Qiskit Global Summer School 2024

Characterizing
noise on quantum
hardware
Haimeng Zhang
Quantum Engineer

Samanthan Barron
Quantum Developer

IBM Quantum
Start with a Quantum circuit
A Quantum circuit consists of unitary gates applied to qubits.

Apply quantum gates

Qubit initialization Basis gates: ECR, ID, RZ, SX, X, Measurement
They form a universal gate set.

Qiskit Global Summer School 2024

Noise in today’s quantum hardware

Quantum circuits are In this lecture:

compiled and run physically
on the quantum hardware; Different sources of noise;

The performance and power Where to read the noise-

of today quantum processors related metrics from the
are still largely limited by IBM Quantum Platform;
noise;
How they are experimentally
We want to understand the characterized.
limitation of quantum
hardware due to noise so we
can make best use of it.

Qiskit Global Summer School 2024

Start with a Quantum circuit Rotation gates generated physically control Hamiltonian gates
by a Hamiltonian
θ
Rx(θ) = exp(−i σx)
2
= exp(−iHdrivet)

where Hdrive = vσx, θ = vt

Hadamard gate:
Measure in the qubit

( )
1 1 1
H= computational basis.
2 1 −1 For a single qubit, measurement
operators P0 = | 0⟩⟨0 | ,
| ψout⟩ = H | ψin⟩ P1 = | 1⟩⟨1 |
puts a qubit from ground produces bitstrings in the classical
state to superposition state form of 0s and 1s

Two-qubit gate generates entanglement.

CNOT ⋅ (H ⊗ I) | 0⟩ | 0⟩ = ( | 0⟩ | 0⟩ + | 1⟩ | 1⟩)/ 2
can put 2 qubits from ground state to max entanglement state

Qiskit Global Summer School 2024

What are the sources of noise?

Incoherent errors: Coherent errors in gates:

Loss of quantum Incorrect Hamiltonian evolution
information in the form of
superposition and
entanglement

State preparation and measurement (SPAM ) errors

Qiskit Global Summer School 2024

Representing a quantum state

A single-qubit quantum state can be represented by a state vector in the 2-dimensional Hilbert state.

(0)
1
Ground state | 0⟩ =

(1)
0
Excited state | 1⟩ =
Qiskit Global Summer School 2024
Representing a quantum state
besides ground and excited states they can also be present in coherent superposition states

(0)
1
Ground state | 0⟩ =

Coherent superposition state:

| ψ⟩ = α | 0⟩ + β | 1⟩, α, β ∈ ℂ
2
| α | : probability in ground state
2
| β | : probability in excited state
2 2
|α| + |β| = 1

(1)
0
Excited state | 1⟩ =
Qiskit Global Summer School 2024
Representing a quantum bit

Density matrix representation

Pure state: Mixed state:

∑
ρ = | ψ⟩⟨ψ | ρ= qi | ψi⟩⟨ψi | ,
i
describes an ensemble of pure states
{qi, | ψi⟩}.

classical mixture of pure states

Qiskit Global Summer School 2024

Representing a quantum bit
Bloch vector representation for single qubit states

(0)
1 1
ρ = (I + v ⃗ ⋅ σ )⃗ Ground state | 0⟩ =
2
where
σ ⃗ = (σx, σy, σz),
v ⃗ = (vx, vy, vz). max superposition states
1
X poles: | ± ⟩ = ( | 0⟩ ± | 1⟩)
Density matrix as a sum of Identity 2
matrix and (pauli matrices* bloch
vector matrices)
Z poles: ground and excited states

I
Center: maximally mixed state ρ =
2

(1)
0
Excited state | 1⟩ =
Qiskit Global Summer School 2024
 Effect of noise

ρ ρ′ = ℰ(ρ)
Noise channel
ℰ

Ideal Noisy

State Pure state Mixed state

⃗ =1
∥v∥ ⃗ <1
∥v∥
vector length

Operations Unitaries Noise channel ℰ

| ψ⟩ → U | ψ⟩ ρ → ℰ(ρ)

Qiskit Global Summer School 2024


SPAM errors
we want to prepare a state initialised in ground state but we
don't have full info of the state and there is a chance that it
might end up in excited state
Example of a mix state: Mathematically described by:

e.g. imperfect state preparation

⃗
p noisy ⃗
= A p ideal

Measurement (readout) error

pi: probability of measuring bitstring i
results in imperfect acquisition
of the classical output of the Aij: the probability of observing state j
quantum circuit. while being prepared in state i.

