Arxiv:2006.03086V2 [Quant-Ph] 12 Nov 2020

Augmented ﬁdelities for single qubit gates

Filip Wudarski,1, 2, ∗ Jeffrey Marshall,1, 2, † Andre Petukhov,1, 3 and Eleanor Rieffel1 1QuAIL, NASA Ames Research Center, Moffett Field, California 94035, USA 2USRA Research Institute for Advanced Computer Science, Mountain View, California 94043, USA 3Google Inc., Santa Barbara, California 93117, USA‡ An average gate fidelity is a standard performance metric to quantify deviation between an ideal unitary gate transformation and its realistic experimental implementation. The average is taken with respect to states uniformly distributed over the full Hilbert space. We analytically (single-qubit) and numerically (two-qubit) show how this average changes if the uniform distribution condition is relaxed, replaced by parametrized distributions – polar cap and von Mises-Fisher distributions – and how the resulting fidelities can differentiate certain noise models. In particular, we demonstrate that Pauli channels with different noise rates along the three axes can be faithfully distinguished using these augmented fidelities.

I. INTRODUCTION less, they are all valuable benchmarking tools that allow researchers to capture and quantify some of the most rel- Impressive progress in quantum technologies has taken evant aspects of the behavior of quantum devices and quantum computing from a theoretical framework to an their building blocks - qubits and gates. experimental playground, where basic proof-of-principle Special attention should be given to the average fidelity concepts can be tested and verified. The most promi- - currently the figure of merit of performance metric. It nent recent example of the latter is the demonstration describes an error between an ideal and experimental re- of quantum advantage [1]: that even the imperfect cur- alization of a gate, and is averaged over all possible states rently available quantum hardware can perform tasks in- uniformly distributed in the Hilbert space (according to tractable for the most powerful supercomputers. Further the Haar measure). Unlike the diamond norm and the advances toward more capable and robust quantum hard- min fidelity, the average gate fidelity can be efficiently ware depend on gaining a better understanding of un- estimated through protocols like randomized benchmark- derlying physical effects, including the characterization ing (RB) [9–12], cross-entropy benchmark [1, 13] or di- of noise in actual quantum hardware. rect fidelity estimation [14–16]. However, being only a In theory, quantum process tomography (QPT) [2] can single parameter the metric cannot distinguish various be used to exhaustively benchmark a quantum device, noise models; it reports only a single element of the pro- identifying all of its imperfections. QPT reconstructs cess matrix, the χ element, which is associated with n n 00 the full process matrix χ (of size 2 × 2 , where n is the depolarizing rate. That is, it effectively identifies all the number of qubits), a matrix that encodes complete channels as depolarizing channels. information about underlying quantum transformation (including unwanted effects caused by noise). Unfortu- In this article, we propose to relax the uniform distri- nately, QPT scales exponentially with system size, be- bution condition and introduce augmented fidelity met- coming impractical for systems larger than a few qubits rics via parametrized distributions. In particular, we an- [3]. Intuitively it seems plausible that well-controlled sys- alytically investigate what information about noise pro- tems will have only few dominating error sources - i.e. the cesses can be extracted from average fidelity with respect χ matrix will be sparse up to some accuracy. Therefore, to a von Mises-Fisher distribution (a normal distribu- lower parameter approximations and associated metrics tion in directional statistics) and a polar cap distribu- and protocols that could assess the performance of quan- tion, i.e. a uniform distribution over a subset of states tum devices, and identify the crucial elements of χ, are parametrized by polar angle (colatitude). This approach arXiv:2006.03086v2 [quant-ph] 12 Nov 2020 promising approaches to noise characterization and quan- augments standard uniform-average fidelity by adding tum hardware benchmarking. Currently, the most com- extra tunable parameters to the metric, that depend on mon figures of merit considered are: diamond norm [4], distribution properties. This work provides a partial so- minimum fidelity [5] or average fidelity [6–8]. All three lution to the problem posed by Nielsen in [7] regarding techniques yield a single value that characterizes devia- gate fidelities over non-uniform distributions. Addition- tions from the ideal transformation. A parameter count ally, we derive the maximal spread in fidelity the value shows that all three methods provide only limited infor- (the difference between minimum and maximum attain- mation about the process matrix for a device. Neverthe- able values) and provide error bars (based on the standard deviation derived from the considered distributions) for processes that share the same depolarizing rate. In particular, we show how to identify noise biases [17, 18] ∗ fi[email protected] in Pauli channels. Our analysis focuses mainly on sin- † jeff[email protected] gle qubit gates, where analytical formulas are derived, ‡ Permanent address but we also open the discussion for similar approaches 2 for multi-qubit gates, in particular we numerically show A. Polar Cap Distribution local distribution for two-qubit gates and how they differ from the uniform-average fidelity. First let us define a single state gate fidelity as

