Unitary designs

Author: Olivia Di Matteo — Posted: 07 September 2021. Last updated: 07 September 2021.


This demo is intended to be a sequel to the demo about the Haar measure. If you are not familiar with the Haar measure, we recommend going through that demo first before exploring this one.

Take a close look at the following mathematical object:


There are many things we can say about it: it consists of seven points and seven lines (the circle counts as a line); each line contains three points, and each point is contained in three lines. Furthermore, any pair of points occur together in exactly one line. This object, called the Fano plane, is an instance of a mathematical structure called a projective plane, which is just one example of a combinatorial design. Designs are sets of objects, and groups of those objects, that satisfy certain balance properties and symmetries. They have been studied for hundreds of years in a huge variety of contexts 1, from error correcting codes, to card games, and even agriculture. So, what about quantum computing?

Designs are actually quite prevalent in quantum computing. You’ve almost certainly come across one before, though you may not have realized it. At the end of the Haar measure demo, we asked a very important question: “do we always need to sample from the full Haar measure?”. The answer to this is “no”, and the reasoning lies in the study of unitary designs.

In this demo, you’ll learn the definition of \(t\)-designs, what it means to generalize them to unitary \(t\)-designs, and you’ll see some canonical examples of designs in quantum computing. You’ll also learn about their connection with the Haar measure, what it means to twirl a quantum channel, and explore how to leverage 2-designs in PennyLane to compute the average fidelity of a noisy channel. You will experience directly a situation where we can use a \(t\)-design as a shortcut over the full Haar measure to greatly improve the efficiency of a task 🎉.

From spheres to unitary \(t\)-designs

Spherical designs

Before diving into unitary designs, let’s look at the sphere for some intuition. Suppose we have a polynomial in \(d\) variables, and we would like to compute its average over the surface of a real, \(d\)-dimensional unit sphere, \(S(R^d)\). We can do so by integrating that function over the sphere (using the proper measure), but that would be a lot of parameters to keep track of.

One could alternatively approximate the average by sampling thousands of points uniformly at random on the sphere, evaluating the function at those points, and computing their average value. That will always work, and it will get us close, but it will not be exact.

In fact, both of those approaches may be overkill in some special cases—if the terms in the polynomial have the same degree of at most \(t\), you can compute the average exactly over the sphere using only a small set of points rather than integrating over the entire sphere. That set of points is called a spherical \(t\)-design. More formally 1, 2:


Let \(p_t: \mathcal{S}(R^d)\rightarrow R\) be a polynomial in \(d\) variables, with all terms homogeneous in degree at most \(t\). A set \(X = \{x: x \in \mathcal{S}(R^d)\}\) is a spherical \(t\)-design if

\[\frac{1}{|X|} \sum_{x \in X} p_t(x) = \int_{\mathcal{S}(R^d)} p_t (u) d\mu(u)\]

holds for all possible \(p_t\), where \(d\mu\) is the uniform, normalized spherical measure. A spherical \(t\)-design is also a \(k\)-design for all \(k < t\).

Now this is a pretty abstract picture, so let’s consider the 3-dimensional sphere. This definition tells us that if we want to take the average of a polynomial over a sphere where all terms have the same degree of at most 2, we can do so using a small, representative set of points called a 2-design, rather than the whole sphere. Similarly, if all terms of the polynomial have the same degree of at most 3, we could use a 3-design.

But what are these representative sets of points? Since we are using these points as a stand-in for averaging over the whole sphere, we’d want the points in the set to be distributed in a way that provides sufficient “coverage”. In the 3-dimensional case, the vertices of some familiar solids form \(t\)-designs 1, 3:


We see from these illustrations that spherical designs are sets of evenly-spaced points. As \(t\) increases, the configurations become increasingly sphere-like. Looking at this in a different way, the more complex a function becomes as its degree increases, the closer the \(t\)-design must be to a sphere; we need to evaluate the function at more points in order to gain sufficient information when a function is varying more quickly due to a higher degree. In 3 dimensions, we can compute the average of a polynomial with degree 2 by evaluating it only at the points of a tetrahedron, despite the fact that it doesn’t look spherical at all. More complex functions require more points and thus more intricate configurations for the design. Spherical designs exist for all \(t\) and dimension \(d\) 1. They are not always unique, and may have varying numbers of points.

