Manuscript
Nonlinear dynamics can make individual trajectories sensitive to small uncertainties in initial conditions, so reproducing attractor dynamics can be more informative than tracking a particular trajectory[1 ] . Ensemble prediction describes how this uncertainty evolves by propagating a probability distribution over possible states[2 ] . The wake behind a circular cylinder at Reynolds number 100 provides a simple test: disturbances grow from an unstable steady flow and saturate into stable periodic vortex shedding[3 , 4 ] . The probability density representing the initial uncertainty obeys a linear transport equation whose drift retains the nonlinear dynamics[5 , 6 ] . The challenge is to preserve these dynamics in low-degree reduced models while controlling the quantum costs of ensemble preparation, propagation and sampling.
Existing quantum approaches address different parts of this task. Near-term studies have demonstrated selected flow processes on quantum processors[7 , 8 ] , and a learned Koopman model implements linear observable evolution in circuits with nonlinear classical encoding and decoding[9 ] . Quantum lattice Boltzmann and Trotterized transport schemes treat prescribed velocity fields in which each component is independent of its own coordinate[10 , 11 ] ; the drift is then divergence-free, which excludes contraction towards attracting equilibria, a mechanism required for the ensemble dynamics targeted here. Reconstructing and re-preparing the distribution between steps[10 , 12 ] interrupts coherent propagation and can erode end-to-end quantum speedups[13 ] .
Carleman linearization and Koopman–von Neumann (KvN) embedding offer fault-tolerant routes to nonlinear drift[5 , 14 ] , with potential exponential advantage over deterministic Eulerian Liouville solvers for sparse KvN Hamiltonians[5 ] . Grid refinement can increase the step count or operator norm of Carleman solvers[15 , 16 ] and the matrix-element scale in sparse KvN Hamiltonian simulation[5 ] . An upwind KvN circuit using a linear combination of Hamiltonian simulations (LCHS–KvN) suppresses parasitic oscillations with about 10⁸ single-target multi-controlled gates for ten steps on a 128-cell grid[17 ] , with the authors noting that implementation requires fault-tolerant quantum hardware.
Bridging the gap between algorithmic advances and end-to-end feasibility in early fault-tolerant quantum computing (EFTQC) requires controlling circuit costs while retaining favourable grid scaling. In this regime, quantum processors run coherent algorithms with quantum error correction (QEC) before large-scale fault-tolerant hardware becomes available[18 ] ; below-threshold error correction has been demonstrated[19 ] , and IBM's roadmap targets a fault-tolerant processor for 2029[20 ] . Logical-qubit budgets constrain encoded coordinates; circuit-depth budgets constrain coherent propagation steps. The budget also covers initial-state preparation and the shots needed for accurate observable estimates.
Reduced-order models (ROMs) support repeated ensemble predictions for flow-control design and parametric studies[3 , 21 , 22 ] and determine both the physical approximation and the circuit: retained coordinates set the register width[5 , 23 ] , while polynomial degree and couplings set the rotations per step. Eliminating unresolved variables generally introduces memory and fluctuating-force terms[24 ] . Model error and initial uncertainty motivate ensemble prediction[25 ] . Jennings et al.[16 ] analyse end-to-end costs for Carleman-linearized spatial lattice-Boltzmann dynamics solved with a quantum linear solver. Reducing the flow before KvN embedding lets us exploit the ROM’s low-degree polynomial drift to compress circuits for ensemble prediction[2 , 26 , 27 , 28 ] .
Low-degree polynomial ROMs further connect physical modelling to efficient quantum propagation. Incompressible Navier–Stokes advection is quadratic, and standard Galerkin projection preserves this polynomial degree[4 ] . Quadratic and cubic models have been developed for wake dynamics[3 ] , reactor oscillations[29 ] , neuron activation and compressible flow[30 ] , and nonlinear structural vibrations[31 ] , while polynomial maps estimate drag and lift from wake-mode amplitudes[32 ] . We fit the reduced equations using sparse identification of nonlinear dynamics (SINDy) with a prescribed low-degree polynomial library[1 , 33 ] . We test whether the model and its closure retain growth, saturation and an amplitude-dependent oscillation frequency[4 , 34 ] . Walsh synthesis exploits the polynomial drift to reduce transport rotation counts.
