FiledSeptember 16, 2026·3 min read·By The Editor

Jacobian vs Gradient

Both are collections of partial derivatives. A gradient is for functions with one output; a Jacobian is for functions with many — and a gradient is simply a Jacobian with one row.

calculusintuition

Both are collections of derivatives. The difference is the shape of what you are differentiating.

A gradient is for functions with one output. A Jacobian is for functions with many. That is it — and a gradient is precisely a Jacobian that happens to have a single row.

The shapes

Figure 1 · it is all the same derivative

The only thing that changes is how many inputs and outputs you have. One of each gives a number. Many inputs, one output gives a column — the gradient. Many of both gives a grid — the Jacobian. A gradient is just a one-row Jacobian.

For a function $f(x_1, x_2, \ldots, x_n)$ producing one number, the gradient collects one partial derivative per input:

\nabla f = \left[ \frac{\partial f}{\partial x_1}, \frac{\partial f}{\partial x_2}, \ldots, \frac{\partial f}{\partial x_n} \right]

In plain English

The ∇ symbol is "nabla" or just "grad". Each entry answers: if I nudge this one input, how much does the output move? Together they point in the direction of steepest increase — which is why gradient descent goes the opposite way.

For a function producing m outputs from n inputs, you need that same row for every output. Stack them and you have the Jacobian, an $m \times n$ grid:

J_{ij} = \frac{\partial f_i}{\partial x_j}

In plain English

Row i, column j answers: how much does output i move when input j is nudged? Row i on its own is the gradient of output i.

Why deep learning mostly says "gradient"

Training minimises a loss, and a loss is always a single number. One output, many inputs — that is the gradient case, every time. Hence the vocabulary: gradient descent, gradient clipping, vanishing gradients.

Jacobians are still there, just not spoken about. Every intermediate layer maps many numbers to many numbers, so each layer has a Jacobian. Backpropagation is a chain of Jacobian multiplications — with one convenient detail:

In plain English

Because the final loss is a single number, backprop propagates a row vector backwards through those Jacobians rather than ever building them. A full Jacobian for a layer with 10,000 inputs and outputs would have 100 million entries. Frameworks compute vector-Jacobian products instead, which is why training is affordable at all.

Where each shows up

Gradient — training loss, optimisation, feature attribution. Anything that scores a model with one number.
Jacobian — normalising flows, sensitivity analysis, robotics, physics simulation. Anywhere a vector maps to a vector and you need every cross-term.
Hessian — worth naming here: the matrix of second derivatives of a one-output function. It describes curvature rather than slope. Second-order optimisers use it; most training does not, because it is $n \times n$ and n is in the billions.

Quick reference

	Inputs	Outputs	Shape	Notation
Derivative	1	1	scalar	$\frac{df}{dx}$
Gradient	many	1	vector (n)	$\nabla f$
Jacobian	many	many	matrix (m×n)	$J$
Hessian	many	1	matrix (n×n)	$H$

The short version

Both are just organised collections of partial derivatives.
Gradient: one output, so one row. Points uphill.
Jacobian: many outputs, so one row each.
A gradient is a one-row Jacobian; the Jacobian is the general case.
Deep learning says "gradient" because a loss is one number — but Jacobians are what backprop multiplies, without ever building them.

Gradient descent, by hand — the gradient put to work, with the chain rule in full.
The dot product — the operation underneath all this matrix arithmetic.
The maths you actually need — the four ideas this is one of.

← Archive— end of essay —Front page →

The shapes

Why deep learning mostly says "gradient"

Where each shows up

Quick reference

The short version

Related