Vector-Jacobian products and automatic differentiation

November 6, 2023

Consider $f : R^{n} \to R^{m}$ . The Jacobian of $f$ , denoted $J_{f}$ , is an $m \times n$ matrix of all the partial derivatives. Pretty basic:

J_{f} = \frac{\partial f}{\partial x} = \frac{\partial f _{1}}{\partial x _{1}} ⋮ \frac{\partial f _{m}}{\partial x _{1}} \dots ⋱ \dots \frac{\partial f _{1}}{\partial x _{n}} ⋮ \frac{\partial f _{m}}{\partial x _{n}}

For $v \in R^{m}$ and $x \in R^{n}$ , let's look at the product of the row vector $v^{⊤}$ and $J_{f} (x)$ . This is known as the vector-Jacobian product (VJP).

v^{⊤} J_{f} (x) = [v_{1} v_{2} \dots v_{m}] \frac{\partial f _{1}}{\partial x _{1}} (x_{1}) ⋮ \frac{\partial f _{m}}{\partial x _{1}} (x_{1}) \dots ⋱ \dots \frac{\partial f _{1}}{\partial x _{n}} (x_{n}) ⋮ \frac{\partial f _{m}}{\partial x _{n}} (x_{n}) = v_{1} \frac{\partial f _{1}}{\partial x _{1}} (x_{1}) + \dots + v_{m} \frac{\partial f _{m}}{\partial x _{1}} (x_{1}) ⋮ v_{1} \frac{\partial f _{1}}{\partial x _{n}} (x_{n}) + \dots + v_{m} \frac{\partial f _{m}}{\partial x _{n}} (x_{n})^{⊤}

The row vectors are pesky, so transposes are taken.

J_{f} (x)^{⊤} v = v_{1} \frac{\partial f _{1}}{\partial x _{1}} (x_{1}) + \dots + v_{m} \frac{\partial f _{m}}{\partial x _{1}} (x_{1}) ⋮ v_{1} \frac{\partial f _{1}}{\partial x _{n}} (x_{n}) + \dots + v_{m} \frac{\partial f _{m}}{\partial x _{n}} (x_{n})

As the "variables" are $f$ , $x$ , and $v$ in $J_{f} (x)^{⊤} v$ , we can define a new operation $vjp (f, x) (v)$ .^[1]

Note that $vjp : (R^{n} \to R^{m}, R^{n}) \to R^{m} \to R^{n}$ .

vjp (f, x) (v) : = J_{f} (x)^{⊤} v

But that still requires us to calculate $\partial f / \partial x_{k}$ , which can be hard depending on $f$ . This is especially the case if $f$ is a composition of multiple functions.^[2]

The chain rule for Jacobians is $J_{f_{k + 1} \circ f_{k}} (x) = J_{f_{k + 1}} (f_{k} (x)) J_{f_{k}} (x)$ . We can define $x_{k} = (f_{k} \circ f_{k - 1} \circ \dots \circ f_{1}) (x)$ as the $k$ th intermediate function value. Reconciling this with VJPs:

vjp (f, x) (v) = J_{f} (x)^{⊤} v = (J_{f_{i} \circ \dots \circ f_{3} \circ f_{2} \circ f_{1}} (x))^{⊤} v = (J_{f_{i} \circ \dots \circ f_{3} \circ f_{2}} (f_{1} (x)) J_{f_{1}} (x))^{⊤} v = (J_{f_{i} \circ \dots \circ f_{3}} (f_{2} (f_{1} (x))) J_{f_{2}} (x_{1}) J_{f_{1}} (x))^{⊤} v = (J_{f_{i}} (x_{i - 1}) \dots J_{f_{2}} (x_{1}) J_{f_{1}} (x))^{⊤} v = J_{f_{1}} (x)^{⊤} J_{f_{2}} (x_{1})^{⊤} \dots J_{f_{i}} (x_{i - 1})^{⊤} v