At fixed model dimension and drift degree, the number of transport rotations per time step grows at most polylogarithmically with the number of cells per reduced coordinate, suggesting potential scaling benefits over deterministic ensemble grids[5 ] . We choose the polynomial degree by balancing model accuracy against KvN circuit cost. We then optimize designs for fault-tolerant execution. For the same task, we compare against a reconstructed block-encoding circuit that loads coordinates into amplitudes to form the drift operator[35 ] , using surface-code estimates[36 , 37 , 38 ] covering preparation, propagation and sampling to 1% relative root-mean-square (RMS) field error against the continuum model. To test whether KvN circuits can describe nonlinear classical dynamics on quantum hardware, we execute a four-qubit circuit for radial transport, omitting the phase register. We mitigate noise using reference circuits whose propagation gates have zero rotation angles. IBM hardware experiments demonstrate nonlinear radial transport through measurements of ensemble observables during approach to an attracting state.
Results
Cylinder-wake dynamics and quantum KvN circuit
An ensemble can spread as disturbances grow away from the unstable steady wake and concentrate as its radial amplitude saturates (Fig. 1a). We retain these dynamics with two oscillatory proper orthogonal decomposition (POD) modes and a shift mode linking steady and mean shedding flows [3 ] . Our prediction task prepares an initial ensemble, propagates it with a quantum circuit derived from the fitted equations, and reconstructs the field from radius and phase measurements. Circuit and shot counts then determine physical qubits and runtime (Fig. 1b–f).
To construct the quantum transport model, we project the computational fluid dynamics (CFD) snapshots in Fig. 1b onto the oscillatory POD pair and shift mode, and SINDy fits a quadratic model constrained by reflection symmetry (Methods). For normalized oscillatory coordinates x , y x,y x , y , we define the squared radial amplitude by r 2 = x 2 + y 2 r^2=x^2+y^2 r 2 = x 2 + y 2 and eliminate the independent normalized shift coordinate z z z using the fitted closure z ≈ r 2 z\approx r^2 z ≈ r 2 , following Brunton et al.¹. Angular averaging the fitted model with this closure gives
d r d t ∗ = μ ( r − r 3 ) , d θ d t ∗ = ω + γ r 2 . (1) \frac{\mathrm{d}r}{\mathrm{d}t^*}=\mu\left(r-r^3\right),\quad\frac{\mathrm{d}\theta}{\mathrm{d}t^*}=\omega+\gamma r^2. \tag*{(1)} d t ∗ d r = μ ( r − r 3 ) , d t ∗ d θ = ω + γ r 2 . (1)
Here r r r is radius normalized by its saturated value, θ \theta θ is oscillation phase and t ∗ t^* t ∗ is nondimensional physical time. The frozen coefficients are μ = 0.092482 \mu=0.092482 μ = 0.092482 , ω = 0.75309 \omega=0.75309 ω = 0.75309 and γ = 0.30173 \gamma=0.30173 γ = 0.30173 , where ω \omega ω is the radius-independent angular rate (Supplementary equation (S6)).
