I am currently working through some papers related with approximations of more general quantum channels such as amplitude and phase damping channels to Pauli channels. The reason to do so is so that the Gottesman-Knill theorem is fulfilled and efficient Monte Carlo simulations can be performed for Quantum Error Correction codes.

In such reading, I reach to some papers that use the so-called Twirling in order to justify this channel approximation, but I have not found much literature specifically talking about the topic, as I have found the information to be scattered and not so well distinguished. My principal doubt is the difference between Pauli and Clifford twirls, and if both of them can be applied to the same channel without losing too much on the approximation. For example:

I am wondering if someone can give me some insight about this topic, or provide specific literature about the topic of Twirling in Quantum Information Theory. Also ideas or references to the use of those techniques as approximations for error correction is interesting for me and welcome.


1 Answer 1



Denoting the Haar measure of some function $f\left(x\right)$ over $d$-dimensional unitaries as $\int_{\mathrm U\left(d\right)}f\left(x\right)d\mu\left(x\right)$, twirling some arbitrary channel $\varepsilon$ can be defined as the operation $\varepsilon \mapsto\int_{\mathrm U\left(d\right)}U^\dagger\varepsilon U dU$, which, when $\varepsilon$ is acting on some density matrix $\rho$ gives $$\rho\mapsto\int_{\mathrm{U}(d)} U^{\dagger} \varepsilon\left(U \rho U^{\dagger}\right) U d U.$$

While the above applies to intergrating over the $d$-dimensional unitary group, the same idea can be applied to (weighted) gatesets $\mathrm D_w$, consisting of elements of unitaries $U_i$ along with associated probabilities $p_i$ such that the weighted gate-set channel is $\mathcal U_{\mathrm D_w}:\rho\mapsto\sum_{U_i\in\mathrm D}p_iU_i\rho U_i^\dagger$. A special case of this are the unweighted gatesets $\mathrm D$ where $p_i = \frac{1}{\left|\mathrm D\right|}\,\forall\, i$. This allows us to define the Twirling operation over a gateset as the operation $\varepsilon\mapsto\sum_{U_i\in\mathrm D}p_iU^\dagger_i\varepsilon U_i$.

The Pauli twirling operation is then defined as $\varepsilon\mapsto\sum_{P_i\in\mathrm P_n}p_iP_i\varepsilon P_i$ where $\mathrm P_n$ is the Pauli group over $n$ qubits, while the Clifford twirl can either be defined as $\varepsilon\mapsto\sum_{U_i\in\mathrm C_n}p_iU^\dagger_i\varepsilon U_i$ or $\varepsilon\mapsto\sum_{U_i\in\mathrm C_1^{\otimes n}}p_iU^\dagger_i\varepsilon U_i$.1

Unitary Designs

There is a difference between what these two twirls do and it lies in the theory of unitary t-designs2: a set $\mathrm D_w$ is a $t$-design if and only if $$\sum_{U\in \mathrm D}\left(U^\dagger\right)^{\otimes t}\otimes \left(U^T\right)^{\otimes t} = \int_{\mathrm U\left(d\right)}\left(V^\dagger\right)^{\otimes t}\otimes \left(V^T\right)^{\otimes t}\,dV$$ and if $\mathrm D_w$ is a $t$-design, it's also a $\left(t-1\right)$-design for any positive integer $t$. This is equivalent to the statement that $$\sum_{U \in D}(U)^{\otimes t} \rho\left(U^{\dagger}\right)^{\otimes t}=\int_{U(d)}(V)^{\otimes t} \rho\left(V^{\dagger}\right)^{\otimes t} d V,$$ which gives the $t^{th}$ statistical moment of $\rho$.

Using Schur's Lemma

For a pure state (density matrix) in 2 dimensions $\rho$, $\int_{\mathrm U\left(d\right)} V\rho V^\dagger\,dV = \frac{1}{2}I$ and so, this fully depolarises (i.e. maximally mixes) the state and as the single qubit Pauli twirl is also the depolarisation channel, the Pauli group is a $1$-design. However, by Schur's lemma, it's not a $2$-design. In other words, as explained in section IIA of Gross, Audenaert and Eisert, twirling essentially writes the state in terms of a sum of irreducible subspaces, which, for a 2D 2-design, are the symmetric $\frac{1}{2}\left(I+SWAP\right)$ and antisymmetric $\frac{1}{2}\left(I-SWAP\right)$ subspaces and this is not what the Pauli twirl does. However, it is what the Clifford twirl does. Furthermore, as shown both here and here, the Clifford group forms a 3-design.

