Neutral-Atom and Rydberg Qubits

Neutral-atom processors separate long-lived internal-state storage from strong, temporary Rydberg interactions. Optical tweezers position and transport the atoms; lasers or microwaves rotate their internal states; state-selective imaging supplies measurement outcomes. The architectural question is how these operations compose into a circuit while preserving coherence and handling loss. Gate-based processing and analog many-body simulation use related apparatus but generally different encodings and operating sequences.

Required background. Hardware Overview distinguishes physical operations from logical performance; Optical Tweezers explains confinement, loading, and transport; Rydberg Blockade derives the interaction that suppresses double excitation.

Atomic Encodings and Functional States

Storage and motion

For $N$ trapped atoms, a useful uncoupled model is

$$H_0 = -\sum_{i=1}^{N} \frac{\hbar\omega_i}{2}Z_i + \sum_{i,\mu} \hbar\nu_{i\mu} \left( a_{i\mu}^{\dagger}a_{i\mu}+\frac12 \right).$$

Here $Z=|0\rangle\langle0|-|1\rangle\langle1|$, with $|0\rangle$ the lower-energy storage state, and $\omega_i>0$ is the splitting after local field and light-shift corrections. $\nu_{i\mu}$ is a motional frequency of atom $i$ along trap axis $\mu$. Unlike ions in a common Coulomb crystal, independently trapped neutral atoms do not ordinarily share collective harmonic modes. Motion nevertheless matters because a position-dependent optical phase, Rabi frequency, Doppler shift, or Rydberg interaction couples the internal and external degrees of freedom.

The durable qubit should live in states with low differential sensitivity to magnetic and optical fields. The strongly interacting Rydberg state is usually a temporary resource, not a memory state. This separation is a central design advantage and a source of control complexity: every entangling operation must enter and leave the interaction manifold without retaining unwanted population, phase, or motion dependence.

Encoding families

Atomic family	Common storage encoding	Useful auxiliary structure	Practical considerations
alkali atoms	two ground-state hyperfine or clock states	Zeeman sublevels and Rydberg levels	mature microwave/Raman control; vector light shifts and Raman scattering
alkaline-earth and alkaline-earth-like atoms	nuclear-spin, clock, or metastable states	narrow intercombination lines, metastable manifolds, Rydberg levels	state-selective control and detection; additional lasers and unwanted transitions
dual-species or dual-isotope arrays	species-dependent hyperfine or nuclear-spin states	one species may cool, measure, or mediate	collision, wavelength, loading, and cross-species calibration complexity

The symbols $\lvert0\rangle$ and $\lvert1\rangle$ are incomplete unless the isotope, hyperfine or electronic manifold, magnetic quantum numbers, bias field, and phase convention are stated. In particular, a clock-like pair can reduce first-order magnetic sensitivity while remaining sensitive to field curvature, differential trap shifts, microwave phase, or Raman scattering.

One atom can play several roles

The same atom may use several distinct internal states:

• $\lvert0\rangle,\lvert1\rangle$ for storage and computation;
• $\lvert r\rangle$ for transient strong interaction;
• a shelving or metastable state for state-selective imaging;
• a bright state for fluorescence;
• a flag manifold whose population reveals a converted erasure.

Reservoir atoms are additional particles available for replacement, not another internal state of a data atom. Mapping a fault into a detectable manifold is valuable only if the mapping has characterized missed-event and false-flag probabilities, and if the decoder receives the location and timing information.

Loading and Rearranging a Target Array

Static and dynamic traps

Large arrays commonly combine two optical functions:

• a spatial light modulator, microlens array, or related element generates a comparatively static set of storage traps;
• one- or two-dimensional acousto-optic deflectors generate moving traps or scanned addressing beams with rapidly programmable positions.

The exact division is implementation-dependent. A static trap can also be updated, and an AOD can synthesize several simultaneous tones. The durable architectural distinction is between a large, stable storage geometry and a smaller set of fast, dynamically routed resources.

Optical power is a system budget. If $P_j$ is the power assigned to trap $j$, then producing more sites at fixed total power generally lowers trap depth or requires additional lasers, modulators, apertures, and objectives. Aberration, diffraction efficiency, ghost traps, polarization, wavelength, and field of view can make the relation nonuniform. A quoted site count should therefore be accompanied by occupancy, depth uniformity, field of view, and the active control region.

Loading is stochastic

Suppose each of $N$ target sites loads independently with probability $p$. Without rearrangement, the probability of a completely filled target is

$$P_{\mathrm{full}}=p^N.$$

Even a respectable single-site probability becomes inadequate at large $N$. For $p=0.75$ and $N=100$,

$$p^N\approx3.2\times10^{-13}.$$

Rearrangement changes the problem. If $M>N$ source sites are loaded, the number of available atoms $K$ is approximately binomial under the independent model,

$$K\sim\operatorname{Binomial}(M,p), \qquad \mathbb E[K]=Mp.$$

Having $K\ge N$ is necessary but not sufficient. The image must classify occupancy, an assignment algorithm must match source atoms to targets, the transport graph must avoid collisions and blocked paths, and transfers between traps must preserve survival and coherence. Correlated loading or spatially varying $p_j$ changes the binomial supply model. Classifier errors additionally make the inferred occupancy differ from the true atom count.

