Robust Characterization of Quantum Processes

Transcription

1 Robust Characterization of Quantum Processes Shelby Kimmel Center for Theoretical Physics, MIT Marcus da Silva, Colm Ryan, Blake Johnson, Tom Ohki Raytheon BBN Technologies JQI Monday Dec. 16

2 Why don t we have a working quantum computer? Too Many rrors

3 Can Improve Operations with Better Characterization of rrors Improvement to Computer Cooling Depolarizing error Improvement to Computer Magnetic Shielding xtra rotation around z-axis

4 Can Improve rror Correcting Codes with Better Characterization of rrors Improvement to rror Correcting Code? Non local, correlated error

5 Standard Techniques Have Problems Need nearly perfect state preparation, measurement and other operations. Otherwise systematic errors give inaccurate or even invalid results. Not robust

6 Robust Techniques Gate Set Tomography Procedures [Stark 13, Blume- Kohout et al. 13, Merkel et al. 12] Characterizes many processes at once Randomized Benchmarking (RB) [merson et al. 05, Knill et al. 08, Magesan et al. 11, 12] Can only characterize 1 parameter of 1 type of process. almost all any Can efficiently test performance of a universal gate set.

7 Outline Background: Issues with standard process characterization Randomized benchmarking framework, challenges of current implementation Our Results: Robust characterization of unital part of any process fficient bound on average fidelity of universal gate set.

8 Quantum Process (Map) Completely positive trace preserving (CPTP) map = any process that takes valid quantum states to valid quantum states..g. unitary, depolarizing process, dephasing process, amplitude damping process nn qubits, OO(16 nn ) free parameters

9 Problem with Standard Process Tomography 0 0, 1 0 Λ 0 Λ 0/1 0, 1

10 Problem with Standard Process Tomography 0 Λ 0 Λ 0/1 0, 1 + Λ + Λ +/ +,

11 Repeated Application Λ 0 0 Λ 0/1 0, 1 Λ 0 0 Λ 0/1 0, 1 Λ 0 0 Λ 0/1 0, 1

12 Repeated Application Λ 0 0 Λ 0/1 0, 1 Λ 0 0 Λ 0/1 0, 1 Λ 0 0 Λ 0/1 0, 1 If eigenstate of, will only see how acts on this state

13 Randomized Benchmarking Λ 0 0 Λ 0/1 0, 1 Λ 0 0 Λ 0/1 0, 1 Λ 0 0 Λ 0/1 0, 1 Randomizing Unitaries Recovery Unitaries

14 Randomized Benchmarking Simulated Randomized Benchmarking xperiment Value of Measurement Observable Sequence Length Decay constant depends on one parameter of

15 Randomized Benchmarking Λ 0 0 Λ 0/1 0, 1 Λ 0 0 Λ 0/1 0, 1 Λ 0 0 Λ 0/1 0, 1 Randomizing Unitaries Have rrors! Recovery Unitaries

16 Two Issues with RB 1. How can we extract more than just 1 parameter? 2. How can we deal with errors on the randomizing operations?

17 Randomizing Operation: Clifford Twirl 1 CC ii CC ii in Cliffords CC ii CC ii ρρ = 1 qq ρρ + qq II dd Result is depolarizing channel (very simple process) that depends on only one parameter of : Average fidelity of to the identity Average fidelity of = dd ψψ ψψ ( ψψ ψψ ) ψψ

18 Randomizing Operation: Clifford Twirl verything inside the Clifford twirl gets simplified to a depolarizing channel CC ii CC ii To implement (approximately), repeat many times, each time randomly choosing CC ii, and average results

19 Randomizing Operation: Clifford Twirl 0 Λ 0 Λ 0/1 0, 1 Randomizing Operations Recovery Unitaries CC CC ii ii CC jj CC jj

20 Randomizing Operations Simulated Randomized Benchmarking xperiment Value of Measurement Observable Sequence Length Decay constant depends on 1 parameter of : Average fidelity of to the identity.

21 1. xtracting More Information CC CC ii ii CC jj CC jj Twirl simplifies too much! no twirl stick additional information inside twirl

22 1. xtracting More Information CC CC xx CC ii ii CC jj CC xx CC jj CC xx is fixed not random. The same CC xx is applied in each twirl.