where U is the investigated gate, EU its CPTP (imper- Consider the fidelity between a state transformed ac- fect) realization and |ψi is a state upon which the gate cording to a given unitary (gate) transformation (ideal acts. Since measuring and computing Eq. (4) for all pos- action) U and the same state transformed with noisy re- sible states is infeasible, one usually reports the average alization of U, which is a completely positive and trace fidelity value, which is taken over |ψi distributed uni- preserving (CPTP) map EU . The average fidelity of the formly in the entire Hilbert space. noisy realization is the average of this fidelity over all Following derivation from [6] we define the restricted initial (pure) states distributed uniformly in the entire ¯ ¯ Hilbert space. Bowdrey et al. [6] introduced a simple for- average gate fidelity FΘ(U, EU ) ≡ FΘ as mula for calculating average gate fidelity for single qubit Z ¯ gate, which was later generalized to multi-qubit gates and FΘ = F|ψihψ|(U, EU )dΩ = connected with entanglement fidelity [7,8]. The average gate fidelity (from now on referred as “uniform-average Z Θ Z 2π 3 1 X σj fidelity”) for n-qubit gates is therefore given by = Tr U c (θ, φ) U † × S(Θ) j 2 θ=0 φ=0 j=0 P22n−1 † † 2n Tr UV U EU (Vk) + 2 3 ¯ ¯ † k=0 k X σk F (U, EU ) = F (U ◦ EU ) = , × E c (θ, φ) sin θdφdθ, (5) 22n(2n + 1) U k 2 (1) k=0 where Vk are traceless unitary matrices forming an or- where S(Θ) := 2π(1 − cos Θ) is the solid angle for thonormal basis with respect to Hilbert-Schmidt inner normalization of the distribution, and cj(θ, φ) are pure † n product (Tr(VkVj ) = 2 δkj, Tr(Vk) = 0 for k = state’s Bloch vector coefficients: c0(θ, φ) = 1, c1(θ, φ) = 2n 1,..., 2 − 1 and V0 = 1). By writing the composed sin θ cos φ, c2(θ, φ) = sin θ sin φ and c3(θ, φ) = cos θ. map These correspond respectively to the identity matrix σ0, and the three x, y, z Pauli matrices σ1,2,3. Note that 22n−1 Eq. (4) assumes coordinate system where θ = 0 corre- † X † E(ρ) ≡ U ◦ EU (ρ) = χklVkρVl , (2) sponds to |0i state, and the polar cap distribution is cen- k,l=0 tered around it. However, transformation to an arbitrary ˜ central state is straightforward via rotation |ψi = UR |ψi. where the χ matrix is called the process matrix for E, it Now performing the integration leaves us with can be demonstrated that the average fidelity only de- 2 Θ 2 pends on the χ00 element corresponding to a “depolariz- 1 (2 + cos Θ) sin X F¯ = + 2 Tr Uσ U †E σ + ing” rate, a unitary invariant element, i.e. Θ 2 12 k U k k=1 2nχ + 1 1 + cos Θ + cos2 Θ F¯(U † ◦ E ) = 00 . (3) + Tr Uσ U †E σ + U 2n + 1 12 3 U 3 1 + cos Θ + Tr Uσ U †E σ . (6) The above formula demonstrates inability to distinguish 8 3 U 0 different noise processes that differ in other χ parameters, and this limitation stems from the averaging procedure It is transparent that for Θ = π one recovers result and properties of the Haar measure. In order to have a Eq. (1) for uniform distribution over the entire Hilbert more sensitive metric, we propose to use several differ- space. Expressing the composed gate-noisy-gate map † ent initial state distributions. In addition to averaging E = U ◦ EU in the form Eq. (2), one can show that over all pure states distributed uniformly, we explore two 2 models: i) uniform distribution parametrized by a po- 1 X Tr σkE σk = χ0,0 − χ3,3, (7) lar angle Θ ∈ [0, π], which we call polar cap distribution 4 k=1 (e.g. for Θ = π/2 we have a distribution over the north- 1 ern hemisphere, while for Θ = π we recover the entire Tr σ Eσ = χ − χ − χ + χ , (8) 2 3 3 0,0 1,1 2,2 3,3 space distribution), and ii) von Mises-Fisher distribution 1 around a state |ψi (without loss of generality we can fix Tr σ3E σ0 = Re(χ0,3), (9) it to |0i) parametrized by “variance” parameter κ. From 8 now on we will focus only on single qubit gates, and will where we used the properties of χ matrix that guarantee refer to the investigated fidelities as augmented fidelities, the CPTP condition, in particular Re(χ0,3) = −Im(χ1,2). leaving extensions to multi-qubit systems to later work. The above equations (6-9) are correct for distributions 3 centered around the North Pole (i.e. state |0i). However, if the center is selected to be one of σ1 or σ2 eigenstates, then Eqs. (6-9) will experience a permutation 3 ↔ 1 or 3 ↔ 2, respectively. For more generic central state the polar cap average fidelity F¯Θ will in principle depend non- trivially on all χ matrix entries, apart from the imaginary parts of χ0,k for k = 1, 2, 3.