The useful loading quantity is the tail probability, not merely the mean:

$$\begin{aligned} P_{\mathrm{supply}}(M,N,p)&=\Pr(K\ge N)\\ &=\sum_{k=N}^{M}\binom{M}{k}p^k(1-p)^{M-k}. \end{aligned}$$

For $N=200$ and $p=0.60$, $M=334$ gives a mean of $200.4$ atoms but only $P_{\mathrm{supply}}\simeq0.5415$. With $M=400$, the shortage probability is about $2.12\times10^{-5}$. These are ideal supply probabilities: missed atoms, false occupancy classifications, inaccessible routes, and transport losses can still prevent successful preparation. Verification imaging can itself cause heating or loss.

Extra source traps suppress the probability of having fewer than $N=200$ atoms. The curves are exact binomial tails for independent, identical loading; they exclude imaging and rearrangement failure. For $p=0.60$, the mean first exceeds the target at $M=334$, yet shortage still occurs in about $46\%$ of attempts.

Assignment and transport remain separate problems. A minimum-distance assignment need not finish fastest when moving tweezers share radio-frequency bandwidth, paths cross, or accelerations must be limited. A lost fresh reservoir atom can often be replaced; losing an atom that carries entangled data requires error correction. This distinction determines which failed moves can be retried.

Initialization and One-Qubit Control

Optical pumping prepares a fiducial internal state. Microwave fields can drive global hyperfine rotations with stable phase; two-photon Raman fields provide faster or more local rotations; narrow optical transitions can address clock or metastable encodings. For a selected near-resonant transition, the rotating-wave control Hamiltonian is

$$\begin{aligned} A_i(t) &= \frac{\hbar\Omega_i(t)}{2} e^{i\phi_i(t)} \lvert1\rangle_i\langle0\rvert, \\ H_i(t) &= A_i(t)+A_i^\dagger(t) + \frac{\hbar\delta_i(t)}{2}Z_i. \end{aligned}$$

Here $\delta_i=\omega_{\mathrm{drive}}-\omega_i$; the sign follows from the storage convention above. The complex drive phase defines the rotation axis in the equatorial plane. The same rotation can be realized by different hardware paths. Those paths have different scattering, phase-reference, addressing, and crosstalk errors.

Global control is naturally parallel but cannot implement arbitrary local circuits by itself. Local control can be supplied by a focused beam, a site-selective light shift followed by a global pulse, frequency selection, transport into an addressing zone, or combinations of these. Each method defines a different concurrency graph.

Rabi, detuning, phase, spectator, and leakage measurements should use the intended parallel pattern: an individually calibrated beam can produce a different light shift when other beams are enabled. Global randomized benchmarking measures an average decay under its protocol and model. It does not certify spatial uniformity, purely coherent errors, local addressing, Rydberg entanglers, or a compiled logical workload.

Rydberg Interactions and Native Gates

Programmable interaction Hamiltonian

Let $\lvert r\rangle_i$ be a chosen Rydberg state coupled from $\lvert1\rangle_i$. In a rotating frame, with $\Delta_i=\omega_{L,i}-\omega_{r1,i}$ including the relevant shifts, a common model is

$$\frac{H}{\hbar} = \sum_i \left[ \frac{\Omega_i}{2} \left( e^{i\phi_i}\lvert r\rangle_i\langle1\rvert +\mathrm{h.c.} \right) -\Delta_i n_i \right] + \sum_{i<j} \frac{V_{ij}}{\hbar}n_i n_j, \qquad n_i=\lvert r\rangle_i\langle r\rvert.$$

For an isolated nonresonant pair, one often writes

$$V_{ij}\approx\frac{C_6(\theta_{ij})}{R_{ij}^{6}}.$$

This scalar form is an approximation. The coefficient can depend strongly on principal quantum number, angular momentum, magnetic field, polarization, and orientation. Near resonant pair channels, multilevel mixing or dipolar $C_3/R^3$ behavior can be important. The calibrated pair potential, not a nominal power law alone, determines the gate.

Define a blockade shift in angular-frequency units by

$$B_{ij}=\frac{|V_{ij}|}{\hbar}.$$

For near-resonant driving in this scalar model, strong blockade requires $B_{ij}\gg\Omega$ and sufficiently narrow distributions of excitation detuning. More generally, inspect the actual pair spectrum and its laser-coupled gaps. Writing the signed shift as $b_{ij}=V_{ij}/\hbar$, the one- and two-excitation energies are $-\hbar\Delta$ and $\hbar(b_{ij}-2\Delta)$ for equal detunings. Thus $b_{ij}\simeq\Delta$ permits resonant addition of a second excitation, and $b_{ij}\simeq2\Delta$ permits two-photon pair resonance. A large bare $|b_{ij}|$ alone does not ensure blockade during a detuned pulse. Multilevel mixing can also leave weakly shifted, optically accessible pair states. In a simple two-level estimate, residual double excitation scales as

$$p_{rr}\sim\left(\frac{\Omega}{B_{ij}}\right)^2.$$

This is a scaling estimate, not a complete gate error model. Pulse shape, finite lifetime, off-resonant intermediate-state scattering, motion, laser noise, and multilevel structure must also be included.

