Vector, Matrix, and Tensor Derivatives

Erik Learned-Miller

The purpose of this document is to help you learn to take derivatives of vectors, matrices,
and higher order tensors (arrays with three dimensions or more), and to help you take
derivatives with respect to vectors, matrices, and higher order tensors.

1 Simplify, simplify, simplify

Much of the confusion in taking derivatives involving arrays stems from trying to do too
many things at once. These “things” include taking derivatives of multiple components
simultaneously, taking derivatives in the presence of summation notation, and applying the
chain rule. By doing all of these things at the same time, we are more likely to make errors,
at least until we have a lot of experience.

1.1 Expanding notation into explicit sums and equations for each
component
In order to simplify a given calculation, it is often useful to write out the explicit formula for
a single scalar element of the output in terms of nothing but scalar variables. Once one has
an explicit formula for a single scalar element of the output in terms of other scalar values,
then one can use the calculus that you used as a beginner, which is much easier than trying
to do matrix math, summations, and derivatives all at the same time.

Example. Suppose we have a column vector ~y of length C that is calculated by forming

the product of a matrix W that is C rows by D columns with a column vector ~x of length
D:
~y = W ~x. (1)
Suppose we are interested in the derivative of ~y with respect to ~x. A full characterization
of this derivative requires the (partial) derivatives of each component of ~y with respect to each
component of ~x, which in this case will contain C × D values since there are C components
in ~y and D components of ~x.
Let’s start by computing one of these, say, the 3rd component of ~y with respect to the
7th component of ~x. That is, we want to compute
∂~y3
,
∂~x7

1
which is just the derivative of one scalar with respect to another.
The first thing to do is to write down the formula for computing ~y3 so we can take its
derivative. From the definition of matrix-vector multiplication, the value ~y3 is computed by
taking the dot product between the 3rd row of W and the vector ~x:
D
X
~y3 = W3,j ~xj . (2)
j=1

At this point, we have reduced the original matrix equation (Equation 1) to a scalar equation.
This makes it much easier to compute the desired derivatives.

1.2 Removing summation notation

While it is certainly possible to compute derivatives directly from Equation 2, people fre-
P
quently make errors when
Q differentiating expressions that contain summation notation ( )
or product notation ( ). When you’re beginning, it is sometimes useful to write out a
computation without any summation notation to make sure you’re doing everything right.
Using “1” as the first index, we have:

~y3 = W3,1~x1 + W3,2~x2 + ... + W3,7~x7 + ... + W3,D ~xD .

Of course, I have explicitly included the term that involves ~x7 , since that is what we are
differenting with respect to. At this point, we can see that the expression for y3 only depends
upon ~x7 through a single term, W3,7~x7 . Since none of the other terms in the summation
include ~x7 , their derivatives with respect to ~x7 are all 0. Thus, we have
∂~y3 ∂
= [W3,1~x1 + W3,2~x2 + ... + W3,7~x7 + ... + W3,D ~xD ] (3)
∂~x7 ∂~x7
∂
= 0 + 0 + ... + [W3,7~x7 ] + ... + 0 (4)
∂~x7
∂
= [W3,7~x7 ] (5)
∂~x7
= W3,7 . (6)

By focusing on one component of ~y and one component of ~x, we have made the calculation
about as simple as it can be. In the future, when you are confused, it can help to try to
reduce a problem to this most basic setting to see where you are going wrong.

