17  Chain rule

Ordinary chain rule

TODO

Multi-variable chain rule

The multi-variable chain rule is used to calculate the derivative of the output of a composite function with respect to the input where the inside function of the composite function is a vector-valued function.

Base case: scalar-valued univariate function c (x) = f (\mathbf{g} (x))

Given a scalar-valued univariate function c (x) = f (\mathbf{g} (x)) that is a composite function of a scalar-valued multivariate function f: \mathbb{R}^{m} \mapsto \mathbb{R} and a vector-valued univariate function \mathbf{g}: \mathbb{R} \mapsto \mathbb{R}^{m} that takes a real-value input, the chain rule for calculating the derivative of c (x) at point x is

\frac{ \mathop{d} c }{ \mathop{d} x } (x) = \sum_{i=1}^{m} \frac{ \partial f }{ \partial g_{i} (x) } (\mathbf{g} (x)) \frac{ \mathop{d} g_{i} }{ \mathop{d} x } (x)

where

• \frac{ \partial f }{ \partial g_{i} (x) } (\mathbf{g} (x)) is the partial derivative of the function f with respect of the dimension i at the point \mathbf{g} (x),

• \frac{ \mathop{d} g_{i} }{ \mathop{d} x } (x) is the derivative of function g_{i} at the point x.

Therefore, the derivative of c (x) is a real number that is calculated using the chain rule as the dot product of the gradient of f (\mathbf{g} (x)) at point \mathbf{g} (x) and the vector of the derivatives of g_{i} (x) at point x

\begin{aligned} \frac{ \mathop{d} c }{ \mathop{d} x } (x) & = \sum_{i=1}^{n} \frac{\partial f}{\partial g_{i} (x)} (\mathbf{g} (x)) \frac{ \mathop{d} g_{i} }{ \mathop{d} x } (x) \\ & = \begin{bmatrix} \frac{\partial f}{\partial g_{1} (x)} (\mathbf{g} (x)) \\ \vdots \\ \frac{\partial f}{\partial g_{m} (x)} (\mathbf{g} (x)) \\ \end{bmatrix}^{T} \begin{bmatrix} \frac{ \mathop{d} g_{1} }{ \mathop{d} x } (x) \\ \vdots \\ \frac{ \mathop{d} g_{m} }{ \mathop{d} x } (x) \\ \end{bmatrix} \\ & = \nabla^{T} f (\mathbf{g} (x)) \frac{ \mathop{d} \mathbf{g} }{ \mathop{d} x } (x). \end{aligned}

Extension 1: scalar-valued multivariate function c (x) = f (\mathbf{g} (\mathbf{x}))

When the input is a vector, the composite function c (\mathbf{x}) = f (\mathbf{g} (\mathbf{x})) is a scalar-valued multivariate function, whose gradient is calculated by applying the multi-variable chain rule to every element of the input vector

\begin{aligned} \nabla^{T} c (\mathbf{x}) & = \begin{bmatrix} \frac{ \partial c }{ \partial x_{1} } (\mathbf{x}) & \dots & \frac{ \partial c }{ \partial x_{n} } (\mathbf{x}) \\ \end{bmatrix} \\ & = \begin{bmatrix} \frac{ \mathop{d} c |_{\mathbf{x}} }{ \mathop{d} x_{1} } (x_{1}) & \dots & \frac{ \mathop{d} c |_{\mathbf{x}} }{ \mathop{d} x_{n} } (x_{n}) \\ \end{bmatrix} \\ & = \nabla^{T} f (\mathbf{g} (\mathbf{x})) \begin{bmatrix} | & & | \\ \frac{ \mathop{d} \mathbf{g} }{ \mathop{d} x_{1} } (x_{1}) & \dots & \frac{ \mathop{d} \mathbf{g} }{ \mathop{d} x_{n} } (x_{n}) \\ | & & | \\ \end{bmatrix} \\ & = \nabla^{T} f (\mathbf{g} (\mathbf{x})) \begin{bmatrix} - & \nabla^{T} g_{1} (\mathbf{x}) & - \\ & \vdots &\\ - & \nabla^{T} g_{m} (\mathbf{x}) & - \\ \end{bmatrix} \\ & = \nabla^{T} f (\mathbf{g} (\mathbf{x})) \mathbf{J} \mathbf{g} (\mathbf{x}). \\ \end{aligned}