23 1. xtracting More Information Simulated Randomized Benchmarking xperiment Value of Measurement Observable Sequence Length Decay constant depends on 1 parameter of : Average Fidelity of to CC xx (can have fast decays)

24 1. xtracting More Information CPTP map: 16 nn 4 nn parameters for nn-qubit map 4 nn To compose two maps, just multiply matrices! 4 nn Vectors VV span a subspace SS Learn inner product between VV and unknown vector uu Can learn projection of uu onto SS Cliffords span unital part Learn inner product between Cliffords and Learn projection of onto unital subspace

25 2. Dealing with rrors CC ii Λ CC CC xx CC ii CC jj Λ CC CC xx CC jj

26 2. Dealing with rrors CC ii Λ CC CC xx CC ii CC jj Λ CC CC xx CC jj

27 2. Dealing with rrors almost complete characterization of Λ CC almost complete characterization of Λ CC = almost complete characterization of All without the systematic errors of previous procedures!

28 xperimental Implementation

29 Negative Witness Test [Moroder et al. 13] To be a valid quantum process, must be trace preserving and completely positive Complete positivity = in Choi representation, all eigenvalues must be positive Negative witness test: Look at value of smallest eigenvalues of reconstructed map in Choi representation. If negative, BAD!

30 Negative Witness Test for Hadamard RBT RBT (No error correction) QPT QPT

31 fficient Fidelity stimate Requires an exponential number of measurement settings with different CC xx Instead, only want to check that your operations are good enough. Want to check implementation of Clifford Gates and T gates = universal gate set

32 fficient Fidelity stimate Λ CC CC xx Average fidelity to any unitary UU of O(log nn) T gates O(poly nn) Cliffords only need to repeat for O(poly nn) different CC xx. Λ CC II If Λ CC is close to Identity, can closely bound the average fidelity of to UU. Can test a universal gate set!

33 Conclusions and Open Questions Can robustly measure unital part of any quantum process Can efficiently and robustly test fidelity of universal quantum gate set operations. xperimentally implemented with superconducting qubit system at BBN What about the non-unital part? Can we extract other information efficiently and robustly (compressed sensing?) How does RB compare to Gate Set Tomography methods?

34 fficient Fidelity stimate Average Fidelity (, UU) tr UU aa xx xx CC xx Unitaries composed of Cliffords and O(log nn) T gates can be written as a linear combination of O(poly nn) Cliffords. Only need to measure O(poly nn) traces, each of which can be done efficiently.

35 fficient Fidelity stimate Average Fidelity (, UU) tr Λ CC UU aa xx xx CC xx Since we haven t characterized Λ CC, we can t get rid of its effect. However, we can measure its average fidelity to the identity, and if it is close to the identity, we can bound its effect

36 fficient Fidelity stimate Λ CC CC xx To get average fidelity to any unitary UU of O(log nn) T gates and O(poly nn) Cliffords, only need to repeat for O(poly nn) overlaps CC xx. Λ CC II If Λ CC is close to Identity, can closely bound the average fidelity of to UU. Can test a universal gate set!

37 What do we measure? Λ CC CC xx We measure tr Λ CC CC xx Λ CC CC xx

38 What can we measure? CPTP map: 16 nn 4 nn parameters for nn qubit map 4 nn 4 nn Choose many different CC xx ss and measure trace. By measuring enough traces can learn unital part We learn: 16 nn 2 4 nn + 1 parameters

39 What do we characterize? CC xx Need to repeat for 16 nn different CC xx 4 nn 4 nn Unital part. (Pauli-Liouville Representation)

40 2. Dealing with rrors CC ii Λ CC CC xx CC ii CC jj Λ CC CC xx CC jj Randomizing Operation = Twirl

41 Can we do better with Randomized Benchmarking? Can we robustly characterize many parameters of any operation? Can characterize We show almost Cliffords all span parameters unital part of of any quantum maps. By learning average quantum map fidelity of to many CC xx ss, can learn projection of into unital subspace. What information can we obtain robustly and efficiently? Can test performance of a universal gate set.