To show that this really works, let’s look at an explicit example. Consider the following polynomial in 3 variables:

\[f(x, y, z) = x^4 - 4 x^3 y + y^2 z^2\]

We can compute the average value of \(f\) by integrating over a unit sphere: the result is \(4/15 \approx 0.26667\). However, this integral is non-trivial to evaluate by hand; the most straightforward way is to convert to polar coordinates, and even then, it involves integrating functions with 4th and 5th powers of trigonometric functions.

Instead, this is a case where we can leverage the fact that all terms in the polynomial have degree 4, and compute the average exactly using only a subset of points that form a 4-design. We choose a dodecahedron for convenience; while this is actually a 5 design, it also forms a 4-design, and is a more familiar shape than the 4-design depicted above.

First, we define the set of points that comprise a dodecahedron:

import numpy as np

# The golden ratio
g = (1 + np.sqrt(5)) / 2

# A dodecahedron has 20 points
dodecahedron = np.array([
    # 8 of them form a cube within the sphere
    [1, 1, 1], [1, 1, -1], [1, -1, 1], [1, -1, -1],
    [-1, 1, 1], [-1, 1, -1], [-1, -1, 1], [-1, -1, -1],

    # 4 of them form a rectangle within the y-z plane
    [0, g, 1/g], [0, g, -1/g], [0, -g, 1/g], [0, -g, -1/g],

    # 4 of them form a rectangle within the x-z plane
    [1/g, 0, g], [1/g, 0, -g], [-1/g, 0, g], [-1/g, 0, -g],

    # 4 of them form a rectangle within the x-y plane
    [g, 1/g, 0],[g, -1/g, 0], [-g, 1/g, 0], [-g, -1/g, 0],

# Normalize the points so they all fit in the unit sphere
dodecahedron = np.array(
   [point / np.linalg.norm(point) for point in dodecahedron]

Now we define our function and compute the average over the dodecahedron:

def f(x, y, z):
    return (x ** 4) - 4 * (x ** 3) * y +  y ** 2 * z ** 2

dodeca_average = np.mean([f(*point) for point in dodecahedron])



This is exactly the value we expect. What happens if we try to do this using only a 3-design, the cube?

# The first 8 points of the dodecahedron are a cube
cube = dodecahedron[:8]

cube_average = np.mean([f(*point) for point in cube])



This clearly differs from the true value. We need a design with \(t=4\) or better in order to compute this average, and when such a design is available, we may save significant computational time.

Unitary designs

We’ve learned now that spherical designs are sets of evenly-spaced points, and saw how they can be used as a shortcut to evaluate the average of a polynomial up to a given degree \(t\). However, there was nothing quantum about this; there weren’t even any complex numbers involved. A unitary design extends this concept from evenly-distributed points to evenly-distributed unitaries. More formally, instead of averaging polynomials over spheres, we consider polynomials that are functions of the entries of unitary matrices 6, 8.


Let \(P_{t,t}(U)\) be a polynomial with homogeneous degree at most \(t\) in \(d\) variables in the entries of a unitary matrix \(U\), and degree \(t\) in the complex conjugates of those entries. A unitary \(t\)-design is a set of \(K\) unitaries \(\{U_k\}\) such that

\[\frac{1}{K} \sum_{k=1}^{K} P_{t,t}(U_k) = \int_{\mathcal{U}(d)} P_{t,t} (U) d\mu(U)\]

holds for all possible \(P_{t,t}\), and where \(d\mu\) is the uniform Haar measure.

We stress again that this expression is exact. The unitaries in a unitary design are a representative set of points that are “evenly spaced” across the unitary group. With just a subset of the full group, we can evaluate complex expressions that would be otherwise intractable.

A surprising result about unitary designs is that they exist for all possible combinations of \(t\) and \(d\) 9. There are some known lower bounds for the number of unitaries required; for example, a 2-design in dimension \(d\) has at least \(d^4 - 2d^2 + 2\) elements 8, 9. However, actually finding the sets (and in particular, finding ones with minimal size), is a challenging problem 10, though very recently some constructions have been put forward 11.