We propagate a Gaussian model of initial radius uncertainty²⁸ under equation (1). Its probability density ρ \rho ρ obeys the Liouville equation ∂ t ρ = − ∇ ⋅ ( F ρ ) \partial_t\rho=-\nabla\cdot(\mathbf{F}\rho) ∂ t ρ = − ∇ ⋅ ( F ρ ) . Its L 2 L^2 L 2 norm can change while total probability is conserved, so direct density evolution can be nonunitary, motivating methods such as LCHS³⁹. The radial drift vanishes at r = 0 , 1 r=0,1 r = 0 , 1 and the angle is periodic. Under these boundary conditions the continuous KvN evolution is unitary⁵,⁶, with the density given by the squared magnitude of the amplitude
∂ t ψ = − F ⋅ ∇ ψ − 1 2 ( ∇ ⋅ F ) ψ , ρ = ∣ ψ ∣ 2 , (2) \partial_t\psi=-\mathbf{F}\cdot\nabla\psi-\frac{1}{2}(\nabla\cdot\mathbf{F})\psi,\quad\rho=\lvert\psi\rvert^2, \tag*{(2)} ∂ t ψ = − F ⋅ ∇ ψ − 2 1 ( ∇ ⋅ F ) ψ , ρ = ∣ ψ ∣ 2 , (2)
where ψ \psi ψ is the amplitude and F \mathbf{F} F the vector field in equation (1). Contraction towards saturation is therefore compatible with unitary propagation. We represent N x = 2 n N_x=2^n N x = 2 n uniform cells on the normalized radius interval [ 0 , 1 ] [0,1] [ 0 , 1 ] with an n n n -qubit register. Each computational basis state labels one cell, and its squared amplitude gives the cell probability. A real skew-symmetric discretization couples neighbouring-cell amplitudes. Each coupling evolves as a two-state R y R_y R y rotation¹², with angle equal to the shared-face drift times the layer duration divided by the cell width. These local amplitude transfers approximate transport of the density by that drift. The even layer acts on pairs ( 0 , 1 ) , ( 2 , 3 ) , … (0,1),(2,3),\ldots ( 0 , 1 ) , ( 2 , 3 ) , … , and the odd layer on ( 1 , 2 ) , ( 3 , 4 ) , … (1,2),(3,4),\ldots ( 1 , 2 ) , ( 3 , 4 ) , … . Rotations commute within each layer because no cell belongs to two pairs. Across layers, pairs share cells, so their application order matters. We approximate their joint evolution using symmetric Trotterization: an even half step, an odd full step and an even half step cancel the leading splitting-error term and give second-order accuracy in time (Methods).
With N x N_x N x radial cells, about N x / 2 N_x/2 N x /2 pairs per layer require O ( N x ) O(N_x) O ( N x ) separate rotations per step. To retain potential quantum grid-scaling benefits⁵, we use Walsh synthesis⁴⁰, expressing each pair’s angle as a weighted sum of pair-index parities. One target bit distinguishes the cells within each pair; the remaining n − 1 n-1 n − 1 bits identify the pair. A Walsh parity is negative when its selected bits contain an odd number of ones and positive otherwise. Each nonzero coefficient requires one target rotation, with its sign set by that parity (Fig. 1c, illustrated for a linear drift). On the uniform grid, the shared-face coordinate is a weighted sum of the pair-index bits plus a constant. For drift degree at most K K K , parities involving more than K K K bits vanish, bounding the rotations per layer by
∑ ℓ = 0 min ( K , n − 1 ) ( n − 1 ℓ ) = O ( n K ) . (3) \sum_{\ell=0}^{\min(K,n-1)}\binom{n-1}{\ell}=O(n^K). \tag*{(3)} ℓ = 0 ∑ m i n ( K , n − 1 ) ( ℓ n − 1 ) = O ( n K ) . (3)
The three layers per step therefore require O ( ( log 2 N x ) K ) O\left((\log_2 N_x)^K\right) O ( ( log 2 N x ) K ) radial transport rotations, with K = 3 K=3 K = 3 for the wake (Supplementary Notes 3 and 4).