Randomised Benchmarking

This is useful and even relatively important from a quantum computing perspective when it comes to benchmarking - it turns out that local random circuits are approximate designs and can be created in polynomial time


This all relates to the papers linked in the question by the process of vectorisation and the property that $|A B C\rangle \rangle=A \otimes C^T|B\rangle \rangle$, which when $\varepsilon$ is written as a sum of Krauss operators gives $$\int_{\mathrm{U}(d)} U^{\dagger} \varepsilon\left(U \rho U^{\dagger}\right) U d U \mapsto\sum_{K} \int_{U(d)}\left(U^{\dagger} \otimes U^T\right) \cdot\left(K \otimes K^{*}\right) \cdot\left(U \otimes U^{*}\right)|\rho\rangle \rangle d U.$$

If we then go a step further and start with a random state, we can effectively integrate over $\rho$ and repeat the vectorisation process to give $$\int_{\mathrm{U}(d)} U^{\dagger} \varepsilon\left(U \rho U^{\dagger}\right) U d U \mapsto\int_{U(d)} U^{\dagger} \otimes U^T \otimes U^T \otimes U^{\dagger} d U \sum_{K}\left|K \otimes K^* \right\rangle \rangle$$ and we now have something that looks more like a 2-design.

Overall, we've gone from twirling over some channel to an operation that looks something similar to 2-design acting on the vectorisation of the Krauss operator form of that channel and finally...

The difference between Pauli and Clifford Twirls

Rewriting $K = \sum_j|u_j^{\left(K\right)}\rangle\langle v_j^{\left(K\right)}|$ gives $K\otimes K^* = \sum_{j, l}|u_j^{\left(K\right)}\rangle|u_l^{*\left(K\right)}\rangle\langle v_j^{\left(K\right)}|\langle v_l^{*\left(K\right)}|$ and so $|K\otimes K^*\rangle\rangle = \sum_{j, l}| v_j^{\left(K\right)}\rangle| v_l^{*\left(K\right)}\rangle|u_j^{\left(K\right)}\rangle|u_l^{*\left(K\right)}\rangle$

From what's above, the Clifford twirl then permutes this to get $\sum_{j, l}|u_l^{*\left(K\right)}\rangle| v_j^{\left(K\right)}\rangle| v_l^{*\left(K\right)}\rangle|u_j^{\left(K\right)}\rangle$, sums the projections onto the symmetric and anti-symmetric subspaces and unpermutes again. Overall, information about the second statistical moment of the noise channel can be learned from repeated measurements.

By comparison, the Pauli twirl generally doesn't do this. It does, in the case of a single qubit, at least give the diagonal elements of the Process matrix, as per the paper they refer to, which happens to be the parameters they're looking for.

1 I'm not sure off the top of my head if there's a diffrerence between $\mathrm C_n$ and $\mathrm C_1^{\otimes n}$, although it seems clear that $\mathrm C_1^{\otimes n}\subset\mathrm C_n$

2 Also known as unitary k-designs

  • $\begingroup$ What justifies the following implication? $\varepsilon \mapsto\int_{\mathrm U\left(d\right)}U^\dagger\varepsilon U dU$, which, when $\varepsilon$ is acting on some density matrix $\rho$ gives $$\rho\mapsto\int_{\mathrm{U}(d)} U^{\dagger} \varepsilon\left(U \rho U^{\dagger}\right) U d U.$$ $\endgroup$
    – Hans
    Commented Oct 31, 2022 at 1:49
  • 2
    $\begingroup$ Just commenting on note 1, there is a difference between $ \mathrm{C}_n $ and $ \mathrm{C}_1^{\otimes} $ because $ \mathrm{C}_1^{\otimes} $ does not contain $ CNOT $ or any other entangling gates, so $ \mathrm{C}_1^{\otimes n} $ for $ n \geq 2 $ is small and only a 1-design like the Pauli group, while $ \mathrm{C}_n $ is large and a 3-design $\endgroup$ Commented Jan 3 at 23:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.