B. von Mises-Fisher Distribution

In directional statistics [19] von Mises-Fisher distribution [20] is a continuous probability distribution on the N-dimensional sphere (see Fig.1), and plays a similar role to a normal distribution on a flat manifolds. Since pure states of qubits live on a Bloch sphere, it is more natural to exploit directional statistics and use von Mises-Fisher distribution as a distribution for initial state preparation than standard normal distribution. For a 2- sphere the probability density function is given by κ p(~x,~µ,κ) = eκ~µ·~x, (10) 4π sinh κ Figure 1. Visualization of von Mises-Fisher distribu- where ~µ,~x are normalized vectors, and κ is similar to tions. Top: 50000 random states selected according to von inverse of the variance - for κ → 0 it converges to a Mises-Fisher distribution around |0i state (red point) with uniform distribution, while for κ → +∞ it is localized left: κ = 10 and right: κ = 100. Bottom: Fidelity between around ~µ, which resembles mean value in the standard |0i state and states |ψi that are drawn from von Mises-Fisher distribution (i.e. between the “red point” state and states rep- normal distribution. Note that Eq. (10) is a special case resented by “blue points”) expressed in the form of histograms of Kent distribution [21]. (orange: κ = 10, red (and inset): κ = 100). If we fix ~µ = (0, 0, 1) (see Fig.1), which corresponds to the distribution around the North Pole, i.e. around the |0i state, the averaging of the fidelity is with respect to maximal and minimal value of fidelity attainable for these the following normalized surface element dΩ processes, and obtain error bars based on the standard Z Z π Z 2π κ sin θ deviation. dΩ = eκ cos θdθdφ = 1. (11) 4π sinh κ The uniform-average fidelity Eq. (3) depends purely θ=0 φ=0 on a single element of the χ (process) matrix of the com- † † Integrating Tr U |ψi hψ| U EU (|ψi hψ|) with respect to posed noisy process E = U ◦ EU . Since the two aug- von Mises-Fisher surface element one arrives at mented fidelities, Eqs. (6), (12), are influenced by χ ma- 2 trix elements present in Eqs. (7-9) (for distributions cen- 1 κ coth κ − 1 X F¯ = + Tr Uσ U †E σ + tered around |0i) it suffices to consider a matrix of the κ 2 4κ2 k U k k=1 following form 2 − 2κ coth κ + κ2 + Tr Uσ U †E σ +  χ ·· χ  4κ2 3 U 3 00 03  · χ11 −iχ03 ·  κ coth κ − 1 † χ = , (13) + Tr Uσ U E σ , (12)  · iχ03 χ22 ·  4κ 3 U 0 χ03 ·· χ33 which depends on the same χ matrix elements as in the polar cap case, i.e. contribution as in Eqs. (7-9). More- where the χij elements are real by hermiticity. The dots over, the same reasoning holds for distributions around in Eq. (13) indicate these elements are arbitrary for our different central states (e.g. |+i). purposes, as they are absent in Eqs. (6) and (12) (up to χ being a genuine process matrix). For simplicity, we set these elements to zero. With this, constraints to impose III. RESULTS P CPTP conditions are k χkk = 1, and χ ≥ 0 which 2 2 translates into χ00χ33 ≥ χ03 and χ11χ22 ≥ χ03. Now we demonstrate through our analytic expressions In order to investigate the spread of fidelities, we need how certain noise channels that are indistinguishable un- to minimize (maximize) the average fidelity, according to √ 2 der the standard average fidelity, Eq. (1), can be dis- Eqs. (6) and (12). First let us introduce p = (1− χ00) . tinguished via augmented fidelities, Eqs. (6), (12). Ad- We can analytically determine the minimal (maximal) √ ditionally, we analytically derive the spread between the value, which is achieved for χ33 = p, χ11 = χ22 = p − p 4