Fun fact

Applying the elements of a unitary design to a fixed pure state produces a set of vectors that form a complex projective design 7. These are much like spherical designs, but they live in a complex vector space. If you’ve ever studied the characterization of quantum systems, you may have come across some special sets of measurements called mutually unbiased bases (MUBs), or symmetric, informationally complete positive operator valued measurements (SIC-POVMs). Both of these sets of vectors are complex projective 2-designs 5.


The vectors of the simplest SIC-POVM in dimension 2, plotted on a Bloch sphere.

Unitary \(t\)-designs in action

Unitary designs come into play in applications that require randomization, or sampling of random unitaries—essentially, they can be used as a stand-in for the Haar measure. The way in which the unitaries are used in the application may place restrictions on the value of \(t\) that is required; arguably the most common is the unitary 2-design.

While in general unitary designs are hard to construct, there are well known results for unitary 1-, 2-, and 3-designs based on familiar objects in quantum computing. Before we see what those are, let’s explore an important use case.

Average fidelity

A key application of unitary 2-designs is benchmarking quantum operations. Suppose we have a noisy quantum channel \(\Lambda\) that should perform something close to the unitary operation \(V\). What can we say about the performance of this channel?

One metric of interest is the fidelity. Consider the state \(|0\rangle\). In an ideal case, we apply \(V\) and obtain \(V|0\rangle\). But applying the channel \(\Lambda\) gives us something a little different. Since it’s noisy, we must consider the state as a density matrix. The action of \(\Lambda\) on our starting state is \(\Lambda(|0\rangle \langle 0|)\). If \(\Lambda\) was perfect, then \(\Lambda(|0\rangle \langle 0|) = V|0\rangle \langle 0|V^\dagger\), and the fidelity is

\[F(\Lambda, V) = \langle 0 | V^\dagger \cdot \Lambda(|0\rangle \langle 0|) \cdot V|0\rangle = 1.\]

In reality, \(\Lambda\) is not going to implement \(V\) perfectly, and \(F < 1\). More importantly though, all we’ve computed so far is the fidelity when the initial state is \(|0\rangle\). What if the initial state is something different? What is the fidelity on average?

To compute an average fidelity, we must do so with respect to the full set of Haar-random states. We usually generate random states by applying a Haar-random unitary \(U\) to \(|0\rangle\). Thus to compute the average over all such \(U\) we must evaluate

\[\bar{F}(\Lambda, V) = \int_{\mathcal{U}} d\mu(U) \langle 0 | U^\dagger V^\dagger \Lambda(U |0\rangle \langle 0| U^\dagger) V U |0\rangle.\]

This is known as twirling the channel \(\Lambda\). Computing the average fidelity in this way would be a nightmare—we’d have to compute the fidelity with respect to an infinite number of states!

However, consider the expression in the integral above. We have an inner product involving two instances of \(U\), and two instances of \(U^\dagger\). This means that the expression is a polynomial of degree 2 in both the elements of \(U\) and its complex conjugates—this matches exactly the definition of a unitary 2-design. This means that if we can find a set of \(K\) unitaries that form a 2-design, we can compute the average fidelity using only a finite set of initial states:

\[\frac{1}{K} \sum_{j=1}^K \langle 0 | U_j^\dagger V^\dagger \Lambda(U_j |0\rangle \langle 0| U_j^\dagger) V^\dagger U_j |0\rangle = \int_{\mathcal{U}} d\mu(U) \langle 0 | U^\dagger V^\dagger \Lambda(U |0\rangle \langle 0| U^\dagger) V U |0\rangle\]

This is great, but a question remains: what is the representative set of unitaries?

The Clifford group

A beautiful result in quantum computing is that some special groups you may already be familiar with are unitary designs:

  • the Pauli group forms a unitary 1-design, and

  • the Clifford group forms a unitary 3-design.

By the definition of designs, this means the Clifford group is also a 1- and 2-design.

