Simple Algorithm
Quantum Computing · Excited State · 22 min

DiVincenzo Criteria & Benchmarks

What a quantum computer needs, and how we grade one.

In 2000 David DiVincenzo wrote down what any quantum computer needs: five requirements, plus two more for sending quantum information between machines. Today the hard questions are quantitative. How good is good enough, and how do you measure a gate's error when your state preparation and measurement are themselves imperfect?

What you’ll learn

  • State DiVincenzo's five criteria and two networking criteria, and attach a measurable figure of merit to each.
  • Define average gate fidelity and process fidelity, convert between them, and compute both for depolarizing noise.
  • Model T1T_1/T2T_2 decoherence and derive the coherence limit on single-qubit gate fidelity.
  • Explain SPAM errors and why randomized benchmarking is insensitive to them; extract an error per Clifford from a decay curve.
  • Say what holistic benchmarks (quantum volume, cross-entropy, layer fidelity) measure and what they hide.

The DiVincenzo criteria

DiVincenzo's 2000 paper turned "build a quantum computer" into a checklist. Five criteria concern computation; two more concern communication, moving quantum states between processors or across a network. Each has since become a set of numbers that experiments report.

#Criterion (DiVincenzo 2000)Modern figures of merit
1A scalable physical system with well-characterized qubitsqubit count and yield; known Hamiltonian, couplings and leakage levels; parameter spread between qubits
2The ability to initialize the state of the qubits to a simple fiducial state, such as ∣000…⟩\ket{000\ldots}reset fidelity and duration, including mid-circuit reset of ancillas
3Long relevant decoherence times, much longer than the gate operation timeT1T_1, T2T_2, gate durations, and the resulting coherence-limited error
4A "universal" set of quantum gatessingle- and two-qubit gate fidelities (e.g. from randomized benchmarking), crosstalk, leakage
5A qubit-specific measurement capabilitysingle-shot readout fidelity and duration; disturbance of neighboring qubits
6The ability to interconvert stationary and flying qubitsefficiency and fidelity of transferring a qubit onto a photon and back
7The ability faithfully to transmit flying qubits between specified locationschannel loss; remote entanglement rate and fidelity
Criteria 6 and 7 matter for networked and modular machines and for quantum communication.
Think first Which DiVincenzo criterion is the main bottleneck for photonic qubits, and which for topological qubits?

Photonic: criterion 4. Photons barely interact, so entangling gates are probabilistic and built from measurements, and loss eats into every operation. Topological: criterion 1. A well-characterized topological qubit has not yet been unambiguously demonstrated. (Superconducting and ion platforms mostly fight criterion 1 at scale: wiring and variability for chips, crowded chains and slow gates for ions.)

Fidelity: how close is close enough?

To grade a gate, compare the operation E\mathcal E that the hardware actually performs, a quantum channel, with the ideal unitary UU. For pure states the fidelity is the overlap ∣⟨ψ∣ϕ⟩∣2\abs{\braket{\psi|\phi}}^2; for gates there are two standard averages.

DerivationDepolarizing noise
  1. E(ρ)=(1−p) ρ+p Id\mathcal E(\rho) = (1-p)\,\rho+p\,\frac{\id}{d}

    The depolarizing channel: with probability pp the state is replaced by the maximally mixed state. Take U=IU = \id.

  2. ⟨ψ|E(∣ψ⟩ ⁣⟨ψ∣)|ψ⟩=(1−p)+pd\mel{\psi}{\mathcal E(\dyad{\psi})}{\psi} = (1-p)+\frac{p}{d}

    The overlap is the same for every pure input, so it equals the average.

  3. Favg=1−d−1d p,r=d−1d pF_{\text{avg}} = 1-\frac{d-1}{d}\,p,\qquad r = \frac{d-1}{d}\,p

    For a qubit r=p/2r = p/2: a state replaced by I/2\id/2 still passes the fidelity test half the time.

  4. (I⊗E)(∣Φ⟩ ⁣⟨Φ∣)=(1−p)∣Φ⟩ ⁣⟨Φ∣+p Id⊗Id(\mathcal I\otimes\mathcal E)(\dyad{\Phi}) = (1-p)\dyad{\Phi}+p\,\frac{\id}{d}\otimes\frac{\id}{d}

    Acting on half of ∣Φ⟩\ket\Phi, whose reduced state is I/d\id/d.

  5. Fpro=1−p+pd2  ⟹  d Fpro+1d+1=1−d−1d p=Favg ✓F_{\text{pro}} = 1-p+\frac{p}{d^2}\;\Longrightarrow\;\frac{d\,F_{\text{pro}}+1}{d+1} = 1-\frac{d-1}{d}\,p = F_{\text{avg}}\ \checkmark

    The theorem checks out.

