Fault-Tolerant Computing With Biased-Noise Superconducting Qubits

We present a universal scheme of pulsed operations for the IBM oscillator-stabilized flux qubit comprising the CPHASE gate, single-qubit preparations and measurements. Based on numerical simulations, we argue that the error rates for these operations can be as low as about .5% and that noise is highly biased, with phase errors being stronger than all other types of errors by a factor of nearly 10^3. In contrast, the design of a CNOT gate for this system with an error rate of less than about 1.2% seems extremely challenging. We propose a special encoding which exploits the noise bias allowing us to implement a logical CNOT gate where phase errors and all other types of errors have nearly balanced rates of about .4%. Our results illustrate how the design of an encoding scheme can be adjusted and optimized according to the available physical operations and the particular noise characteristics of experimental devices.

After years of painstaking labor, superconducting qubits [1] are taking shape as viable elements for the construction of a scalable quantum computer.Since the initial demonstrations of coherent quantum dynamics in superconducting qubits [2,4,3,5], it has been recognized that these systems have great potential versatility [6,7,8], so that one can genuinely envision a quantum-computing integrated circuit emerging from this research.However, no clear way forward has been announced, owing largely to one undeniable feature of large-scale quantum computation: it will require a very high degree of fidelity in the execution of quantum operations, much higher than has been reported in any present experiments.
How high a fidelity, or how low an error rate, will be needed?On the basis of fundamental early theoretical work [9,10,11], lip service is frequently paid to a necessary universal set of operations containing the twoqubit controlled-σ x (cnot) gate, and a necessary "threshold" error rate in the 10 −5 −10 −4 range.Some recent modeling for superconducting qubits [12,13] suggests that such noise levels could conceivably be reached in the lab; however, in current experimental practice the ability even to reliably detect such small error rates, let alone to achieve them, is in fact very questionable.
In this paper, we will consider the possibility of constructing a universal set of operations for the IBM "Koch qubit" [14,8] 1 .Although our set of operations will not contain the cnot gate, we will propose an encoding scheme for implementing logical cnot gates which can then be used for fault-tolerant quantum computation.With the combination of our encoding scheme and other improvements in the theory of quantum fault-tolerance [15,16,17], error rates for our elementary operations in the .1%−.5% range are expected to be tolerable for practical quantum computation.We estimate using numerical simulations that error rates in this range are possible for the Koch qubit.Though these estimates are far from the error rates that have been currently measured in experiments, we hope our results will motivate and stimulate the research in superconducting qubits in the IBM lab and elsewhere.More generally and even beyond superconducting qubits, our encoding scheme illustrates how techniques of quantum error correction and fault tolerance can be tailored to the available physical operations and the particular noise characteristics of experimental devices.Figure 1: On the left, a cnot gate is applied between two qubits.In the middle, the cnot gate is simulated by using one cphase and two H gates. On the right, another simulation where the two H gates are implemented by "teleportation" using ancilla qubits and measurements.Vertical lines with dots on both ends denote cphase gates.A triangle pointing to the right denotes the preparation of a qubit in the state |+ .A triangle pointing to the left denotes a measurement of σ x .Conditioned on the measurement outcomes, σ x correction operators may be necessary as shown.
Our proposal involves a synthesis of recent experimental and theoretical developments.On the experimental side [14,8], pulsed operations for the Koch qubit were discussed in [18]; the set of operations considered in [18] included the controlled-σ z (cphase) gate, the Hadamard (H) gate, single-qubit "diagonal" rotations of the form exp(iθσ z ), the preparation of a qubit in the state |+ = 1 √ 2 (|0 + |1 ), and the measurement of σ x (where σ x , σ y , and σ z are the Pauli spin operators).For these operations, [18] numerically estimated the error rates by considering all physical sources of noise that we presently know how to model for this system-1/f flux noise, instrumental jitter in pulse timing and amplitude, and Johnson noise from resistances in the circuit.For the cphase gate it estimated an error rate of about .45%,for the H gate about .4%, for diagonal rotations about .001%for any θ, for the preparation of the state |+ about .3%, and for the measurement of σ x about .2% 2 .
In contrast, no direct implementation of a cnot gate with error rate close to even 5% could be devised for the Koch qubit.Alternatively, a cnot gate can be implemented indirectly by composing other elementary operations.Fig. 1 shows two possible methods; the first indirect implementation uses one cphase and two H gates, while the second one uses three cphase gates, two qubits prepared in the state |+ , and two measurements of σ x .Therefore, we estimate that the first indirect implementation would have an error rate of about 1.25%, and the second one about 2.3%.
Based on the modeling in [18], it was further observed that noise in the cphase gate acted predominantly as dephasing; i.e., it affected the relative phases in the wavefunction giving rise to "phase" errors which can be expressed as diagonal matrices in the computational basis.All other errors, including errors due to relaxation and also "leakage" errors associated with transitions to states outside the computational space, were observed to be about an order of magnitude weaker.The same observation also applies to single-qubit diagonal rotations but, on the other hand, no similar bias was observed for the noise in the H gate.