Blockade gates

A blockade-controlled phase protocol transiently excites one or both qubits so that the interaction changes the cyclic phase accumulated by one computational basis state. After local phase corrections, the desired unitary may be written

$$U_{\mathrm{CZ}} = \operatorname{diag}(1,1,1,-1).$$

Conjugating the target by Hadamard gates yields a controlled-NOT. Depending on the pulse protocol and available levels, Rydberg interactions can also realize controlled phases of other angles, exchange-like gates, simultaneous parallel CZ layers, or native multiqubit controlled operations.

A useful phase check is

$$\Phi_{\mathrm{ent}} =\phi_{00}-\phi_{01}-\phi_{10}+\phi_{11}\pmod{2\pi}.$$

For a diagonal operation that returns all population to the storage subspace, local phase corrections cannot change $\Phi_{\mathrm{ent}}$. A CZ requires $\Phi_{\mathrm{ent}}=\pi$; return of the four populations alone is insufficient. The elementary control-$\pi$, target-$2\pi$, control-$\pi$ blockade sequence gives $\operatorname{diag}(1,-1,-1,-1)$ in one common laser-phase convention; applying $Z\otimes Z$ gives the stated CZ. Rydberg Blockade derives that sequence and its finite-interaction corrections. Native gates may instead use global shaped pulses, so their local phases, loss, leakage, and parallel context must be characterized for the actual protocol. The reviews by Saffman, Walker, and Mølmer and Morgado–Whitlock compare these gate families.

Dressing and exchange

Off-resonant Rydberg dressing admixes a small Rydberg amplitude into a storage state. It can generate a softer effective interaction while reducing the instantaneous Rydberg population. The price is a weaker interaction and a longer exposure to laser phase noise, scattering, and technical drift.

Near resonant pair channels can generate excitation exchange. Such dynamics are useful for simulation and specialized gates, but they require a clear definition of the participating pair states and angular dependence. The words “Rydberg interaction” do not specify whether the platform uses blockade, dressing, resonant exchange, or an engineered combination.

Blockade, decay, and position uncertainty

Three compact estimates organize much of the trade space.

First, finite blockade contributes a leakage scale

$$\epsilon_B\sim\left(\frac{\Omega}{B}\right)^2.$$

Second, for independent Markovian decay channels with a common population lifetime $\tau_r$, the expected number of decay events is

$$\mu_{\mathrm{decay}} = \frac{1}{\tau_r} \int_0^{T_g} P_r(t)\,dt,$$

where $P_r(t)=\sum_i\langle n_i\rangle$ is evaluated in the open dynamics. For rare events, the probability of at least one decay is approximately $\mu_{\mathrm{decay}}$, and ideal coherent populations give its leading-order estimate. It is not a universal gate-infidelity floor: decay branches, re-excitation, detected erasures, and conditioning matter. State-dependent rates require $\int\sum_i\Gamma_i\langle n_i\rangle\,dt$ instead. Third, a position fluctuation $\delta R$ changes a van der Waals interaction by

$$\frac{\delta V}{V} \approx -6\frac{\delta R}{R}$$

for a small radial change at fixed $C_6$ and orientation. Angular fluctuations add derivatives of the anisotropic coefficient. Faster driving can reduce decay exposure but worsen finite blockade and off-resonant excitation. Increasing principal quantum number can strengthen interactions while changing lifetime, field sensitivity, level density, laser requirements, and blackbody coupling. There is no single knob that improves every term.

Geometry, Reachability, and Parallelism

Interaction graphs are state- and protocol-dependent

For a chosen geometry and gate protocol, define a reachability graph $G_R=(V,E_R)$ whose vertices are active atoms and whose edges are calibrated pairs. An edge may require moving its atoms into an interaction zone. Thus a complete edge does not imply a complete array of fixed, simultaneous interactions.

For analog dynamics, the interaction matrix is weighted and generally dense,

$$J_{ij}\propto\frac{C_6(\theta_{ij})}{R_{ij}^{6}},$$

subject to blockade constraints and multilevel corrections. For gate-based operation, the scheduler often chooses a sparse subset of pairs and suppresses or refocuses the rest.

Pair reachability is not gate concurrency

A legal parallel layer is more restrictive than a matching of disjoint pairs. Two gates can have distinct atoms yet conflict because their blockade volumes overlap, their addressing beams share tones, their moving paths cross, or the same control waveform cannot calibrate both pair geometries. A conflict graph $G_C$ can represent these constraints: one vertex for each requested gate and an edge between incompatible gates. A parallel gate set must then be an independent set of $G_C$. Pairwise conflicts are not always sufficient: three mutually compatible gates can still exceed a shared optical-power or tone-count limit. Such constraints require an additional capacity bound or a higher-order conflict model.

If all $N$ atoms can eventually be paired, the abstract maximum number of disjoint two-qubit gates is

$$N_{\mathrm{pair}}^{\max} = \left\lfloor\frac{N}{2}\right\rfloor.$$

The physical maximum may be lower. Claims of all-to-all or nonlocal connectivity should state whether they mean direct interaction in one static geometry, coherent transport to a common zone, or compilation through swap operations; they should also give parallel width and route latency.

Interaction range does not create a sharp graph

Defining a blockade radius by $|V(R_b)|=\hbar\Omega$ gives

$$R_b = \left( \frac{|C_6|}{\hbar\Omega} \right)^{1/6}$$

for the scalar van der Waals approximation. This is a useful scale, not a hard boundary. The interaction remains nonzero outside $R_b$, and atoms inside it can experience different shifts. Gate layout must budget both inadequate target blockade and unwanted spectator interaction.