Reconstructing the velocity field requires the mean oscillatory coordinates and the mean shift coordinate. Under the shift closure, the normalized modal state is ( x , y , z ) T = ( r cos θ , r sin θ , r 2 ) T (x,y,z)^{\mathsf{T}}=(r\cos\theta,r\sin\theta,r^2)^{\mathsf{T}} ( x , y , z ) T = ( r cos θ , r sin θ , r 2 ) T . Integrating the angular equation in equation (1) gives θ = θ c + Φ \theta=\theta_c+\Phi θ = θ c + Φ , with θ c = θ 0 + ω ( t ∗ − 40 ) \theta_c=\theta_0+\omega(t^*-40) θ c = θ 0 + ω ( t ∗ − 40 ) , where θ 0 = atan2 ( y , x ) \theta_0=\operatorname{atan2}(y,x) θ 0 = atan2 ( y , x ) is taken from the CFD state projected at t ∗ = 40 t^*=40 t ∗ = 40 and Φ \Phi Φ accumulates along each radial trajectory. The accumulated phase, the vector of ensemble averages (moments) m \mathbf{m} m and the mean coordinates are
Φ ( t ∗ ) = γ ∫ 40 t ∗ r ( s ) 2 d s , m = ( ⟨ r cos Φ ⟩ , ⟨ r sin Φ ⟩ , ⟨ r 2 ⟩ ) T , \Phi(t^*)=\gamma\int_{40}^{t^*}r(s)^2\,\mathrm{d}s,\qquad\mathbf{m}=\left(\langle r\cos\Phi\rangle,\langle r\sin\Phi\rangle,\langle r^2\rangle\right)^{\mathsf{T}}, Φ ( t ∗ ) = γ ∫ 40 t ∗ r ( s ) 2 d s , m = ( ⟨ r cos Φ ⟩ , ⟨ r sin Φ ⟩ , ⟨ r 2 ⟩ ) T ,
⟨ x ⟩ = m 1 cos θ c − m 2 sin θ c , ⟨ y ⟩ = m 1 sin θ c + m 2 cos θ c , ⟨ z ⟩ = m 3 . (4) \langle x\rangle=m_1\cos\theta_c-m_2\sin\theta_c,\quad\langle y\rangle=m_1\sin\theta_c+m_2\cos\theta_c,\quad\langle z\rangle=m_3. \tag*{(4)} ⟨ x ⟩ = m 1 cos θ c − m 2 sin θ c , ⟨ y ⟩ = m 1 sin θ c + m 2 cos θ c , ⟨ z ⟩ = m 3 . (4)
Angle brackets denote ensemble averages. Two phase qubits coherently accumulate Φ \Phi Φ through radius-controlled updates around each radial step, without intermediate measurement (Fig. 1d). For this Gaussian input, matrix product state (MPS) preparation has a gate count logarithmic in radial cell count at fixed bond dimension, layer budget and compilation precision[41 ] . Classical compilation is offline; each shot executes the preparation circuit, N t N_t N t propagation steps and radius and phase readouts. Equation (4) then gives the modal averages for field reconstruction (Fig. 1e; Methods). We call the resulting circuit Walsh–KvN. Its 33-step benchmark uses nine logical qubits; end-to-end surface-code estimates place this task within the EFTQC window in Fig. 1f.
Ensemble simulation of vortex shedding with Walsh–KvN
Walsh–KvN recovers the alternating transverse-velocity lobes of the finite-element method (FEM) reference (Figure 2 a), using 64 radial cells and 33 propagation steps from t ∗ = 40 t^*=40 t ∗ = 40 to 49.829. For Figure 2 b, we divide training data into four consecutive time blocks, fit each candidate on three and predict the fourth. We average field errors against CFD projected onto the retained POD modes over four tests. Rotation bounds sum one Walsh layer for each of three modal coordinates, with 64 cells per coordinate (Methods). We retain the 12-coefficient quadratic fit on the Pareto frontier, beyond which error reductions per added rotation diminish. Exact modal expectations from the ideal circuit follow the FEM spiral (Figure 2 c). Near saturation, the FEM trajectory approaches the shift-closure relation z ≈ r 2 z\approx r^2 z ≈ r 2 (Figure 2 d).
Figure 2. Wake reconstruction and ROM selection. a, Transverse velocity at Reynolds number 100; rows show times and columns compare FEM and Walsh–KvN. b, Model validation. Filled symbols mark the Pareto frontier; the circle marks the adopted quadratic fit. We sum rotation bounds for one Walsh layer in each of three modal coordinates. c,d, Modal trajectories and shift closure from exact circuit expectations. e, Errors for the three-mode ROM (3-mode), its reconstruction after shift closure and angular averaging (Radial), and the reconstruction from circuit readouts (Circuit). Grey/red errors are divided by the FEM field norm; blue is noiseless bias divided by the matched continuum Radial ensemble field norm. The dotted 1% RMS target also includes sampling and implementation errors. f, Drag (top) and lift (bottom). Blue lines show exact circuit expectations; open points use 815 separate ideal shots shared per time. Shading marks the circuit interval.
fig:2e compares the three-mode ROM (3-mode), the field after shift closure and angular averaging (Radial), and circuit reconstruction (Circuit). For one initial state, 3-mode differs from FEM by 6.42% at the final time, including projection and reduced-dynamics errors. Radial differs from 3-mode by 0.380% above zero and below 0.60% throughout the interval. Both comparisons are normalized by the FEM field norm. For the initial ensemble, the blue Circuit vs Radial curve gives 0.107% noiseless bias after 33 steps.
