Rydberg Array Frontiers

Status: coherent storage arrays containing thousands of neutral atoms, programmable Rydberg simulators containing hundreds of atoms, parallel two-qubit entangling operations, logical encodings, erasure-aware circuits, and several ingredients of fault-tolerant operation are established. Combining all of those capabilities at once, for many repeated correction cycles and at useful algorithmic scale, is active. A general-purpose fault-tolerant neutral-atom computer, practical optimization advantage, and classically unverifiable analog computation with a complete validation story are not established.

Last reviewed: 26 July 2026. Array size, gate fidelity, logical-error suppression, and publication status are date-sensitive. In particular, a large array that demonstrates storage and transport is not automatically an equally large Rydberg-gated processor.

Core Question

Can a reconfigurable array of individually controlled atoms preserve its microscopic advantages while scaling from calibrated pairs to useful many-body and logical workloads?

Rydberg arrays combine three resources:

1. long-lived internal states for storage;
2. optical tweezers for imaging, transport, and geometry control; and
3. transient Rydberg excitation for strong, switchable interactions.

The frontier lies in composing them. A platform must load and rearrange atoms, maintain coherence, deliver spatially uniform gates or analog couplings, identify loss, execute repeated measurement and feedforward, and validate the result. Improving only one entry in that chain can leave the end-to-end workload unchanged.

A compact way to state the problem is

$$\begin{aligned} \text{useful scale} &\neq \text{atom count alone}, \\ \text{useful scale} &= f\!\left( N,\, F_{\mathrm{SPAM}},\, F_1,\, F_2,\, T_2,\, \mathcal P,\, t_{\mathrm{cycle}},\, \eta_{\mathrm{duty}},\, \mathcal V \right), \end{aligned}$$

where $\mathcal P$ denotes parallelism and connectivity, while $\mathcal V$ denotes the strength of the validation evidence. The function $f$ depends on the task. An analog quench, a logical memory, and a maximum-independent-set instance do not consume the same resources or tolerate the same errors.

Why It Matters

Reconfigurability changes the hardware question

Neutral atoms are identical by construction. Their interaction graph is not fixed by fabricated wiring: tweezers can arrange atoms into chains, ladders, two-dimensional graphs, and selected three-dimensional geometries. Mobile atoms can also mediate connectivity between code blocks or move qubits between storage and interaction zones.

This does not make connectivity free. Rearrangement consumes time, may heat atoms, can introduce loss and phase shifts, and must be scheduled around measurement and cooling. The scientifically useful statement is therefore not that an array has arbitrary connectivity, but that a specified graph can be realized with measured transport fidelity, latency, and crosstalk.

One platform supports several computational modes

The same laboratory can use Rydberg interactions in distinct ways:

• a calibrated pulse sequence can realize digital entangling gates;
• a global drive can realize an interacting spin Hamiltonian;
• local detunings can encode graph weights or spatially varying fields;
• rearrangement can change the graph between circuit stages; and
• auxiliary electronic manifolds can convert some leakage or loss into detectable erasures.

This digital–analog range is scientifically valuable because it permits cross-checks. It is also a source of ambiguity: a device optimized for a single analog ramp should not be judged by the same metric as a logical processor, and a high-fidelity two-atom gate does not certify a many-body Hamiltonian.

The platform tests ideas beyond computation

Rydberg arrays are laboratories for constrained dynamics, quantum scars, nonequilibrium phase transitions, frustrated magnets, gauge constraints, topological signatures, and doped quantum magnetism. Their local imaging can measure strings and high-order correlations that are difficult to obtain in bulk matter.

The strongest scientific use is not merely to reproduce a known phase diagram. It is to expose a controlled finite quantum system to measurements that discriminate models, quantify imperfections, and reveal dynamics in regimes where analytic methods and classical numerics become limited.

Error location can be a resource

An unlocated Pauli error and a located erasure are not equally damaging. Alkaline-earth-like atoms provide several electronic manifolds and nuclear-spin states, making it possible to flag some decay, leakage, or loss events without fully measuring the stored qubit. That possibility connects atomic physics directly to decoder design.

The benefit is conditional. Erasure conversion is useful only when the flag is reliable, the conversion itself does not add larger errors, correlations are characterized, and the code and decoder actually use the location information.

Canonical Boundary

This page is a dated assessment of research directions and evidence. It does not repeat the platform derivations.

• Rydberg Atoms Basics owns high-principal-quantum-number structure, scaling laws, quantum defects, and pair interactions.
• Rydberg Atoms owns excitation schemes, field compensation, trapping, lifetime, detection, pair calibration, preparation, and the platform error budget.
• Rydberg Blockade owns the two-atom blockade Hamiltonian, finite-blockade corrections, superatoms, blockade gates, constrained dynamics, and the derivation of the effective array Hamiltonian.
• Optical Tweezers owns stochastic loading, imaging, rearrangement, transport, and motional control.
• Neutral-Atom and Rydberg Qubits owns the durable architecture-level synthesis of encodings, zones, digital gates, analog operation, loss, throughput, and error-correction fit.
• Surface Code owns the code construction and decoding concepts used below.

Here the canonical questions are comparative and time dependent: what has actually been demonstrated, what scales together, which claims depend on postselection or model assumptions, and what evidence would change the assessment.

Minimal Background

Storage states and interaction states play different roles

In a common encoding, $|0\rangle$ and $|1\rangle$ are long-lived hyperfine, nuclear-spin, or clock-related states. A short pulse couples one of those states to a Rydberg level $|r\rangle$. The Rydberg level supplies the strong interaction but is not ordinarily the long-term memory.

This separation is useful:

$$\begin{array}{c|c|c} \text{role} & \text{typical state} & \text{desired property} \\ \hline \text{storage} & |0\rangle, |1\rangle & \text{long coherence and low crosstalk} \\ \text{interaction} & |r\rangle & \text{large, fast conditional shift} \\ \text{flag or ancilla} & |a\rangle & \text{selective detection or transport} \end{array}$$

The distinction prevents a common counting error. An experiment can demonstrate a large coherent register using storage states without simultaneously demonstrating Rydberg entangling control across the entire register.

The soft Ising model is a calibrated approximation

For a two-level ground–Rydberg description, a widely used model is

$$\frac{H}{\hbar} = \sum_i \left( \frac{\Omega_i}{2}\sigma_i^x -\Delta_i n_i \right) + \sum_{i<j}V_{ij}n_i n_j, \qquad n_i=|r_i\rangle\langle r_i|.$$

The parameters are measured quantities, not labels supplied by the apparatus: $\Omega_i$ can vary spatially, $\Delta_i$ includes light and interaction shifts, and $V_{ij}$ depends on pair state, angle, distance, and electric field. Spontaneous decay, dephasing, motion, leakage, and readout then turn this Hamiltonian into an open-system model.