Check your understanding

A two-qubit gate has process infidelity 1−Fpro=1.0%1-F_{\text{pro}} = 1.0\%. What is its average gate error rr?

  1. 1.0%1.0\%
  2. 1.25%1.25\%
  3. 0.5%0.5\%
  4. 0.8%0.8\%
Show the answer

D. 0.8%0.8\%

r=dd+1(1−Fpro)r = \tfrac{d}{d+1}(1-F_{\text{pro}}) with d=4d = 4: r=45×1.0%=0.8%r = \tfrac45\times1.0\% = 0.8\%. Always check which fidelity a paper quotes: for two qubits the two differ by 20%.

Decoherence and the coherence limit

Two processes dominate idle errors. Amplitude damping at rate 1/T11/T_1 relaxes ∣1⟩→∣0⟩\ket1\to\ket0. Pure dephasing at rate 1/Tφ1/T_\varphi randomizes the phase without exchanging energy. Together they set the decay rate 1/T21/T_2 of the coherence ∣ρ01∣\abs{\rho_{01}}:

1T2=12T1+1Tφ\frac{1}{T_2} = \frac{1}{2T_1}+\frac{1}{T_\varphi}
Hence T2≤2T1T_2\le2T_1, with equality when there is no pure dephasing. On the Bloch vector: rz↦1−(1−rz)e−t/T1r_z\mapsto1-(1-r_z)e^{-t/T_1} and rx,y↦rx,y e−t/T2r_{x,y}\mapsto r_{x,y}\,e^{-t/T_2}.

Measured in a Ramsey experiment (two π/2\pi/2 pulses separated by a wait), the coherence time is called T2∗T_2^* and includes slow, quasi-static frequency fluctuations. A Hahn echo (a π\pi pulse halfway through the wait) refocuses those and gives the longer T2echoT_2^{\text{echo}}. Run the two standard experiments:

DerivationCoherence-limited gate fidelity
  1. r⃗ ′=Mr⃗+c⃗,M=diag⁡(e−t/T2,e−t/T2,e−t/T1),c⃗=(0,0,1−e−t/T1)\vec r\,' = M\vec r+\vec c,\qquad M = \diag\big(e^{-t/T_2},e^{-t/T_2},e^{-t/T_1}\big),\quad\vec c = \big(0,0,1-e^{-t/T_1}\big)

    During a gate of duration tt, model the noise as T1T_1/T2T_2 decay of the Bloch vector (the ideal rotation is undone in the comparison).

  2. ⟨ψ|ρ′|ψ⟩=12(1+r⃗⋅r⃗ ′)=12(1+r⃗⋅Mr⃗+r⃗⋅c⃗)\mel{\psi}{\rho'}{\psi} = \tfrac12\big(1+\vec r\cdot\vec r\,'\big) = \tfrac12\big(1+\vec r\cdot M\vec r+\vec r\cdot\vec c\big)

    The fidelity of a pure input (Bloch vector r⃗\vec r, ∣r⃗∣=1\abs{\vec r} = 1) with the output.

  3. ∫dψ  rirj=13δij,∫dψ  ri=0\int d\psi\;r_ir_j = \tfrac13\delta_{ij},\qquad\int d\psi\;r_i = 0

    Average over the sphere of pure states.

  4. Favg=12+Tr⁡M6=3+e−t/T1+2e−t/T26F_{\text{avg}} = \frac12+\frac{\Tr M}{6} = \frac{3+e^{-t/T_1}+2e^{-t/T_2}}{6}

    The exact result for this noise model.

  5. r≈t6T1+t3T2(t≪T1,T2)r\approx\frac{t}{6T_1}+\frac{t}{3T_2}\qquad(t\ll T_1,T_2)

    To first order in tt. With T2=2T1T_2 = 2T_1 this is r≈t/(3T1)r\approx t/(3T_1).

Predict before you look

A single-qubit gate lasts t=50t = 50 ns on a qubit with T1=T2=100T_1 = T_2 = 100 μs. What is its coherence-limited error rr, in units of 10−410^{-4}?

Show the answer

2.5 × 10⁻⁴

t6T1=506×105≈0.83×10−4\tfrac{t}{6T_1} = \tfrac{50}{6\times10^5}\approx0.83\times10^{-4} and t3T2=503×105≈1.67×10−4\tfrac{t}{3T_2} = \tfrac{50}{3\times10^5}\approx1.67\times10^{-4}, so r≈2.5×10−4r\approx2.5\times10^{-4}. Measured errors above this limit come from control imperfections, leakage or crosstalk, which tells you what to fix next.