On the theoretical side, the findings of [18] motivated us to revisit a longstanding problem in the theory of fault-tolerant quantum computation: how to formulate an encoding scheme that exploits large asymmetries in the structure of noise in elementary operations.Here and in our companion paper [19], we present such a scheme based on encoding the noisy qubits using a repetition code, and we show how fault-tolerant logical cnot gates can be implemented for this code.The intuition we gained from [18] is that highly biased noise, with phase errors being much stronger than all other types of errors, is possible for gates such as the cphase, but it should not be physically expected for generic operations.Furthermore, an encoding scheme has to address the problem that a noise bias can easily be lost as elementary operations are composed together.
Our solution is to choose a universal set of elementary operations whose implementation induces noise that is biased towards dephasing, and where all gates commute with σ z so that the noise bias is maintained.The operations we will use are the preparation of the state |+ , the controlled-σ z (cphase) gate, and measurements of observables in the equator of the Bloch sphere, i.e., of the form exp(iθσ z )σ x for certain angles 3 .Noise for the preparation of the state |+ is naturally biased since |+ is an eigenstate of σ x .The 2 These numbers are obtained by using the results in [18] and the definition of an error rate in the Supplementary Material.In [18], the entanglement fidelity and not the error rate is used as a measure of the noise strength.
3 Alternatively, we may restrict to measurements of σx, if we add to our set of elementary operations the preparation of the states |+i = exp(i π 4 σz)|+ and |T = exp(i π 8 σz)|+ ; this is the universal set considered in [19].For the Koch qubit, it is more natural to move all rotations of the form exp(iθσz) before the measurements since, as we will discuss, corrective single-qubit structure of noise for the cphase gate depends on the physical implementation and it can be engineered to be biased; here, we propose such a biased-noise implementation for the Koch qubit based on adiabatic pulses so that transitions between the computational-basis states are strongly suppressed.Finally, noise in a single-qubit measurement can be described by errors preceding the ideal measurement which need not have any specific structure since measurement has a classical output.
These theoretical ideas may have implications for the experiments with the Koch qubit, and solid-state qubits in general.They suggest that a profitable focus of experiments could be to design qubits with long relaxation time T 1 .Provided this is achieved, dephasing, which is the dominant source of noise, can be much more effectively suppressed by using our encoding scheme.Furthermore, the implementation of low-noise cnot or H gates is not necessary; it suffices to implement cphase gates with highly biased noise, together with single-qubit preparations and measurements.For the Koch qubit, eliminating the need to implement the H gate allows us to re-examine the pulse schemes in [18], and to find a simpler implementation of the cphase gate for which noise is much more biased than in [18].Similar simplifications may be possible in other types of superconducting qubits.In particular, a standard experimental approach to suppressing dephasing noise is to intersperse quantum gates with additional corrective "hardware" operations such as spin echo pulses.Our encoding scheme can be seen as a complementary solution where the hardware operations are restricted to a much smaller set and dephasing is suppressed at the "software" level by using error correction applied on blocks of several noisy qubits.
We will now discuss the details of our proposal.The central question is how to devise a strategy for fault-tolerant quantum computation that effectively exploits a bias in the noise of the cphase gate.To develop some intuition, let us assume for the moment that there is no leakage, that independent phase errors occur with probability ε, and that all other types of errors occur with probability ε ε.Then a clear strategy for the lowest-level coding of quantum data is to use an n-qubit repetition code.One logical qubit is encoded in a "block" of n physical qubits, a logical |+ state is represented by all n qubits being in the state |+ , and a logical |− by all n qubits being in the state |− .Since the logical σ z operator is a σ z acting on all n qubits in the block, phase errors on more than half of the qubits in the block are necessary for a logical error to occur after error correction; taking n odd, and assuming that for each qubit there are t time steps where a phase error may occur, the probability of a logical error is approximately On the other hand, the logical σ x operator is a σ x acting on any qubit in the block, so that even a single error different than a phase error on any qubit and at any time step cannot be corrected; thus, the probability of a logical error is approximately If the noise "bias" ε/ε is large, we can choose some large n and obtain a significant reduction of the logical error rate; e.g., if the bias is 10 3 and even for t • ε = 5%, we find that ε L ≈ ε L < 3.5 × 10 −4 by setting n = 7.This sends an optimistic message, since several schemes are known for effectively implementing fault-tolerant quantum computation with error rates of order 10 −4 [20].As we will discuss below, our pulsed implementation of a cphase gate can in fact be optimized to lead to a noise bias of order 10 3 .But first, we must examine more closely whether there is actually a gain in using the repetition code as our naive argument so far indicates.The main concern is that a large intrinsic noise asymmetry in our elementary operations might be spoiled as these operations are composed together.