lows for larger spread. Therefore this analysis provides a trade-off between faithful state preparation (more localized distribution) and sensitivity to noise manifested by fidelities. As we have emphasized, while standard benchmarking techniques (such as randomized benchmarking (RB) [9]) probe the uniform-average fidelity, identifying the χ00 element, but do not detect properties of noise pres- ence encoded elsewhere in the χ matrix. Additionally, it is well known that a twirling protocol (see for example [9, 16] and references therein), which is a mathemati- cal justification for RB methods, is insensitive to initial state distribution; since it averages over the entire unitary group, it transforms each channel into a depolarizing one. Thus, after performing a twirling protocol, one is left with a contribution stemming only from the χ00 element (i.e. depolarizing rate). It is attractive, yet in- correct, to equate every noise process with depolarizing noise, simplifying the entire noise analysis to this averaged case. For each distribution one can also determine variance σ2(F ) of the fidelity Z Z 2 2 2 σ (F ) = F|ψihψ|(U, EU ) dΩ − F|ψihψ|(U, EU )dΩ , Figure 2. Analytical augmented fidelities for processes (14) that share the same χ00 = 0.985 element (depolarizing where F|ψihψ|(U, EU ) is given by Eq. (4) and dΩ is a sur- rate) corresponding to 99% fidelity. Red (dashed) and face element related to the underlying distribution of blue (dotted) curves bound the region between the minimal initial states |ψi. In [22] the formula for the uniform and maximal fidelities of a noisy process. Green region (in- side) corresponds to diagonal χ matrices (i.e. Pauli channels). distribution over the entire space was provided. It is Top: augmented fidelity over polar cap distribution, Eq. (6), also straightforward to calculate this in the case of polar parametrized by polar angle Θ, bottom: augmented fidelity, cap and von Mises-Fisher distributions (see Appendix), Eq. (12), with respect to von Mises-Fisher distribution (nor- which depends not only on the θ and κ parameters, but in mal distribution in directional statistics) as a function of κ general on all χ matrix elements. The variance becomes parameters (corresponding to the inverse of variance in stan- smaller for more localized distributions (i.e. as κ → ∞ dard statistics for normal distribution). Error bars indicate and Θ → 0) and reaches its largest value for distributions standard deviation of fidelities (color coded) for minimal and close to uniform over all states. maximal values. Black (dashed) line represents the depolar- In Fig.2 we report standard deviation error bars (i.e. izing channel that would be identified through RB. from Eq. (14)) for the minimal (blue) and maximal (red) channels. This corresponds to the spread in fidelity values over the pure states of the distributions, for these ex- and χ03 = −χ11 (respectively χ03 = χ11). A similar treme channels, and depends on all elements of Eq. (13). analysis can be performed for a Pauli channel, i.e. with Our analysis shows that we have two independent diagonal χ matrix. In that case, the minimal fidelity sources of fidelity deviations. One related to a statis- values are achieved for χ11 = 1 − χ00 (or χ22 = 1 − χ00) tical distribution, and the second one to noise process. and maximal for χ33 = 1 − χ00. In principle they can either benefit (increase fidelity) or The spread between minimal and maximal fidelities hamper (decrease) the performance. It is important to for noise models with the uniform-average fidelity of 99% properly identify their impact and origins. (corresponding to χ00 = 0.985) is depicted in Fig.2. In the limit of κ → 0 and Θ = π we recover results for average over all states. Note, that in that case spread A. Noise bias completely disappears and it is impossible to differentiate between various noise models (i.e. sharing the same Recently the problem of noise bias in Pauli channels, χ00 element, but otherwise having distinct elements of i.e. having diagonal χ matrix, has attracted consider- the χ matrix) solely based on fidelity Eq. (1). On the able attention, especially in the field of error correction other end, when κ → ∞ and Θ → 0 both distribu- [17, 18, 23]. The noise bias (in Z direction) is defined tions tend to a localized state |0i and fidelities display for Pauli channels, as ηZ = χ33/(χ11 + χ22), and informs the largest spread. Moreover, the spread increases with us which Pauli error is more prominent. Here, we pro- decreasing χ00, i.e. higher uniform-average infidelity al- pose to identify Pauli errors by looking at the average 5 gate fidelity either with polar cap or von Mises-Fisher yield distinct values. Note that changing the center is distribution centered around eigenstates of X,Y and Z due to a special type of single-qubit rotation transforma- to the eigenvalue +1. In TableI we report values of χkk tion, and therefore could be also used as a benchmark for elements for two different Pauli channels (which we call single qubit rotation gates (the profile of the curves are PC1 and PC2) that take the same value of χ00 = 0.985, qualitatively the same as in Fig.3). i.e. corresponding to uniform-average fidelity of 99%, and the same noise bias ηZ = 1/14. The results for these B. 2-qubit Case Table I. We consider special class of two-qubit fidelities with ID χ11 χ22 χ33 ηZ 1 F (U, E ) = PC1 0.012 0.002 0.001 14 U 1 PC2 0.010 0.004 0.001 † 14 Tr U |ψ1i hψ1| ⊗ |ψ2i hψ2| U EU (|ψ1i hψ1| ⊗ |ψ2i hψ2|) , (15) channels are displayed in Fig.3. Note, that for distri- and now taking the average only over the local distributions Z F¯(U, EU ) = F (U, EU )dΩ1dΩ2, (16)