Zone-Based Processing

A zone architecture separates incompatible tasks spatially. A representative pipeline uses:

• a loading or reservoir zone for fresh atoms;
• a storage zone optimized for coherence and density;
• one or more interaction zones optimized for Rydberg gates;
• a readout and reset zone whose scattered photons are isolated from data;
• dynamic tweezers that move selected atoms or logical blocks between zones.

This separation can protect stored data from resonant light and concentrate expensive addressing resources. It also turns movement into part of the instruction set. For a move of duration $T_m$, the relevant record includes survival, coherence, motional excitation, phase accumulation, transfer error, and path conflicts. A transport fidelity inferred from a return experiment may not separately identify those mechanisms.

Let one repeated logical cycle contain routing, gates, measurement, reset, and classical decision. A serial latency model is

$$T_{\mathrm{cycle}} = T_{\mathrm{route}} +T_{\mathrm{gate}} + T_{\mathrm{read}} +T_{\mathrm{reset}} + T_{\mathrm{classical}}.$$

With separate zones, some terms can overlap across batches. The asymptotic cadence is then bounded below by the slowest occupied pipeline stage, but route conflicts, finite reservoirs, dynamic branches, and feedback barriers can prevent ideal overlap.

Zone-based designs also permit qubit reuse: a measured atom can be reset and returned, or a lost atom can be replaced by one from the reservoir. This is not automatic fault tolerance. Reuse must preserve the logical state carried by the remaining atoms, report the replacement location to the decoder, and avoid contamination from measurement light or transport.

Measurement, Reset, Loss, and Erasure

Presence and internal state are different observables

Fluorescence detects atoms only in states that scatter under the imaging sequence. Presence detection therefore requires a suitable state-insensitive sequence or a known initial manifold. Qubit readout additionally uses a state-dependent mapping, such as shelving, bright/dark fluorescence, or spin-to-position conversion. If internal state and absence are separately resolved, the outcome alphabet can be

$$y\in\{0,1,\varnothing\},$$

where $\varnothing$ denotes detected absence. A three-outcome confusion matrix is therefore more informative than a single assignment fidelity:

$$M_{y|x}=P(y\mid x), \qquad x\in\{0,1,\mathrm{lost}\}.$$

Blow-away followed by a single presence image does not generally supply these three outcomes: an atom intentionally removed from $|1\rangle$ and an atom lost before measurement both appear absent. An additional measurement or state-mapping sequence is needed to distinguish them. The confusion matrix describes classical populations, not measurement backaction on superpositions or neighboring atoms.

A destructive method can have excellent terminal readout while being unsuitable for repeated syndrome extraction. Mid-circuit operation asks for low disturbance to spectators, low atom loss, controlled reset, and bounded latency in addition to classification accuracy.

Loss can be useful information

Atom loss leaves the computational Hilbert space. If it is detected with a known location, the event is an erasure rather than an unlocated error. Suppose a mechanism causes loss with probability $p_{\ell}$ and the experiment identifies it with efficiency $\eta_e$. For this one loss mechanism,

$$\begin{aligned} p_{\mathrm{located}} &=\eta_e p_{\ell}, \\ p_{\mathrm{unlocated}} &=(1-\eta_e)p_{\ell}. \end{aligned}$$

False erasure flags, delayed detection, and faults occurring after the last check must be added. The same physical process can therefore produce a mixture of located erasures, unlocated leakage, and ordinary Pauli error.

Alkaline-earth-like atoms offer additional manifolds that can convert some decay or leakage events into detectable population outside the qubit pair. This erasure conversion changes the information supplied to the decoder; it does not erase the physical fault. Its benefit depends on conversion efficiency, flag fidelity, correlations, reset cost, and the code’s response to located faults.

Large arrays amplify small loss probabilities

If each of $N$ active atoms independently survives one cycle with probability $s$, then

$$P_{\mathrm{no\ loss}}(K) = s^{NK}$$

is the probability that no atom is lost in $K$ cycles. This exponential is a warning against postselecting only shots in which every atom survives. At scale, a useful architecture must detect, tolerate, replace, or correct loss rather than demand that it never happen.

Independence is only a baseline. Vacuum events may be approximately local, whereas laser excursions, trap-power changes, image-light errors, or path collisions can produce correlated loss. Reporting only average survival hides the distinction.

Analog Quantum Simulation

From geometry to an interacting spin model

Analog ground–Rydberg simulation instead uses the pair $|g\rangle,|r\rangle$ as the spin basis; this is distinct from the two storage levels used above. Define $Z=|g\rangle\langle g|-|r\rangle\langle r|$ and $X=|g\rangle\langle r|+|r\rangle\langle g|$. A common laser phase gives

$$\frac{H}{\hbar} = \frac{\Omega(t)}{2}\sum_i X_i -\Delta(t)\sum_i n_i + \sum_{i<j} \frac{V_{ij}}{\hbar}n_i n_j.$$

Here $n_i=|r\rangle_i\langle r|=(1-Z_i)/2$ is the Rydberg population projector.