For example, consider the problem of implementing a logical cnot gate between two blocks of encoded qubits.The logical cnot gate for the repetition code can be implemented by bitwise cnot gates, and each cnot gate can be simulated by using cphase gates as in Fig. 1.But then due to the H gates, phase errors that occur during the simulation of each cnot gate are converted into errors of other types; e.g., a σ z error during the implementation of a H gate will be converted to some linear combination of a σ z , a σ x , and a σ y error.So, this construction of a logical cnot gate destroys the asymmetric structure of the noise and nullifies the effectiveness of the repetition code.Avoiding this interconversion of noise is possible, but it requires a circuit of greater complexity.At the heart of this circuit construction is the identity operation in (a) (b) Figure 2: (a) On the left, an identity circuit which consists of preparing a qubit in the state |+ , measuring σ z ⊗ σ z on this qubit and the input qubit, and finally measuring σ x on the input qubit.After applying the correction operators shown conditioned on the measurement outcomes, the state of the input qubit is "teleported" to the output qubit.On the right, the measurement of σ z ⊗ σ z is implemented by using an "ancilla" qubit (shown in green) which interacts with the other two "data" qubits (shown in red) via cphase gates before it is measured.(b) The circuit in (a) where the data qubits are encoded in the n-bit repetition code (here n = 3); the ancilla qubit remains unencoded, but the subcircuit enclosed by the dashed curve must be repeated sequentially several times, and the majority of the measurement outcomes must be taken in order to correct errors (the repetitions are not shown).As in (a), logical σ z or σ x correction operators (not shown) may be necessary on the output data block.Fig. 2(a); it consists of preparing a qubit in the state |+ , measuring σ z ⊗ σ z on this qubit and the input qubit, and finally measuring σ x on the input qubit.
The basic idea is now to encode the qubits in the repetition code.We have already mentioned that the preparation of a logical |+ involves simply preparing every qubit in the block in the state |+ .A measurement of the logical σ x is also very simple and robust; it can be implemented by measuring σ x on each qubit in the block, and then taking the majority of the outcomes to correct errors.It remains to discuss how to implement a measurement of the logical σ z ⊗ σ z on the two blocks.The difficulty is that the measurement must be robust against the dominant phase errors, and it must also respect the biased structure of the noise.Fig. 2(b) shows our solution: A single unencoded "ancilla" qubit is prepared in the state |+ , and then cphase gates are applied between this ancilla qubit and all other qubits in the two blocks.Finally, a measurement of σ x is performed on the ancilla, and a logical σ x correction may be necessary on the output block conditioned on the measurement outcome.Since even a single phase error on the ancilla can cause an error in the measurement outcome, the measurement of the logical σ z ⊗ σ z must be repeated sequentially several times; by taking the majority of the outcomes we can suppress the probability of a logical error at the output.
Using the building blocks just discussed, it is possible to construct a fault-tolerant logical cnot gate that respects the physical-level bias of the noise; see Fig. 3.This circuit implements a logical cnot gate between the two blocks encoded in the repetition code, while at the same time the logical state of each block is teleported to a new block and phase errors are corrected.Because of the use of teleportation, this circuit has the additional feature that it largely prevents the propagation of leakage errors.In Fig. 3, the output data qubits always interact with an ancilla qubit prior to any interaction with the input data qubits; therefore, there is no possibility for leakage to propagate from the input to the output qubits.And furthermore, if we consider implementing a sequence of logical cnot gates, we observe that every qubit is eventually measured after only a small, fixed number of time steps, and the measurement effectively converts leakage to regular qubit errors [21].
In addition, it is possible to construct with the same building blocks a fault-tolerant preparation of the logical |0 and |+ states, and a measurement of the logical σ z and σ x for the repetition code.The logical cnot gate, the preparation of logical |0 and |+ , the measurement of logical σ z and σ x , and single-qubit measurements of exp(iθσ z )σ x for θ = π/4 and θ = π/8 suffice for implementing universal fault-tolerant quantum computation; the full scheme is discussed in our companion paper [19].
Since our building blocks only use cphase gates, |+ preparations, and σ x measurements, and since phase Figure 3: A logical cnot gate between two blocks encoded in the repetition code (here again, the 3-bit code).Ancilla qubits (shown in green) are prepared in the state |+ , they interact via cphase gates with the data qubits (shown in red), and they are measured along the eigenbasis of σ x .As in Fig. 2, the measurements with ancilla qubits must be repeated sequentially several times so that errors can be corrected by taking the majority of the outcomes (the repetitions are not shown).Finally, σ x is measured on each qubit in the two input blocks, and the majority is taken on each block.Conditioned on the results of the majorities, logical correction operators (not shown) may be necessary on the output blocks [19].