The derivation and validity conditions belong to Rydberg Blockade. Frontier claims should report how the parameters and omitted terms were bounded.

A scaling vector is more informative than a qubit count

For a physical array, report at least

$$\mathbf m_{\mathrm{phys}} = \left( N_{\mathrm{site}}, N_{\mathrm{atom}}, f_{\mathrm{fill}}, T_1, T_2, F_{\mathrm{image}}, F_1, F_2, \mathcal P, t_{\mathrm{cycle}}, \eta_{\mathrm{duty}} \right).$$

Here $N_{\mathrm{site}}$ is capacity, $N_{\mathrm{atom}}$ is demonstrated occupancy, $f_{\mathrm{fill}}=N_{\mathrm{atom}}/N_{\mathrm{site}}$ is the raw or rearranged filling fraction, and $\mathcal P$ states how many comparable operations can run in parallel. The lifetime $T_1$ should specify what is lost: a trapped atom, a computational-state population, or a Rydberg excitation. The coherence time $T_2$ should identify the pulse sequence and noise filtering.

For a logical experiment, add

$$\mathbf m_{\mathrm{log}} = \left( N_L,\, d,\, n_{\mathrm{round}},\, p_L,\, \text{acceptance},\, \text{decoder},\, \text{feedforward} \right),$$

where $N_L$ is the number of logical qubits, $d$ is code distance when applicable, and $p_L$ is a logical error defined for a stated circuit. A logical-state demonstration with postselection is valuable, but it is not the same resource as repeated in-circuit correction.

Local fidelity does not equal global success

If $G$ independent operations each have fidelity $F$, the probability that no operation fails in the simplified stochastic model is

$$P_0=F^G,$$

and the expected number of faults is

$$\mathbb E[K]=G(1-F).$$

These formulas are diagnostics, not a full error model. Coherent, context-dependent, and correlated errors can accumulate differently. Nevertheless, the exponential makes a central scaling point: impressive local metrics must improve as the workload grows.

The same issue appears in imaging. If each of $N$ atoms survives one image with probability $s$, independent survival would give

$$P_{\mathrm{all}}=s^N.$$

A high per-atom survival probability can coexist with a modest probability that an entire large register remains unchanged. Fault-tolerant operation must therefore detect, replace, route around, or correct local events instead of demanding a perfect global shot.

Erasure information changes decoding

An ideal erasure channel announces where an error occurred:

$$\mathcal E_{\epsilon}(\rho) =(1-\epsilon)\rho +\epsilon |e\rangle\langle e|,$$

where $|e\rangle$ lies outside the computational space and is distinguishable from $|0\rangle$ and $|1\rangle$. Known locations reduce decoder uncertainty. Real experiments must also include missed flags, false flags, imperfect reset, and correlations:

$$\epsilon \longrightarrow \left( \epsilon_{\mathrm{flag}}, \epsilon_{\mathrm{miss}}, \epsilon_{\mathrm{false}}, \epsilon_{\mathrm{corr}} \right).$$

Erasure and Loss Channels develops the channel language. A platform should not call all atom loss a correctable erasure unless the location and timing are available to the protocol.

Below-threshold behavior is a scaling claim

A common heuristic for a family of distance-$d$ codes is

$$p_L(d) \approx A \left( \frac{p}{p_{\mathrm{th}}} \right)^{(d+1)/2}, \qquad p<p_{\mathrm{th}}.$$

The exponent and prefactor depend on the code, noise, syndrome circuit, and decoder. Evidence for below-threshold operation is therefore stronger than a single logical fidelity: increasing protection must reduce an operationally defined logical error under comparable conditions.

Repeated rounds matter. A state that is prepared, checked once, and postselected can demonstrate logical ingredients without demonstrating a memory whose lifetime improves under sustained correction.

Analog validation is a chain, not one comparison curve

A persuasive analog-simulation claim establishes several links:

$$\begin{aligned} \text{controls} &\longrightarrow \text{Hamiltonian parameters} \\ &\longrightarrow \text{state preparation} \\ &\longrightarrow \text{measurement model} \\ &\longrightarrow \text{observable and uncertainty} \\ &\longrightarrow \text{physical inference}. \end{aligned}$$

Validation can include exact calculations for small systems, tensor-network or quantum Monte Carlo comparisons where controlled, parameter sweeps, symmetry and sum-rule checks, independent observables, calibration holdouts, and finite-size scaling. No single method covers every regime.

Agreement on small instances calibrates a device but does not by itself prove the large-instance result. Conversely, classical intractability does not make an answer trustworthy without internal consistency checks.

Maximum independent set illustrates analog optimization

Let a graph $G=(V,E)$ be encoded by atoms at its vertices. In the hard constraint regime, a cost Hamiltonian can be written

$$H_{\mathrm{MIS}} = -\Delta\sum_{i\in V}n_i +U\sum_{(i,j)\in E}n_i n_j, \qquad U>\Delta>0.$$

The penalty disfavors occupying adjacent vertices, while the negative detuning rewards larger independent sets. A time-dependent Rydberg drive can attempt to prepare a low-energy configuration.

This encoding does not establish an algorithmic advantage. A fair benchmark must include graph embedding, auxiliary atoms, calibration, annealing time, shot count, classical parameter search, readout mitigation, and comparison against strong classical heuristics at matched solution quality and confidence.

What Is Known

Thousands of coherently controlled storage qubits are established

Manetsch et al. reported more than 6,100 neutral atoms in an array of about 12,000 optical tweezers, with a hyperfine coherence time of $12.6(1)\,\mathrm s$, a room-temperature trapping lifetime near $23\,\mathrm{min}$, imaging survival of $99.98952(1)\%$, imaging fidelity above $99.99\%$, and global single-qubit gate fidelity of $99.9834(2)\%$ Nature 647, 60–67 (2025).

Established: large coherent storage arrays, high-fidelity imaging, global single-qubit control, and transport operations at a scale beyond earlier Rydberg processors.

Not established by that result: 6,100 simultaneously entangling Rydberg-controlled qubits, a 6,100-qubit error-corrected processor, or an algorithmic quantum advantage. The distinction between atoms held coherently and atoms participating in a validated entangling workload is essential.