where dΩk corresponds to the surface element associated with qubit k. If both |ψ1i and |ψ2i are uniformly distributed over the entire space, then Eq. (16) yields

1 F¯ = 1 + 8χ + 2(χ + χ + χ + 9 00,00 01,01 02,02 03,03 (17) χ10,10 + χ20,20 + χ30,30) ,

where we use map in the form

3 † X U ◦ EU (ρ) = χkl,mn σk ⊗ σl ρ σm ⊗ σn . (18) k,l,m,n=0

Note that Eq. (17) depends not only on the χ00,00 element as in the case of full space uniform distribution, but also on the other (diagonal) elements of the χ matrix. For local distributions of polar cap and von Mises- Fisher type we perform numerical analysis based on 1673 random process matrices [24] with the same unitarily invariant element χ00,00 = 0.985 (corresponding to 99% unifrom-average fidelity). The results are displayed in Fig.4, which show variations in the augmented fidelities through changing the distribution parameters (here two Figure 3. Augmented fidelity of Pauli channels. Differ- angles Θ1,2, or von Mises-Fisher inverse variance κ1,2). ent channels (line styles) characterized by values in TableI with polar cap (top panel) and von Mises-Fisher (bottom) distributions. Black dashed line corresponds to the depolarizing channel with fidelity 99%. Color coded are distribution IV. CONCLUSIONS centered around different initial states: (red) |+i state, (blue) |y+i (eigenvector of Pauli Y to eigenvalue +1, and (green) |0i. In this article we examined and calculated augmented fidelities of noisy single qubit gates by averaging over dif- butions centered around |0i (green color in Fig.3) we ferent initial state distributions. In particular we focused see only a single line style. This is due to the fact that on two models – polar cap and von Mises-Fisher distri- augmented fidelity fails to discriminate Z bias in this pro- butions, parametrized by a polar angle Θ and variance- tocol. However, if one changes the center of the distribu- related parameter κ, respectively. The introduced meth- tion, the difference becomes clear and two Pauli channels ods augment the uniform-average fidelity (strongly based 6