1.2.1 Completing the derivative: the Jacobian matrix

Recall that our original goal was to compute the derivatives of each component of ~y with
respect to each component of ~x, and we noted that there would be C × D of these. They

2
can be written out as a matrix in the following form:
 ∂~y1 ∂~y1 ∂~y1 ∂~
y1 
∂~x1 ∂~
x2 ∂~
x3
... ∂~
xD
 
 ∂~y 
 2 ∂~y2 ∂~y2 . . . ∂~
y2 
 ∂~x1 ∂~x2 ∂~x3 ∂~
xD 
 . .. .. .. .. 
 .. . . . . 
∂~
yC ∂~
yC ∂~
yC ∂~
yC
∂~x1 ∂~
x2 ∂~
x3
... ∂~
xD

In this particular case, this is called the Jacobian matrix, but this terminology is not too
important for our purposes.
Notice that for the equation
~y = W ~x,
the partial of ~y3 with respect to ~x7 was simply given by W3,7 . If you go through the same
process for other components, you will find that, for all i and j,
∂~yi
= Wi,j .
∂~xj

This means that the matrix of partial derivatives is

 ∂~y1 ∂~y1 ∂~y1 ∂~
y1 
∂~x1 ∂~x2 ∂~x3
. . . ∂~xD
 
  W1,1 W1,2 W1,3 . . . W1,D
 2 ∂~y2 ∂~y2 . . . ∂~y2   W2,1 W2,2 W2,3 . . . W2,D 
 ∂~y   
 ∂~x1 ∂~x2 ∂~x3 ∂~
xD = . .. .. ... .. 
  ..
 
 .
 .. .
.. .
.. . .. .
..  . . . 
∂~
yC ∂~
yC ∂~
yC ∂~
yC WC,1 WC,2 WC,3 . . . WC,D .
∂~x1 ∂~x2 ∂~x3
. . . ∂~
xD

This, of course, is just W itself.

Thus, after all this work, we have concluded that for

~y = W ~x,

we have
d~y
= W.
d~x

2 Row vectors instead of column vectors

It is important in working with different neural networks packages to pay close attention to
the arrangement of weight matrices, data matrices, and so on. For example, if a data matrix
X contains many different vectors, each of which represents an input, is each data vector a
row or column of the data matrix X?
In the example from the first section, we worked with a vector ~x that was a column
vector. However, you should also be able to use the same basic ideas when ~x is a row vector.

3
2.1 Example 2
Let ~y be a row vector with C components computed by taking the product of another row
vector ~x with D components and a matrix W that is D rows by C columns.

~y = ~xW.

Importantly, despite the fact that ~y and ~x have the same number of components as before,
the shape of W is the transpose of the shape that we used before for W . In particular, since
we are now left-multiplying by ~x, whereas before ~x was on the right, W must be transposed
for the matrix algebra to work.
In this case, you will see, by writing
D
X
~y3 = ~xj Wj,3
j=1

that
∂~y3
= W7,3 .
∂~x7
Notice that the indexing into W is the opposite from what it was in the first example.
However, when we assemble the full Jacobian matrix, we can still see that in this case as
well,
d~y
= W. (7)
d~x

3 Dealing with more than two dimensions

Let’s consider another closely related problem, that of computing
d~y
.
dW
In this case, ~y varies along one coordinate while W varies along two coordinates. Thus, the
entire derivative is most naturally contained in a three-dimensional array. We avoid the term
“three-dimensional matrix” since it is not clear how matrix multiplication and other matrix
operations are defined on a three-dimensional array.
Dealing with three-dimensional arrays, it becomes perhaps more trouble than it’s worth
to try to find a way to display them. Instead, we should simply define our results as formulas
which can be used to compute the result on any element of the desired three dimensional
array.
Let’s again compute a scalar derivative between one component of ~y , say ~y3 and one
component of W , say W7,8 . Let’s start with the same basic setup in which we write down
an equation for ~y3 in terms of other scalar components. Now we would like an equation that
expresses ~y3 in terms of scalar values, and shows the role that W7,8 plays in its computation.