Continuous replenishment can stabilize large arrays

Chiu et al. demonstrated continuous operation of an array containing more than 3,000 atoms for over two hours. Their architecture reloaded atoms into tweezers at a rate of about $3\times10^5\,\mathrm{s^{-1}}$ and initialized qubits at about $3\times10^4\,\mathrm{s^{-1}}$, while preserving stored subarrays during replenishment Nature 646, 1075–1080 (2025).

Established: persistent refilling and coherent preservation of stored qubits can be combined, removing the assumption that an experimental shot must terminate when atoms are lost.

Important qualification: maintaining an array for two hours does not mean that the entire array occupied one unbroken coherent superposition for two hours. Storage coherence, atom presence, replacement, and global availability are different observables.

Continuous operation is a systems result. Its importance grows when replenishment, reset, cooling, routing, syndrome extraction, and computation are scheduled together without corrupting active code blocks.

High-fidelity parallel entangling gates are established

Evered et al. demonstrated Rydberg-mediated entangling gates with reported two-qubit fidelity of $99.5\%$, executed in parallel on up to 60 atoms, along with low-error three-qubit operations Nature 622, 268–272 (2023).

The result established that blockade gates need not be restricted to isolated pairs. It also exposed the next metric: the distribution and correlations of errors across a parallel layer. A mean pair fidelity does not specify whether residual laser noise, atomic motion, Rydberg decay, crosstalk, and calibration drift are independent.

At larger scale, the relevant object is a layer or circuit error model:

$$\mathcal N_{\mathrm{layer}} \neq \bigotimes_{g=1}^{G}\mathcal N_g$$

unless factorization has been tested. Characterizing nonlocal tails is particularly important for decoders that assume local stochastic noise.

Reconfigurable logical processors are established

Bluvstein et al. used up to 280 atoms to encode as many as 48 logical qubits in surface-code and color-code arrangements, with code distances up to seven. The experiment implemented transversal logical operations, logical algorithms, mid-circuit readout, feedforward, and atom rearrangement Nature 626, 58–65 (2024).

Established: reconfigurable neutral-atom hardware can operate on many logical blocks and use movement to restructure logical connectivity.

Qualification: not every demonstrated state-preparation and circuit step was fault tolerant, and error detection with postselection is not the same as indefinitely repeated correction. The paper was a logical-processor milestone, not a claim of a completed universal fault-tolerant machine.

Sales Rodriguez et al. subsequently demonstrated logical magic-state distillation using distance-three and distance-five color codes. The output logical magic states had higher fidelity than the inputs, establishing a key resource-conversion step for universal logical computation Nature 645, 620–625 (2025).

The scientific benchmark for distillation is end to end: input error, acceptance probability, output error, circuit faults, and resource overhead must all be reported. A conditional output of high fidelity can be valuable while still having a low production rate.

A broader fault-tolerant architecture has been demonstrated in components

Bluvstein et al. reported experiments using up to 448 neutral atoms, including surface-code operation, logical entanglement through transversal gates and lattice surgery, and a route to universality based on teleportation with three-dimensional $[[15,1,3]]$ codes. A four-round characterization showed a reported suppression factor of $2.14(13)$ below threshold, and mid-circuit atom reuse increased cycle rates by two orders of magnitude Nature 649, 39–46 (2026). The journal published a correction in January 2026.

Established: a wide set of fault-tolerant ingredients can coexist on one neutral-atom architecture, including repeated syndrome information, feedforward, logical entanglement, qubit reuse, and universal logical teleportation primitives.

Not established: a general-purpose machine running arbitrary, large-scale, useful algorithms with all components continuously below threshold. “Foundations for scalable fault tolerance” is a more accurate description than “completed fault-tolerant computer.”

The next decisive evidence will combine deeper correction cycles, more logical qubits, non-Clifford resource production, real-time decoding, atom replacement, and application circuits in one accounting of logical error per unit wall-clock time.

Erasure conversion has moved from proposal to hardware primitive

Scholl et al. demonstrated erasure conversion in a Rydberg quantum simulator, identifying a large fraction of spontaneous-decay events so they could be treated as located errors Nature 622, 273–278 (2023). This result established that atomic level structure can reshape the error channel, rather than merely lower its total probability.

Two 2026 experiments extended the idea with $^{171}\mathrm{Yb}$:

• Zhang et al. encoded nuclear-spin qubits through a metastable manifold, used mid-circuit erasure measurements and adaptive circuits, demonstrated a $[[4,2,2]]$ logical code with erasure-aware decoding, and teleported logical information between blocks with adaptive ancilla selection Nature Physics 22, 910–916 (2026).
• Senoo et al. coherently mapped entanglement among Rydberg, nuclear-spin, and optical-clock degrees of freedom, transferred an ordered Greenberger–Horne–Zeilinger state of up to 20 atoms into a metastable nuclear-spin register, and reported an error-detected two-qubit gate fidelity of $99.78(4)\%$ Nature Physics 22, 903–909 (2026).

Established: multiple electronic manifolds can support interaction, storage, and erasure-detection roles in one atomic species.

Active: showing that the full conversion-and-detection stack lowers logical error after accounting for missed events, false flags, extra circuit depth, correlated decay, and finite acceptance.

Analog Rydberg simulation has reached hundreds of atoms

The progression from pair physics to many-body dynamics is established across several complementary experiments:

• Bernien et al. observed coherent dynamics and many-body revivals in a 51-atom chain Nature 551, 579–584 (2017).
• Bluvstein et al. controlled driven dynamics in arrays from 3 to 200 atoms, studying scar-related revivals and driven stabilization Science 371, 1355–1359 (2021).
• Ebadi et al. mapped ordered phases and critical behavior in programmable arrays of 64 to 256 atoms, including a $16\times16$ geometry Nature 595, 227–232 (2021).
• Scholl et al. studied two-dimensional antiferromagnetic correlations in arrays containing hundreds of atoms Nature 595, 233–238 (2021).

These experiments established programmable preparation, evolution, and local readout for interacting spin models at scales beyond routine exact diagonalization. They did not make finite-size effects, imperfect preparation, or open-system corrections disappear.

Constrained dynamics can probe scars, gauge structure, and topology

The blockade constraint realizes Hilbert spaces that differ sharply from unconstrained spin systems. Revivals from selected product states provided evidence for weak ergodicity breaking associated with many-body scar structure. Driven protocols showed that such dynamics can be enhanced or stabilized.

Semeghini et al. used 219 atoms in a frustrated geometry and measured string operators and other signatures associated with a topological spin-liquid regime Science 374, 1242–1247 (2021). The careful statement is that the experiment observed multiple topological signatures in a controlled finite system. Whether a finite, imperfect preparation realizes a phase in the thermodynamic sense depends on the observable set, gap, correlations, finite-size scaling, and model fidelity.