ment between theory and experiment, especially in case when the underlying distribution displays anisotropic properties. As demonstrated in Figs.2,3, the greatest spread and smallest error bars in our augmented fidelity metrics between noise channels occurs for a point distribution (i.e. at Θ = 0 = κ−1). As discussed in the previous para- graph, this is not a realistic scenario to probe, and this effectively sets a lower bound on Θ and κ−1 based on experimental capabilities. As Θ and κ−1 are increased however, the ability to discriminate channels decreases. Therefore it is important to be able to determine in experiment what reasonable lower bounds are on these quantities. Lastly we mention that the introduced figures of merit can be measured experimentally with current technology as a slight modification to current techniques. Indeed, similar to the uniform-average fidelity (see [6]), one may restrict to performing state tomography along 6 initial states, i.e. in the ±x, ±y, ±z directions on the Bloch sphere. This would of course mean that state preparation and measurement errors are included in EU , though certain techniques may allow one to mitigate these effects �� (see for example [25, 26]). However, the number of measurements required in this case is larger than the number Figure 4. Heatmaps of two-qubit augmented fideli- of measurements for process tomography (6 (states) × 3 ties for local polar cap distribution (top panel) and (measurements per state) =18 for the first method and von Mises-Fisher distribution (bottom). The left/right 12 for QPT). Therefore, we leave construction of a more panel(s) show the minimum/maximum augmented fidelity en- efficient protocol in the single qubit and higher dimen- velope over with all process matrices considered. Note, that sional cases to future work. in the limit Θ → π and κ → 0, the augmented fidelity 1,2 1,2 One promising research direction is to explore met- fails to reach 99% fidelity, which is related to Eq. (17) and the additional contribution of other diagonal elements. rics corresponding non-uniform distributions in higher dimensions, investigating their capabilities to discriminate different noise channels. In this work, we restricted to distribution defined on the product states, where even on Haar invariance property), and carry additional infor- local uniform distribution can provide additional insight mation about underlying noise process. This informa- about the noise process. Additionally, we explored (nu- tion is manifested in larger possible spread in observed merically) von Mises-Fisher and polar cap distributions fidelities, that also allows to identify noise biases in Pauli for product states, demonstrating that conjugation of Channels. Because the uniform-average fidelity and as- these techniques with other benchmarking methods can sociated protocols probe only the depolarizing character improve our understanding of the device imperfections. of the noise, they can under or overestimate the deteri- The generalization to nonlocal case is not straightfor- orating effect of device’s miscalibration. Therefore, it is ward, and would need to take into account non-trivial imperative to have additional tools for assessing perfor- geometrical structure of higher-dimensional pure quan- mance of quantum devices. In particular, efficient meth- tum states [27]. The goal would be an efficient (in num- ods that reliably infer more χ matrix elements could im- ber of measurements) and scalable (in number of qubits) prove functionality of next generations of quantum hard- protocol that can reliably probe the proposed metrics. ware. An auspicious direction is to examine protocols inspired Since it is impossible to perfectly prepare an arbitrary by RB technique for local distributions, which would be pure state, this method may also lend itself to probing sampled independently, hence offering scaling similar to state preparation errors. single-qubit case. This we leave as an open problem for Any reasonably effective pure state preparation will future research. Another aspect worth exploring is to likely resemble some possibly skewed distribution cen- use these types of distribution over a subspace of the full tered around the target state. This means that comput- Hilbert space, such as is done in Refs. [28, 29] for the ing the fidelity for a particular state is similar to using our standard uniform-average fidelity metric. This is of par- non-uniform distribution fidelity metrics. Note that the ticular interest as it is known higher system levels can Kent distribution (which generalizes von Mises-Fisher) play a dominant role in the projected two-level dynam- could in principle also be used to achieve a greater agree- ics of a qubit system [30], and could in principle help us 7 to identify leakage errors with better accuracy. There- Appendix: Variance fore having methods to better distinguish noise processes acting on the full d-level (qudit) system could have im- The variance of a fidelity is computed with respect to mediate implications for hardware design. its distribution as

Z Z 2 2 V ar(F ) = F|ψihψ|(U, EU ) dΩ− F|ψihψ|(U, EU )dΩ , (A.1) where

h † i F|ψihψ|(U, EU ) = Tr U |ψi hψ| U EU (|ψi hψ| , (A.2) ACKNOWLEDGMENTS and dΩ is a surface element related to the underlying distribution of initial states |ψi. Taking an arbitrary rep- We appreciate fruitful discussion with Robin Blume- resentation of a noise process (i.e. characterized by a χ Kohout. We are grateful for support from NASA Ames matrix, see Eq. (2)) results in a formula that depends Research Center, the AFRL Information Directorate un- non-trivially on all elements of χ matrix. For our pur- der grant F4HBKC4162G001. FW and JM are thankful poses it suffices to consider noise processes that minimize for support from NASA Academic Mission Services, Con- (maximize) fidelities, which means of the form Eq. (13). tract No. NNA16BD14C. AP is grateful to his QuAIL In that case polar cap distribution variance for minimal colleagues for kind hospitality during his work at NASA. fidelity is