SPAM: when the ruler is bent

Every characterization experiment starts with state preparation and ends with measurement, and errors in either (SPAM errors) contaminate the result. For readout the basic object is the assignment matrix Mij=P(read i∣prepared j)M_{ij} = P(\text{read } i\mid\text{prepared } j), often summarized by the assignment fidelity

FRO=1−12[P(0∣1)+P(1∣0)].F_{\text{RO}} = 1-\tfrac12\big[P(0\mid1)+P(1\mid0)\big].

Readout errors are usually asymmetric: a qubit in ∣1⟩\ket1 can relax during the measurement window, so P(0∣1)>P(1∣0)P(0\mid1)>P(1\mid0).

Think first How could you see a 0.1%0.1\% gate error through a 2%2\% readout error?

Make the gate error accumulate while SPAM happens only once: apply the gate (or a sequence of gates) many times before measuring. After mm gates the gate contribution grows roughly like m×0.1%m\times0.1\%, while SPAM stays fixed, so fitting the rate of decay versus mm separates the two. That is the idea behind randomized benchmarking; gate set tomography takes the other route and estimates everything self-consistently.

Randomized benchmarking

Randomized benchmarking (RB) turns "repeat many times" into a robust protocol:

  1. Pick mm gates uniformly at random from the Clifford group.
  2. Append the one Clifford that inverts the whole sequence, so that the ideal circuit is the identity.
  3. Prepare ∣0…0⟩\ket{0\ldots0}, run the sequence, and record whether you measure 0…00\ldots0 again (the survival probability).
  4. Average over many random sequences, for several lengths mm, and fit the decay.
F(m)=A pm+B,r=(d−1)(1−p)dF(m) = A\,p^m+B,\qquad r = \frac{(d-1)(1-p)}{d}
SPAM errors only affect AA and BB; the decay parameter pp gives the average error per Clifford rr (Magesan, Gambetta & Emerson 2011).
Check your understanding

Single-qubit RB yields p=0.998p = 0.998. What is the average error per Clifford?

  1. 0.1%0.1\%
  2. 0.2%0.2\%
  3. 0.15%0.15\%
  4. 0.4%0.4\%
Show the answer

A. 0.1%0.1\%

r=(d−1)(1−p)/d=(1−0.998)/2=0.1%r = (d-1)(1-p)/d = (1-0.998)/2 = 0.1\%. To quote an error per native gate, divide by the average number of native gates per Clifford, which only makes sense if all native gates have similar errors.

python
import numpy as np
from scipy.optimize import curve_fit

# Sequence lengths and survival probabilities, averaged over random sequences
m = np.array([1, 10, 25, 50, 100, 200, 400, 800])
F = np.array([0.979, 0.971, 0.957, 0.934, 0.893, 0.822, 0.716, 0.597])

def decay(m, A, p, B):
    return A * p**m + B

(A, p, B), cov = curve_fit(decay, m, F, p0=[0.5, 0.99, 0.5])
d = 2  # one qubit
r = (d - 1) * (1 - p) / d
print(f"p = {p:.4f}, error per Clifford r = {r:.1e}")

Illustrative data following A=0.48A = 0.48, B=0.5B = 0.5, p=0.998p = 0.998 (rounded), so the fit should recover r≈1×10−3r\approx1\times10^{-3}. Note that F(1)≈0.98F(1)\approx0.98: the 2% shortfall at short lengths is SPAM, and it does not bias pp.

Holistic benchmarks, and going deeper

Gate-level numbers miss what happens when everything runs at once: crosstalk, compilation overheads, limited connectivity, drift. Holistic benchmarks run whole circuits instead.

BenchmarkWhat it reportsMain caveat
Randomized benchmarkingaverage error per gate, free of SPAMan average over a gate set; hides coherent-error structure, leakage, and crosstalk unless run simultaneously
Quantum volume2n2^n for the largest nn-qubit, depth-nn random circuits whose "heavy" outputs appear more than 2/32/3 of the timeneeds classical simulation of every circuit, so it cannot certify very large nn
Cross-entropy benchmarking (XEB)fidelity of sampling from random circuits; about 0.2%0.2\% for Google's 53-qubit, 20-cycle circuits in 2019also needs classical simulation; can mislead when noise is not well described as uniform depolarization
Layer fidelityfidelity of full layers of simultaneous two-qubit gates across many qubitsspecific to the chosen layers and gate set

