Every quantum computation is a race against noise. Three fundamental channels corrupt qubit states in different ways — understanding which one dominates your hardware determines which mitigation strategy works.
A perfect qubit holds any point on the Bloch sphere indefinitely. Real qubits drift, shrink, and collapse. The trajectory depends on which noise channel dominates.
The qubit's phase drifts randomly — the x-y plane shrinks while z stays fixed. Like a spinning top wobbling. Only causes Z-type phase errors. This is why Tuna-9 shows 3–5x more Z errors than X errors.
Ice cube on a table — T₂ is like a crowd of clocks that started synchronized but gradually drift apart. Each clock still ticks, but they lose their shared rhythm. The average of all the clock hands shrinks to zero.
Every qubit has two lifetimes. T₁ measures how long it holds energy (|1⟩ → |0⟩). T₂ measures how long it holds phase coherence. T₂ is always ≤ 2T₁ — you can't maintain phase if you've already lost the energy.
On Tuna-9, T₂/T₁ ≈ 0.6 — well below the 2T₁ limit. This means pure dephasing (not relaxation) is the dominant decoherence mechanism. That's why Z-errors outnumber X-errors 3–5x in our Bell experiments.
Not all errors are equal. On Tuna-9, we proved through systematic experiments (ZNE gate folding, readout calibration) that 80% of error comes from measurement — not from the quantum gates themselves.
Measuring |1⟩ as |0⟩ (or vice versa). Asymmetric on Tuna-9: 8.5% for |1⟩→|0⟩ vs 0.7% for |0⟩→|1⟩.
Qubit state decays during the circuit. Worse for longer circuits and qubits with short T₂.
Imperfect rotations and entangling operations. Our ZNE experiment showed adding 4 extra CNOTs only added ~1 kcal/mol.
Qubit doesn't start perfectly in |0⟩. Small but systematic — affects every shot the same way.
Faulty speedometer — Readout error is like a speedometer that reads 55 when you're going 60. The car (quantum state) is fine — it's the measurement that's wrong. That's why readout error mitigation (calibrating the "speedometer") works so well, giving us a 119x improvement on IBM with TREX.
Quantum computing is a race: finish the computation before the qubit forgets. The ratio of coherence time to gate time determines how many operations you can perform.
| Platform | 1q gate | 2q gate | T₂ | Ops before T₂ | RB fidelity |
|---|---|---|---|---|---|
| Tuna-9 | 20 ns | 40 ns | 22 μs | ~550 | 99.82% |
| IQM Garnet | 32 ns | 72 ns | 30 μs | ~420 | 99.7% |
| IBM Torino | 30 ns | 68 ns | 150 μs | ~2200 | 99.5% |
IBM Torino has 4x the operations budget of Tuna-9 (2200 vs 550 before T₂), but its raw Bell fidelity is lower (86.5% vs 93.5%). Why? Because IBM uses default qubit placement, while we hand-picked the best Tuna-9 pair. Knowing your chip beats having more qubits.
We ran identical Bell-state circuits on all three platforms. The dominant noise channel determines both the error rate and which mitigation works.
Z-errors 3–5x more likely than X-errors. Readout asymmetric (|1⟩→|0⟩ dominant).
Similar Z-bias. Best Bell 98.4% (pair-dependent). CZ native gate.
More balanced X/Y/Z error rates. Longer coherence but complex crosstalk at scale.
Three kitchens, one recipe — Running a circuit on different chips is like cooking the same recipe in three kitchens. One has a hot oven (depolarizing — burns everything equally). Another has a drafty door (dephasing — only the soufflé collapses). The mitigation is different: you don't fix a draft with oven mitts.
Our key finding: mitigation strategy must match the dominant error source. TREX (readout correction) achieves 119x improvement when readout dominates. ZNE (gate noise extrapolation) fails completely when gates aren't the bottleneck.
| Noise Source | Best Mitigation | Why It Works | Improvement |
|---|---|---|---|
| Readout errors | TREX / REM | Calibrates measurement, doesn't touch circuit | 119x |
| Dephasing (T₂) | Post-selection | Catches parity violations from phase flips | 3.7x |
| Gate errors | ZNE | Amplify gate noise, extrapolate to zero | 1.3x* |
| Decoherence (T₁) | Dynamical decoupling | Refocusing pulses during idle time | ~2x |
*ZNE 1.3x on IBM kicked Ising (gate-noise-dominated circuit). For our shallow VQE circuit, ZNE gives only 2x because gates aren't the bottleneck.
Quantum noise is the primary obstacle to useful quantum computing. Every real qubit interacts with its environment, causing errors that accumulate over the course of a computation. The two fundamental timescales are T₁ (energy relaxation — how long before |1⟩ decays to |0⟩) and T₂ (dephasing — how long superposition phase information survives).
Different noise channels affect the Bloch sphere differently. Amplitude damping (T₁) shrinks the sphere toward the north pole. Pure dephasing (T₂) collapses it to the z-axis. Depolarizing noise contracts it uniformly toward the center. Understanding which channel dominates on a given processor determines which error mitigation strategies will work.
This explorer shows noise channel dynamics on the Bloch sphere with real T₁/T₂ values from IBM Quantum, Quantum Inspire Tuna-9, and IQM Garnet processors.