1 h V ar(F ) = − 40 − 12χ (cos(Θ) + 1) + χ + χ − 2χ (2 cos(Θ) + cos(2Θ)) − 6χ + 3χ + χ − 2 2 Θ 5760 0,3 1,1 2,2 3,3 0,0 1,1 2,2 3 − − 1920χ χ sin4(Θ) + 1920χ2 cos(Θ) − 1 − 1202χ χ + χ (χ + χ − 2χ ) cos(4Θ) cos(Θ) − 1 0,1 1,3 0,0 0,2 2,3 0,3 1,1 2,2 3,3 2 2 2 2 2 2 + 3 3χ1,1 + 2χ2,2χ1,1 + 3χ2,2 + 8χ3,3 + 4 χ1,2 − 4 χ1,3 + χ2,3 − 8 (χ1,1 + χ2,2) χ3,3 cos(5Θ) 2 2 2 2 2 2 2 2 + 30 96χ0,2 + 64χ0,3 + 20χ1,2 + 8χ3,3 + 5 3χ1,1 + 2χ2,2χ1,1 + 3χ2,2 + 16 χ1,3 + χ2,3 2 2 2 2 2 2 + 8 (χ1,1 + χ2,2) χ3,3 cos(Θ) + 5 − 64χ0,2 + 128χ0,3 − 20χ1,2 + 24χ3,3 − 5 3χ1,1 + 2χ2,2χ1,1 + 3χ2,2 2 2 + 16 χ1,3 + χ2,3 + 8 (χ1,1 + χ2,2) χ3,3 cos(3Θ) + 480 (2χ0,2χ2,3 + χ0,3 (χ1,1 + χ2,2 + 2χ3,3)) cos(2Θ) Θ − 5120χ2 sin4 (cos(Θ) + 2) + 160χ − 24χ sin2(Θ) + (χ + χ ) (9 cos(Θ) − cos(3Θ) − 8) 0,1 2 0,0 0,3 1,1 2,2 2 2 2 + 2χ3,3(3 cos(Θ) + cos(3Θ) − 4) − 384χ3,3 − 8 320χ0,2 + 90χ2,3χ0,2 + 320χ0,3 + 45χ0,3 (χ1,1 + χ2,2) 2 2 2 2 2 i + 16 3χ1,1 + 2χ2,2χ1,1 + 3χ2,2 + 4 χ1,2 + χ1,3 + χ2,3 − 16 (75χ0,3 + 16 (χ1,1 + χ2,2)) χ3,3 ,

and for von Mises-Fisher distribution 8

1 h V ar(F ) = 3χ2 κ4 − χ2 κ4 − χ2 κ4 + 3χ2 κ4 + 2χ κ4 − 2χ χ κ4 + 2χ κ4 + 2χ χ κ4 + 2χ χ κ4 κ 4κ4 0,0 1,1 2,2 3,3 1,1 1,1 2,2 2,2 1,1 3,3 2,2 3,3 4 4 3 3 3 3 3 2 2 − 2χ3,3κ − κ + 8χ0,3κ − 8χ0,3χ1,1κ − 8χ0,3χ2,2κ + 32χ0,2χ2,3κ − 8χ0,3χ3,3κ + 16(κ coth(κ) − 1)χ0,1κ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 − 16χ0,2κ + 16χ0,3κ + 4χ1,1κ + 16χ1,2κ − 80χ1,3κ + 4χ2,2κ − 80χ2,3κ − 4csch (κ) 2 2 2 2 2 2 2 2 · (2κχ0,3 + χ1,1 + χ2,2 − 2χ3,3) κ + 24χ3,3κ + 4χ1,1κ − 8χ1,1χ2,2κ + 4χ2,2κ − 12χ1,1χ3,3κ 2 2 2 − 12χ2,2χ3,3κ − 8χ3,3κ + 2χ0,0 − κ + 4(κ coth(κ) − 1)χ0,3 + (κ + 2 coth(κ)) (χ1,1 + χ2,2) 2 2 + (3κ − 4 coth(κ))χ3,3 κ − 2χ1,1 − 2χ2,2 + 4χ3,3 κ + 32χ0,3χ1,1κ + 32 κ − 3 coth(κ)κ + 3 χ0,1χ1,3κ 2 2 2 2 2 2 2 2 2 2 2 + 32χ0,3χ2,2κ + 96χ0,2χ2,3κ − 64χ0,3χ3,3κ + 4 coth(κ) 4χ0,2κ + χ1,1κ + 4χ1,3κ + χ2,2κ + 4χ2,3κ − χ1,1κ 2 2 2 2 2 2 2 + 2χ1,1χ2,2κ − χ2,2κ − 24χ0,2χ2,3κ + 2χ0,3 −κ + κ − 2 χ1,1 + κ − 2 χ2,2 + κ + 4 χ3,3 κ − 7χ1,1 2 2 2 2 2 2 2 2 2 − 7χ2,2 − 2 κ + 8 χ3,3 − 2χ1,1χ2,2 − 12 χ1,2 − 4 χ1,3 + χ2,3 + 2κ − κ − 16 χ1,1 − κ − 16 χ2,2 χ3,3 κ

2 2 2 2 2 2 i + 32χ1,1 + 48χ1,2 − 192χ1,3 + 32χ2,2 − 192χ2,3 + 80χ3,3 + 16χ1,1χ2,2 − 80χ1,1χ3,3 − 80χ2,2χ3,3 .