In practice, it is difficult to tell Challenging to characterize at scale:

them apart. (2^n)^2 elements.
n= no. of qubits
Reported on backend qubit by qubit.

Could be correlated between qubits.

Scalable Mitigation of Measurement Errors on Quantum Computers, PRX

Quantum 2, 040326 (2021)
A matrix for five qubits
SPAM errors

Example of a mix state: Mathematically described by:

e.g. imperfect state preparation

⃗
p noisy ⃗
= A p ideal

Measurement (readout) error

pi: probability of measuring bitstring i
results in imperfect acquisition
of the classical output of the Aij: the probability of observing state j
quantum circuit. while being prepared in state i.

In practice, it is difficult to tell Challenging to characterize at scale:

them apart. (2^n)^2 elements.

Reported on backend qubit by qubit.

Could be correlated between qubits.

Scalable Mitigation of Measurement Errors on Quantum Computers, PRX

Quantum 2, 040326 (2021)
 Noise channels

incoherent errors
ρ ρ′ = ℰ(ρ)
Noise channel
ℰ

Mathematical representation (Kraus):

† †
∑ ∑
ℰ(ρ) = Ki ρKi Kj Kj =I
i i

Linear these props ensure that these maps map a physically relevant
Complete positive density matrix for a density matrix
Trace preserving

Further reading: Quantum Channels, John Watrous, IBM Quantum Learning


Incoherent errors
Action of an amplitude damping channel on the Bloch sphere.

Energy relaxation: decay from excited state to ground state

Superconducting qubits operate at low temperatures, typically

around 20 mK with qubit frequency between 4-6 GHz.

Hence there is much stronger decay from the excited state | 1⟩ to

the ground state | 0⟩ compared with the excitation from | 0⟩ to
| 1⟩.

Energy relaxation is characterized the timescale T1;

−t/T1
The probability of a relaxation error is given by 1 − e for circuit
evolution time t.

Amplitude damping channel: Equivalently:

1 0 | 0⟩ ↦ | 0⟩ with probability 1 Ground state becomes the steady state of the

(0 1 − p)
K0 = , | 1⟩ ↦ | 0⟩ with probability p superconducting qubit, largely unaffected by relaxation errors.
ground state is largely unaffected by relaxation errors; other states are
K1 = p | 0⟩⟨1 | . affected by the t1 limit
takes excited state to ground state with a probability p 14
Incoherent errors
Action of a phase damping channel on the Bloch sphere.
shrinks size of x & y component of
Dephasing: coherent state loses its phase info Bloch vector

Superposition states become classical mixtures

Dephasing is characterized by the timescale T2,

The probability of a dephasing error is given by 1 − e −t/T2 for circuit

evolution time t.

Both relaxation and dephasing error turn quantum information

classical!

Phase flip channel:

K0 = 1 − p I,
K1 = p Z.

15
Incoherent errors
depolarizing channel
A special case when px = py = pz same probabilities
I
ℰ(ρ) = p + (1 − p)ρ maximally next state and with (1-p) maps
with prob p it maps the state to the
2
3p p
ℰ(ρ) = (1 − )ρ + (XρX + YρY + ZρZ)
to the excited state

4 4

Pauli noise channel

- applies single qubit Pauli operators to qubit state at each given probability
- Used to model noisy quantum processors in practice

∑
ℰ(ρ) = pj Pj ρPj
j
= pI IρI + pX XρX + pYYρY + pZ ZρZ

Pj: Pauli operators

pj: the error rate associated with Pj

Sparse Pauli noise channel is used to model noisy quantum

processors in practice [1]

[1] E. Berg et al., Probabilistic error cancellation with sparse Pauli-Lindblad models on
noisy quantum processors. Nature Physics, pages 1–6, 2023. Pauli noise channels are unital, meaning that it maps the maximally
mixed state to itself. - steady state is maximally mixed state
unital: maps an identity to an identity matrix and maximally mixed state to itself
16
Coherent errors
- associated with imperfect gate implementation

Gates are generated by turning on certain Hamiltonians that we Coherent errors can be modeled by
can control and engineer. unwanted unitary gates in the circuit.

