)^{k-j}B_{kj}\frac{\det M^{kj}}{\det M^{kk}} + B_kk \\ &= \frac{\sum_{j=1}^k (-1)^{j+k} B_{kj}\det M^{kj}}{\det M^{kk}} \\ &= \frac{\det B}{\det M^{kk}} \\ &\neq 0. \end{aligned}

Finally, we apply the implicit function theorem for one variable for the $y_k$ that remains.     ▢

References

• Gerald B. Folland, Advanced Calculus, Prentice Hall, Upper Saddle River, NJ, 2002, pp. 114–116, 420–422.

Tags:
• #### Why the 1700 Japanese could have discovered calculus and modern science by 2200 or 2700

Originally published at 狗和留美者不得入内. You can comment here or there. Some thoughts on Greek, Persian, Arab, Indian, and Chinese science In the…

• #### 知乎被留美王八蛋把控了，连我的内容为数学推导的评论都得删

没错，我不小心用了“傻逼”这个词，可是内容基本完全是数学。因为违规，又被禁了一天

• #### Selberg on Ramanujan

昨晚，我想起了大挪威数学家 Atle Selberg。他是个很有趣的人，好像在 1943 年挪威被纳粹德国占领时获得了奥斯陆大学的博士。好像他还参军了，也坐牢了。对于他的工作我知道他和 Erdos…

• Post a new comment

#### Error

Anonymous comments are disabled in this journal

default userpic