Consequently the Ising coupling is $V_{ij}/4$ and the local $Z_i$ coefficient is $\hbar\Delta/2-\sum_{j\ne i}V_{ij}/4$. Uniform laser detuning need not produce a uniform effective field at an open boundary. Geometry sets $R_{ij}$, the selected Rydberg state and fields set $V_{ij}$, and laser waveforms set $\Omega(t)$ and $\Delta(t)$. In the strong-blockade limit, nearby simultaneous excitations are energetically suppressed, leading to constrained models such as the PXP model. Rydberg Blockade derives the constrained Hamiltonian and the conditions for neglecting off-resonant pair excitation.

Analog is not gate-based digital computation

An analog experiment programs a Hamiltonian and observes its evolution. A digital experiment compiles an abstract circuit into calibrated gates. The same apparatus can support both, but a successful analog simulation does not establish accurate digital gates:

Question	Digital mode	Analog mode
programmed object	gate sequence	geometry and time-dependent Hamiltonian
primary calibration	process or gate instruction	$\Omega_i$, $\Delta_i$, $V_{ij}$, geometry
dominant comparison	circuit or logical output	observables, correlations, phase structure, dynamics
validation	randomized circuits, tomography, logical tests	limits, symmetries, size scaling, independent methods
typical failure	coherent gate accumulation, leakage, routing	Hamiltonian mismatch, inhomogeneity, state-preparation bias

Hundreds of interacting atoms in an analog simulation do not imply a universal hundreds-qubit digital computer. Conversely, gate benchmarks do not establish that a many-body Hamiltonian is known accurately across every pair.

Calibration of geometry, drive, detuning, interactions, and observation all enter the comparison with a target many-body model. Agreement with one observable can miss a wrong phase or coupling. Verification of Quantum Simulation explains complementary observables, controlled limits, and uncertainty-aware tests.

Noise and Error Mechanisms

Neutral-atom errors are not captured by one coherence time or one average gate fidelity.

Mechanism	Physical origin	Observable signature	How to reduce it
storage dephasing	magnetic noise, differential light shifts, motion through inhomogeneous traps	Ramsey decay, spatially varying phase	clock states, magic conditions, echo, trap calibration
population relaxation	optical pumping, Raman scattering, blackbody or environmental coupling	state-dependent $T_1$, leakage	detuning, filtering, state choice, leakage checks
trap loss	background collisions, heating, shallow traps, imaging and transport	empty site, survival decay	better vacuum, cooling, reservoirs, erasure-aware decoding
motional error	finite temperature, acceleration, transfer between traps, recoil	contrast or gate error versus temperature and trajectory	cooling, smooth transport, composite pulses
Rydberg decay	spontaneous and blackbody-driven transitions	loss, leakage, dephasing during gates	faster optimized pulses, state choice, erasure conversion
finite blockade	interaction shift too small or inhomogeneous	double excitation, coherent phase error	geometry, pulse shaping, pair calibration
laser noise	phase, frequency, intensity, pointing, wavefront	coherent rotation error, drift, site dependence	stabilization, common-mode design, local monitors
addressing crosstalk	finite beam waist, shared AOD tones, scattered light	spectator rotations or Stark shifts	cancellation, spacing, scheduling, calibration in parallel context
electric-field sensitivity	dc and ac fields, surface charge, ion production	Rydberg detuning and pair-state drift	electrodes, compensation, shielding, frequent spectroscopy
readout backaction	resonant photons, imperfect shelving, detector confusion	spectator error, loss, false erasure	spatial zones, hiding, non-destructive mapping, decoder model
correlated faults	common lasers, trap power, camera, waveform or vacuum events	spatial and temporal clusters	correlation-aware diagnostics, interlocks, code-aware scheduling

Memory coherence is conditional

An echo or dynamical-decoupling time is measured under a specific pulse train, trap depth, temperature, and field environment. It is not necessarily the available algorithmic idle time. During computation, atoms may move, traps may change depth, Rydberg beams may illuminate spectators, and phase frames may be updated. Useful memory benchmarks reproduce those conditions.

Loss and leakage need time stamps

Finding an empty site at final readout does not reveal when the atom left. A decoder may need to know whether the loss occurred before or after a given entangling gate. Intermediate presence checks, circuit cancellation rules, and probabilistic loss-time models can supply this information. The decoder and hardware schedule are therefore coupled.

Parallel operation can reveal new errors

A high-fidelity isolated pair gate may degrade in a dense layer because of shared optical power, AOD intermodulation, spectator interactions, wavefront variation, or common laser transients. Characterization should therefore include the intended simultaneous patterns and not only one pair at a time.

Infrastructure and Control

Vacuum and atom source

Vacuum lifetime sets a basic hazard rate for loss. Large arrays magnify even a small per-atom rate, so loading architecture, differential pumping, source contamination, and local pressure matter. A long single-atom lifetime should be translated into array-level no-loss probabilities at the relevant cycle duration.

Cryogenic and Vacuum Infrastructure develops that translation from local gas loads and species through collision hazards, whole-array survival, diagnostics, source cycles, and availability.

Repeated or continuous operation also needs an atom supply that does not disturb stored qubits. A reservoir can be spatially separated from storage, but cooling and imaging light, magnetic transients, and moving traps still need isolation and scheduling.