More recent work observed string-breaking dynamics in an emergent two-dimensional gauge-theory setting Nature (2025). Such experiments are valuable because local imaging can follow both endpoints and the connecting string. The mapping to a gauge theory remains an effective description whose constraints and violations should be measured.

Doped quantum magnetism is experimentally accessible

Qiao et al. realized mobile holes in a two-dimensional Rydberg-tweezer antiferromagnet and investigated a finite-system analogue of doped strong-correlation physics Nature 644, 889–895 (2025).

This development matters because atom motion is not only a compilation tool. It can become part of the simulated Hamiltonian. The interpretation must separate coherent tunnelling, spin exchange, preparation defects, and atom loss, because all can change measured spin–hole correlations.

The experiment opens a route toward $t$–$J$-type questions in programmable geometries. It does not by itself settle the thermodynamic phase diagram of the doped Hubbard model.

Graph optimization has been demonstrated without practical advantage

Ebadi et al. implemented maximum-independent-set optimization on Rydberg arrays with up to 289 atoms and compared quantum evolution with classical methods on families of unit-disk graphs Science 376, 1209–1215 (2022).

Kim et al. introduced Rydberg quantum wires that use auxiliary atoms to embed higher-degree and nonplanar graph connections Nature Physics 18, 755–759 (2022). The technique expands the graph family at the cost of physical overhead and additional error channels.

Weighted graph optimization has been demonstrated using local light shifts to program vertex-dependent detunings, with closed-loop adjustment of the annealing schedule PRX Quantum 6, 010301 (2025). That study explicitly estimated that systems containing thousands of atoms would be needed before the tested setting could plausibly enter an advantage regime.

A public dataset covering maximum-independent-set experiments on king-lattice graphs up to 141 atoms and hundreds of thousands of shots provides a useful target for reproducible benchmarking Scientific Data 11, 111 (2024).

Established: programmable Rydberg arrays can encode and sample nontrivial graph-optimization instances, including weighted and embedded problems.

Not established as of this review: practical computational advantage over the best comparable classical workflows. Better-than-one-baseline performance or favorable scaling over a narrow instance range is not sufficient.

The evidence is capability specific

The following table deliberately keeps unlike achievements in separate rows.

Capability	Representative result	What is established	What is not implied
coherent array scale	more than 6,100 atoms in about 12,000 sites	storage, imaging, transport, and global control at large scale	6,100 validated Rydberg gates or logical qubits
persistent operation	more than 3,000 atoms maintained for over two hours	replenishment while preserving stored subarrays	one array-wide coherent state lasting two hours
parallel entangling gates	$99.5\%$ reported pair fidelity on up to 60 atoms	many simultaneous blockade gates	independent errors at arbitrary scale
logical processing	up to 48 logical qubits in a 280-atom system	reconfigurable logical circuits, readout, and feedforward	universal sustained fault tolerance
fault-tolerant ingredients	experiments using up to 448 atoms and four-round below-threshold characterization	several logical primitives and suppression with protection	a useful general-purpose fault-tolerant computer
analog simulation	programmable dynamics with hundreds of atoms	calibrated finite-system many-body experiments	automatic thermodynamic or classically unverifiable truth
optimization	maximum-independent-set and weighted-graph demonstrations	programmable encodings and quantum sampling	practical quantum advantage
erasure-aware control	detected decay, manifold mapping, and adaptive circuits	hardware can expose error location	all loss becomes a perfect erasure

What Is Actively Developing

Scaling entangling control with storage scale

The gap between the largest coherent storage arrays and the largest validated entangling workloads is now a primary engineering and physics target. Relevant developments include:

• flatter optical intensity and phase across larger fields of view;
• local beam steering with lower crosstalk;
• position and Doppler control during Rydberg pulses;
• electric-field stabilization across extended arrays;
• automated pair-interaction calibration;
• higher parallelism without correlated laser errors; and
• compilation that limits how often atoms enter lossy interaction states.

Progress should be reported as a joint distribution. A two-qubit fidelity measured on selected central pairs is not representative unless edge sites, transported atoms, simultaneous layers, and temporal drift are sampled.

Continuous fault-tolerant cycles

Persistent operation suggests a path away from destructive experimental shots. A full cycle would include

$$\begin{aligned} \text{detect loss} &\longrightarrow \text{supply atom} \longrightarrow \text{cool and initialize} \\ &\longrightarrow \text{route} \longrightarrow \text{verify} \longrightarrow \text{rejoin code}. \end{aligned}$$

Each step has latency and failure probability. The replacement process must not dephase neighboring data qubits or saturate control resources. Real-time decoding and scheduling must keep pace with syndrome generation.

The decisive metric is not reservoir loading rate alone, but logical throughput:

$$R_L = \frac{\text{accepted logical operations}} {\text{wall-clock time}},$$

reported with the logical error, code distance, decoder latency, and replacement overhead.

Erasure-biased architectures

Alkaline-earth-like species offer nuclear-spin memory, metastable manifolds, optical-clock transitions, and Rydberg interactions. Active work asks whether those degrees of freedom can produce a favorable bias between flagged erasures and unflagged Pauli errors.

The relevant figure is not the erasure fraction alone. One useful accounting is

$$p_{\mathrm{eff}} = w_e p_e +w_u p_u +w_f p_f,$$

where $p_e$ is correctly flagged erasure, $p_u$ is unflagged error, and $p_f$ is false flag. The weights depend on the code, decoder, and cost of discarding or repairing a block. A high erasure fraction can still be harmful if flags arrive too late or are strongly correlated.

Hybrid storage, interaction, and optical interfaces

Coherent mapping between electronic manifolds suggests specialized roles: nuclear spin for storage, Rydberg states for gates, clock states for spectroscopy or flags, and optical transitions for networking. Active questions include:

• whether mapping errors remain below the error avoided by specialization;
• how phase references are maintained across manifolds;
• how atom loss is distinguished from state leakage;
• whether photons can connect distant array modules efficiently; and
• whether modular boundaries simplify or complicate fault tolerance.

No single internal state needs to optimize every operation. The architecture must still optimize the complete round trip.

Digital–analog and dynamically reconfigured algorithms

Rydberg hardware can alternate analog evolution, local rotations, measurements, and physical rearrangement. This permits protocols that are neither standard gate circuits nor one fixed anneal.

