Reconfigurable Atom Arrays
Moving qubits mid-circuit and what it enables.
In most quantum processors a qubit's neighbors are fixed when the chip is made. In a neutral-atom array they are not: optical tweezers can pick up entangled atoms mid-circuit, carry them across the array and set them down next to new partners without disturbing their quantum states. Connectivity becomes choreography.
What you’ll learn
- Quantify stochastic loading and rearrangement, and the vacuum-lifetime limit on array size.
- Explain how acousto-optic deflectors move tweezers, and the constraints this places on parallel moves.
- Derive the condition for transporting an atom without motional excitation.
- Prove that transversal CNOT is a logical CNOT for CSS codes, and explain why moving code blocks makes it native.
- Describe, as reported, what the zoned logical processor of Bluvstein et al. (2024) demonstrated, and what remains open.
Assembly as an optimization problem
Load tweezer sites, each independently occupied with probability , and keep the atoms needed for an -site target pattern. Three questions govern the process: how many sites to load, which atom goes where, and how the total time compares with the atoms' lifetime in the trap.
Which atom goes where is an assignment problem: minimize the total move distance, or the number of moves, between loaded sites and targets. The Hungarian algorithm solves it exactly in time; heuristics that move whole rows or columns in parallel are faster and match the hardware's constraints (next step). Each move drops an atom with a small probability, so the number of moves matters as well as their length.
Background-gas collisions eject atoms independently at rate , with of tens of seconds to minutes in room-temperature chambers and longer in cryogenic ones.
For an array of atoms held for a cycle of duration .
Once , a loss-free cycle is the exception, however perfect the rearrangement. Large arrays therefore need fast cycles, repeated rearrangement, reservoirs of spare atoms and, increasingly, continuous reloading of fresh atoms while the computation runs, which recent experiments have demonstrated.
An array of atoms has a vacuum lifetime s. A cycle of imaging, rearranging and computing takes s. How many atoms do you expect to lose per cycle?
Show the answer
4 atoms
, so a loss-free cycle has probability only . At this scale loss is not an occasional defect but a steady flux that must be detected and refilled.
Moving atoms coherently
This hardware shapes which moves are cheap. A crossed-AOD grid moves rows and columns together, so parallel moves must preserve the grid's order: rows cannot pass through one another, and a set of atoms can move in parallel only if it forms a sub-grid. Atoms are handed between static tweezers (from a spatial light modulator) and mobile AOD tweezers by ramping intensities. Within these rules, one large parallel move takes about as long as moving a single atom.
Because moving the trap barely perturbs the internal state, hyperfine coherence survives transport. Beugnon et al. (2007) moved a single atomic qubit between tweezers coherently, and Bluvstein et al. (2022) moved whole arrays of entangled atoms mid-circuit, preparing graph states and small error-correcting-code states. Moves take tens to hundreds of μs: slow compared with gates, but parallel across the array.
Zones and transversal logic
Movable qubits invite a zoned layout: a storage zone where idle qubits wait away from the Rydberg light; an entangling zone where one global Rydberg pulse applies CZ gates to every pair of atoms brought within blockade range; and a readout zone where atoms can be imaged without scattering light onto the rest. A program becomes a sequence of "move these blocks, then pulse".
In an atom array a transversal gate is one move and one pulse: bring block A on top of block B, atom next to atom , and fire the global Rydberg pulse to apply all CZs at once (with Hadamards on the target block for CNOT). Try a minimal version with the two-qubit repetition code, and .
CNOT q0→q2 followed by CNOT q0→q3 produces exactly the same state from this input. Why is it nevertheless the wrong way to implement the logical gate?
- It is not a unitary operation
- It works on this input only by coincidence and fails on other code states
- One physical qubit interacts with two qubits of block B, so a single fault on q0 can spread into two errors in block B
- It uses more gates than the transversal version
Show the answer
C. One physical qubit interacts with two qubits of block B, so a single fault on q0 can spread into two errors in block B
On the code space q0 and q1 always agree, so both circuits act as the same logical CNOT. The difference appears with faults: an error on q0 before the gates propagates to both q2 and q3, and on block B is a logical error for the two-qubit repetition code. Transversal gates confine each fault to one qubit per block.
