Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
The XZZX surface code
Benjamin J. Brown
arXiv:2009.07851
Z X
Z
X
Z
X
Z
X
Z
X
Z
X
Z
(a)
(b)
(d)(c)
(e)
(f)
(g)
X Z
Z X
X
X
X
Z
Z
Z
Z
Z
Z
Z
Z
Z
Y
Y
Y
X
t{∆(a)
(b)
(c)
(d)
Z
MM
M
Z
X
With thanks to my coauthors:J. Pablo Bonilla Ataides,David K. Tuckett,Stephen D. Bartlettand Steven T. Flammia
Scaling quantum computersTo scale a quantum computer we must:
Control a large number of qubits below threshold as they performrepeated stabilizer measurements1,2.
Furthermore, there is a huge overhead cost to performfault-tolerant quantum logic operations.3
1M. Takita et al., Phys. Rev. Lett. 119, 180501 (2017)2C. Kraglund Andersen et al, Nat. Phys. 16, 875 (2020)3C. Gidney and A. Fowler, Quantum 3, 135 (2019)
Scaling quantum computersCat qubits 4, 5 (that experience highly biased noise) have beenproposed as high-quality qubits (with bias preserving gates 6, 7) forsurface code implementations8.
=
= DC flux-bias for the SNAIL
= out-of-plane interconnectthrough which all the microwave drives (such as the two-photon drives, coupling drives for CX gates and drives for single-qubit gates) are applied
= resonator for qubit readout
SNAIL
4 μm
d
a
2 cm
Time
c
b
e f
4A. Grimm et al. Nature 584, 205 (2020)5R. Lescanne et al., Nat. Phys. 16, 509 (2020)6J. Guillaud and M. Mirrahimi, Phys. Rev. X 9, 041053 (2019)7S. Puri et al. Sci. Adv. 6, eaay5901 (2020)8C. Chamberland et al. arXiv:2012.04108 (2020)
The XZZX surface code
The XZZX surface code9 is locally equivalent to the standardsurface code10
Z X
Z
X
XX
Z
Z
X Z
Z X
X X
X X
Z Z
Z Z
9X.-G. Wen, Phys. Rev. Lett. 90 16803 (2003)10A. Kitaev, Ann. Phys. 303, 2 (2003)
The XZZX surface code
. . . except every other qubit is rotated by a hadamard gate.(the green qubits are rotated)
Z X
Z
X
XX
Z
Z
X Z
Z X
X X
X X
Z Z
Z Z
The XZZX surface codeThe XZZX code demonstrates:
I Thresholds that match the hashing bound for all uniformsingle-qubit Pauli noise channels.
I Exceptionally high fault-tolerant thresholds for biased noisemodels
I Reduced overheads (by a factor O(1/ηd/4
)) at low error rates
and high bias η
I We also argue that we can maintain these advantages whileperforming computation
Z X
Z
X
XX
Z
Z
X Z
Z X
X X
X X
Z Z
Z Z
Error correction with the surface code
We use stabilizers to measure (Pauli) errors E .
SE |ψ〉 = (−1)E |ψ〉
On the surface code, errors appear as strings, and defects appearin pairs at their endpoints11.
This can be viewed as a defect parity conservation law12,13.11E. Dennis et al., J. Math. Phys. 43, 4452 (2002)12A. Kitaev, Ann. Phys. 303, 2 (2003)13BJB and D. J. Williamson, Phys. Rev. Research 2, 013303 (2020)
Symmetries and conservation lawsMWPM is possible because the surface code respects a defectconservation symmetry14,15
Mathematically, the symmetry is apparent in the stabilizer group∏v
Av = 1 ⇒∏v
av = 1
(av = ±1 eigenvalues of stabilizers Av . )
⇒ #v with av = −1 is even ≡ defect conservation (mod2)
14E. Dennis et al., J. Math. Phys. 43, 4452 (2002)15BJB and D. J. Williamson, Phys. Rev. Research 2, 013303 (2020)
A symmetry with respect to Pauli-Z errors16, 17
We find a richer space of symmetries if we restrict our error model.
(a)
(b)
(c)
X Z
Z X
Z
ZZ
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Pauli-Z errors commute with the product of stabilizers along thediagonal and therefore respect a 1D symmetry on the XZZX code.
Decoding Z errors is therefore equivalent to decoding therepetition code (50% threshold).
16BJB and D. J. Williamson, Phys. Rev. Research 2, 013303 (2020)17D. K. Tuckett, Phys. Rev. Lett. 124, 130501 (2020)
Symmetries for all Pauli errors18, 19
The XZZX code has many symmetries for other error models.
(a)
(b)
(c)
(d)
(e)
X Z
Z X
X
X
X
Z
ZZ
Z
Z
Z
Z
Y
Y
Y
The XZZX code has symmetries with respect to Pauli-X errorsalong perpendicular diagonals to those for Pauli-Z errors.