Optical system

A neutral-atom processor may require trapping, cooling, imaging, pumping, Raman, clock, Rydberg, shelving, and repumping wavelengths. The high-numerical- aperture objective must combine field of view, resolution, wavefront quality, chromatic performance, and collection efficiency. Scaling site count can stress laser power and aberration correction; scaling parallel control can stress modulator bandwidth, tone count, and beam steering.

The optical system should be described as a calibrated transfer function from commanded waveform to field at each atom. Camera images alone do not capture phase, polarization, or fast intensity transients.

Electric and magnetic fields

Rydberg levels have large polarizabilities and can be sensitive probes of stray electric fields. Photocharging, ions generated during operation, electrode drift, and nearby surfaces can shift the resonance. Bias magnetic fields define quantization and separate Zeeman components but also create spatial detuning if gradients are uncontrolled. Pair-state spectroscopy and field compensation belong in the routine calibration graph.

Timing and throughput

Static patterns, transport, drive pulses, images, classifiers, and decoder decisions must share a timing model. For a serial experiment,

$$T_{\mathrm{shot}}=T_{\mathrm{load}}+T_{\mathrm{image}} +T_{\mathrm{rearrange}}+T_{\mathrm{cool}} +T_{\mathrm{prepare}}+T_{\mathrm{circuit}} +T_{\mathrm{read}}+T_{\mathrm{classical}}.$$

Reservoirs and reusable atoms can amortize loading, while independent zones can overlap tasks on different batches. Feedback within the same logical block can still impose a causal barrier. If each shot is accepted with probability $q$ and has fixed duration $T_{\mathrm{shot}}$, the mean accepted-shot rate is $q/T_{\mathrm{shot}}$; gate duration alone does not determine it.

Calibration also has physical dependencies. Trap depth changes temperature and differential light shifts; geometry changes interaction; both alter gate performance. Control, Readout, and Calibration develops how those dependencies affect pulse and measurement models.

Error-Correction Fit

Architectural opportunities

Neutral-atom arrays offer several features relevant to error correction:

• atoms are naturally uniform, so fabrication spread does not set their bare internal spectrum;
• arrays can be rearranged into code-specific geometries;
• coherent movement can provide nonlocal pairings or transversal operations without long swap chains;
• loss can often be detected directly and supplied as erasure information;
• readout, reset, and reservoir zones can remove entropy and replenish atoms;
• blockade can support parallel controlled phases and some multiqubit operations.

These features do not select a unique code. Surface codes favor repeated local checks and strict measurement cadence. Quantum LDPC and transversal-gate schemes may exploit longer-range transport or block movement. The right choice depends on the measured movement, gate, loss, readout, reset, and correlation model, not connectivity alone.

Repeated correction is the decisive test

A logical state preparation or postselected logical circuit is evidence of encoding, not yet of sustained error correction. Stronger demonstrations ask whether increasing code distance suppresses the logical error per round at comparable physical noise, and whether repeated rounds preserve a logical state better than an appropriate uncorrected baseline. At fixed distance, more rounds generally accumulate failure probability; they do not by themselves suppress it. The decoder and acceptance policy are part of the comparison. Claims should separate:

• error detection from active or frame-based correction;
• one round from repeated rounds;
• postselected from unconditional performance;
• physical-qubit count from active and logical-qubit counts;
• a memory experiment from universal logical operations;
• below-threshold scaling in a tested finite regime from a complete fault-tolerant computer.

Loss-aware decoding is especially important. If a loss is detected only at the end, its possible time locations must be propagated through the circuit. If a qubit is replaced, the reset and reinsertion operation belongs in the circuit-level error model.

For current experimental records and open problems, see Quantum Hardware Scaling and Rydberg Array Frontiers. The enduring comparisons are storage versus interaction scale, isolated versus concurrent gates, conditional versus unconditional performance, and logical errors at comparable code distance and elapsed time.

Worked Loading, Survival, and Timing Example

Consider a notional zone-based processor with 320 storage sites, 256 active data and ancilla atoms, a reservoir of 64 atoms, 32 simultaneous interaction pairs, and per-active-atom survival $s=0.9995$ per repeated cycle.

1. Check array-level survival

Under the independent model, the probability that all 256 active atoms survive one cycle is

$$P_1 = (0.9995)^{256} \approx 0.880.$$

After 100 cycles, demanding that every atom survive throughout gives

$$\begin{aligned} P_{100} &= (0.9995)^{25600} \\ &\approx e^{-12.8} \approx 2.8\times10^{-6}. \end{aligned}$$

The architecture cannot scale by accepting only trajectories with no loss. It needs timely loss detection, erasure-aware decoding, replacement, or logical tolerance.

2. Separate gate width from qubit count

The processor has 256 active atoms but only 32 interaction pairs per gate layer. A circuit layer containing 96 compatible two-qubit gates therefore requires at least

$$\left\lceil\frac{96}{32}\right\rceil=3$$

gate batches before routing conflicts or local corrections are included. The active-qubit count alone would miss this factor.