Active research includes variational schedules, measurement-assisted state preparation, graph changes between layers, subsystem evolution, and digital corrections to analog Hamiltonians. Claims should state the control budget: the number of calibrated waveforms, local-addressing operations, measurements, optimization iterations, and classical feedback steps.

An ansatz can be expressive while being impractical to train. Conversely, a restricted ansatz can be scientifically useful if symmetries and conserved quantities make its validation transparent.

Validation beyond easy classical regimes

As array size grows, exact simulation ceases to be available. Active validation strategies include:

• overlapping regimes where exact diagonalization and tensor networks remain controlled;
• randomized parameter instances with hidden calibration checks;
• conserved quantities, Ward-like identities, and sum rules;
• local consistency between reduced density matrices;
• forward–reverse or echo protocols;
• multiple preparation paths to the same observable;
• cross-platform comparisons; and
• scaling collapse with explicit finite-size corrections.

The goal is not to attach one universal confidence number. It is to build a redundant evidence network that makes specific failure modes visible.

Optimization benchmarks with complete resource accounting

Optimization performance depends strongly on graph ensemble and objective. Active work is moving from hand-selected instances toward preregistered or public datasets, held-out instances, matched classical tuning budgets, and wall-clock accounting.

A useful report includes

$$\left( P_{\mathrm{opt}}, \frac{E-E_{\mathrm{opt}}}{|E_{\mathrm{opt}}|}, t_{\mathrm{wall}}, N_{\mathrm{shots}}, t_{\mathrm{cal}}, t_{\mathrm{classical}}, N_{\mathrm{physical}} \right).$$

$P_{\mathrm{opt}}$ is the probability of obtaining an optimum, while the relative energy or objective gap measures near-optimal output. Time-to-solution must specify the target confidence and include repetitions.

The graph presented to the quantum device may be much larger than the logical instance because quantum wires and auxiliaries consume atoms. That overhead belongs in the comparison.

Many-body dynamics with motion and dissipation

Most canonical Rydberg spin models assume fixed atom positions. Newer experiments increasingly use movement intentionally or study times on which motion, loss, and dissipation affect the dynamics.

This creates opportunities for doped magnets, quantum cellular automata, state-dependent transport, and dissipative phase preparation. It also requires models that join internal and motional degrees of freedom rather than treating motion only as dephasing.

The open problem is controlled complexity: adding motion or dissipation is scientifically useful only when the added terms are measured well enough to support the claimed effective theory.

What Is Debated or Uncertain

What should count as a qubit?

Capacity, occupied sites, coherently controlled atoms, simultaneously gated atoms, logical qubits, and algorithmically active qubits are all legitimate numbers. They answer different questions.

The debate becomes unproductive when one number silently substitutes for another. A trustworthy claim uses a noun:

• tweezer sites for hardware capacity;
• loaded atoms for occupancy;
• coherent storage qubits for characterized memory;
• entangling qubits for a validated gate layer;
• logical qubits for encoded degrees of freedom; and
• active logical qubits for those used in the reported circuit.

Is analog or digital operation the more scalable route?

Analog evolution uses the native interaction efficiently and can reach large many-body systems with shallow control. Digital circuits offer modular error analysis and a clearer route to general fault tolerance, but incur gate overhead.

There need not be one winner. Analog simulation may remain the strongest tool for selected Hamiltonians while digital logical processing develops on the same atomic platform. Hybrid protocols may exploit both. The right comparison is task specific and must include calibration and validation, not circuit depth alone.

When does optimization become quantum advantage?

No practical optimization advantage is established by the evidence reviewed here. Several harder questions remain:

• Which graph ensembles reflect useful problems rather than hardware-friendly constructions?
• Are classical baselines given comparable tuning and preprocessing?
• Does the quantum scaling survive embedding overhead and finite precision?
• Is time measured at the pulse level or from instance input to verified output?
• Does a claimed speedup hold at fixed solution quality and confidence?

Polynomial or exponential fits over a limited size range can be suggestive, but they are not asymptotic evidence. Instance distributions can also hide easy structure. Public data and independent classical reanalysis are therefore central.

What establishes a many-body phase in a finite array?

Finite systems do not possess thermodynamic phases in exactly the infinite system sense. Experiments infer phase structure from order parameters, correlation lengths, gaps, strings, entanglement proxies, response functions, and finite-size scaling.

Topological and gauge-theory claims require particular care because local observables alone may be insufficient. Multiple nonlocal diagnostics can strengthen the case, but preparation bias and model mismatch must still be bounded.

The uncertainty is scientific rather than rhetorical: a finite-system observation can be important and reproducible while the thermodynamic classification remains active.

Does atomic motion make connectivity effectively all-to-all?

Movement can connect distant qubits without fixed couplers, but every move has duration, heating, loss, collision-avoidance constraints, and control contention. A compiler may also need empty sites and routing space.

Connectivity should therefore be represented by a costed, time-dependent graph. “Any atom can meet any other atom” is a capability statement; it is not a zero-cost circuit model.

How robust is the erasure advantage?

Erasure-aware thresholds can exceed thresholds for unlocated errors under favorable noise assumptions. Real atomic errors may violate those assumptions:

• one decay event can scatter light onto neighboring atoms;
• a flag may be delayed until after an error propagates;
• state-dependent loss can bias logical information;
• false flags can consume code capacity; and
• global laser noise can correlate otherwise local gates.

The relevant test is logical suppression under the measured mixed noise model. Physical erasure conversion is a promising ingredient, not a standalone proof.

Can large analog results remain independently checkable?

At the frontier, the device may produce distributions that no classical method can compute exactly. Verification then becomes partial. One can test local marginals, symmetries, low-order moments, limiting cases, and cross-platform consistency without reconstructing the full distribution.

How much partial evidence is enough depends on the claim. A qualitative observation of domain growth needs less than a precision critical exponent or a claim of computational advantage. The validation standard should scale with the specificity and consequence of the conclusion.

Key Experimental Platforms

Platform mode	Typical resource	Natural strength	Principal bottleneck	Frontier use
alkali hyperfine arrays	ground-state hyperfine memory plus Rydberg excitation	mature trapping, fast control, large arrays	Rydberg decay, Doppler error, laser uniformity	digital gates, analog spin models, logical processors
alkaline-earth-like arrays	nuclear spin, metastable and clock manifolds, Rydberg states	state specialization and erasure flags	more complex lasers, mapping overhead, manifold leakage	erasure-aware logic, hybrid memory and gates
global analog arrays	global Rabi drive and detuning over programmable geometry	large native many-body evolution	Hamiltonian disorder and limited error correction	phases, quenches, scars, gauge dynamics
locally addressed arrays	site-resolved shifts, rotations, and measurement	graph weights and spatial protocols	beam crosstalk and calibration load	weighted optimization, digital–analog control
mobile-atom architectures	rearrangement between interaction and storage zones	dynamic connectivity and replenishment	transport time, heating, scheduling	logical routing, atom reuse, persistent operation
modular optical interfaces	atom–photon mapping between separated arrays	distance-independent links in principle	collection efficiency and interface fidelity	distributed logical systems and networking

No row is a complete machine. Current architectures combine several modes, and the interfaces between them often determine system performance.

