Why Qubits Are Fragile
A qubit is a physical object in a delicate quantum state. Heat, stray electromagnetic fields, imperfect control pulses and measurement all disturb it. Two numbers summarise the problem:
- Coherence time: how long a qubit keeps its state before the environment scrambles it.
- Gate error rate: the probability a single operation goes wrong. Good devices today report errors somewhere around 0.1% to 1% for two-qubit gates, depending on technology and vendor. That sounds small, but a circuit with a million gates would almost certainly fail.
Machines in this regime are often called NISQ (noisy intermediate-scale quantum) devices. They run short circuits and return statistical answers, useful for research and experiments, not for breaking cryptography.
Qubit Technologies at a Glance
| Technology | Strengths | Challenges |
|---|---|---|
| Superconducting circuits | Very fast gates; mature chip fabrication | Needs dilution-refrigerator cooling near absolute zero; wiring and connectivity limits |
| Trapped ions | Very high gate fidelity; all-to-all connectivity within a trap | Slower gates; scaling the number of ions per trap |
| Neutral atoms | Large arrays of identical atoms; reconfigurable layouts | Slower cycle times; atom loss during operation |
| Photonic | Operates at room temperature in parts; networking-friendly | Photon loss; probabilistic gates |
| Spin qubits in silicon | Compatible with semiconductor manufacturing | Early stage; fidelity and uniformity at scale |
No technology has clearly won. Different vendors pursue different ones, and which scales best to millions of components is an open engineering question.
Physical Qubits vs Logical Qubits
A physical qubit is the noisy hardware object. A logical qubit is an error-protected qubit encoded across many physical ones, with continuous checking and correction. Vendor qubit counts almost always refer to physical qubits. An algorithm like Shor on RSA-2048 needs thousands of logical qubits, each costing hundreds to thousands of physical qubits depending on the code and the hardware error rate.
Published resource estimates keep falling as algorithms and codes improve. A 2025 analysis by Craig Gidney estimated that RSA-2048 could be factored with under a million noisy physical qubits in under a week, down from earlier estimates of tens of millions. Treat any number as a moving estimate and check the latest research before quoting it.
Why Redundancy Works: A Runnable Demo
Classical error correction repeats a bit and takes a majority vote. Quantum states cannot be copied, so real quantum codes are subtler, but the core intuition (spread information over many noisy components and check for disagreement) carries over. This simulation encodes one bit in n copies, flips each copy with probability p, and measures how often the majority vote is wrong:
import numpy as np
rng = np.random.default_rng(1)
def logical_error_rate(p, n, trials=200_000):
"""n physical bits repeat one logical bit; each flips with probability p; decode by majority vote."""
flips = rng.random((trials, n)) < p
return float(np.mean(flips.sum(axis=1) > n // 2))
print(f"{'p':>6} {'n=1':>9} {'n=3':>9} {'n=5':>9} {'n=7':>9}")
for p in (0.2, 0.05, 0.01, 0.001):
print(f"{p:>6} " + " ".join(f"{logical_error_rate(p, n):>9.5f}" for n in (1, 3, 5, 7)))Output from our run:
p n=1 n=3 n=5 n=7
0.2 0.19941 0.10371 0.05732 0.03308
0.05 0.04958 0.00685 0.00111 0.00014
0.01 0.00995 0.00033 0.00001 0.00000
0.001 0.00105 0.00001 0.00000 0.00000Reading across a row shows errors falling as redundancy grows. Reading down shows the crucial condition: when the physical error rate is high (0.2), adding copies helps only slowly, but at low rates (0.01 and below) a handful of copies suppresses errors dramatically. Real quantum codes behave the same way with a threshold: below a certain physical error rate, bigger codes make the logical qubit better; above it, adding qubits makes things worse. For the leading surface code that threshold is commonly quoted at around 1% under idealised noise models, with practical hardware needing to be comfortably better than that.
What Makes Quantum Error Correction Harder
- No cloning: you cannot copy an unknown quantum state, so codes entangle data qubits with helper qubits instead.
- Two kinds of error: bit flips and phase flips both occur, and the code must catch both.
- Measure without looking: checks (called syndromes) reveal where an error occurred without revealing the encoded data, which would collapse it.
- Fast classical decoding: the correction must be computed in real time, a serious classical-computing and latency problem in its own right.
- Logical operations: performing gates on encoded qubits costs extra overhead, especially for non-Clifford gates, which usually require resource-hungry procedures such as magic-state distillation.
Milestones Worth Understanding
The most important result in recent years is demonstrating that a larger error-correcting code can perform better than a smaller one, meaning hardware is below threshold. Google Quantum AI reported such behaviour with its Willow chip in late 2024, and other groups have reported logical-qubit experiments on ion and neutral-atom platforms. These are vendor and academic reports of specific experiments, generally peer-reviewed, but they demonstrate scaling behaviour on a few logical qubits, not a machine that runs useful large programs.
Several vendors have published roadmaps targeting early fault-tolerant machines with on the order of tens to hundreds of logical qubits around the end of the decade. Treat those as targets, not commitments. Historically, roadmaps in this field have slipped as often as they have been met, and the hardest scaling problems (wiring, cooling, decoding speed, manufacturing yield) are engineering problems with uncertain timelines.
A Checklist for Reading Quantum Headlines
- Physical or logical? A headline qubit count without logical-qubit context says little.
- Which error rate? Look for two-qubit gate fidelity and measurement error, not just qubit count.
- Which benchmark? "Quantum advantage" claims are about specific tasks, often contrived, and are sometimes later matched by improved classical algorithms.
- Is it peer-reviewed or a press release? Prefer papers, and check whether independent groups reproduced the result.
- What is the circuit depth? Results at shallow depth do not extend to the million-operation circuits real algorithms need.
- Is the timeline a target or a delivery? Separate shipped hardware from planned hardware.
What This Means for Your Planning
Nobody can say when a cryptographically relevant quantum computer will exist, and credible experts disagree. That uncertainty is exactly why security teams plan for it now: if your data must stay confidential for 10 or more years, a machine arriving in 10 to 15 years is already your problem. The practical steps are in Post-Quantum Cryptography Migration. For everything else, quantum computing is a technology to watch and learn, using simulators and cloud access, not one to build production systems around today.
Frequently Asked Questions
What is the difference between physical and logical qubits?
A physical qubit is the actual noisy hardware element. A logical qubit is an error-protected qubit encoded across many physical qubits with continuous error detection and correction. Useful algorithms need logical qubits.
Which qubit technology will win?
It is unclear. Superconducting circuits, trapped ions, neutral atoms, photonics and silicon spin qubits each have different strengths in speed, fidelity and scalability, and vendors are pursuing several of them in parallel.
What does below threshold mean in quantum error correction?
It means the physical error rate is low enough that making the error-correcting code larger reduces the logical error rate. Above threshold, adding more qubits makes performance worse.
When will a quantum computer break RSA?
No one knows. It needs a large error-corrected machine, and estimates of the required physical qubits have fallen to around a million or fewer for RSA-2048, but building such hardware is an unsolved engineering challenge. Planning for it early is prudent because encrypted data can be stored now and decrypted later.
How should I judge a quantum advantage claim?
Check what benchmark was used, whether the result is peer-reviewed and independently reproduced, whether classical algorithms have since matched it, and whether the qubits are physical or logical. Advantage on a contrived task does not imply a useful speed-up.