A qubit is a fiction the hardware agrees to maintain. Physicists talk about two states, a crisp 0 and 1, but almost every real qubit is carved out of a system with more energy levels than that. A transmon is an anharmonic oscillator with a whole ladder of rungs. A trapped ion has a forest of atomic states. Engineers pick two of them, label them the computational states, and spend enormous effort keeping everything else off-limits. Leakage is what happens when that agreement breaks and the qubit quietly climbs to a rung it was never supposed to occupy.
Why the extra rungs matter
Consider a superconducting transmon. Its lowest state is 0, the next one up is 1, and the one above that, usually written as 2, is the first leakage state. The whole reason a transmon is engineered with weak anharmonicity is to keep the 1-to-2 transition at a slightly different frequency than the 0-to-1 transition, so a microwave pulse tuned to flip the qubit doesn't also push it up the ladder. But the gap is not large. A pulse that is too fast, too strong, or slightly off frequency can spill a little population into the 2 state. Once there, the qubit is no longer a qubit. It is something the rest of the machine does not know how to talk to.
That is what makes leakage nastier than an ordinary bit-flip or phase-flip error. Standard errors keep the qubit inside its two-state world, just pointing the wrong way. Error-correcting codes like the surface code are built to catch exactly those. Leakage steps outside the code's vocabulary entirely. A stabilizer measurement expects to compare qubits that are all either 0 or 1. Feed it a qubit sitting in the 2 state and the parity checks return garbage, and worse, the leaked qubit can corrupt its neighbors during two-qubit gates, spreading the damage before anyone notices.
The stubbornness problem
Leakage also lingers. A normal error might get flipped back on the next round of correction, but a leaked qubit tends to stay leaked. It does not naturally fall back into the computational space on any useful timescale, and because the decoder cannot even identify it as leaked, it keeps poisoning cycle after cycle. A single leakage event can therefore do far more harm than a single ordinary error, which is why teams building fault-tolerant prototypes treat it as a first-class threat rather than a rounding error.
Pushing qubits back down the ladder
The fix is to actively herd leaked qubits back into line. The most direct approach is a leakage-reduction operation, sometimes a dedicated reset that drains the 2 state back down to 0 or 1 without disturbing qubits that are behaving. Google demonstrated a version of this in its surface-code experiments, inserting operations that convert leakage into ordinary, correctable errors so the decoder can cope. Some designs use an auxiliary structure that couples specifically to the leakage state and dumps its energy away, acting like a pressure valve on the top of the ladder.
Gate design carries a lot of the load too. Pulse-shaping techniques deliberately sculpt a control pulse so its frequency content avoids exciting the 1-to-2 transition, buying speed without spilling population upward. The DRAG technique, short for derivative removal by adiabatic gate, is the workhorse here, adding a carefully tuned quadrature component to the drive that cancels the leakage a naive pulse would cause.
Not just a superconducting worry
Every platform has its own version of the ladder. Neutral-atom and trapped-ion qubits can be knocked into states outside the chosen pair by stray light or imperfect pulses. Spin qubits in silicon can leak into valley states or nearby charge configurations. The details differ, but the lesson is the same: the two-level qubit is an idealization sitting on top of a richer system, and the richness is always waiting to reassert itself.
This is why leakage shows up on hardware roadmaps as an explicit budget line, not an afterthought. A machine that wants to run millions of error-correction cycles cannot tolerate qubits wandering off into states the code cannot see. Keeping the fiction of two states intact, cycle after cycle, turns out to be one of the quieter but essential jobs in building a machine that actually corrects itself.