Therefore, the gradient is a vector

\begin{aligned} \nabla c (\mathbf{x}) = \mathbf{J}^{T} \mathbf{g} (\mathbf{x}) \nabla f (\mathbf{g} (\mathbf{x})). \end{aligned}

Extension 2: vector-valued univariate function \mathbf{c} (x) = \mathbf{f} (\mathbf{g} (x))

When the output is a vector, the composite function \mathbf{c} (x) = \mathbf{f} (\mathbf{g} (x)) is a vector-valued univariate function, whose derivative is a vector

\begin{aligned} \frac{ \mathop{d} \mathbf{c} }{ \mathop{d} x } (x) & = \begin{bmatrix} \frac{ \mathop{d} c_{1} }{ \mathop{d} x } (x) \\ \vdots \\ \frac{ \mathop{d} c_{n} }{ \mathop{d} x } (x) \\ \end{bmatrix} \\ & = \begin{bmatrix} - & \nabla^{T} f_{1} (\mathbf{g} (x)) & - \\ & \vdots & \\ - & \nabla^{T} f_{n} (\mathbf{g} (x)) & - \\ \end{bmatrix} \begin{bmatrix} \frac{ \mathop{d} g_{1} }{ \mathop{d} x } (x) \\ \vdots \\ \frac{ \mathop{d} g_{m} }{ \mathop{d} x } (x) \\ \end{bmatrix} \\ & = \mathbf{J} \mathbf{f} (\mathbf{g} (x)) \frac{ \mathop{d} \mathbf{g} }{ \mathop{d} x } (x). \\ \end{aligned}

Extension 3: vector-valued multivariate function \mathbf{c} (\mathbf{x}) = \mathbf{f} (\mathbf{g} (\mathbf{x}))

When both input and output are vectors, the composite function \mathbf{c} (\mathbf{x}) = \mathbf{f} (\mathbf{g} (\mathbf{x})) is a vector-valued multivariate function, whose derivative is a Jacobian matrix

\begin{aligned} \mathbf{J} \mathbf{f} (\mathbf{x}) & = \begin{bmatrix} \frac{\partial f_{1}}{x_{1}} (\mathbf{x}) & \dots & \frac{\partial f_{1}}{x_{n}} (\mathbf{x}) \\ \vdots & \ddots & \vdots \\ \frac{\partial f_{m}}{x_{1}} (\mathbf{x}) & \dots & \frac{\partial f_{m}}{x_{n}} (\mathbf{x}) \\ \end{bmatrix} \\ & = \begin{bmatrix} - & \nabla^{T} f_{1} (\mathbf{g} (\mathbf{x})) & - \\ & \vdots & \\ - & \nabla^{T} f_{n} (\mathbf{g} (\mathbf{x})) & - \\ \end{bmatrix} \begin{bmatrix} | & & | \\ \frac{ \mathop{d} \mathbf{g} }{ \mathop{d} x_{1} } (x_{1}) & \dots & \frac{ \mathop{d} \mathbf{g} }{ \mathop{d} x_{n} } (x_{n}) \\ | & & | \\ \end{bmatrix} \\ & = \begin{bmatrix} - & \nabla^{T} f_{1} (\mathbf{g} (\mathbf{x})) & - \\ & \vdots & \\ - & \nabla^{T} f_{n} (\mathbf{g} (\mathbf{x})) & - \\ \end{bmatrix} \begin{bmatrix} - & \nabla^{T} g_{1} (\mathbf{x}) & - \\ & \vdots & \\ - & \nabla^{T} g_{m} (\mathbf{x}) & - \\ \end{bmatrix} \\ & = \mathbf{J} \mathbf{f} (\mathbf{g} (\mathbf{x})) \mathbf{J} \mathbf{g} (\mathbf{x}). \\ \end{aligned}