Let’s look at the single-qubit case:

- if we want to build Misaligned rotation
θ
RX(θ) = exp( − i σx) = exp(−iHt)
an X gate; that
involves 180 deg
2 rotation around
H = vσx, θ = 2vt
X̃ = Yϵ X
X-axis; we turn on the
hamiltonian sigma x
What if we engineered the wrong Hamiltonian? and we calibrate the
desired rotaion axis noisy x gate y- gate ideal x gate
and rotation angle
Example: incorrect phase in a microwave generator
Over/under rotation
H = vσx + ϵσy.

Example: poorly-calibrated pulse durations. X̃ = Xϵ X

θ = 2v(t + t′)
x gate
with
over/under
rotation
angle

 Two-qubit errors (crosstalk)
ECR gates are way longer than single qubit gates; thus 2- qubits are considered to have more errors

Echoed cross-resonance (ECR) gate is equivalent to a

=
CNOT up to single-qubit rotations.

It implements the following Hamiltonian:

IX IZ ZI ZX ZZ
He = ωix + ωiz + ωzi + ωzx + ωzz
2 2 2 2 2
only ZX part is required; rest are unwanted; which can lead to errors

The ZZ static capacitive coupling results in unwanted

crosstalk errors to the connected qubits

An order of magnitude longer gate durations compared with

single-qubit gates.
ff
Coherent noise summary
Single-qubit over rotation for d cycles

Coherent errors often times A number of solutions for

result in oscillations in coherent errors are know:
signals. dynamical decoupling,
Twirling (random compiling).

It can build up much more

rapidly than incoherent
errors.
ideal oscillation will be either in 1 or -1; rest are unwanted

Example in Qiskit Global Summer School 2023, lecture by Dr. Zlatko Minev
 - qubit2 has very low coherence time
- ECR gates should not function parallely ideally

Consider a chain of physically connected qubits:

- in order to avoid crosstalk b/w 2 physically connected qubits, we must implement a barrier b/w their gates
- for decoherence, we can delay the scheduling
Crosstalk Decoherence

As-late-as-possible (ALAP)
scheduling

P. Murali et al., Software Mitigation of Crosstalk on Noisy Intermediate-Scale Quantum Computers, ASPLOS ‘20
 How noise propagates in time and space?
Many applications of near-term quantum computing involve layered circuits
Brickwork random circuit on a 1D chain

ℰ
two-qubit random gate

Local noise

implementing layers of 2-qubit circuits

1. The output of the circuit converges exponentially fast to the maximally mixed state,
2. Noise in the deep depth limit can be characterized by the global depolarizing noise.
K. Tsubouchi et al., Universal cost bound of quantum error mitigation based on quantum estimation theory, arXiv: 2208.09385 (2023)
𝒰
 How noise propagates in time and space?

ℰ Lightcone: noise in only a specific region of

the circuit can contribute to final expectation
values

⟨O⟩

In the case of estimating expectation values, only errors in the backwards lightcone (shaded pink)
contribute.

M. Tran, K, Sharma, K, Temme, Locality and Error Mitigation of Quantum Circuits, arXiv:2303.06496, 2023
𝒰
 Noise characterization
we impose random gates to impose certain symmetry on the
quantum channel

costly process first we prepare 4^n input states; then

we pass it through quantum vhannel
that we want to process; then we
measure 4^n states

ρ ρ′ = ℰ(ρ)
Quantum state tomography Noise channel
Learn an unknown quantum state ρ from experiments ℰ

For a single qubit, we can expand ρ as:

tr(ρ)I + tr(Xρ)X + tr(Yρ)Y + tr(Zρ)Z

ρ=
2

Measure the expectation value of the unknown state with all elements of a complete basis
Needs many copies of the unknown state
Needs to do 4n − 1 different measurements for n qubits -1 is beacuse we know that the 1 is rho's measurement
Exponentially hard; active area of research to make this easier.

 Noise characterization

ρ ρ′ = ℰ(ρ)
Quantum process tomography Noise channel
Learn an unknown quantum process ℰ from experiments ℰ

Prepare 4n input states: | n⟩ Measure ℰ(ρ) in 4n complete basis

| m⟩
| + ⟩ = ( | n⟩ + | m⟩)/ 2
| − ⟩ = ( | n⟩ + i | m⟩)/ 2
ℰ has d 4 − d 2 independent real parameters
n
16 experiments, even harder than state tomography; active
area of research to make it scalable.

How noise is characterized: randomized benchmarking (RB)
Can we quantify how much error happens per gate in a circuit?