errors commute with cphase gates, it is simple to estimate the error rates for the logical cnot gate given the physical-level error rates.In [19], we give upper bounds for the logical error rates in closed form as a function of the block size n of the repetition code and the number of repetitions k of the measurements executed with ancilla qubits.The outcome is qualitatively as in Eqs. ( 1) and (2) with t = c k for some constant c ≈ 2 or 3.It follows from this analysis that if the noise bias in our elementary operations is about 10 3 or greater, encoding in the repetition code is effective and logical errors are significantly weaker than the physical-level phase errors for ε of order .1% or smaller; see Fig. 4 4 .The noise bias reported in [18] was about a factor of 10, which has motivated us to re-examine whether greater levels of bias are conceivable for the Koch qubit.As we will now discuss, the noise bias can, in fact, be dramatically improved at the expense of a very minor increase in the rate of phase errors.Fig. 5 shows the structure of the Koch qubit [14,8].It is nominally a "flux" qubit, meaning that the computational basis states |0 and |1 are quantum states corresponding to distinct circulating-current orientations; see Fig. 5.The mode of operation is highly tunable via an external "control flux" Φ c threading the small loop; see Fig. 6.At small Φ c , the degeneracy of the two circulating-current states is lifted by the "flux bias" corresponding to the flux difference in the two large loops, and the two basis states are easily detected and initialized but not very phase-coherent.As Φ c is increased, the barrier decreases until it eventually disappears, and tunnel splitting between the basis states turns on rapidly.At even larger Φ c , we enter an almost harmonic single-well regime, but before this another essential element of the Koch qubit enters the picture.
The flux-qubit states are strongly coupled to the fundamental mode of a superconducting transmissionline resonator, so that |0 and |1 pass via an avoided level crossing to | 0 and | 1 corresponding to the 0and 1-photon modes of the oscillator.The superconducting transmission line is highly phase-coherent [8].In the IBM experiments we have seen circa 50,000 Ramsey fringes associated with these states corresponding to T 2 = 2.5µsec [22]; we expect much longer T 2 times to be possible.This fact has led us in [18] to the following strategy for implementing low-noise operations.When not being acted upon, quantum information Figure 4: Upper bounds on the total probability ε cnot L of logical errors for the logical cnot gate as a function of the probability ε of physical-level phase errors for a bias of 10 3 and 10 4 .To obtain our bounds, we have optimized over the block size n of the repetition code, and the number k of repetitions of measurements with ancilla qubits; the optimal choice is to have n = k, while the optimal value depends on ε and the bias as shown.The straight line with slope unity serves as a guide to the eye. is stored in the highly-coherent oscillator energy levels; we then say that the qubit is "parked."All needed operations are done by adiabatic pulsing out of parking.Each flux qubit has a fixed, untuneable two-qubit coupling to a set of nearby flux qubits in order to implement two-qubit gates.To assure that the effective coupling between parked qubits is negligible, these couplings are to be only between qubits with different resonator frequencies, so that resonant transfer of photons between different resonators does not take place in the parked state.Below we examine the simplest scheme with just two resonator frequencies, 3.1GHz and 3 4 • 3.1GHz (the commensurability of these frequencies aids in the maintenance of rotating frames for these qubits).Correspondingly, we have two species of qubits, A and B respectively, and two-qubit couplings exist only between qubits of different species-this is not a severe restriction since, in our encoding scheme, the only interactions are between data qubits (red qubits in the figures) and ancilla qubits (green qubits).Fig. 7(a) shows the energy levels of the coupled two-qubit system as a function of the control flux-assumed equal for the two qubits in the figure-with the flux bias held on the "symmetric line" = 0 on which the effective qubit potential U has a reflection symmetry as in Fig. 6.In the "parking" region Φ c ≥ 1.46Φ 0 , the effective interaction between the two qubits is strongly suppressed.The energy levels are essentially those of the unperturbed transmission line resonators and, because to very high accuracy we have energy additivity there is no conditional phase accumulation in the two-qubit state.Out of parking, this energy additivity is violated and a conditional phase shift can accumulate.If only one qubit is pulsed out of parking, single-qubit "diagonal" rotations of the form exp(−iθσ z ) can be effected (in the frame of reference rotating at the resonator frequency).If two coupled qubits are pulsed out of parking, a cphase gate can be implemented by choosing the pulse timing appropriately; at the same time, known single-qubit phase shifts accumulate which can be compensated by corrective diagonal rotations on each qubit [18].
There is, however, one point of concern which reveals itself more clearly when we plot the energy levels of the same two-qubit system as a function of time during the implementation of a cphase gate; see Fig. 7(b).At the energy-level crossing marked C, the avoided-crossing gap between the state | 1 | 1 and a state outside the computational space becomes small, and this may potentially lead to increased leakage errors.One possible strategy to avoid this problem would be to increase the avoided-crossing gap by changing the flux bias to = 0 in order to move away from the symmetric line; this solution was adopted in [18].However, departing from the symmetric line has its disadvantages.The lowering of symmetry makes the system less protected from low-frequency noise, and it also causes the effective relaxation time T 1 to be shortened making it harder to achieve a high bias in the noise.