More shots reduce sampling fluctuations but leave systematic field bias. Accounting for both and implementation-error bounds, 815 accepted shots meet the 1% relative RMS field-error target after 33 steps at a per-attempt implementation-error budget of 10 − 4 10^{-4} 1 0 − 4 (Methods; Supplementary Note 7).
We fit FEM drag and lift data with polynomials of modal coordinates[32 ] and average them directly from circuit readouts, without field reconstruction. Force readout measures phase correlations in different settings, so we assess it separately with 815 ideal shots per time for both forces. Their relative RMS errors against FEM are 0.21% and 10.5%, respectively (Figure 2 f; Methods). The larger lift error partly reflects greater phase sensitivity; drag is dominated by constant and radial terms. Lift’s FEM RMS denominator is only 1/7.9 of drag’s, further magnifying the relative error.
End-to-end fault-tolerant resource estimates
Physical-qubit count and total task time are coupled: minimizing one need not make the other feasible. We minimize each separately and assess both designs against the early fault-tolerant hardware window[42 ] . The processor combines a core that executes the logical circuit with magic-state factories (MSFs; Figure 3 a). MSFs distil and supply T states for non-Clifford gates through the core buffer, which stores them before consumption. The core’s data block holds and processes logical qubits encoded in surface-code patches.
Figure 3. Architecture, magic-state supply and joint selection. a , Surface-code core and multilevel magic-state factory (MSF). Each Pauli π / 8 \pi/8 π /8 rotation consumes one T state. Factory output accounts for cultivation, distillation and logical faults. b , Core and final-level factory timing. The QEC round duration and classical reaction time are t r o u n d t_{\mathrm{round}} t round and t r e a c t t_{\mathrm{react}} t react , respectively. With supply enforced, τ T \tau_{\mathrm{T}} τ T equals the nominal core interval τ c o r e \tau_{\mathrm{core}} τ core ; with waiting allowed, insufficient final-level units set τ T \tau_{\mathrm{T}} τ T to the mean supply interval τ M S F \tau_{\mathrm{MSF}} τ MSF . c , We construct the architecture frontier by Pareto filtering data-block layouts, code distances and MSF configurations using physical-qubit count Q Q Q , effective core error per T state e c o r e ( T ) e_{\mathrm{core}}^{(\mathrm{T})} e core ( T ) , factory-output error e M S F e_{\mathrm{MSF}} e MSF and T-state consumption interval τ T \tau_{\mathrm{T}} τ T . Joint selection evaluates retained architectures and synthesis precisions for arbitrary rotations, minimizing qubit count or total task time under the per-attempt error budget.
Fast data blocks use more qubits than compact blocks to shorten T-state consumption intervals[37 ] ; under a qubit cap, insufficient factory supply can limit this gain (Supplementary Figure S10 ). The supply policy also matters: we compare requiring mean supply to meet core demand (supply enforced) with permitting waits (waiting allowed; Figure 3 b). At physical error probability 5 × 10 − 4 5 \times10^{-4} 5 × 1 0 − 4 , waiting reduces minimum qubits by 27% while increasing time per attempt by 3.2% (Supplementary Figure S11 ).
Supplementary Figure S10. Coherent propagation capacity and factory supply. a, Maximum admissible step count for compact, intermediate and fast data blocks across physical error probabilities 2 × 10 − 4 2 \times10^{-4} 2 × 1 0 − 4 –9 × 10 − 4 9 \times10^{-4} 9 × 1 0 − 4 , under physical-qubit caps of 10 5 10^{5} 1 0 5 and 10 6 10^{6} 1 0 6 . b, Ratio of nominal core interval to T-state consumption interval at the corresponding capacity-maximizing designs. Unity denotes core-limited execution; values below unity denote factory-limited execution. Capacity uses a 10 − 4 10^{-4} 1 0 − 4 per-attempt error budget and a 24-hour budget divided by 815 accepted shots. It does not certify field accuracy after longer evolutions.