3. Distinguish cadence from completion time

Suppose independent stages require $a=180\,\mu\mathrm{s}$ for movement, $b=1.5\,\mu\mathrm{s}$ for the gate, and $c=285\,\mu\mathrm{s}$ for readout, reset, and classical processing. Assume one server per stage, sufficient buffers, identical batches, and no cross-batch feedback barrier. Starting empty, three batches complete as follows:

Batch	Movement ends (μs)	Gate ends (μs)	Final stage ends (μs)
1	180	181.5	466.5
2	360	361.5	751.5
3	540	541.5	1036.5

The steady completion interval is $\max(a,b,c)=285\,\mu\mathrm{s}$, whereas the total three-batch completion time is

$$\begin{aligned} T_3&=a+b+c+2\max(a,b,c)\\ &=1036.5\,\mu\mathrm{s}. \end{aligned}$$

The gate time is not added to every steady interval when it has its own overlapping resource. If instead the same batch must finish measurement and feedback before its next movement, the serial interval is $466.5\,\mu\mathrm{s}$. Neither model follows from the existence of zones alone: the dependency and resource schedule determines which applies.

4. Test reservoir sufficiency

The expected number of losses per cycle is

$$\mathbb E[L] = 256(1-s) = 0.128.$$

A 64-atom reservoir matches the mean number of replacements over 500 cycles in this model. That statement is not a reliability guarantee: losses fluctuate, replacement can fail, the reservoir itself can lose atoms, and reload operations may occur in fixed batches. A full design uses the loss distribution and replenishment cadence, not only its mean.

Common Pitfalls

Equating sites, atoms, and logical qubits. An empty trap is not a qubit; an encoded storage atom need not participate in the active gate layer. State the subset and parallel width when comparing performance.

Treating a blockade radius as a sharp cutoff. The interaction has tails, and detuning can restore pair resonance even at a large bare interaction.

Conditioning away loss. A high fidelity on surviving shots can coexist with low unconditional success. Loss locations and approximate times are useful decoder information, but a detected loss remains a physical fault.

Equating replacement with state recovery. A fresh atom supplies a known state, not the unknown quantum information carried by the atom it replaces. Recovering that information requires redundancy and a recovery operation.

Exercises

1. Size a source array

A target requires 200 atoms. Source sites load independently with probability $0.60$. What is the smallest integer $M$ for which the mean number loaded is at least 240? Why does this not guarantee a 200-atom target?

Solution

The mean is $0.60M$, so

$$0.60M\ge240 \quad\Longrightarrow\quad M\ge400.$$

Thus the smallest integer by the stated mean criterion is $M=400$. The actual number is a random variable with standard deviation

$$\begin{aligned} \sqrt{Mp(1-p)} &= \sqrt{400(0.60)(0.40)} \\ &= \sqrt{96} \approx9.80. \end{aligned}$$

The mean exceeds 200 by about four standard deviations, but success also depends on correlated loading, image errors, route feasibility, movement loss, and target verification. A reliability requirement should use a tail probability and measured nonidealities, not only the mean.

2. Estimate finite-blockade leakage

A gate uses $\Omega/2\pi=4\,\mathrm{MHz}$ and a calibrated blockade shift $B/2\pi=40\,\mathrm{MHz}$. Estimate $(\Omega/B)^2$. What is omitted?

Solution

The ratio is independent of whether both quantities are quoted as angular or ordinary frequencies, provided the convention is consistent:

$$\left(\frac{\Omega}{B}\right)^2 = \left(\frac{4}{40}\right)^2 = 10^{-2}.$$

This one-percent scale is only the elementary finite-blockade estimate. It omits pulse-shape coefficients, coherent phase error, decay, intermediate- state scattering, Doppler and position effects, multilevel pair states, laser noise, preparation, measurement, and loss classification.

3. Estimate rare Rydberg decay

During a $1.0\,\mu\mathrm{s}$ gate, the total integrated Rydberg population is

$$\int_0^{T_g}P_r(t)\,dt=0.45\,\mu\mathrm{s}.$$

For effective lifetime $\tau_r=150\,\mu\mathrm{s}$, estimate the decay probability to first order.

Solution

For a small probability,

$$p_{\mathrm{decay}} \approx \frac{0.45\,\mu\mathrm{s}}{150\,\mu\mathrm{s}} = 3.0\times10^{-3}.$$

The estimate says nothing about whether decay produces a Pauli error, leakage, loss, or a detected erasure. Those branches have different consequences for a logical circuit.

4. Compare target and spectator interactions

Under an isotropic $C_6/R^6$ approximation, a target pair is separated by $5\,\mu\mathrm{m}$ and a spectator is $10\,\mu\mathrm{m}$ from one member. Find the magnitude ratio $|V_{\mathrm{spectator}}/V_{\mathrm{target}}|$.

Solution

The coefficient cancels:

$$\begin{aligned} \left| \frac{V_{\mathrm{spectator}}}{V_{\mathrm{target}}} \right| &= \left(\frac{5}{10}\right)^6 \\ &= \frac{1}{64} \approx1.56\times10^{-2}. \end{aligned}$$

The spectator interaction is not zero. Whether it is acceptable depends on the pulse, detuning, duration, allowed coherent phase, angular dependence, and other nearby atoms.

5. Separate reachability from parallel width

A 48-atom array can route any selected pair to an interaction zone. The zone supports at most eight simultaneous pairs. A circuit layer requests 20 disjoint pair gates. Give the abstract pair reachability count and a lower bound on the number of batches.

Solution

If every pair is reachable, the pair graph has

$$\binom{48}{2} = 1128$$

edges. The 20 requested gates require at least

$$\left\lceil\frac{20}{8}\right\rceil =3$$

batches. Route conflicts, pair-dependent calibration, and local corrections can increase the actual number. The 1,128 reachable edges do not describe simultaneous connectivity.