Loss, erasure and the logical clock
Atom loss is the platform's signature error. A lost atom is not a Pauli error but leakage out of the qubit space; left undetected, it is harmful, because later gates act on an empty site. Detected, it becomes an erasure: an error at a known location, which codes handle far better.
This motivates erasure conversion (Wu, Kolkowitz, Puri & Thompson, 2022): choose the qubit levels so that the dominant failure modes, such as decay from the Rydberg state, end in states that can be detected mid-circuit without disturbing the other qubits, turning would-be Pauli errors into erasures.
The price of moving atoms is time. A round of error correction needs moves (tens to hundreds of μs), gates (below a μs) and readout (around a millisecond or more), so logical clock rates are well below those of superconducting processors, whose correction cycles take about a microsecond. Neutral atoms trade speed for parallelism, reconfigurable all-to-all connectivity and large numbers of identical qubits. That connectivity also makes high-rate quantum LDPC codes, whose checks are long-range, natural candidates (Xu et al., 2024).
Key takeaways
- Loading is binomial: a load overhead of plus an margin suffices, and the vacuum lifetime caps array size through .
- AODs move whole rows and columns, so parallel moves must preserve order; a moving tweezer is a forced oscillator, and smooth profiles move atoms coherently.
- Transversal CNOT is a logical CNOT for CSS codes and spreads no errors within a block; in atom arrays it costs one move and one global Rydberg pulse.
- Bluvstein et al. (2024) reported a zoned logical processor with up to 280 physical qubits, largely using error detection; repeated real-time correction at scale is the next step.
- Detected loss is an erasure (toric-code threshold 50%), but moves and readout make the logical clock slow compared with superconducting circuits.
Check yourself
In the interactive lesson these are scored with your confidence, and what you miss comes back for review.
For a fixed failure probability, how does the number of extra sites you must load beyond scale with the target size ?
- Linearly in
- Like
- Like
- It is independent of
Show the answer
C. Like
The Chernoff bound requires , which grows like : binomial fluctuations scale as the square root of the mean.
Why can a crossed-AOD tweezer grid not freely permute atoms in a single parallel move?
- Rows and columns move together, so a parallel move must preserve the grid's ordering; rows cannot cross each other
- AODs can only move one tweezer at a time
- AOD tweezers cannot hold atoms while moving
- Moving tweezers always destroys qubit coherence
Show the answer
A. Rows and columns move together, so a parallel move must preserve the grid's ordering; rows cannot cross each other
Each tweezer sits at a (row tone, column tone) pair. Shifting tones moves whole rows or columns, and two rows cannot swap without passing through each other.
Transversal CNOT between two blocks of the same CSS code is a logical CNOT because
- CNOT commutes with every Pauli operator
- transversal gates always act as the logical identity
- CSS codes have no -type stabilizers
- conjugation by the transversal CNOT maps each stabilizer to a product of stabilizers, and the logical Paulis transform as under a CNOT
Show the answer
D. conjugation by the transversal CNOT maps each stabilizer to a product of stabilizers, and the logical Paulis transform as under a CNOT
and keep the joint stabilizer group intact, while and .
Why is turning atom loss into a detected erasure so valuable for error correction?
- An erasure can be undone by reloading the same atom
- Knowing where errors occurred lets the decoder tolerate far higher rates: 50% loss for the toric code, versus about 10% for unlocated Pauli errors
- Erasures never affect the logical qubit
- Erasure detection removes the need for syndrome measurements
Show the answer
B. Knowing where errors occurred lets the decoder tolerate far higher rates: 50% loss for the toric code, versus about 10% for unlocated Pauli errors
A located error carries no uncertainty about where to correct. Logical failure then requires erased qubits to form a percolating path across the code.
Further reading
- M. Endres et al., "Atom-by-atom assembly of defect-free one-dimensional cold atom arrays", Science (2016).
- D. Barredo, S. de Léséleuc, V. Lienhard, T. Lahaye & A. Browaeys, "An atom-by-atom assembler of defect-free arbitrary two-dimensional atomic arrays", Science (2016).
- D. Bluvstein et al., "A quantum processor based on coherent transport of entangled atom arrays", Nature (2022).
- D. Bluvstein et al., "Logical quantum processor based on reconfigurable atom arrays", Nature (2024).
- T. M. Stace, S. D. Barrett & A. C. Doherty, "Thresholds for topological codes in the presence of loss", Physical Review Letters (2009).