Single/Two-qubit randomized benchmarking

C1 C2 Cm Cm+1

Ci is a single/two-qubit random gate sampled from a finite Clifford gate set;

Cm+1 is the performed to make the total sequence equal to identity operation

Measure the probabilities to get back to the ground state at the end of the sequence.

Vary sequence length, fit the fidelity decay to an exponential curve to report Error Per Clifford.

IBM Quantum 25
Extract error rate per gate (EPG) in RB noise model follows Depolarising noise model

Fit the survival probability to an exponential

m
Aα + B

Report average gate error: alpha= average gate error

d−1
ϵ= (1 − α)
d
Average gate fidelity
dFp + 1
Fg =
d+1
In RB, we report the average gate error per Clifford gate
ϵg = 1 − Fg

Limitation of RB: it does not capture coherent or

crosstalk error
Example from Rhys. Rev. Lett. 127.130501 (2021)
RB characteristic is how we characterise single qubit gate error and ECR gate error

26
Reported gate errors

27
Layer fidelity: error per layer gate (EPLG)
limitation of randomised benchmarking protocol: insensitive to coherent errors due to randomisation

Layer fidelity expands on N connected qubits

randomized benchmarking A set of connecting gates on that device
from the entangling Clifford gates
Layer structures are the building
blocks for a lot of circuits we We ask, what is the process fidelity of this
care about layer of gates?

Includes crosstalk

Relates to other metrics to infer

the error mitigation overhead

D. McKay et al., Benchmarking Quantum Processor Performance at Scale, arXiv: 2311.05933 (2023)

IBM Quantum 28
Steps
Layer fidelity circuits 3 1Q/2Q Simultaneous, Direct RB with Barriers

Define layer structure over Separating the layer into disjoint

1 physically connected qubits 2 sets and measure with RB

Layer 1 Layer 2

29
EPLG vs RB (EPG)
n2q is no. of 2 qubit gates
1/n2q
EPLG = 1 − LF

Eagle processor Heron processor

suffers less from crosstalk error

dark blue: layered

light blue: isolated
Quantile plot of the individual gate errors measured from the best
100 qubit chain from
simultaneous direct RB (“layered”) versus the backend reported
gate errors (“isolated”)

30
EPLG metric reported on
backends

Lab session: layer fidelity experiment via Qiskit Runtime led by Samantha Barron.

https://docs.quantum.ibm.com/run/system-information#system-configuration-values
31
How noise is characterized: randomized benchmarking
Can we quantify how much error happens per gate in a circuit?

Single-qubit randomized benchmarking

C1 C2 Cm Cm+1
Use the idea of randomization

For sequence length m

Step 1: fix m ≤ M − 1 and generate Km
sequences consisting of m + 1 quantum
operations;
Step 2: For each of the Km sequences,
measure the survival probability
Step 3: Average over the Km random
realizations to find the averaged
sequence fidelity Ci is a single-qubit random gate sampled from a finite Clifford gate set;
Cm+1 is the performed to make the total sequence equal to identity
Repeat Steps 1 through 3 for difference operation.
values of m, and fit for the model

IBM Quantum 33
Navigating the IBM Quantum Platform

Noise-related metrics
T1
T2
Readout assignment error
ID error
sx error
Pauli-x error
ECR error
Characterize noise by layer: layer fidelity
EPLG
Can we quantify how much error happens per layer in a circuit?

Edge-coloring problem of the coupling graph:

Assign colors to edges of the graph so that no two incident
edges have the same color.

IBM Quantum 35
Quantum hardware is
a device that computes an output from input data
using fundamental rules of quantum mechanics. readout signal

microwave pulse schedules

room temperature

quantum circuits
20 mK

quantum algorithms
Quantum information is stored in a quantum
bit, made of superconducting circuits

Qiskit Global Summer School 2024

Noise can be suppressed using single-qubit control gates:
dynamical decoupling
A sequence of pulses applied with the goal to remove unwanted system-bath interactions

Example: a pure dephasing system-bath coupling

z z
HSB = σ ⊗ B
Apply a time-dependent control on the system:
x
HS = λ(t)σ
Ideal pulse: δ → 0, λ → ∞
fτ −iτHSB
Free evolution: fτ ≡ e
−iδλσ x −i π2 σ x x
Pulse: X ≡ e ⊗ IB = e ⊗ IB = − iσ ⊗ IB