Key Theoretical Tools

Effective Hamiltonians and open-system reductions

The theoretical starting point is a hierarchy:

$$\text{multilevel atom plus field} \longrightarrow \text{effective drive} \longrightarrow \text{pair interaction} \longrightarrow \text{spin model} \longrightarrow \text{noise channel}.$$

Adiabatic Elimination and related effective-Hamiltonian methods quantify when intermediate electronic states can be removed. Master equations, trajectory methods, and filter functions connect measured noise to dynamics.

The central discipline is uncertainty propagation: the effective model must retain error bars or nuisance parameters inherited from the calibration layer.

Optimal control and robust pulse design

Blockade gates depend on pulse phase, amplitude, detuning, interaction shift, and motion. Gradient methods, composite pulses, shortcuts to adiabaticity, and robust optimization can reduce sensitivity to known variations.

An optimized pulse is evidence only after testing outside its training model. Holdout positions, detunings, interaction strengths, and dates help detect overfitting to a calibration snapshot.

Randomized benchmarking and process diagnostics

Fidelity is not one number. Randomized benchmarking estimates average error under specific assumptions; interleaved variants isolate a target gate; simultaneous benchmarking probes crosstalk; cycle benchmarking targets circuit layers; and state or process tomography gives richer information at smaller scale.

Leakage, loss, and erasure require extensions to standard trace-preserving models. Report whether fidelity is unconditional, conditioned on survival, or corrected for state preparation and measurement.

Error-correcting codes and erasure-aware decoders

Surface codes, color codes, low-density parity-check codes, subsystem codes, and teleportation-based constructions trade locality, transversal gates, measurement depth, and decoder complexity differently.

Neutral-atom movement can relax a fixed-nearest-neighbor constraint, but a code analysis must charge for movement. Erasure-aware decoders should include missed and false erasures, not only ideal flags.

Classical many-body methods

Exact diagonalization, Krylov evolution, tensor networks, semiclassical methods, cluster expansions, and quantum Monte Carlo provide overlapping validation regimes. Their limitations depend on dimension, entanglement, frustration, sign structure, and evolution time.

The strongest studies compare several methods where they overlap and state where extrapolation begins. Failure of one classical method is not evidence that all classical analysis is impossible.

Finite-size scaling and statistical inference

Critical exponents, correlation lengths, and phase boundaries require finite-size models that include preparation and detection errors. Bayesian or likelihood-based inference can propagate uncertainty and compare alternative Hamiltonians.

Data splitting is useful: one subset calibrates parameters, another tests predictions. Reusing the same observables for tuning and validation can make agreement look stronger than it is.

Classical optimization and complexity-aware baselines

Maximum independent set admits exact solvers, branch-and-bound methods, message passing, tensor-network approaches, local search, simulated annealing, and graph-specific heuristics. Baselines should exploit the same known graph structure available to the quantum method.

Time-to-solution at confidence $q$ is commonly estimated from a per-shot success probability $p_s$:

$$N_{\mathrm{shot}} \ge \frac{\ln(1-q)}{\ln(1-p_s)}.$$

The wall-clock total also includes state preparation, measurement, reset, calibration, and classical optimization. Reporting only coherent evolution time gives an incomplete comparison.

Representative Papers and Reviews

Platform and architecture

1. 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.
2. M. Morgado and S. Whitlock, “Quantum simulation and computing with Rydberg-interacting qubits,” AVS Quantum Science 3, 023501 (2021), doi:10.1116/5.0036562.
3. S. J. 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.
4. H. J. Manetsch et al., “A tweezer array with 6,100 highly coherent atomic qubits,” Nature 647, 60–67 (2025), doi:10.1038/s41586-025-09641-4.
5. N.-C. Chiu et al., “Continuous operation of a coherent 3,000-qubit system,” Nature 646, 1075–1080 (2025), doi:10.1038/s41586-025-09596-6.

Logical and erasure-aware operation

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.
7. P. Sales Rodriguez et al., “Experimental demonstration of logical magic state distillation,” Nature 645, 620–625 (2025), doi:10.1038/s41586-025-09367-3.
8. D. Bluvstein et al., “A fault-tolerant neutral-atom architecture for universal quantum computation,” Nature 649, 39–46 (2026), doi:10.1038/s41586-025-09848-5.
9. 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.
10. B. Zhang et al., “Logical qubits with erasure conversion using metastable neutral atoms,” Nature Physics 22, 910–916 (2026), doi:10.1038/s41567-026-03309-0.
11. A. Senoo et al., “High-fidelity entanglement and coherent multi-qubit mapping in an atom array,” Nature Physics 22, 903–909 (2026), doi:10.1038/s41567-026-03258-8.

Many-body dynamics

12. H. Bernien et al., “Probing many-body dynamics on a 51-atom quantum simulator,” Nature 551, 579–584 (2017), doi:10.1038/nature24622.
13. D. Bluvstein et al., “Controlling quantum many-body dynamics in driven Rydberg atom arrays,” Science 371, 1355–1359 (2021), doi:10.1126/science.abg2530.
14. S. Ebadi et al., “Quantum phases of matter on a 256-atom programmable quantum simulator,” Nature 595, 227–232 (2021), doi:10.1038/s41586-021-03582-4.
15. P. Scholl et al., “Quantum simulation of 2D antiferromagnets with hundreds of Rydberg atoms,” Nature 595, 233–238 (2021), doi:10.1038/s41586-021-03585-1.
16. G. Semeghini et al., “Probing topological spin liquids on a programmable quantum simulator,” Science 374, 1242–1247 (2021), doi:10.1126/science.abi8794.
17. D. González-Cuadra et al., “Observation of string breaking on a $(2+1)$D Rydberg quantum simulator,” Nature (2025), doi:10.1038/s41586-025-09051-6.
18. M. Qiao et al., “Realization of a doped quantum antiferromagnet in a Rydberg tweezer array,” Nature 644, 889–895 (2025), doi:10.1038/s41586-025-09377-1.

