The DiVincenzo Criteria: What It Takes to Build a Quantum Computer

Every few years a research group announces a new kind of qubit: a different atom, a different defect in diamond, a new superconducting circuit. The press release sounds exciting, but how do you tell whether the newcomer is a plausible quantum computer or a physics curiosity? In April 2000, physicist David DiVincenzo published a paper that answered that question with unusual clarity. His list of requirements — now called the DiVincenzo criteria — has become the standard engineering rubric for judging every quantum computing platform, and it still holds up a quarter of a century later.

At the time, the field was small but ambitious. Shor's factoring algorithm and Grover's search algorithm existed on paper, the basic principles of quantum error correction had been worked out, and the first tiny experiments — liquid-state nuclear magnetic resonance systems running algorithms on a handful of qubits — were appearing. What the field lacked was a shared definition of the finish line. DiVincenzo, then at IBM, supplied one: a short list of things any physical system must do before it can run quantum algorithms at scale, published as “The Physical Implementation of Quantum Computation” in Fortschritte der Physik (48:771–783). Its criteria are still the first thing a serious hardware proposal gets checked against.

The five criteria for quantum computation

The core of the paper is five requirements. They are deliberately technology-agnostic: they say nothing about superconducting circuits, trapped ions, or photons, because DiVincenzo wanted a test any platform could be measured against. Here they are, with what each one actually demands in practice.

1. A scalable physical system with well-characterized qubits

A good qubit is not enough; you need a system that can hold many of them without the control infrastructure growing out of control. “Well-characterized” is doing a lot of work here: it means you know the system's Hamiltonian, the transition frequencies of each qubit, how long each qubit stays coherent, and how each qubit couples to its neighbours. Without that knowledge, you cannot model errors, calibrate gates, or run error correction. A platform that demonstrates one beautiful qubit in isolation but has no plausible path to a hundred identical ones fails this criterion — and history is full of such platforms.

This is also the criterion that separates demonstrations from architectures. Superconducting chips, ion traps, and neutral-atom arrays all have credible scaling stories. Liquid-state NMR, the platform behind some of the earliest quantum algorithm demonstrations in the late 1990s, famously does not: the signal strength drops as the molecule gets larger, so it could never scale beyond a dozen or so qubits.

2. The ability to initialize qubits to a fiducial state

Every quantum algorithm starts from a known input, conventionally the all-zeros state |00…0⟩. If you cannot reliably prepare that starting state, the computation begins with corrupted data and nothing downstream can fix it. Think of it as the reset button on a classical computer: unglamorous, but nothing runs without it.

In practice, initialization takes different forms: superconducting qubits relax to their ground state in a millikelvin fridge; trapped-ion systems use optical pumping with a laser. What matters is initialization fidelity — modern platforms reset correctly well above 99 percent of the time — because every residual error feeds into the total that error correction must handle.

3. Long decoherence times, much longer than the gate-operation time

Decoherence — the loss of quantum information to the environment — is the enemy. Two timescales describe it: T₁, the time for a qubit to relax from its excited state (energy loss), and T₂, the time over which the qubit's phase remains well-defined (dephasing). DiVincenzo's requirement is not that these times be long in absolute terms, but that they be long relative to how fast you can operate.

The useful figure of merit is the number of gate operations you can fit inside one coherence time:

N_gates = T2 / t_gate

For quantum error correction to work, this ratio needs to be on the order of 10,000 or more per logical qubit. To see how platforms compare: IBM's 127-qubit Eagle processor (ibm_kyiv) reports median T₁ of 288 microseconds and T₂ of 127 microseconds, with two-qubit gates taking well under a microsecond — a ratio in the hundreds of thousands. Neutral-atom systems do even better on the raw coherence side, with T₁ beyond 100 seconds and T₂ ranging from 100 milliseconds to 20 seconds, though their gate operations are slower. The ratio, not the raw numbers, is what the criterion tests.