It would therefore be desirable to perform the operation at the symmetric line.Our new observation is that the avoided-crossing gap is actually very small; it is never larger than 100kHz even taking parameter shifts due to low-frequency noise into account.So for our pulse profiles, the Landau-Zener tunneling probability is extremely close to one, meaning that we can pulse through the avoided-crossing gap and suffer essentially no leakage.After trying many pulse designs, and checking the noise bias they produce, we find the pair of pulses in Fig. 7(c) for the two species of qubits to work the best.We observe that in this pulsing scheme the low-frequency qubit B is unparked far more deeply (point P b ) than the high-frequency qubit A (point P a ); this choice is optimal because, if the two species of qubits were unparked symmetrically, the same level of dephasing noise could only be maintained by making the pulse duration significantly longer increasing the noise due to relaxation and leakage.
To estimate the effect of noise in the implementation of our cphase gate, we have performed multiple computer simulations of the evolution of the two coupled qubits during the gate where, in each simulation, low-frequency noise is added to the ideal dynamics.To model low-frequency noise, in each simulation we take the actual applied flux and the actual timing in a pulse to deviate from the ideal by an amount which is chosen from a Gaussian distribution with zero mean and variance 6µΦ 0 and 6psec respectively [18].From each simulation we extract a unitary which describes the noisy implementation of the cphase gate; by averaging over a large number of simulations which corresponds to integrating over the Gaussian fluctuations, we obtain a superoperator describing the noisy implementation of the gate.By expressing this superoperator as N • CZ where CZ is the ideal superoperator when there is no noise, we can define an "error rate" in terms of the norm of E = N − Î where Î is a trace-decreasing superoperator proportional to the identity The lowest energy levels of a single Koch qubit on the "symmetric line" corresponding to = 0, as a function of Φ C .For small Φ C , Φ C = 1.4Φ 0 (Φ 0 = h/2e), the ground state is doubly degenerate; the degeneracy may be lifted by setting = 0, allowing for qubit measurement and preparation.The two basis states |0 and |1 correspond to the two states |L and |R of a double-well potential with a very large potential barrier between them as sketched, which describe the two different orientations of the persistent current I P in the flux qubit; the rescaled dynamical variable f is explained in [23].For both |0 and |1 , the superconducting transmission line is in its "vacuum" 0-photon state.As Φ C is increased to around 1.45Φ 0 , the barrier height drops, leading to a rapid increase in the tunnel splitting between the two lowest states.
When the tunnel splitting equals the transmission-line resonator frequency ω T , there is an avoided level crossing with splitting 2g.For larger Φ C , the qubit is "parked"; now, the two lowest energy states | 0 and | 1 correspond to the 0-and 1-photon states of the transmission line respectively, their energies are independent of Φ C , and their effective potential is highly harmonic as sketched.In this regime, for both | 0 and | 1 , the flux-qubit state is the symmetric state superoperator.Although this rate does not necessarily correspond to the probability of an error, it is possible to formulate a fault-tolerance analysis similar to the case of stochastic noise.In particular, we may expand E = E phase + E other where E phase contains the terms of E which are diagonal in the computational basis, and E other contains all other terms.Then, Eqs. ( 1) and ( 2) remain unchanged if ε and ε are re-interpreted as the norms of E phase and E other respectively.In the Supplementary Material, we give more details about this definition and about the derivation of the error rates that we discuss below.Based on our modeling, we find the following hierarchy of expected error levels for the cphase gate: The rate for phase errors is 1.96 × 10 −3 for qubit A and 4.6 × 10 −3 for qubit B. For all other types of errors, the dominant contribution is due to relaxation to the ground state during the 35.1nsec duration of the cphase gate; using our previous results based on the Caldeira-Leggett model which give an estimate of T 1 ≈ 10msec [23,18], we expect their rate to be around 3.5 × 10 −6 for both qubits A and B. Therefore, as desired, there is a large contrast between the rates for phase errors and all other types of errors.We must also finally assess the magnitude of leakage.Since leakage errors cannot in general be corrected by our repetition code, it is crucial that leakage is suppressed to a very large extent; in fact, we have found that the rate for leakage errors is approximately 3.5 × 10 −6 .