Key takeaways

  • DiVincenzo's five criteria (plus two for networking) are now quantitative: reset fidelity, coherence relative to gate time, gate and readout fidelities, and scalability.
  • Favg=(dFpro+1)/(d+1)F_{\text{avg}} = (dF_{\text{pro}}+1)/(d+1); depolarizing noise with parameter pp has r=(d−1)p/dr = (d-1)p/d.
  • 1/T2=1/(2T1)+1/Tφ1/T_2 = 1/(2T_1)+1/T_\varphi, and decoherence alone limits a single-qubit gate of duration tt to r≈t/(6T1)+t/(3T2)r\approx t/(6T_1)+t/(3T_2).
  • SPAM errors contaminate one-shot characterization; randomized benchmarking extracts a SPAM-free average error from F(m)=A pm+BF(m) = A\,p^m+B, with r=(d−1)(1−p)/dr = (d-1)(1-p)/d.
  • Holistic benchmarks (quantum volume, XEB, layer fidelity) capture crosstalk and compilation, but every single number hides structure.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

A platform has excellent gates, but its only readout destroys the entire register at once. Which criterion does it fail, and why does error correction care?

  1. Criterion 5 (qubit-specific measurement): error correction must repeatedly measure ancilla qubits mid-circuit without disturbing the data qubits
  2. Criterion 2 (initialization), because measuring resets the qubits
  3. Criterion 3 (coherence), because measurement causes decoherence
  4. None: any measurement suffices as long as it happens at the end
Show the answer

A. Criterion 5 (qubit-specific measurement): error correction must repeatedly measure ancilla qubits mid-circuit without disturbing the data qubits

Error correction extracts syndromes by measuring ancillas every cycle while the data qubits keep their encoded state. A readout that cannot address individual qubits, or that disturbs the rest, fails criterion 5.

Question 2

The single-qubit depolarizing channel ρ↦(1−p)ρ+p I/2\rho\mapsto(1-p)\rho+p\,\id/2 has average gate error

  1. pp
  2. 3p/43p/4
  3. p/2p/2
  4. p/4p/4
Show the answer

C. p/2p/2

r=d−1d p=p/2r = \tfrac{d-1}{d}\,p = p/2 for d=2d = 2: a state replaced by I/2\id/2 still overlaps the input with probability 12\tfrac12.

Question 3

A qubit has T1=100T_1 = 100 μs and T2=50T_2 = 50 μs. What is its pure-dephasing time TφT_\varphi?

  1. 5050 μs
  2. 200200 μs
  3. 100100 μs
  4. about 6767 μs
Show the answer

D. about 6767 μs

1/Tφ=1/T2−1/(2T1)=1/50−1/200=3/2001/T_\varphi = 1/T_2-1/(2T_1) = 1/50-1/200 = 3/200 per μs, so Tφ≈67T_\varphi\approx67 μs.

Question 4

Why is the RB decay parameter pp insensitive to state-preparation and measurement errors?

  1. Because RB uses only Clifford gates, which cannot produce SPAM errors
  2. Because SPAM occurs once per sequence, independently of mm, so it changes only AA and BB
  3. Because the final inverting gate corrects SPAM errors
  4. Because RB measures in a random basis each time
Show the answer

B. Because SPAM occurs once per sequence, independently of mm, so it changes only AA and BB

The survival probability is Tr⁡[E Λdepm(ρ)]\Tr[E\,\Lambda_{\text{dep}}^m(\rho)] with an imperfect ρ\rho and EE. Expanding gives A pm+BA\,p^m+B, where AA and BB depend on ρ\rho and EE, but pp depends only on the gate noise.

Question 5

Device A has 50 qubits with 1% two-qubit error; device B has 20 qubits with 0.1% error and all-to-all connectivity. Which has the larger quantum volume, and why?

  1. A, because quantum volume grows with qubit count
  2. Neither: quantum volume ignores error rates
  3. B, because quantum volume requires depth-nn circuits on nn qubits to succeed, and A's errors cap nn at a handful of qubits
  4. They tie, because both have more than 16 qubits
Show the answer

C. B, because quantum volume requires depth-nn circuits on nn qubits to succeed, and A's errors cap nn at a handful of qubits

An n×nn\times n model circuit contains roughly n2n^2 two-qubit gates. With 1% error per gate, circuits wider than about 7 qubits lose too much fidelity, even before routing overheads; with 0.1% error, all 20 of B's qubits can pass. Quality beats quantity.

Go deeper

Further reading

  • D. P. DiVincenzo, "The physical implementation of quantum computation", Fortschritte der Physik 48, 771 (2000).
  • M. A. Nielsen, "A simple formula for the average gate fidelity of a quantum dynamical operation", Physics Letters A 303, 249 (2002).
  • E. Magesan, J. M. Gambetta & J. Emerson, "Scalable and robust randomized benchmarking of quantum processes", Physical Review Letters 106, 180504 (2011).
  • J. Eisert, D. Hangleiter, N. Walk, I. Roth, D. Markham, R. Parekh, U. Chabaud & E. Kashefi, "Quantum certification and benchmarking", Nature Reviews Physics 2, 382 (2020).