6. Calculate survival scaling

For $N=500$ active atoms and per-cycle survival $s=0.9998$, find the probability of no loss in one cycle and in 50 cycles under independence.

Solution

For one cycle,

$$P_1 =(0.9998)^{500} \approx e^{-0.10} \approx0.905.$$

For 50 cycles,

$$\begin{aligned} P_{50} &=(0.9998)^{25000} \\ &\approx e^{-5.00} \\ &\approx6.74\times10^{-3}. \end{aligned}$$

Thus excellent per-atom survival is compatible with frequent array-level loss. Detected erasures, replacement, and logical tolerance become system requirements.

7. Expand the analog interaction term

Using $n_i=(1-Z_i)/2$, expand $V_{ij}n_i n_j$. Identify the constant, single-spin, and Ising-coupling pieces.

Solution

Direct expansion gives

$$V_{ij}n_i n_j = \frac{V_{ij}}{4} \left( 1-Z_i-Z_j+Z_iZ_j \right).$$

The first term is a constant energy offset. The middle terms contribute local longitudinal fields when summed over pairs. The last is an Ising coupling of strength $V_{ij}/4$ under this Pauli convention. Dropping the constant is safe for dynamics, but the induced local fields must be combined consistently with the detuning term.

8. Split located and unlocated loss

A gate has loss probability $p_{\ell}=3.0\times10^{-3}$. The loss-detection efficiency is $\eta_e=0.92$, and false erasure flags occur with probability $2.0\times10^{-4}$ per gate. Compute the simple located and missed-loss probabilities, and state why the false flag is separate.

Solution

The located part is

$$\begin{aligned} p_{\mathrm{located}} &=(0.92)(3.0\times10^{-3}) \\ &=2.76\times10^{-3}, \end{aligned}$$

and the missed part is

$$\begin{aligned} p_{\mathrm{unlocated}} &=(0.08)(3.0\times10^{-3}) \\ &=2.4\times10^{-4}. \end{aligned}$$

The $2.0\times10^{-4}$ false-flag probability occurs when the decoder is told an erasure happened even though the specified loss did not. It is not part of $p_{\ell}$ and must be represented as a separate observation error. Correlated flags and timing ambiguity would require a richer model.

References

• H. Bernien et al., “Probing many-body dynamics on a 51-atom quantum simulator,” Nature 551, 579–584 (2017), doi:10.1038/nature24622.
• D. Bluvstein et al., “A quantum processor based on coherent transport of entangled atom arrays,” Nature 604, 451–456 (2022), doi:10.1038/s41586-022-04592-6.
• D. Bluvstein et al., “Logical quantum processor based on reconfigurable atom arrays,” Nature 626, 58–65 (2024), doi:10.1038/s41586-023-06927-3.
• A. Browaeys and T. Lahaye, “Many-body physics with individually controlled Rydberg atoms,” Nature Physics 16, 132–142 (2020), doi:10.1038/s41567-019-0733-z.
• H. Evered et al., “High-fidelity parallel entangling gates on a neutral-atom quantum computer,” Nature 622, 268–272 (2023), doi:10.1038/s41586-023-06481-y.
• D. Jaksch et al., “Fast quantum gates for neutral atoms,” Physical Review Letters 85, 2208–2211 (2000), doi:10.1103/PhysRevLett.85.2208.
• A. M. Kaufman and K.-K. Ni, “Quantum science with optical tweezer arrays of ultracold atoms and molecules,” Nature Physics 17, 1324–1333 (2021), doi:10.1038/s41567-021-01357-2.
• H. Levine et al., “Parallel implementation of high-fidelity multiqubit gates with neutral atoms,” Physical Review Letters 123, 170503 (2019), doi:10.1103/PhysRevLett.123.170503.
• H. Levine et al., “Dispersive optical systems for scalable Raman driving of hyperfine qubits,” Physical Review A 105, 032618 (2022), doi:10.1103/PhysRevA.105.032618.
• M. Morgado and S. Whitlock, “Quantum simulation and computing with Rydberg-interacting qubits,” AVS Quantum Science 3, 023501 (2021), doi:10.1116/5.0036562.
• M. A. Norcia et al., “Seconds-scale coherence on an optical clock transition in a tweezer array,” Science 366, 93–97 (2019), doi:10.1126/science.aay0644.
• M. A. Norcia et al., “Mid-circuit qubit measurement and rearrangement in a $^{171}\mathrm{Yb}$ atomic array,” Physical Review X 13, 041034 (2023), doi:10.1103/PhysRevX.13.041034.
• M. Saffman, T. G. Walker, and K. Mølmer, “Quantum information with Rydberg atoms,” Reviews of Modern Physics 82, 2313–2363 (2010), doi:10.1103/RevModPhys.82.2313.
• M. Saffman, “Quantum computing with atomic qubits and Rydberg interactions: Progress and challenges,” Journal of Physics B 49, 202001 (2016), doi:10.1088/0953-4075/49/20/202001.
• P. Scholl et al., “Erasure conversion in a high-fidelity Rydberg quantum simulator,” Nature 622, 273–278 (2023), doi:10.1038/s41586-023-06516-4.