Evaluate Xfτ Xfτ at time t = 2τ:

λ: amplitude −iτHSB −iτHSB −iτσx HSBσx −iτHSB +iτHSB −iτHSB
Xfτ Xfτ = σxe σxe =e e =e e =I
δ: duration
Bath has no effect on the system at the instant t = 2τ!
37
Noise can be suppressed using single-qubit control gates:
dynamical decoupling
In Qiskit, DD is added to circuit evolution by a transpilation pass
Example in Qiskit Docs: https://docs.quantum.ibm.com/api/qiskit/qiskit.transpiler.passes.DynamicalDecoupling

38
Noise can be suppressed using single-qubit control gates:
dynamical decoupling
In practice, pulse has finite duration
Example in Qiskit Docs: https://docs.quantum.ibm.com/api/qiskit/
qiskit.transpiler.passes.DynamicalDecoupling

39
(Readout pulses omitted)
Summary

H = HS + HB + HSB Describes thermalization

SYSTEM
HSB: system-bath coupling
BATH

Qiskit Global Summer School 2024

τ = ψ⟩⟨ψ An ensemble of pure quantum states:

∑
ρ= qi | ψi⟩⟨ψi |
i
This is called the density matrix

Maximally mixed state:

1
2
0
ρ= 1
0 2

Pure state
Qiskit Global Summer School 2024
Mixed state
Source: 1. Lorem source name
Source: 2. Lorem second source name 41
Visualizing a quantum state

Bloch sphere representation

1
ρ = (I + v ⃗ ⋅ σ )⃗
2 ⃗ = 1,
For pure states, ∥ v ∥
where ⃗ <1
For mixed states, ∥ v ∥
σ ⃗ = (σx, σy, σz),
v ⃗ = (vx, vy, vz).

Qiskit Global Summer School 2024

How to model a
quantum computer?

H = HS + HB + HSB Describes thermalization

SYSTEM
HSB: system-bath coupling
BATH

Qiskit Global Summer School 2024

Quantum map
Transform one quantum state to another Pauli transfer matrix

Example:

Φ : ρ ↦ ρ′

Qiskit Global Summer School 2024


 Example: the dephasing
map
1

{ ρ w/ prob. 1 − p
I w/ prob. p
ρ ↦ ρ′ = 2

Equivalently,
I
ρ′ = p + (1 − p)ρ
2

Write in the form of Kraus operators,

3
K0 = 1 − pI,
4
p
Ki = σi, for i = 1,2,3
4

Qiskit Global Summer School 2024

 Noise channel :
Quantum map that transforms one quantum Fidelity of a state passing through a noise channel
state to another

Example:
Φ : ρ ↦ ρ′

↦
𝒩
Qiskit Global Summer School 2024

 Example: the
depolarizing map
1

{ ρ w/ prob. 1 − p
I w/ prob. p The depolarizing map is a unital map
ρ ↦ ρ′ = 2

Equivalently,
I
ρ′ = (1 − p)ρ + p
2

Write in the form of Kraus operators,

3
K0 = 1 − pI,
4
p
Ki = σi, for i = 1,2,3
4

Qiskit Global Summer School 2024

 Example: amplitude
damping map
| 0⟩ ↦ | 0⟩ with probability 1 The amplitude damping map is NOT a unital map
| 1⟩ ↦ | 0⟩ with probability p

Equivalently,
I
ρ′ = p + (1 − p)ρ
2

Write in the form of Kraus operators,

1 0
(0 1 − p)
K0 = ,

K1 = p | 0⟩⟨1 | .

Qiskit Global Summer School 2024


Summary of noise
examples Kraus operator
representation

Bit flip

Phase flip

Amplitude damping

Depolarizing

Pauli

Qiskit Global Summer School 2024

How do quantum computation fail

Coherent errors:
Incorrect Hamiltonian evolution

Incoherent errors:
Loss of quantum information in
the form of superposition and
entanglement

1 2 3 4
Qubits have finite lifetime Gates and their errors Noise channel examples Noise characterization

Qubit T1 time Single-qubit coherent noise Process tomography

Qubit T2 time Two-qubit crosstalk Randomized benchmarking
Hardware connectivity:
heavy-hex lattice

The connectivity dictations how The coupling map of the

noise propagates spatially device:
Nodes: qubits
Edges: native two-qubit
entangle gates are allowed

Qiskit Global Summer School 2024