The \(n\)-qubit Pauli group, \(\mathcal{P}(n)\), is the set of all tensor products of Pauli operations \(X\), \(Y\), \(Z\), and \(I\). The \(n\)-qubit Clifford group, \(\mathcal{C}(n)\), is the normalizer of the Pauli group. In simpler terms, the Clifford group is the set of operations that send Paulis to Paulis (up to a phase) under conjugation i.e.,

\[C P C^\dagger = \pm P^\prime, \quad \forall P, P^\prime \in \mathcal{P}(n), \quad C \in \mathcal{C}(n).\]

The Clifford group has some profoundly interesting properties and countless uses across quantum computing, from circuit compilation to error correcting codes. For a single qubit, the group is built from just two operations. One is the Hadamard:

\[H X H^\dagger = Z, \quad H Y H^\dagger = -Y, \quad H Z H^\dagger = X.\]

This clearly maps Paulis to Paulis (up to a phase). The other is the phase gate \(S\):

\[S X S^\dagger = Y, \quad S Y S^\dagger = -X, \quad S Z S^\dagger = Z.\]

If both \(H\) and \(S\) map Paulis to Paulis, then products of them do as well. In group theory terms, the single-qubit Clifford group is generated by \(H\) and \(S\). For example, consider the action of \(HS\):

\[(HS) X (HS)^\dagger = -Y, \quad (HS) Y (HS)^\dagger = -Z, \quad (HS) Z (HS)^\dagger = X.\]

Since \(Y = iXZ\), it is enough to specify Clifford operations by how they act on \(X\) and \(Z\). For a particular Clifford, there are 6 possible ways it can transform \(X\), namely \(\pm X, \pm Y\), or \(\pm Z\). Once that is determined, there are four remaining options for the transformation of \(Z\), leading to 24 elements total.

It takes some work, but you can take combinations of \(H\) and \(S\) and evaluate their action on \(X\) and \(Z\) (or look at their matrix representations) until you find all 24 unique elements. The results of this endeavour are expressed below as strings:

single_qubit_cliffords = [
 'H', 'S',
 'HS', 'SH', 'SS',
 'HSH', 'HSS', 'SHS', 'SSH', 'SSS',
 'HSHS', 'HSSH', 'HSSS', 'SHSS', 'SSHS',

To see for yourself how this set of unitaries is evenly distributed, try applying each of the Cliffords to the initial state \(|0\rangle\), and plot the resulting states on the Bloch sphere. You’ll find they are symmetric and evenly spaced; in fact, they are all eigenstates of \(X\), \(Y\), and \(Z\). Furthermore, under the full group action, the result is balanced in the sense that each eigenstate is obtained the same number of times.

The multi-qubit Clifford group can also be specified by only a small set of generators (in fact, only one more than is needed for the single-qubit case). Together, \(H\), \(S\), and CNOT (on every possible qubit or pair of qubits) generate the \(n\)-qubit group. Be careful though—the size of the group increases exponentially. The 2-qubit group alone has 11520 elements! The size can be worked out in a manner analogous to that we used above in the single qubit case: by looking at the combinatorics of the possible ways the gates can map Paulis with only \(X\) and \(Z\) to other Paulis.

An experiment

The whole idea of unitary designs may sound too good to be true. Can we really compute the exact average fidelity using just 24 operations? In this section, we put them to the test: we’ll compute the average fidelity of an operation first with experiments using a large but finite amount of Haar-random unitaries, and then again with only the Clifford group.

import pennylane as qml

# Scipy allows us to sample Haar-random unitaries directly
from scipy.stats import unitary_group

# set the random seed

# Use the mixed state simulator
dev = qml.device("default.mixed", wires=1)

Let’s set up a noisy quantum channel. To keep things simple, assume it consists of applying SX, the square-root of \(X\) gate, followed by a few different types of noise. First, write a quantum function for our ideal experiment:

def ideal_experiment():
    return qml.state()

Next, we apply some noise. We do so by making use of a relatively new feature in PennyLane called quantum function transforms. Such transforms work by modifying the underlying, low-level quantum tapes which queue the quantum operations. Suppose the noisy channel is composed of the following:

def noisy_operations(damp_factor, depo_factor, flip_prob):
    qml.AmplitudeDamping(damp_factor, wires=0)
    qml.DepolarizingChannel(depo_factor, wires=0)
    qml.BitFlip(flip_prob, wires=0)