Optimization

19. S. Ebadi et al., “Quantum optimization of maximum independent set using Rydberg atom arrays,” Science 376, 1209–1215 (2022), doi:10.1126/science.abo6587.
20. M. Kim et al., “Rydberg quantum wires for maximum independent set problems,” Nature Physics 18, 755–759 (2022), doi:10.1038/s41567-022-01629-5.
21. J. Byun et al., “Demonstration of weighted-graph optimization on a Rydberg-atom array using local light shifts,” PRX Quantum 6, 010301 (2025), doi:10.1103/PRXQuantum.6.010301.
22. K. Kim et al., “Quantum computing dataset of maximum independent set problem on king lattice of over hundred Rydberg atoms,” Scientific Data 11, 111 (2024), doi:10.1038/s41597-024-02926-9.

Reference lists are evidence maps, not endorsements of every interpretation. For fast-moving results, consult the version of record, corrections, data availability statement, and supplementary calibration methods.

Common Misconceptions

“The largest atom count is the processor size”

An atom count may describe occupied storage sites. Processor size for a specific claim should count the atoms that participate in a characterized gate layer, analog evolution, or logical circuit. State the role and metric.

“Rydberg blockade forbids double excitation”

Blockade is finite. Double excitation, phase error, weak interaction channels, motion, and decay remain. Rydberg Blockade quantifies the approximation.

“Identical atoms eliminate calibration”

Atoms share intrinsic spectra, but they experience different trap depths, laser intensities, phases, fields, temperatures, and neighbors. Identical particles reduce fabrication spread; they do not eliminate inhomogeneity.

“Rearrangement gives free all-to-all connectivity”

Rearrangement gives programmable connectivity with measurable transport time, heating, loss, and scheduling cost. Those resources must appear in a logical circuit estimate.

“A reported gate fidelity predicts an algorithm”

A pair-averaged fidelity may omit correlated errors, leakage, drift, and measurement overhead. Circuit performance depends on layer depth, parallelism, error structure, and correction.

“Postselection is fault tolerance”

Postselection can reveal logical suppression and prepare valuable states. Scalable fault tolerance also requires acceptable yield, repeated syndrome extraction, real-time decisions, and errors that fall with increasing protection.

“Every detected atom loss is a correctable erasure”

The protocol needs the location and useful timing of the event. Missed flags, false flags, correlated scattering, and reset errors can remove the expected decoder advantage.

“Hundreds of atoms imply quantum advantage”

Problem structure matters more than atom count alone. Classical algorithms can exploit geometry, low entanglement, symmetries, or approximation tolerance. Advantage requires a complete, reproducible comparison at matched task quality.

“A simulator observation proves the thermodynamic phase”

A finite array provides finite-size evidence. Strong claims combine multiple observables, scaling, model calibration, and uncertainty. Topological signatures and gauge constraints deserve especially explicit validation.

“Analog and digital results can be ranked by one fidelity”

Analog evolution is judged by Hamiltonian accuracy and observable inference; digital gates by channels and circuit performance; logical systems by suppression and throughput. Their metrics overlap but are not interchangeable.

Exercises

Exercise 1: Local imaging survival versus global yield

An array contains $N=6100$ atoms. Suppose each atom survives one image with probability

$$s=0.9998952,$$

and treat survival events as independent for this estimate.

1. Compute the probability that all atoms survive one image.
2. Compute the probability that all atoms survive ten images.
3. Explain why the result does not make high imaging survival unimportant.

Solution

For one image,

$$P_1=s^N =(0.9998952)^{6100} \approx 0.528.$$

For ten images, there are $10N$ atom-image opportunities:

$$P_{10}=s^{10N} \approx 1.68\times10^{-3}.$$

Thus an excellent local survival probability does not guarantee that a large register remains globally unchanged. The local metric is still crucial: it sets the rate of events that replacement, erasure detection, routing, or error correction must handle. Scalable operation is built from high local fidelity plus protocols that tolerate the remaining local events, rather than from demanding a perfect global shot.

Exercise 2: Parallel gate layers

A layer contains 30 disjoint two-qubit gates, each with fidelity $F_2=0.995$. Use an independent stochastic-fault model.

1. What is the probability that the layer contains no gate fault?
2. What is the probability that a 100-layer circuit contains no gate fault?
3. What is the expected number of gate faults in the 100-layer circuit?

Solution

For one layer,

$$P_{\mathrm{layer}} =(0.995)^{30} \approx 0.860.$$

There are $G=30\times100=3000$ gates in the full circuit, so

$$P_{\mathrm{circuit}} =(0.995)^{3000} \approx 2.95\times10^{-7}.$$

The expected number of faults is

$$\mathbb E[K] =G(1-F_2) =3000(0.005) =15.$$

This is not a prediction for a fault-tolerant circuit, which detects and corrects faults, nor does it describe coherent or correlated errors. It shows why local gate fidelity, code performance, and error structure must be reported together.

Exercise 3: Occupancy is not entangling scale

A system has $12{,}000$ tweezer sites and $6100$ loaded atoms.

1. Compute the occupied-site fraction.
2. Suppose a demonstrated entangling layer acts on 60 atoms. What fraction of the loaded atoms participate in that layer?
3. Why would it be misleading to quote either fraction as a universal device utilization?

Solution

The occupied-site fraction is

$$f_{\mathrm{fill}} =\frac{6100}{12000} \approx 0.5083.$$

The illustrative entangling participation fraction is

$$f_{\mathrm{ent}} =\frac{60}{6100} \approx 9.84\times10^{-3} =0.984\%.$$

These numbers refer to different published capabilities and should not be combined as though they came from one simultaneous workload. More generally, occupancy, coherent control, and entangling participation answer different questions. Utilization depends on the task, geometry, parallel layer, and whether idle storage qubits are part of the algorithm.

Exercise 4: Why the graph Hamiltonian encodes independence

Consider

$$H =-\Delta\sum_{i\in V}n_i +U\sum_{(i,j)\in E}n_i n_j, \qquad U>\Delta>0.$$

Show that no ground-state bit string can occupy both endpoints of an edge. Then show that among independent sets, a ground state has maximum cardinality.

Solution

Suppose a configuration occupies both endpoints of at least one edge. Choose an occupied vertex $i$ that has $k\ge1$ occupied neighbors and set $n_i$ from one to zero. Removing the vertex loses the reward $-\Delta$, increasing the first term by $\Delta$, but removes $k$ penalties, decreasing the second term by $kU$. The energy change is

$$\Delta E =\Delta-kU \le \Delta-U <0.$$

The original configuration was therefore not a ground state. Every ground state is an independent set.