We have also designed pulses for the preparation of a qubit in the state |+ , for single-qubit diagonal rotations, and for the measurement of σ x ; these pulses are only slightly modified from [18].Since the diagonal rotations commute with the cphase gate, they can always be moved to occur immediately before the measurements of σ x ; the combined operation is a measurement of an operator of the form exp(iθσ z )σ x .The rate for phase errors in a preparation of |+ is 2.75 × 10 −3 for both qubits A and B, and the rate for all other types of errors is 3.5 × 10 −7 .However, because a preparation of |+ is performed by a non-adiabatic pulse, leakage is significantly larger than in the cphase gate; the leakage error rate is 3.77 × 10 −7 for qubit Variations of the eigenlevels of the two-qubit system for the optimal implementation of a cphase gate.To very good approximation, the quantum evolution adiabatically follows the four states moment in the evolution is at the crossing C (and its time-reversed image in the second half of the pulse), when | 1 | 1 crosses a state outside the computational space.Leakage to this state is suppressed because the avoided-crossing gap is very small (the inset shows the dispersion at C magnified by 200,000x), and because the angle of crossing is made larger by delaying the onset of the A pulse relative to the B pulse.Apart from this state and the four computational-basis states, all other eigenstates can be entirely left out of our numerical modeling because of their large distance from the | 1 | 1 and | 1 | 0 states.(c) The optimal control-flux pulses for A and B. Pulses are constructed as sums of tanh functions in order to have a standard smooth shape.It is found preferable to unpark A much less deeply (to P a ) than B (to P b ).Although P a and P b in (a) do not exactly correspond to the points P a and P b in (c) because Φ A C and Φ B C are not equal throughout the optimal pulses, they provide a good illustration of the relatively large energy difference (∼ 1.5GHz) corresponding to the small change in Φ C (∼ .004Φ0 ) between P a and P b in (c).
A and 1.5 × 10 −5 for qubit B. For a measurement of exp(iθσ z )σ x , we can describe noise in terms of effective errors of no specific structure preceding the ideal measurement, and their rate is 1.83 × 10 −3 essentially independent of θ.
Given these physical-level error rates, we can obtain upper bounds on the error rates for the logical cnot gate by using the equations in [19].In this analysis, we use the error rates we have computed by averaging over a large number of simulations as fixed, constant error rates, and we will ignore any correlations between the fluctuations around these average values in space and time.The analysis must also consider leakage errors which are not discussed in [19].Our method for analyzing leakage is based on the following observations.First, as we have already noted, the fact that teleportation is used to implement the logical cnot gate prevents leakage errors from propagating across multiple logical gates.In the worst case, one leakage error can propagate from the logical gate where it occurs to the following one.But this can be prevented by inserting a logical teleportation preceding every logical gate; i.e., the logical state of each output block from a logical gate is teleported to a new block as in Fig. 2 before the next logical gate is applied.Then, a leakage error that occurs in a logical gate or in the logical teleportations preceding it can only affect this logical gate and no other.
In our analysis of the logical cnot gate, we have optimized over the block size n of the repetition code, and the number k of repetitions of measurements with ancilla qubits.For our physical-level error rates, the optimal choice is (n, k) = (5, 7), where k is larger than n because qubits of species B (which are chosen as the ancilla qubits) are more noisy that qubits of species A (which are chosen as the data qubits).With this choice, we find that the logical cnot gate has nearly balanced rates for phase errors and all other types of errors of at most 4.62 × 10 −3 and 3.98 × 10 −3 respectively.This is an improvement by a factor of about 3 over the best alternative method we have for implementing a cnot gate without the encoding in the repetition code; as we have already discussed, we have found no direct implementation of a cnot gate for the Koch qubit with an error rate better than 5%, and simulating the cnot gate indirectly as in Fig. 1 leads to balanced rates for phase errors and all other types of errors of about 1.25%.Since our error-rate upper bounds also apply to the other logical operations for the repetition code needed for universality [19], our analysis indicates that our estimated physical-level error rates for the Koch qubit are tantalizingly close to those needed for effective fault-tolerant operation; in the literature, the highest estimated error thresholds are of order 1%, and the best proven thresholds are of order .1% [15,16,17,20].
To conclude, we have discussed an encoding scheme which has been especially tailored to the physical operations that are naturally available in the IBM qubit (cphase gates, qubit preparations and measurements), and the noise characteristics that can be expected for this system according to our theoretical modeling.We believe that the basic principles underlying our proposal-primarily, the implementation of adiabatic cphase gates with highly biased noise, and the suppression of phase errors by encoding in the repetition code-can be successfully adapted to apply to other promising systems besides flux qubits such as, e.g., superconducting phase qubits [4,7].
Our estimated error rates incorporate contributions from all sources of noise which are understood in experiments at present.However, the experimental reality today is that noise is dominated by T 1 processes which are not fully understood, and coherence times are significantly below the values we have obtained from our calculation [18].Our results should therefore be seen as preliminary and suggestive.Certainly, simplifications in modeling low-frequency 1/f noise have been made, so that noise fluctuations have been assumed to be constant during the execution of each gate, and also noise correlations and flux drifts across multiple gates have been ignored; see [18].
