Mathematica tips for numerical linear algebra (4)

Transpose and Flatten

So now we know how to multiply matrices using Mathematica’s Dot function. This also works for higher-rank tensors (more indices) such as $m_{abc} n_{cde} = (mn)_{abde}$, which we can implement as

m = RandomComplex[{}, {10, 20, 30}];
n = RandomComplex[{}, {30, 40, 50}];

m // Dimensions
(* {10,20,30} *)
n // Dimensions
(* {30,40,50} *)

m.n // Dimensions
(* {10,20,40,50} *)

Here we used Dimensions to see that check that $m_{abc}$ is indeed a rank-three tensor of size {10,20,30}, and that Dot contracts the adjacent indices in m.n, leaving the remaining indices in the same order.

What if the indices we want to contract aren’t adjacent? The trick is to shuffle the indices of our tensors using either Transpose or Flatten.

Transpose

Let’s start with Transpose. Given a tensor $m$, the two simplest uses of this function are Transpose[m] and Transpose[m, k<->l]. The first of these transposes the first two indices of $m$:

m = RandomComplex[{}, {10, 20, 30}];

m // Dimensions
(* {10,20,30} *)
Transpose[m] // Dimensions
(* {20,10,30} *)

The second form of the function allows us to specify which indices are transposed. For example, taking k=1 and l=3 will transpose the 1st and 3rd indices of $m$:

m = RandomComplex[{}, {10, 20, 30}];

m // Dimensions
(* {10,20,30} *)
Transpose[m, 1 <-> 3] // Dimensions
(* {30,20,10} *)

Using this second form, we can contract tensors such as $m_{abc} p_{dec}$:

m = RandomComplex[{}, {10, 20, 30}];
p = RandomComplex[{}, {40, 50, 30}];

m.Transpose[p, 1 <-> 3] // Dimensions
(* {10,20,50,40} *)

What if we want the indices of the resulting tensor to be ordered as {10,20,40,50} instead of {10,20,50,40}? Obviously, we could do the calculation in two steps, first constructing the product as above, and then taking a further transpose on the 3rd and 4th indices. Fortunately, Transpose can do this in one step. The result of Transpose[m, {k1, k2, k3}] is a tensor such that the 1st index of $m$ becomes the k1-th index, the 2nd index of $m$ becomes the k2-th index, and so on. For example, swapping the first two indices of $m$ is implemented as

m = RandomComplex[{}, {10, 20, 30}];

Transpose[m, {2, 1, 3}] // Dimensions
(* {20,10,30} *)

which we interpret as sending the 1st index to position 2, the 2nd index to position 1, and leaving the 3rd index unchanged.

We can then perform the contraction of $m$ and $p$ as above but with the 3rd and 4th indices swapped as

m = RandomComplex[{}, {10, 20, 30}];
p = RandomComplex[{}, {40, 50, 30}];

m.Transpose[p, {3, 2, 1}] // Dimensions
(* {10,20,50,40} *)

The key to remember how this form of Transpose works is that you specify where the indices end up.

Flatten

We can also shuffle indices with the function Flatten. This function also has other uses – such as letting us contract more than one index at a time – but we’ll focus on using it to transpose for the moment.

The key with Flatten is that you specify where the indices came from. For example, compare

m = RandomComplex[{}, {10, 20, 30, 40}];

Transpose[m, {1, 4, 2, 3}] // Dimensions
(* {10,30,40,20} *)

Flatten[m, {{1}, {4}, {2}, {3}}] // Dimensions
(* {10,40,20,30} *)

Using Transpose with the argument {1, 4, 2, 3}, we leave the 1st index unchanged, send the 2nd index to 4th position, send the 3rd index to 2nd position, and send the 4th index to 3rd position. Using Flatten with the argument {{1}, {4}, {2}, {3}}, we leave the 1st index unchanged, set the 2nd index of the result to be the 4th index of the original tensor, set the 3rd index of the result to be the 2nd index of the original tensor, and finally set the 4th index of the result to be the 3rd index of the original tensor. These operations are not the same! And the difference is whether you are thinking about where to send the indices to or where the indices came from.

If we want the same result from both, we would use

m = RandomComplex[{}, {10, 20, 30, 40}];

Transpose[m, {1, 4, 2, 3}] // Dimensions
(* {10,30,40,20} *)

Flatten[m, {{1}, {3}, {4}, {2}}] // Dimensions
(* {10,40,20,30} *)

Transpose[m, {1, 4, 2, 3}] == Flatten[m, {{1}, {3}, {4}, {2}}]
(* True *)

Note that if you just need to transpose a tensor, Tranpose is often faster than Flatten. For example, increasing the dimensions of $m$ somewhat:

m = RandomComplex[{}, {10, 200, 300, 400}];

Transpose[m, {1, 4, 2, 3}] // Dimensions // RepeatedTiming
(* {0.60360695, {10,300,400,200}} *)

Flatten[m, {{1}, {3}, {4}, {2}}] // Dimensions // RepeatedTiming
(* {4.8638016, {10,300,400,200}} *)

Flatten is 8 times slower than Tranpose in this case! If you have to repeat this operation many thousands of times, this could become a huge bottleneck!

Anthony Ashmore
Anthony Ashmore
Marie Skłodowska-Curie Fellow