For an independent set $S$, the penalty term vanishes and

$$E(S)=-\Delta |S|.$$

Minimizing energy is equivalent to maximizing $|S|$. Thus the ground-state configurations encode maximum independent sets. The argument establishes the static encoding, not the probability that a finite-time quantum evolution finds a ground state.

Exercise 5: Interpreting a below-threshold suppression factor

Use the heuristic

$$p_L(d) =A \left( \frac{p}{p_{\mathrm{th}}} \right)^{(d+1)/2}.$$

Suppose an experiment reports that the physical operating point is a factor $2.14$ below threshold, so $p/p_{\mathrm{th}}\approx1/2.14$. Ignoring changes in $A$, estimate the ratio $p_L(7)/p_L(3)$.

Solution

The exponent is two for $d=3$ and four for $d=7$. Therefore

$$\frac{p_L(7)}{p_L(3)} \approx \left( \frac{p}{p_{\mathrm{th}}} \right)^2 = \left( \frac{1}{2.14} \right)^2 \approx 0.218.$$

The heuristic predicts about a factor $4.58$ reduction. This estimate is not a reanalysis of the cited experiment. Finite rounds, circuit details, decoder choice, leakage, erasures, and correlated errors can change both exponent and prefactor. A below-threshold claim should be based on measured scaling, not only this formula.

Exercise 6: Missed erasures

Each of $n=100$ operations produces an erasure with probability $\epsilon=0.002$. A detector flags a true erasure with efficiency $q=0.98$. Neglect false flags and correlations.

1. What is the probability of at least one physical erasure?
2. What is the per-operation probability of an unflagged erasure?
3. What is the probability of at least one unflagged erasure?

Solution

The probability of at least one erasure is

$$P_{\ge1} =1-(1-\epsilon)^n =1-(0.998)^{100} \approx 0.181.$$

The unflagged probability per operation is

$$\epsilon_{\mathrm{miss}} =\epsilon(1-q) =0.002(0.02) =4.0\times10^{-5}.$$

Hence

$$P_{\mathrm{miss},\ge1} =1-(1-\epsilon_{\mathrm{miss}})^{100} \approx 3.99\times10^{-3}.$$

Efficient flagging converts a common unknown-location event into a much rarer unflagged event, but it does not remove the flagged erasures. The code still needs enough erasure tolerance, and a complete model must add false flags and correlations.

Exercise 7: Design an advantage benchmark

A Rydberg experiment reports a favorable scaling of maximum-independent-set success probability relative to simulated annealing. List at least six pieces of evidence needed before interpreting this as practical quantum advantage.

Solution

A credible study should include at least:

1. a specified graph ensemble and a rule preventing instance cherry-picking;
2. the physical embedding overhead, including wire and auxiliary atoms;
3. strong classical baselines with comparable tuning and preprocessing;
4. success probability or objective gap at a fixed confidence target;
5. complete wall-clock accounting for loading, calibration, evolution, readout, repetitions, and classical feedback;
6. uncertainty intervals and enough instances to separate scaling from finite-size fluctuations;
7. held-out or public data that permit independent reanalysis;
8. evidence that the favorable trend persists beyond a crossover caused by hardware overhead; and
9. a clear distinction between coherent evolution time, quantum-processing time, and end-to-end time.

No single checklist proves asymptotic advantage. It makes the claim falsifiable and ensures that the compared workflows solve the same task.

Connections to Other Volumes

• Neutral-Atom and Rydberg Qubits provides the processor architecture that this dated evidence assessment audits.
• Neutral Atoms develops decay, dephasing, loss, measurement, and trajectory models for the platform.
• Erasure and Loss Channels provides the canonical channel description for located loss.
• Transverse-Field Ising Model develops a central effective model realized by driven arrays.
• Lattice Models Overview places Rydberg spin models among Hubbard, Heisenberg, and constrained lattice systems.
• Entanglement and Quantum Information connects Bell states, multipartite entanglement, channels, and operational resources.
• Surface Code develops stabilizers, code distance, syndrome extraction, and decoding.
• Quantum Gates supplies conventions for the controlled and entangling operations used in digital protocols.
• Fidelity distinguishes state, process, entanglement, and average gate fidelity.
• Quantum Information Roadmap gives a path from composite systems through channels and error correction to hardware applications.
• AMO Physics Roadmap connects atomic structure, light–matter interaction, trapping, and metrology to neutral-atom platforms.

What Changed This Year

This section records changes since the previous annual literature baseline, not every paper published in 2026.

Fault-tolerant architecture entered the version of record

The neutral-atom fault-tolerance architecture reported online in late 2025 appeared in the January 2026 issue of Nature. It combined up to 448 atoms, surface-code experiments, logical entanglement, teleportation-based universal primitives, repeated syndrome information, and atom reuse. A January correction is part of the record and should be consulted with the article.

The update strengthens the evidence that many fault-tolerant components can coexist. It does not change the assessment that sustained, useful, general-purpose fault-tolerant computation remains active.

Ytterbium experiments joined erasure conversion to logical operation

Two papers published on 12 June 2026 demonstrated complementary $^{171}\mathrm{Yb}$ capabilities. Zhang et al. used erasure-aware logical circuits and adaptive teleportation; Senoo et al. mapped entanglement among Rydberg, nuclear-spin, and clock-related degrees of freedom and demonstrated error-detected gates.

Together they move multi-manifold atomic encodings from a conceptual advantage toward an experimentally integrated architecture. The next test is whether the extra mapping and detection operations lower logical error at competitive throughput.

Optimization work is broadening beyond one annealing protocol

Continuous-time quantum-walk ansätze were demonstrated on neutral-atom hardware in July 2026 Physical Review A 114, 012425 (2026). The result adds another experimentally accessible optimization family and motivates scaling tests.

It does not establish practical quantum advantage. The same end-to-end requirements apply: instance distributions, embedding, tuning, shots, wall-clock time, solution quality, and strong classical baselines.

Assessment

The 2026 change is architectural integration, not the disappearance of scaling constraints. Large storage, continuous operation, parallel gates, logical primitives, erasure information, and analog many-body control are now all experimentally real. The frontier is to make them simultaneous, repeatable, and quantitatively superior for a clearly defined scientific or computational task.

The next scheduled review should check:

• logical error versus code distance over more rounds;
• decoder and replacement latency in continuous operation;
• correlated-error measurements for large parallel gate layers;
• unconditional versus postselected logical throughput;
• analog validation beyond classically tractable sizes; and
• preregistered, end-to-end optimization benchmarks.