Furthermore, we should emphasize that there are several possibilities for complementing the methods of "software" error correction we have described here with "hardware" error correction which is done directly at the physical level; e.g., systematic noise correlations and flux drifts over long time scales be suppressed by periodically recalibrating qubits off-line before they are re-used, or superconducting qubits could be designed to have physical error-correction properties as in [24] which uses the same set of elementary operations as ours.Finally, we should note that the topology of interactions for our repetition-code scheme is not attainable with short-range interactions on a square lattice.We expect however that a greater but limited range of interactions where each qubit is coupled to 10 or 20 other nearby qubits would suffice.Various possibilities exist for implementing such interactions for superconducting-qubit layouts where crossovers, multiple couplers, and indirect couplings via transmission lines are all available.In our analysis, we model each noisy elementary operation by a superoperator.Therefore, we ignore temporal or spatial correlations between different operations, such as the correlations in time which are inevitably present for 1/f noise.We consider two sources of noise.First, 1/f noise in the control parameters-the applied fluxes, and the pulse synchronization-during the execution of each operation; 1/f noise leads primarily to dephasing and it will be modeled by a superoperator N 1/f .Secondly, thermal relaxation noise which is continuously present; relaxation is the primary source of errors different than phase errors and it will be modeled by an amplitude-damping superoperator N T1 .
Leakage errors depend both on the implementation of our operations and also on 1/f noise.We recall that during qubit preparation and measurement, which are implemented by non-adiabatic pulses, leakage can arise from Landau-Zener transitions to excited levels in the flux-qubit potential; during the execution of a cphase gate, leakage primarily occurs at the energy-level crossing marked C in Fig. 7(b).Since the relevant parameter is the minimum energy gap between states inside and outside the computational space during the implementation of an operation, leakage errors can occur even in the absence of 1/f noise.When 1/f noise is present, the minimum gaps are shifted from their ideal values when there is no noise, which in turn has an effect on leakage.
We model a noisy preparation of a qubit in the state |+ as the ideal preparation followed by the noise superoperator where • denotes composition.We model a noisy cphase gate as the ideal gate followed by a superoperator as in Eq. ( 3), where N 1/f is now supported on both qubits acted upon by the gate, and N T1 is assumed to act independently on each of the two qubits.Finally, we model a noisy measurement of exp(iθσ z )σ x as the ideal measurement preceded by a superoperator as in Eq. ( 3).The noise superoperator for each elementary operation can be expressed in terms of a discrete set of Kraus operators (at most d 2 where d is the dimension of its support) [25].If the identity is one of the Kraus operators, the noise model corresponds to local stochastic noise; in this case, we may define the phase error rate as the probability of all non-identity Kraus operators which are diagonal matrices in the computational basis.The probability of all other non-identity Kraus operators then defines the rate for all other types of errors.For this noise model, we can perform a fault-tolerance analysis to determine upper bounds on the probabilities of logical errors as in [19].
However, as we will discuss below, the superoperators in our modeling do not have this property so that noise cannot be simply described in terms of probabilistic errors.Nonetheless, we may use an alternative definition of an error rate, and with this definition the analysis in [19] remains essentially unchanged.The idea is to express N as the sum of an ideal and an erroneous part, N = Î + E, where Î is a trace-decreasing superoperator which is proportional to the identity superoperator [26].We may then define a generalized error rate or error strength in terms of the distance between N and Î, where || • || is the diamond norm [27].If the superoperator E has an n-qubit input, ||E|| = ||I n ⊗ E|| 1 = max X :||X||tr=1 ||(I n ⊗E)(X)|| tr , where I n is the identity superoperator on n qubits, and ||•|| tr is the standard trace norm, i.e., ||A|| tr = Tr √ A † A. In the remainder of this Supplementary Material, we give the details about the error rates for our elementary operations; the results are summarized in the end in Table 1.To simplify our calculations, we have estimated the various norms by only varying among a few possible inputs.Even though we have not performed a rigorous maximization, we believe that the inputs we have chosen are close to the worst case.k=0 I k XI † k is trace-decreasing and proportional to the identity superoperator, E l contains all terms with at least one insertion of a leakage error {M k,l }, i.e., E ¬d contains all remaining terms with at least one insertion of a non-dephasing error {M k,¬d }, i.e., E ¬d (X) = (N 1/f − E l )(X) − Table 1: Error-rate estimates for our elementary operations.For preparations and cphase gates, ε is the rate for phase errors, ε is the rate for all other types of errors, and ε l is the rate for leakage errors.For measurements, the rate ε includes errors from all sources.
To obtain the error rate for the remaining measurement process, we follow the evolution of the basis states | + and | − by performing a numerical simulation similar to the cases already discussed.If the initial state is | + , we calculate the probability that the final state before the projection is orthogonal to |0 , in which case we assume an error in the measurement outcome always occurs; and similarly for | − , we calculate the probability that the final state is orthogonal to |1 .We define this probability as the error rate for the measurement; for both qubits A and B, we find ε ≈ 1.83 × 10 −3 .