4
 However, what we see is that W7,8 plays no role in the computation of ~y3 , since

~y3 = ~x1 W1,3 + ~x2 W2,3 + ... + ~xD WD,3 . (8)

In other words,
∂~y3
= 0.
∂W7,8
However, the partials of ~y3 with respect to elements of the 3rd column of W will certainly
be non-zero. For example, the derivative of ~y3 with respect to W2,3 is given by
∂~y3
= ~x2 , (9)
∂W2,3
as can be easily seen by examining Equation 8.
In general, when the index of the ~y component is equal to the second index of W , the
derivative will be non-zero, but will be zero otherwise. We can write:
∂~yj
= ~xi ,
∂Wi,j
but the other elements of the 3-d array will be 0. If we let F represent the 3d array
representing the derivative of ~y with respect to W , where
∂~yi
Fi,j,k = ,
∂Wj,k
then
Fi,j,i = ~xj ,
but all other entries of F are zero.
Finally, if we define a new two-dimensional array G as

Gi,j = Fi,j,i

we can see that all of the information we need about F can be stored in G, and that the
non-trivial portion of F is really two-dimensional, not three-dimensional.
Representing the important part of derivative arrays in a compact way is critical to
efficient implementations of neural networks.

4 Multiple data points

It is a good exercise to repeat some of the previous examples, but using multiple examples of
~x, stacked together to form a matrix X. Let’s assume that each individual ~x is a row vector
of length D, and that X is a two-dimensional array with N rows and D columns. W , as in
our last example, will be a matrix with D rows and C columns. Y , given by

Y = XW,

5
will also be a matrix, with N rows and C columns. Thus, each row of Y will give a row
vector associated with the corresponding row of the input X.
Sticking to our technique of writing down an expression for a given component of the
output, we have
D
X
Yi,j = Xi,k Wk,j .
k=1

We can see immediately from this equation that among the derivatives
∂Ya,b
,
∂Xc,d

they are all zero unless a = c. That is, since each component of Y is computed using only
the corresponding row of X, derivatives of components between different rows of Y and X
are all zero.
Furthermore, we can see that
∂Yi,j
= Wk,j (10)
∂Xi,k
doesn’t depend at all upon which row of Y and X we are comparing.
In fact, the matrix W holds all of these partials as it is–we just have to remember to
index into it according to Equation 10 to obtain the specific partial derivative that we want.
If we let Yi,: be the ith row of Y and let Xi,: be the ith row of X, then we see that

∂Yi,:
= W,
∂Xi,:

which is a simple generalization of our previous result from Equation 7.

5 The chain rule in combination with vectors and ma-

trices
Now that we have worked through a couple of basic examples, let’s combine these ideas with
an example of the chain rule. Again, assuming ~y and ~x are column vectors, let’s start with
the equation
~y = V W ~x,
and try to compute the derivative of ~y with respect to ~x. We could simply observe that the
product of two matrices V and W is simply another matrix, call it U , and therefore
d~y
= V W = U.
d~x
However, we want to go through the process of using the chain rule to define intermediate
results, so that we can see how the chain rule applies in the context of non-scalar derivatives.

6
 Let us define the intermediate result

m
~ = W ~x.

Then we have that

~y = V m.
~
We can then write, using the chain rule, that
d~y d~y dm
~
= .
d~x dm
~ d~x
To make sure that we know exactly what this means, let’s take the old approach of
analyzing one component at a time, starting with a single component of ~y and a single
component of ~x:
d~yi d~yi dm
~
= .
d~xj dm~ d~xj
But how exactly should we interpret the product on the right? The idea with the chain
rule is to multiply the change in ~yi with respect to each scalar intermediate variable by the
change in the scalar intermediate variable with respect to ~xj . In particular, if m~ has M
components, then we write
M
d~yi X d~yi dm
~k
= .
d~xj k=1
dm~ k d~xj
Recall from our previous results about derivatives of a vector with respect to a vector
that
d~yi
dm~k
is just Vi,k and that
dm~k
d~xj
is just Wk,j . So we can write
M
d~yi X
= Vi,k Wk,j ,
d~xj k=1

which is just the component expression for V W , our original answer to the problem.
To summarize, we can use the chain rule in the setting of vector and matrix derivatives
by

• Clearly stating intermediate results and the variables used to represent them,

• Expressing the chain rule for individual components of the final derivatives,

• Summing appropriately over the intermediate results within the chain rule expression.