Vector-Jacobian Products (VJPs) with jax.vjp

Reverse-mode automatic differentiation is the workhorse behind gradient descent in deep learning. It efficiently computes the gradient of a scalar output (like a loss function) with respect to potentially millions of inputs (model parameters). While jax.grad provides a user-friendly way to get these gradients, understanding the underlying mechanism, the Vector-Jacobian Product (VJP), unlocks more advanced capabilities. jax.vjp is the JAX function that computes VJPs.

The Essence of Reverse Mode: Vector-Jacobian Products

Recall that for a function $f: \mathbb{R}^n \rightarrow \mathbb{R}^m$, its Jacobian matrix $J_f(x)$ at a point $x \in \mathbb{R}^n$ contains all the partial derivatives $\frac{\partial f_i}{\partial x_j}$. The Jacobian is an $m \times n$ matrix.

Reverse-mode AD computes the product of a cotangent vector $v \in \mathbb{R}^m$ with the Jacobian matrix, yielding $v^T J_f(x)$. This product is the Vector-Jacobian Product (VJP). The result is a vector in $\mathbb{R}^n$ (or technically, a $1 \times n$ row vector).

Why is this useful? Consider a typical machine learning scenario where we have a composition of functions leading to a scalar loss $L$. Let $f$ be the last function in this chain, mapping from some intermediate representation $y \in \mathbb{R}^m$ to the final scalar loss $L \in \mathbb{R}$. So, $L = f(y)$. The gradient $\nabla_y L$ is a row vector $\frac{\partial L}{\partial y} \in \mathbb{R}^{1 \times m}$. If $y = g(x)$ where $x \in \mathbb{R}^n$, the chain rule tells us $\nabla_x L = \nabla_y L \cdot J_g(x)$. This is exactly a VJP! The vector $v^T$ is the gradient (or cotangent) flowing backward from the output, and $J_g(x)$ is the Jacobian of the function $g$ whose input gradients we want.

Reverse mode works by propagating these cotangent vectors backward through the computation graph, starting with the initial cotangent $\frac{dL}{dL} = 1.0$ for a scalar loss function $L$.

Computing VJPs with jax.vjp

JAX provides direct access to VJP computation through jax.vjp. Its basic signature is:

y, vjp_fun = jax.vjp(fun, *primals)

Let's break this down:

1. fun: The Python callable (function) to differentiate.
2. *primals: The input arguments (x, y, etc.) at which to evaluate the function and compute the VJP. These are the points where the Jacobian is implicitly calculated.
3. y: The output of the function evaluated at the primals, i.e., y = fun(*primals). This is the result of the forward pass.
4. vjp_fun: A new function (a closure) that computes the VJP. This function takes one argument, the cotangent vector v, which must have the same structure (shape and dtype) as the output y. When called as vjp_fun(v), it returns the VJP: $v^T J_{\text{fun}}(\text{primals})$. The result is a tuple containing the computed gradients with respect to each primal input.

Example: Scalar Function

Let's start with a simple function $f(x, y) = x^2 \sin(y)$.

import jax
import jax.numpy as jnp

def f(x, y):
  return x**2 * jnp.sin(y)

# Define primal inputs
x_primal = 3.0
y_primal = jnp.pi / 2.0

# Compute VJP
y_out, vjp_fun = jax.vjp(f, x_primal, y_primal)

print(f"Forward pass output y: {y_out}")

# The output y is scalar, so the cotangent v is also scalar.
# Let's use v = 1.0, which corresponds to finding d(1.0 * y) / dx and d(1.0 * y) / dy
# This is equivalent to finding the gradient of y.
cotangent_v = 1.0
primal_grads = vjp_fun(cotangent_v)

print(f"Gradient wrt x: {primal_grads[0]}") # Should be 2 * x * sin(y) = 2 * 3.0 * sin(pi/2) = 6.0
print(f"Gradient wrt y: {primal_grads[1]}") # Should be x^2 * cos(y) = 3.0^2 * cos(pi/2) = 0.0

# Compare with jax.grad
grad_f = jax.grad(f, argnums=(0, 1)) # Get gradients wrt both x and y
grads_via_grad = grad_f(x_primal, y_primal)

print(f"Gradients via jax.grad: {grads_via_grad}")

In this case, calling vjp_fun(1.0) gives the same result as jax.grad(f, argnums=(0, 1))(x, y). This highlights the fundamental relationship: jax.grad for scalar functions is essentially a convenience wrapper around jax.vjp that automatically provides the initial cotangent of 1.0.

Example: Non-Scalar Function

Now consider a function with a vector output: $g(x) = (x_0^2, x_1^3)$, where $x = (x_0, x_1)$.

import jax
import jax.numpy as jnp

def g(x):
  return jnp.array([x[0]**2, x[1]**3])

# Define primal input
x_primal = jnp.array([2.0, 3.0])

# Compute VJP
y_out, vjp_fun = jax.vjp(g, x_primal)

print(f"Forward pass output y: {y_out}") # Should be [4., 27.]

# The output y is a vector of shape (2,), so the cotangent v must also have shape (2,)
cotangent_v = jnp.array([10.0, 20.0]) # An arbitrary cotangent vector

# Compute the VJP: v^T J
primal_grad = vjp_fun(cotangent_v)

print(f"VJP result (gradient wrt x): {primal_grad}")
# Expected:
# Jacobian J = [[2*x0, 0], [0, 3*x1^2]] = [[4, 0], [0, 27]]
# v^T J = [10, 20] * [[4, 0], [0, 27]] = [10*4, 20*27] = [40, 540]
# jax.vjp returns a tuple, even for a single primal, so result is ([40., 540.],)

Here, vjp_fun takes the vector cotangent_v and multiplies it by the Jacobian of g evaluated at x_primal to produce the resulting gradient contributions back to x. The shape of the cotangent must match the shape of the function's output y_out.

Why Use jax.vjp Directly?

While jax.grad suffices for standard gradient descent where you need the gradient of a scalar loss, jax.vjp offers more control:

1. Efficiency for Non-Scalar Gradients: If you need gradients of multiple outputs simultaneously or want to manipulate intermediate gradients, VJPs can sometimes be structured more efficiently than repeated grad calls.
2. Custom Gradient Rules: As you'll see later in this chapter, jax.custom_vjp allows you to define bespoke VJP rules for specific functions. Understanding jax.vjp is necessary for this.
3. Implementing Complex AD Logic: For tasks requiring intricate gradient manipulations beyond standard backpropagation (e.g., certain optimization algorithms, advanced automatic differentiation research), direct VJP access is indispensable.
4. Pedagogical Understanding: Working with jax.vjp clarifies how reverse-mode AD operates and how jax.grad is constructed.

In summary, jax.vjp provides a lower-level interface to JAX's powerful reverse-mode automatic differentiation system. It computes Vector-Jacobian Products, which represent the core operation of backpropagation. While jax.grad handles the common case of scalar loss gradients, mastering jax.vjp gives you the tools to tackle more complex differentiation tasks and understand the system more deeply.