Pauli-Y errors have the same symmetry as the CSS surface code(we exploited these symmetries in earlier work)
18BJB and D. J. Williamson, Phys. Rev. Research 2, 013303 (2020)19D. K. Tuckett, Phys. Rev. Lett. 124, 130501 (2020)
High threshold error ratesThe XZZX code has a threshold that matches the zero-ratehashing bound for all uniform single-qubit Pauli channels.
X
Y
Z0.1
0.2
0.3
0.4
0.5
Threshold
errorrate:pc
X
Y
Z0.1
0.2
0.3
0.4
0.5
Threshold
errorrate:pc
0.1
0.2
0.3
0.4
0.5
pc
XZZX CSS
R ≡ k/n ≥ 1− H(~p), (1)
H(~p) - Shannon entropy.
~p = (pX , pY , pZ ) with pO the probability that Pauli error O occurs.
R - rate with R = 0 the zero-rate hashing bound.
Numerics exceeding hashing
At high bias, our numerics demonstrate a threshold above thezero-rate hashing bound
100 101 102 1030.1
0.2
0.3
0.4
0.5
Bias: η
Threshold
errorrate:pc
XZZX
CSS
Hashing
∞13,17,21,25
17,21,25,29
21,25,29,33
25,29,33,37
29,33,37,41
33,37,41,45
37,41,45,49
41,45,49,53
45,49,53,57
49,53,57,61
53,57,61,65
57,61,65,69
61,65,69,73
65,69,73,77
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Code distances used in pc estimationpc−
ph.b
.
η = 30
η = 100
η = 300
η = 1000
See also other work on exceeding hashing for instance20,21,22
20D. DiVincenzo et al., Phys. Rev. A 57, 830 (1998)21G. Smith and J. Smolin, Phys. Rev. Lett. 98, 030501 (2007)22J. Bausch and F. Leditzky, arXiv:1910.00471 (2019)
Exceeding hashing
A zero-rate code with a threshold above hashing implies we cansend information at non-zero rate with p above hashing.
We need the following:
I A finite rate code Rout = Kout/Nout > 0
I An inner code of constant size Nin with threshold pth. > ph.b.
Concatenating gives a family of codes with rate:
R ′ = Rout/Nin > 0. (2)
using qubits with ph.b. < p < pth. for some constant Nin.
Numerics exceeding hashing
We also exceed hashing with a suboptimal matching decoder
100 101 102 1030.1
0.2
0.3
0.4
0.5
Bias: η
Thresholderrorrate:pc
XZZX
CSS
Hashing
∞
X
Z
Z
X ZZXdX
dZ
X X
Z
X
Z
X
(a)
(b)
(c)
We achieved this using XZZX codes with different boundaryconditions.
The high-speed matching decoder allows us to probe very largesystem sizes
Fault-tolerant thresholds with biased noise
Matching decoders generalise readily to the fault-tolerantsetting.23, 24, 25
X
t{∆(a)
(b)
(c)
(d)
Z
MM
M
Z
X
100 101 102 1030.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
Bias: ηThresholderrorrate:pc
XZZX
CSS
Standard decoder
∞
We find exceptional fault-tolerant thresholds for the XZZX codeunder phenomenological biased noise.
23E. Dennis et al., J. Math. Phys. 43, 4452 (2002)24BJB and D. J. Williamson, Phys. Rev. Research 2, 013303 (2020)25D. K. Tuckett, Phys. Rev. Lett. 124, 130501 (2020)
Logical failure rates below thresholdWe need a low logical failure rate Pwith a small number of qubits n.
At low error rate p we can generically achieve
P ∼ pd/2 (3)
To leading order, this is the probability of d/2 errors occurring.
Z ZZ
ZZ Z
Z
ZX
XXX
Z Z Z
The surface code uses n = O(d2) qubits.
Logical failure rates below threshold
We can choose an XZZX code with a weight n Pauli-Z logical.
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
At infinite bias26, we can expect a logical failure rate like
P ∼ pn/2. (4)
. . . but what happens at finite bias?
26D. Tuckett et al. Phys. Rev. X 9, 041031 (2019)
Logical failure rates below threshold
In practice we find errors ofd/4 low rate errorsand d/4 high rate errors.