Relaxation
We model relaxation noise by the amplitude-damping superoperator N T1 acting independently on each qubit; with Q either A or B, the two Kraus operators are where we have already defined I Q and σ Q z , and σ Q x = I ⊗ σ x according to the tensor-product structure in Eq. (5).
Since only M 1 is non-diagonal in the computational basis, we define the rate for non-dephasing errors as ||M 1 || = γ, where M 1 (X) = M 1 XM † 1 .We also define the rate for phase errors due to the operator M 0 as ||M 0 − Î|| , where M 0 (X) = M 0 XM † 0 and Î(X) = cX with 0 ≤ c ≤ 1.If we take c = 1 + √ 1 − γ 2 /4, we find ||M 0 − Î|| ≈ γ/2.For a cphase gate, we use the worst-case estimate T 1 = 10msec, and we assume T 1 can be treated as approximately constant during the execution of the gate; then, γ = t/T 1 = 3.5 × 10 −6 where t = 35nsec is the duration of the gate.For diagonal rotations which are executed with pulses of duration t ≤ 5nsec, γ is very small and can be neglected.Finally, for preparation and measurement, T 1 changes as a function of the control flux and we calculate γ = t 0 ds s T1(s) = 3.5 × 10 −7 .

Figure 5 :
Figure5: Physical layout of two Koch qubits[8] coupled as needed to implement a cphase gate.Qubits A and B differ only in the resonant frequencies of the transmission line resonators attached to them.These frequencies are controlled by the physical length of the transmission lines, and are in the ratio 3:4.The singlequbit device Hamiltonians can be varied independently by controlling the "control flux" Φ C corresponding to the flux through the small loop, and the "flux bias" corresponding to the flux difference in the two large loops.The SQUIDs perform quantum measurements on the states of the individual qubits.The qubits are coupled via a fixed mutual inductance M to a superconducting loop.For small values of Φ C , the basis states |0 and |1 of each qubit correspond to different orientations, counterclockwise and clockwise, of the persistent current I P .

Figure 6 :
Figure6: The lowest energy levels of a single Koch qubit on the "symmetric line" corresponding to = 0, as a function of Φ C .For small Φ C , Φ C = 1.4Φ 0 (Φ 0 = h/2e), the ground state is doubly degenerate; the degeneracy may be lifted by setting = 0, allowing for qubit measurement and preparation.The two basis states |0 and |1 correspond to the two states |L and |R of a double-well potential with a very large potential barrier between them as sketched, which describe the two different orientations of the persistent current I P in the flux qubit; the rescaled dynamical variable f is explained in[23].For both |0 and |1 , the superconducting transmission line is in its "vacuum" 0-photon state.As Φ C is increased to around 1.45Φ 0 , the barrier height drops, leading to a rapid increase in the tunnel splitting between the two lowest states.When the tunnel splitting equals the transmission-line resonator frequency ω T , there is an avoided level crossing with splitting 2g.For larger Φ C , the qubit is "parked"; now, the two lowest energy states | 0 and | 1 correspond to the 0-and 1-photon states of the transmission line respectively, their energies are independent of Φ C , and their effective potential is highly harmonic as sketched.In this regime, for both | 0 and | 1 , the flux-qubit state is the symmetric state |S ≡ 1

Figure 7 :
Figure 7: cphase gate modeling.(a) The energy levels of two coupled qubits for the special case Φ A C = Φ B C .(b) Variations of the eigenlevels of the two-qubit system for the optimal implementation of a cphase gate.To very good approximation, the quantum evolution adiabatically follows the four states | 0 | 0 , | 0 | 1 , | 1 | 0 and | 1 | 1 .A crucial moment in the evolution is at the crossing C (and its time-reversed image in the second half of the pulse), when | 1 | 1 crosses a state outside the computational space.Leakage to this state is suppressed because the avoided-crossing gap is very small (the inset shows the dispersion at C magnified by 200,000x), and because the angle of crossing is made larger by delaying the onset of the A pulse relative to the B pulse.Apart from this state and the four computational-basis states, all other eigenstates can be entirely left out of our numerical modeling because of their large distance from the | 1 | 1 and | 1 | 0 states.(c) The optimal control-flux pulses for A and B. Pulses are constructed as sums of tanh functions in order to have a standard smooth shape.It is found preferable to unpark A much less deeply (to P a ) than B (to P b ).Although P a and P b in (a) do not exactly correspond to the points P a and P b in (c) because Φ A C and Φ B DDV and BMT have been partly supported by IARPA under ARO Contract No. W911NF-04-C-0098.JP is supported in part by DoE under Grant No. DE-FG03-92-ER40701, NSF under Grant No. PHY-0456720, and NSA under ARO Contract No. W911NF-05-1-0294.Supplementary Material on Error Rates P. Aliferis, F. Brito, D. P. DiVincenzo, J. Preskill, M. Steffen, and B. M. Terhal