Qiskit Global Summer School 2024

Execution on noisy
quantum hardware
Fighting errors before fault tolerance

Pedro Rivero
Quantum Algorithm Engineering
Technical Lead
IBM Quantum
Qiskit Patterns
The anatomy of a quantum algorithm

Step 1 Step 2 Step 3 Step 4

Map classical inputs to a Optimize problem for Execute using Qiskit Analyze result
quantum problem quantum execution. Runtime Primitives. in classical format.

PassManager([UnitarySynthesis(),
BasisTranslator(),
Sampler 000101...,
110110...
EnlargeWithAncilla(),
AISwap(), ⃗
circuit(𝜃) bit-strings
Collect1qRuns(),
Optimize1qGates(),
Collect2qBlocks(),
ConsolidateBlocks()])

Estimator ⟨𝑂⟩
⃗
circuit(𝜃) + ̂ expectation
observable 𝑂 value

IBM Quantum / © 2024 IBM Corporation 3

Qiskit Patterns
The anatomy of a quantum algorithm

Step 1 Step 2 Step 3 Step 4

Map classical inputs to a Optimize problem for Execute using Qiskit Analyze result
quantum problem quantum execution. Runtime Primitives. in classical format.

PassManager([UnitarySynthesis(),
BasisTranslator(),
Sampler 000101...,
110110...
EnlargeWithAncilla(),
AISwap(), ⃗
circuit(𝜃) bit-strings
Collect1qRuns(),
Optimize1qGates(),
Collect2qBlocks(),
ConsolidateBlocks()])

Estimator ⟨𝑂⟩
⃗
circuit(𝜃) + ̂ expectation
observable 𝑂 value

IBM Quantum / © 2024 IBM Corporation 4

Noise in quantum systems

Quantum computers are noisy “in every way possible”

Fault tolerance is still unfeasible today

We need interim ways to recover a better signal:

1. Limit the amount of noise
2. Clean the signal by filtering the noise out

This is accomplished by:

1. Run modified noisy quantum computations
2. Process collected outputs on a classical computer
3. Compute an improved result

IBM Quantum 5
Fighting noise in Suppression • Reduce or avoid the impact of errors
quantum systems • Before or during execution (typically)
• Requires additional classical resources

Mitigation • Filter errors out after they occur

• After or during execution (typically)
• Requires additional quantum resources

Correction • Detect and fix errors as they occur

• During execution
• Requires additional quantum and
classical resources

Source:
https://www.ibm.com/quantum/blog/quantum-error-
IBM Quantum suppression-mitigation-correction 6
Sources of noise

• SPAM errors: Readout errors

related to state preparation Environmental noise
and measurement/readout
• Gate errors:
imperfect operations on
qubits

• Environmental noise:
even if there are no
operations on qubits, these
are exposed to errors
coming from interaction
with the environment
Gate errors

IBM Quantum 7
Qiskit Runtime

• Sampler options: https://docs.quantum.ibm.com/api/qiskit-ibm-

runtime/qiskit_ibm_runtime.options.SamplerOptions
• Estimator options: https://docs.quantum.ibm.com/api/qiskit-ibm-
IBM Quantum runtime/qiskit_ibm_runtime.options.EstimatorOptions 8
Error suppression

Dynamical decoupling (DD)

Environmental noise: DD
Inserts sequences of gates in
idling qubits to avoid the
effects of cross-talk

Pauli twirling (PT)

Executes an ensemble of
equivalent quantum circuits
to alter the structure of the Gate errors: PT
observed noise

IBM Quantum 9
Dynamical decoupling (DD)

• Activity on neighboring
qubits can induce noise
while idling (i.e. cross-talk)
• Having gates applied to
qubits can help suppress
this effect

• The introduced gates need

to add up to the identity to
preserve the underlaying
unitary

• These gates will also

introduce errors, so there is
a balance to be found

IBM Quantum 10
Dynamical decoupling (DD)

• Activity on neighboring
qubits can induce noise
while idling (i.e. cross-talk)
• Having gates applied to
qubits can help suppress
this effect

• The introduced gates need

to add up to the identity to
preserve the underlaying
unitary

• These gates will also

introduce errors, so there is
a balance to be found

IBM Quantum 11
Dynamical decoupling (DD)

Dynamical decoupling options: https://docs.quantum.ibm.com/api/qiskit-ibm-

IBM Quantum runtime/qiskit_ibm_runtime.options.DynamicalDecouplingOptions 12
Randomized compiling (twirling)

• Used to convert arbitrary

noise channels into other
forms of noise
• Executing statistical
ensembles of unitarily
equivalent circuits

• Pauli Twirling (PT) converts

any quantum channel into
a Pauli channel.
• Suppresses the impact of
coherent noise.
• Noisy results degrade in
more predictable ways
(useful for ZNE).
IBM Quantum 13
Randomized compiling (twirling)

• Used to convert arbitrary

noise channels into other
forms of noise
• Executing statistical
ensembles of unitarily
equivalent circuits

• Pauli Twirling (PT) converts

any quantum channel into
a Pauli channel.
• Suppresses the impact of
coherent noise.
• Noisy results degrade in
more predictable ways
(useful for ZNE).
IBM Quantum 14
Randomized compiling (twirling)

• Used to convert arbitrary

noise channels into other
forms of noise
• Executing statistical
ensembles of unitarily
equivalent circuits

• Pauli Twirling (PT) converts

any quantum channel into
a Pauli channel.
• Suppresses the impact of
coherent noise.
• Noisy results degrade in
more predictable ways
(useful for ZNE).
IBM Quantum 15
Randomized compiling (twirling)