Noting that $J_{f_{i}} (x)^{⊤} v = vjp (f_{i}, x) (v)$ , we can express $vjp (f, x) (v)$ with a composition of $vjp$ s.

vjp (f, x) (v) = J_{f_{1}} (x)^{⊤} \dots J_{f_{i - 1}} (x_{i - 2})^{⊤} J_{f_{i}} (x_{i - 1})^{⊤} v = J_{f_{1}} (x)^{⊤} \dots J_{f_{i - 1}} (x_{i - 2})^{⊤} vjp (f_{i}, x_{i - 1}) (v) = J_{f_{1}} (x)^{⊤} \dots vjp (f_{i - 1}, x_{i - 2}) (vjp (f_{i}, x_{i - 1}) (v)) = (vjp (f_{1}, x) \circ \dots \circ vjp (f_{i}, x_{i - 1})) (v)

Which brings us to our conclusion:

vjp (f, x) (v) = (vjp (f_{1}, x) \circ \dots \circ vjp (f_{i}, x_{i - 1})) (v)

The value of this representation of VJPs lies in the fact that it enables reverse mode automatic differentiation.

Automatic differentiation

Automatic differentiation (autodiff) follows from the fact that we can obtain the VJP of a composite function algorithmically by evaluating the VJPs of its constituent functions. Since even the most complicated of functions are made up of a composition of elementary functions, and the VJP of elementary functions is trivial, we can build up VJPs for "complicated" (i.e. lots of variables, deep composition, etc.) functions step-by-step. We require numeric values for $v$ and $x$ which sets this method apart from symbol differentiation.

Reverse mode autodiff refers to the fact that two passes -- forward and backward -- are required to calculate VJPs. Calculating the VJP of $f_{k}$ requires us to have calculated the VJP of $f_{k + 1}$ as well as $x_{k - 1}$ . The forward pass as we evaluate $f$ from $f_{1}$ to $f_{i}$ gives us the intermediate value $x_{k - 1}$ . The backwards pass as we build up the VJPs from $f_{i}$ to $f_{1}$ calculates the VJP of $f_{k + 1}$ .

Wait, are VJPs actually doing differentiation though? Sure, we can calculate the product of a vector and Jacobian, but differentiation would imply something like a gradient or the entire Jacobian. Turns out VJPs allow us to calculate both. Using $f : R^{n} \to R$ and $v = 1_{1, 1}$ , we obtain the VJP:

vjp (f, x) (1_{1, 1}) = J_{f} (x)^{⊤} 1_{1, 1} = \frac{\partial f _{1}}{\partial x _{1}} (x_{1}) ⋮ \frac{\partial f _{1}}{\partial x _{n}} (x_{n}) = \nabla f (x)

So the VJP of a scalar function can be used to calculate its gradient at a point. Similarly, we can calculate the entire Jacobian of a function row-by-row. Choosing $v$ as one-hot encoded vectors results in the VJP being the row of the Jacobian encoded by $v$ .

J_{f} (x) = vjp (f, x) (δ_{i = 0})^{⊤} ⋮ vjp (f, x) (δ_{i = m})^{⊤}

It is worth noting that building the entire Jacobian with VJPs requires $m$ passes for $f : R^{n} \to R^{m}$ with an $m \times n$ Jacobian. We can assume that this method is more efficent the smaller $m$ is compared to $n$ . In a machine learning context, $m = 1 ≪ n$ is common as the outputs are scalar loss values and the inputs are model weights. This makes VJPs suitable for machine learning backpropagation.

Vector-Jacobian products can be built up from VJPs and produce gradients and Jacobians. This enables reverse mode automatic differentiation, which is efficient for machine learning. Everything falls into place.

We could use $vjp (f, x, v)$ , but currying $v$ out simplifies later calculations. ↩
$f_{k}$ was used as the $k$ th component of $f$ before, but we use it to denote different functions from here on. ↩