4. A “universal” set of quantum gates

A quantum computer must be able to run any algorithm, which means its gate set must be universal: able to approximate any unitary operation to arbitrary precision. Fortunately, universality is cheap. You do not need a huge menu of gates — arbitrary single-qubit rotations plus one entangling two-qubit gate (such as CNOT or CZ) are enough. The Solovay–Kitaev theorem guarantees that a finite gate set can approximate any unitary efficiently, so a typical universal set looks like {H, CNOT, T} or {H, CNOT, arbitrary Z-rotation}.

The physics of implementing those gates differs wildly by platform: microwave pulses for superconducting qubits, lasers for trapped ions and neutral atoms (via the Rydberg blockade), photon interactions for optical systems. The criterion does not care how you implement them. It cares that you can implement both ingredients — local rotations and entanglement — with high fidelity, and that the fidelity is good enough to compose many gates without the computation drowning in noise.

5. A qubit-specific measurement capability

A computation you cannot read out is useless. This criterion demands that you can measure individual qubits without disturbing their neighbours — which requires the readout signal to be distinguishable per qubit, either by spatial separation (the qubits are physically far apart) or by spectral distinguishability (each qubit answers on its own frequency).

Readout errors are typically the largest single error source in today's systems, often 1–5 percent per measurement versus 0.1–1 percent for two-qubit gates. The criterion itself is binary in spirit: can you point at one qubit and ask “what is your state?” and get a reliable answer? If measurement always perturbs the neighbours, the platform cannot run the syndrome measurements that error correction depends on.

Two more criteria for quantum communication

DiVincenzo added two further conditions, not for computation but for quantum communication — the exchange of quantum information between locations:

6. The ability to interconvert stationary and flying qubits. A qubit sitting in a chip is “stationary”; a photon carrying quantum information down a fibre is a “flying” qubit. Quantum networks need to convert between the two.
7. The ability to faithfully transmit flying qubits between specified locations. Once encoded into a photon, the quantum information must survive the trip without decohering — a hard problem in optical fibre (attenuation) and free space (atmospheric decoherence).

These were written with quantum key distribution in mind, but they have become relevant to computing too: the leading proposals for large-scale quantum computers are modular, linking smaller processors together. That makes the communication criteria part of the computing story as well.

How today's platforms stack up

The criteria are best read as a rubric, not a checklist. Every current platform satisfies them imperfectly at scale; the interesting question is at what qubit count and fidelity each criterion begins to strain.

Superconducting qubits (IBM, Google) lead on scale and gate speed: hundreds of qubits, sub-microsecond gates, and the best-documented coherence times at that scale. Their challenges are fabrication variation (qubits are never quite identical) and the wiring overhead of the dilution refrigerator.
Trapped ions (Quantinuum, IonQ) lead on fidelity: the longest coherence times relative to gate speed and all-to-all connectivity, since any ion can be entangled with any other. Their challenges are gate speed (slower than superconducting gates) and the engineering of scaling beyond tens of ions per trap.
Neutral atoms trap hundreds of atoms in optical tweezers and entangle them via the Rydberg blockade. Excellent coherence, flexible geometry; the frontier is gate fidelity and measurement speed.
Photonic systems naturally satisfy the communication criteria — their qubits are already flying — but building deterministic two-qubit gates with photons remains hard.
Silicon spin qubits promise compatibility with semiconductor fabrication, but are earlier on the scaling curve.

Notice what the criteria don't require: they don't demand that a single chip hold a million qubits, and they don't demand fault tolerance. Any exotic proposal — topological qubits, for instance — gets evaluated the same way: show scalability, initialization, coherence, universality, and measurement, or explain which one you're trading away and why.

Why the criteria still matter

Twenty-five years on, the list's power is its refusal to pick a winner. It turned “build a quantum computer” from a slogan into an engineering specification with testable requirements, and it gave the field a shared vocabulary for comparing technologies that share almost nothing physically. When you read that a new platform achieved some record, run it through DiVincenzo's five questions: does it scale, does it initialize, does it stay coherent long enough, is the gate set universal, can you measure one qubit at a time? If the answer to any of them is “not yet,” you know exactly where the real work remains.