• Used to convert arbitrary

noise channels into other
forms of noise
• Executing statistical
ensembles of unitarily
equivalent circuits

• Pauli Twirling (PT) converts

any quantum channel into
a Pauli channel.
• Suppresses the impact of
coherent noise.
• Noisy results degrade in
more predictable ways
(useful for ZNE).
IBM Quantum 16
Randomized compiling (twirling)

• Used to convert arbitrary

noise channels into other
forms of noise
• Executing statistical
ensembles of unitarily
equivalent circuits

• Pauli Twirling (PT) converts

any quantum channel into
a Pauli channel.
• Suppresses the impact of
coherent noise.
• Noisy results degrade in
more predictable ways
(useful for ZNE).
IBM Quantum 17
Randomized compiling (twirling)

• Used to convert arbitrary

noise channels into other
forms of noise
• Executing statistical
ensembles of unitarily
equivalent circuits

• Pauli Twirling (PT) converts

any quantum channel into
a Pauli channel.
• Suppresses the impact of
coherent noise.
• Noisy results degrade in
more predictable ways
(useful for ZNE).
IBM Quantum 18
Pauli twirling (PT)

Twirling options: https://docs.quantum.ibm.com/api/qiskit-ibm-

IBM Quantum runtime/qiskit_ibm_runtime.options.TwirlingOptions 19
Error mitigation

Twirled readout error Readout errors: TREX

extinction (TREX)

Diagonalizes the readout

error transfer matrix,
calibrates it, and applies its
inverse in post-processing

Zero noise extrapolation

(ZNE)

Measures the effects of

increased noise to infer what Gate errors: ZNE
the results would look like in
the absence of noise
IBM Quantum 20
Readout errors

• Readouts errors cause the

wrong states to be
measured
• This can be modeled as a
classical noise channel
• Readout error can be
measured per qubit and
the full error matrix
reconstructed as a tensor
product

• The inverse matrix can be

used for error mitigation
when efficiently calculable

IBM Quantum 21
Twirled readout error extinction (TREX)

• Diagonalizes the readout-

error transfer matrix via
measurement twirling
• Such diagonal matrix is
learned by running identity
calibration circuits

• Finally, one can trivially

invert the diagonal matrix
and apply it to the target
results

• Only valid for expectation-

value problems

IBM Quantum 22
Twirled readout error extinction (TREX)

• Diagonalizes the readout-

error transfer matrix via
measurement twirling
• Such diagonal matrix is
learned by running identity
calibration circuits

• Finally, one can trivially

invert the diagonal matrix
and apply it to the target
results

• Only valid for expectation-

value problems

IBM Quantum 23
Twirled readout error extinction (TREX)

Resilience options (V2): https://docs.quantum.ibm.com/api/qiskit-ibm-

IBM Quantum runtime/qiskit_ibm_runtime.options.ResilienceOptionsV2 24
Zero noise extrapolation (ZNE)

• Divided in two phases

1. Noise amplification:
Exact
the original circuit
unitary is executed at

Expectation value
different levels of noise
2. Extrapolation:
the zero-noise limit is
inferred from the noisy
expectation-value
results

• Needs careful attention but

exhibits great potential
Noise factor
• Only valid for expectation-
value problems
IBM Quantum 26
Zero noise extrapolation (ZNE)

• Divided in two phases

1. Noise amplification:
Exact
the original circuit
unitary is executed at

Expectation value
different levels of noise
2. Extrapolation:
the zero-noise limit is
inferred from the noisy
expectation-value
results

• Needs careful attention but

exhibits great potential
Noise factor
• Only valid for expectation-
value problems
IBM Quantum 27
Zero noise extrapolation (ZNE)

• Divided in two phases

1. Noise amplification:
Exact
the original circuit
unitary is executed at

Expectation value
different levels of noise
2. Extrapolation:
the zero-noise limit is
inferred from the noisy
expectation-value
results

• Exhibits great potential but

needs careful attention
Noise factor
• Only valid for expectation-
value problems
IBM Quantum 28
Noise amplification (ZNE)

• Pulse stretching commonly

requires costly pulse level
calibration of the hardware
• Gate folding is largely a
heuristic approach but
offers a good trade-off
between result quality and
resource requirements

• Probabilistic error
amplification (PEA)
requires learning circuit-
specific noise but has
general applicability and
strong theoretical backing

IBM Quantum 29
Extrapolation (ZNE)

• Theoretical/experimental
results predict exponential
decay in observed
expectation values

• Exponential extrapolation
mitigates aggressively but
is unstable, since the scale
is unknown

• Polynomial extrapolation is
stable but mitigates worse,
since it retains the scale of
the noisy data

• Needs careful attention but

exhibits great potential

IBM Quantum 30
Zero noise extrapolation (ZNE)

Resilience options (V2): https://docs.quantum.ibm.com/api/qiskit-ibm-

IBM Quantum runtime/qiskit_ibm_runtime.options.ResilienceOptionsV2 31
Probabilistic error amplification (PEA)

• Noise amplification
technique for ZNE

• Executing statistical
ensembles of circuits

• Two tasks per layer:

1. Noise learning
2. Noise injection
• General applicability and
strong theoretical backing

IBM Quantum 32
Probabilistic error amplification (PEA)

Layer noise learning options: https://docs.quantum.ibm.com/api/qiskit-ibm-

IBM Quantum runtime/qiskit_ibm_runtime.options.LayerNoiseLearningOptions 33
Fighting noise before error correction

• Different types of noise Readout errors: TREX

need different suppression Environmental noise: DD
and mitigation techniques
• Different techniques can be
combined

Gate errors: ZNE + PT

IBM Quantum 34
Combining techniques

IBM Quantum 35
Resilience levels

IBM Quantum 36