In both formula χi,j for i 6= j correspond to real part of real elements (only present in the ﬁrst row and column) the χ matrix, the imaginary part that is distinct from the has no contribution to the variance.

[1] F. Arute et al., Nature 574, 505 (2019). 494 (John Wiley & Sons, 2009). [2] M. Mohseni, A. T. Rezakhani, and D. A. Lidar, Phys. [20] R. Fisher, Proceedings of the Royal Society A: Math- Rev. A 77, 032322 (2008). ematical, Physical and Engineering Sciences 217, 295 [3] Y. S. Weinstein et al., The Journal of Chemical Physics (1953). 121, 6117–6133 (2004). [21] J. T. Kent, Journal of the Royal Statistical Society. Series [4] G. Benenti and G. Strini, Journal of Physics B: Atomic, B (Methodological) 44, 71 (1982). Molecular and Optical Physics 43, 215508 (2010). [22] E. Magesan, R. Blume-Kohout, and J. Emerson, Phys- [5] Y. Lu, J. Y. Sim, J. Suzuki, B.-G. Englert, and H. K. ical Review A 84 (2011), 10.1103/PhysRevA.84.012309, Ng, “Direct estimation of minimum gate fidelity,” (2020), arXiv: 0910.1315. arXiv:2004.02422 [quant-ph]. [23] P. Aliferis et al., New Journal of Physics 11, 013061 [6] M. D. Bowdrey et al., Physics Letters A 294, 258 (2002). (2009). [7] M. A. Nielsen, Physics Letters A 303, 249 (2002). [24] A random χ matrix was generated by fixing the χ00,00 [8] M. Horodecki, P. Horodecki, and R. Horodecki, Physical element, drawing randomly a vector of positive num- Review A 60, 1888 (1999). bers that sum up to 0.015 as diagonal elements, and [9] J. Emerson, R. Alicki, and K. Życzkowski, Journal of Op- then drawing randomly off-diagonal elements between [- tics B: Quantum and Semiclassical Optics 7, S347 (2005). min{diagonal elements}, min{diagonal elements} and re- [10] A. Erhard et al., Nature Communications 10, 5347 jecting samples when χ < 0. From total 6000 samples, we (2019). accepted 1673 positive definite matrices that yield CPTP [11] E. Magesan, J. M. Gambetta, and J. Emerson, Physical condition of the map. Review Letters 106, 180504 (2011), arXiv: 1009.3639. [25] B. I. Bantysh, D. V. Fastovets, and Y. I. Bogdanov, in In- [12] J. J. Wallman and S. T. Flammia, New Journal of Physics ternational Conference on Micro- and Nano-Electronics 16, 103032 (2014). 2018 , Vol. 11022, edited by V. F. Lukichev and K. V. [13] S. Boixo et al., Nature Physics 14, 595 (2018), arXiv: Rudenko, International Society for Optics and Photonics 1608.00263. (SPIE, 2019) pp. 697 – 708. [14] S. T. Flammia and Y.-K. Liu, Physical Review Letters [26] M. Sun and M. R. Geller, “Efficient characterization 106, 230501 (2011). of correlated spam errors,” (2018), arXiv:1810.10523 [15] M. P. da Silva, O. Landon-Cardinal, and D. Poulin, [quant-ph]. Physical Review Letters 107, 210404 (2011). [27] I. Bengtsson and K. Życzkowski, Geometry of quantum [16] O. Moussa, M. P. da Silva, C. A. Ryan, and R. Laflamme, states: an introduction to quantum entanglement (Cam- Physical Review Letters 109, 070504 (2012). bridge university press, 2017). [17] D. K. Tuckett, S. D. Bartlett, and S. T. Flammia, Phys. [28] L. H. Pedersen, N. M. Møller, and K. Mølmer, Physics Rev. Lett. 120, 050505 (2018). Letters A 367, 47 (2007). [18] D. K. Tuckett et al., Phys. Rev. X 9, 041031 (2019). [29] L. H. Pedersen, N. M. Møller, and K. Mølmer, Physics [19] K. V. Mardia and P. E. Jupp, Directional statistics, Vol. Letters A 372, 7028 (2008). [30] C. Neill et al., Science 360, 195 (2018).