Let’s create a transform that applies this noise to any quantum function after the original operations, but before the measurements. We use the convenient qfunc_transform() decorator:

def apply_noise(tape, damp_factor, depo_factor, flip_prob):
    # Apply the original operations
    for op in tape.operations:

    # Apply the noisy sequence
    noisy_operations(damp_factor, depo_factor, flip_prob)

    # Apply the original measurements
    for m in tape.measurements:

We can now apply this transform to create a noisy version of our ideal quantum function:

# The strengths of various types of noise
damp_factor = 0.02
depo_factor = 0.02
flip_prob = 0.01

noisy_experiment = apply_noise(damp_factor, depo_factor, flip_prob)(ideal_experiment)

The last part of the experiment involves applying a random unitary matrix before all the operations, and its inverse right before the measurements. We can write another transform here to streamline this process:

def conjugate_with_unitary(tape, matrix):
    qml.QubitUnitary(matrix, wires=0)

    for op in tape.operations:

    qml.QubitUnitary(matrix.conj().T, wires=0)

    for m in tape.measurements:

Finally, in order to perform a comparison, we need a function to compute the fidelity compared to the ideal operation.

from scipy.linalg import sqrtm

def fidelity(rho, sigma):
    # Inputs rho and sigma are density matrices
    sqrt_sigma = sqrtm(sigma)
    fid = np.trace(sqrtm(sqrt_sigma @ rho @ sqrt_sigma))
    return fid.real

Let’s now compute the average fidelity, averaging over 50000 Haar-random unitaries:

n_samples = 50000

fidelities = []

for _ in range(n_samples):
    # Select a Haar-random unitary
    U = unitary_group.rvs(2)

    # Apply transform to construct the ideal and noisy quantum functions
    conjugated_ideal_experiment = conjugate_with_unitary(U)(ideal_experiment)
    conjugated_noisy_experiment = conjugate_with_unitary(U)(noisy_experiment)

    # Use the functions to create QNodes
    ideal_qnode = qml.QNode(conjugated_ideal_experiment, dev)
    noisy_qnode = qml.QNode(conjugated_noisy_experiment, dev)

    # Execute the QNodes
    ideal_state = ideal_qnode()
    noisy_state = noisy_qnode()

    # Compute the fidelity
    fidelities.append(fidelity(ideal_state, noisy_state))

fid_mean = np.mean(fidelities)
print(f"Mean fidelity = {fid_mean}")


Mean fidelity = 0.9867904283557791

Now let’s repeat the procedure using only Clifford group elements. First, we write a quantum function that performs a Clifford operation (or its inverse) based on its string representation.

def apply_single_clifford(clifford_string, inverse=False):
    for gate in clifford_string:
        if gate == 'H':
            sign = -1 if inverse else 1
            qml.PhaseShift(sign * np.pi/2, wires=0)

Next, we write a transform that applies a Clifford in the context of the full experiment, i.e., apply the Clifford, then the operations, followed by the inverse of the Clifford.

def conjugate_with_clifford(tape, clifford_string):
    apply_single_clifford(clifford_string, inverse=False)

    for op in tape.operations:

    apply_single_clifford(clifford_string, inverse=True)

    for m in tape.measurements:

You may have noticed this transform has exactly the same form as conjugate_with_unitary from above. Only the input type has changed, since the application of Cliffords here is specified by their string representation.

It’s now time to run the experiments:

fidelities = []

for C in single_qubit_cliffords:
    conjugated_ideal_experiment = conjugate_with_clifford(C)(ideal_experiment)
    conjugated_noisy_experiment = conjugate_with_clifford(C)(noisy_experiment)

    ideal_qnode = qml.QNode(conjugated_ideal_experiment, dev)
    noisy_qnode = qml.QNode(conjugated_noisy_experiment, dev)

    ideal_state = ideal_qnode()
    noisy_state = noisy_qnode()

    fidelities.append(fidelity(ideal_state, noisy_state))

Let’s see how our results compare to the earlier simulation:

clifford_fid_mean = np.mean(fidelities)

print(f"Haar-random mean fidelity = {fid_mean}")
print(f"Clifford mean fidelity    = {clifford_fid_mean}")


Haar-random mean fidelity = 0.9867904283557791
Clifford mean fidelity    = 0.9867892194401247

Incredible 🤯 🤯 🤯 We were able to compute the average fidelity using only 24 experiments. Furthermore, the mean fidelity obtained from the Clifford experiments is exact; even with 50000 Haar-random experiments, we see deviations starting a few decimal places in. Consider the resources that would be saved if you were actually implementing this in a lab! It’s not hard to see why the Clifford group plays such an important role in characterization procedures.