X
X
X
XZ
Z
Z
P ∼ pd/4︸︷︷︸d/4 highrate errors
×(p
η
)d/4
︸ ︷︷ ︸d/4 low
rate errors
=
(1
η
)d/4
︸ ︷︷ ︸factorof
improvement
pd/2. (5)
Logical failure rates below thresholdWe test the ansatz at low p using the splitting method27,28
10−4 10−310−27
10−22
10−17
10−12
10−7
Physical error rate: p
Logicalerrorrate:P
d = 5
d = 7
d = 9
d = 11
d = 13
d = 15
P ∼ pd/4︸︷︷︸d/4 highrate errors
×(p
η
)d/4
︸ ︷︷ ︸d/4 low
rate errors
=
(1
η
)d/4
︸ ︷︷ ︸factorof
improvement
pd/2. (6)
27C. Bennett, J. Comput. Phys. 22, 245 (1976)28Bravyi and Vargo, Phys. Rev. A 88,062308 (2013)
Logical failure rates below thresholdFor small n and high bias we identify signatures of pd
2logical
failure rate scaling
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
7 9 11 13 15
10−3
10−2
Code distance: d
Logicalfailure
rate:P
−1.4 −1.3 −1.2
−0.3
−0.25
−0.2
log(p/(1− p))
LineGradients
We observe this when
pd2 �
(1
η
)d/4
pd/2 (7)
(in other words, at moderate p and small-ish n)
Logical failure rates below thresholdIt is favourable to find codes with open boundary conditions.
Changing the lattice geometry means we have logical failure rateslike
PX ∼(p
η
)dX /2
and PZ ∼ pdZ/2, (8)
at low p.
X
Z
Z
X ZZXdX
dZ
X X
Z
X
Z
X
(a)
(b)
(c)
We can therefore save resources by choosing small dXwithout compromising PZ .
Fault-tolerant quantum computation
We can perform computations using code deformations29,30, e.g.,
Braiding twists31,32
Z Z
Z Z
Z Z
Z Z
Z Z
Z Z
Z Z
Lattice surgery33
29R. Raussendorf et al. Phys. Rev. Lett. 98, 190504 (2007)30H. Bombin and M. A. Martin-Delgado, J. Phys. A 42, 095302 (2009)31H. Bombin, Phys. Rev. Lett. 105, 030403 (2010)32BJB et al., Phys. Rev. X 7, 021029 (2017)33C. Horsman et al., New J. Phys. 14, 123011 (2012)
Fault-tolerant quantum computation
We can maintain these advantages undergoing fault-tolerantquantum computation34.
Z
XX
Z
Y
Z
XX
Z
X
Z
Z
X
X
Z
Z
X
X
Z Z Z Z Z
X
Z Z
XdX
dZ
(c)
(d)
(f)
(e)
(a)
(b)
(top left)
(bottom left)
(right)
The one-dimensional symmetries we need for high thresholds aremaintained under initialisation and XZZX codes with twists.
34D. Litinski, Quantum 3, 128 (2019)
Forthcoming work:XZZX thresholds with cat qubit circuits
1.61.41.21.00.8 10-4
4.03.53.02.52.0 10-44.5
1-photon cat
6.25-photon cat
10-1
10-2
10-3
10-4
10-1
10-2
10-3
10-4
10-5
Log
ical
err
or r
ate
Log
ical
err
or r
ate
(a)
(b)
10-4
10-3
10-2
10-1
10-2
10-1
0.007 0.008 0.009 0.010 0.011
0.007 0.008 0.009 0.010 0.011
Log
ical er
ror
pro
bab
ility
0.016 0.0220.018 0.020
0.016 0.0220.018 0.020
Total CX error probability
(a)
(b)
Biased noise CX
Total CX error probability
Logic
al er
ror
pro
bab
ility
Ordinary CX
work with A. Darmawan, A. Grimsmo, S. Puri and D. Tuckett.
Outlook
We have shown the XZZX code:
I has threshold error rates that match hashing
I has exceptionally high fault-tolerant thresholds
I significantly reduced resource costs below threshold
I can maintain its performance during computational operations
Z X
Z
X
XX
Z
Z
X Z
Z X
X X
X X
Z Z
Z Z
Outlook
Our work raises questions in several areas, such as:
I Have we really exceeded the hashing bound?
I How do we specialise codes for a given noise model in general?
I What is the potential for non-CSS codes?
Z X
Z
X
Z
X
Z
X
Z
X
Z
X
Z
(a)
(b)
(d)(c)
(e)
(f)
(g)
X Z
Z X
X
X
X
Z
Z
Z
Z
Z
Z
Z
Z
Z
Y
Y
Y
X
t{∆(a)
(b)
(c)
(d)
Z
MM
M
Z
X
Backup slide: Different geometriesPrevious work35 has shown negligible difference in logical errorrates against bit-flip noise for modest p as a function of number ofqubits n.
Z ZZ
ZZ Z
Z
ZX
XXX
Z Z Z
y
x
x
y
p
√n
pth
50
Monte Carlo
B.V. method
Low
est-order
approxim
ation
Generalised path-counting model
0.05 0.06 0.07 0.08 0.09 0.10p
-4
-3
-2
-1
log10P
5 6 7 8 9 10
0.9951.0001.0051.0101.0151.0201.025
p/%P
◆/P
■
35M. Beverland et al. J. Stat. Mech.:Theo. Exp. 2019, 073404 (2019)