# Quantum gradients¶

## Automatic differentiation¶

Derivatives and gradients are ubiquitous throughout science and engineering. In recent years, automatic differentiation has become a key feature in many numerical software libraries, in particular for machine learning (e.g., Theano, Autograd, Tensorflow, or Pytorch).

Generally speaking, automatic differentiation is the ability for a software library to compute the derivatives of arbitrary numerical code. If you write an algorithm to compute some function \(h(x)\) (which may include mathematical expressions, but also control flow statements like `if`

, `for`

, etc.), then automatic differentiation provides an algorithm for \(\nabla h(x)\) with the same degree of complexity as the original function.

*Automatic* differentiation should be distinguished from other forms of differentiation. *Manual differentiation*, where an expression is differentiated by hand — often on paper — is extremely time-consuming and error-prone. In *numerical differentiation*, such as the finite-difference method familiar from high-school calculus, the derivative of a function is approximated by numerically evaluating the function at two infinitesimally separated points. However, this method can sometimes be imprecise due to the constraints of classical floating-point arithmetic.

## Computing gradients of quantum functions¶

Variational circuits are parameterized functions \(f(x;\bm{\theta})\) which can be evaluated by measuring a quantum circuit. If we can compute the gradient of a quantum function, we can then use this information in an optimization or machine learning algorithm, tuning the quantum circuit via gradient descent to produce a desired output. While numerical differentiation is an option, PennyLane is the first software library to support **automatic differentiation of quantum circuits** [1].

How is this accomplished? It turns out that the gradient of a quantum function \(f(x;\bm{\theta})\) can in many cases be expressed as a linear combination of other quantum functions. In fact, these other quantum functions typically use the same circuit, differing only in a shift of the argument.

Making a rough analogy to classically computable functions, this is similar to how the derivative of the function \(f(x)=\sin(x)\) is identical to \(\frac{1}{2}\sin(x+\frac{\pi}{2}) - \frac{1}{2}\sin(x-\frac{\pi}{2})\). So the same underlying algorithm can be reused to compute both \(\sin(x)\) and its derivative (by evaluating at \(x\pm\frac{\pi}{2}\)).
This intuition holds for many quantum functions of interest: *the same circuit can be used to compute both the quantum function and the gradient of the quantum function* [2].

## A more technical explanation¶

Circuits in PennyLane are specified by a sequence of gates. The unitary transformation carried out by the circuit can thus be broken down into a product of unitaries:

Each of these gates is unitary, and therefore must have the form \(U_{j}(\gamma_j)=\exp{(i\gamma_j H_j)}\) where \(H_j\) is a Hermitian operator which generates the gate and \(\gamma_j\) is the gate parameter. We have omitted which wire each unitary acts on, since it is not necessary for the following discussion.

Note

In this example, we have used the input \(x\) as the argument for gate \(U_0\) and the parameters \(\bm{\theta}\) for the remaining gates. This is not required. Inputs and parameters can be arbitrarily assigned to different gates.

### A single parameterized gate¶

Let us single out a single parameter \(\theta_i\) and its associated gate \(U_i(\theta_i)\). For simplicity, we remove all gates except \(U_i(\theta_i)\) and \(U_0(x)\) for the moment. In this case, we have a simplified quantum circuit function

For convenience, we rewrite the unitary conjugation as a linear transformation \(\mathcal{M}_{\theta_i}\) acting on the operator \(\hat{B}\):

The transformation \(\mathcal{M}_{\theta_i}\) depends smoothly on the parameter \(\theta_i\), so this quantum function will have a well-defined gradient:

The key insight is that we can, in many cases of interest, express this gradient as a linear combination of the same transformation \(\mathcal{M}\), but with different parameters. Namely,

where the multiplier \(c\) and the shift \(s\) are determined completely by the type of transformation \(\mathcal{M}\) and independent of the value of \(\theta_i\).

Note

While this construction bears some resemblance to the numerical finite-difference method for computing derivatives, here \(s\) is finite rather than infinitesimal.

### Multiple parameterized gates¶

To complete the story, we now go back to the case where there are many gates in the circuit. We can absorb any gates applied before gate \(i\) into the initial state: \(|\psi_{i-1}\rangle = U_{i-1}(\theta_{i-1}) \cdots U_{1}(\theta_{1})U_{0}(x)|0\rangle\). Similarly, any gates applied after gate \(i\) are combined with the observable \(\hat{B}\): \(\hat{B}_{i+1} = U_{N}^\dagger(\theta_{N}) \cdots U_{i+1}^\dagger(\theta_{i+1}) \hat{B} U_{i+1}(\theta_{i+1}) \cdots U_{N}(\theta_{N})\).

With this simplification, the quantum circuit function becomes

and its gradient is

This gradient has the exact same form as the single-gate case, except we modify the state \(|x\rangle \rightarrow |\psi_{i-1}\rangle\) and the measurement operator \(\hat{B}\rightarrow\hat{B}_{i+1}\). In terms of the circuit, this means we can leave all other gates as they are, and only modify gate \(U(\theta_i)\) when we want to differentiate with respect to the parameter \(\theta_i\).

Note

Sometimes we may want to use the same classical parameter with multiple gates in the circuit. Due to the product rule, the total gradient will then involve contributions from each gate that uses that parameter. PennyLane handles this automatically.

### Pauli gate example¶

Consider a quantum computer with parameterized gates of the form

where \(\hat{P}_i=\hat{P}_i^\dagger\) is a Pauli operator.

The gradient of this unitary is

Substituting this into the quantum circuit function \(f(x; \bm{\theta})\), we get

where \([X,Y]=XY-YX\) is the commutator.

We now make use of the following mathematical identity for commutators involving Pauli operators [mitarai2018quantum]:

Substituting this into the previous equation, we obtain the gradient expression

Finally, we can rewrite this in terms of quantum functions:

### Gaussian gate example¶

For quantum devices with continuous-valued operators, such as photonic quantum computers, it is convenient to employ the Heisenberg picture, i.e., to track how the gates \(U_i(\theta_i)\) transform the final measurement operator \(\hat{B}\).

As an example, we consider the Squeezing gate. In the Heisenberg picture, the Squeezing gate causes the quadrature operators \(\hat{x}\) and \(\hat{p}\) to become rescaled:

and

Expressing this in matrix notation, we have

The gradient of this transformation can easily be found:

We notice that this can be rewritten this as a linear combination of squeeze operations:

where \(s\) is an arbitrary nonzero shift [3].

As before, assume that an input \(y\) has already been embedded into a quantum state \(|y\rangle = U_0(y)|0\rangle\) before we apply the squeeze gate. If we measure the \(\hat{x}\) operator, we will have the following quantum circuit function:

Finally, its gradient can be expressed as

Note

For simplicity of the discussion, we have set the phase angle of the Squeezing gate to be zero. In the general case, Squeezing is a two-parameter gate, containing a squeezing magnitude and a squeezing angle. However, we can always decompose the two-parameter form into a Squeezing gate like the one above, followed by a Rotation gate.

Footnotes

[1] | This should be contrasted with software which can perform automatic differentiation on classical simulations of quantum circuits, such as Strawberry Fields. |

[2] | In situations where no formula for automatic quantum gradients is known, PennyLane falls back to approximate gradient estimation using numerical methods. |

[3] | In physical experiments, it is beneficial to choose \(s\) so that the additional squeezing is small. However, there is a tradeoff, because we also want to make sure \(\frac{1}{2\sinh(s)}\) does not blow up numerically. |