In this demo, we’ve barely scratched the surface of designs and their applications in quantum computing. While benchmarking is a key application area, there are many others. The Pauli group as a unitary 1-design has applications in the construction of private quantum channels 12. Approximate unitary \(t\)-designs (where the equality in the definition is replaced by approximately equal up to some finite precision) are also of interest, as there ways to construct them that are more efficient than those of exact designs 6. In particular, it has been shown that approximate complex projective 4-designs have applications to the state discrimination problem 4.

Furthermore, unitary designs are not the only designs that you’ll encounter in quantum computing. The familiar Hadamard gate is just a 2-dimensional example of a broader family of Hadamard designs, on which there has been extensive research 13. Some sets of mutually orthogonal Latin squares have a direct correspondence with mutually unbiased bases, which are optimal quantum measurements 14, as well as complex projective designs; and Latin squares themselves have direct correspondence with affine and projective planes, bringing us full circle back to the Fano plane from which we began.


An affine plane, Hadamard matrix, and a depiction of mutually orthogonal Latin squares.



C. J. Colbourn and J. H. Dinitz (2006) Handbook of Combinatorial Designs, Second Edition. Chapman & Hall/CRC.


P. Delsarte, J.M. Goethals, J.J. Seidel (1977) Spherical Codes and Designs. Geometriae Dedicata 6 363-388.


R. H. Hardin and N. J. A. Sloane (1992) New spherical 4-designs. Discrete Mathematics, 106-107 (255-264). (PDF).


A. Ambainis and J. Emerson (2007) Quantum t-designs: t-wise independence in the quantum world. Twenty-Second Annual IEEE Conference on Computational Complexity 129-140. (arXiv).


A. Klappenecker and M. Roetteler (2005) Mutually unbiased bases, spherical designs, and frames. Proceedings of SPIE Vol. 5914.


C. Dankert, R. Cleve, J. Emerson, and E. Levine (2009) Exact and Approximate Unitary 2-Designs: Constructions and Applications. Phys. Rev. A 80, 012304. (arXiv).


C. Dankert (2005) Efficient Simulation of Random Quantum States and Operators. MSc Thesis, University of Waterloo. (arXiv).


D. Gross, K. Audenaert, and J. Eisert (2007) Evenly distributed unitaries: on the structure of unitary designs. J. Math. Phys. 48, 052104. (arXiv).


A. Roy and A. J. Scott (2009) Unitary designs and codes. Des. Codes Cryptogr. 53 13-31. (arXiv).


E. Bannai, M. Nakahara, D. Zhao, and Y. Zhu (2019) On the explicit constructions of certain unitary t-designs. J. Phys. A: Math. Theor. 52 495301. (arXiv).


Y. Nakata et al. (2021) Quantum circuits for exact unitary t-designs and applications to higher-order randomized benchmarking. (arXiv).


A. Ambainis, M. Mosca, A. Tapp, and R. de Wolf (2000) Private Quantum Channels. Proc. 41st FOCS, 547-553. (PDF).


J. Seberry and M. Yamada (1992) Hadamard matrices, sequences, and block designs. Contemporary Design Theory – A Collection of Surveys (D. J. Stinson and J. Dinitz, Eds.), John Wiley and Sons, 431-560. (PDF).


M. Gaeta, O. Di Matteo, A. B. Klimov, and H. de Guise (2014) Discrete phase-space approach to mutually orthogonal Latin squares. J. Phys. A: Math. Theor. 47 (43) 435303. (arXiv).

About the author

Olivia Di Matteo

Olivia Di Matteo

Olivia is an Assistant Professor in the Department of Electrical and Computer Engineering at The University of British Columbia. Her research interests are quantum compilation, software, and algorithms.

Total running time of the script: ( 3 minutes 40.181 seconds)

Gallery generated by Sphinx-Gallery