Supplementary Figure S11. Circuit-specific optimization of factory supply. a, Physical qubits and runtime per attempt versus core code distance for the wake circuit with 7,392 native T/T† gates. b, Relative changes in selected physical qubits and runtime when waits are allowed. c, Runtime changes across physical error probabilities and native T counts, keeping 4,803 arbitrary rotations fixed. d, Corresponding core-distance sweep for the 200,000-native-T control. The control tests circuit-cost sensitivity rather than an additional flow simulation. In a and d, blue circles use the left physical-qubit axis and orange squares use the right runtime axis; axis colours match the corresponding curves.
We vary data-block layouts, code distances that set error protection, distillation levels and parallel factory units. Pareto filtering keeps architectures for which no alternative improves qubit count, T-state timing, core error or magic-state error without worsening another measure. Tighter rotation synthesis reduces approximation error but requires more T states. We then jointly select a retained architecture and rotation-synthesis precision under the error budget to minimize qubits or total task time (Figure 3 c; Methods). Across physical error probabilities 2 × 10 − 4 2 \times10^{-4} 2 × 1 0 − 4 –9 × 10 − 4 9 \times10^{-4} 9 × 1 0 − 4 , we use a 10 − 4 10^{-4} 1 0 − 4 per-attempt implementation-error budget. At the reference probability, the minimum-qubit radial design needs 53,382 physical qubits and 20.07 h; the fastest needs 559,942 qubits and 1.62 h (Table 1 ).
Table 1. Fastest estimated fault-tolerant designs.
To assess how tailoring circuits to low-degree polynomial drift affects feasibility within the illustrative EFTQC window (Figure 4 a)[18 , 43 , 44 , 45 ] , we compare Walsh–KvN with two reference circuits (BE–KvN and LCHS–KvN). BE–KvN[35 ] uses the same centred-difference generator, initial distribution, grid and interval. Each block-encoding call costs O ( K ( log 2 N x ) 2 ) O\left(K\left(\log_{2}N_x\right)^2\right) O ( K ( log 2 N x ) 2 ) gates for N x N_x N x cells and drift degree K K K . Its oracle uses normalized coordinate inputs x i x_i x i satisfying ∑ i x i 2 = 1 \sum_i x_i^2 = 1 ∑ i x i 2 = 1 . Loading our fixed-interval grid requires rescaling these coordinates; restoring the physical drift scale gives the N x 1 + K / 2 N_x^{1+K/2} N x 1 + K /2 call-count factor at fixed duration and tolerance in this reconstruction (Supplementary equations (S81)–(S84)). Enforcing normalization through grid spacing would change the interval; normalizing the full generator instead requires proportionally longer evolution. Walsh–KvN is about three orders of magnitude faster for the minimum-qubit and fastest designs; both circuits fit the window’s qubit range, but only Walsh–KvN finishes within one day (Figure 4 b,Figure 4 c).
Figure 4. Physical resources for nonlinear ensemble prediction. a a a , Published hardware settings and estimates; shading marks the illustrative early fault-tolerant window. b b b , Minimum-qubit (open) and fastest (filled) designs under a 10 − 4 10^{-4} 1 0 − 4 per-attempt error budget. Curves exceeding one million qubits are clipped. c c c , Total task time at physical error probability 5 × 10 − 4 5 \times10^{-4} 5 × 1 0 − 4 ; inset, expected time per accepted shot. Walsh–KvN and matched BE–KvN need 815 and 813 accepted shots for 1% relative RMS field error. LCHS–KvN uses its published sine-drift task; total time assumes 815 accepted shots for comparison, without a matched field-error target. Its fastest design uses 2.98 million physical qubits and appears only in c c c . Axis breaks indicate separate time scales. Supplementary Notes 8–11 specify the model and search.
LCHS–KvN[17 ] uses nonunitary upwind transport to suppress parasitic oscillations. Its optimized bound has a logarithmic grid factor whose exponent is independent of drift degree under the assumptions in Supplementary Note 3. Its published ten-step benchmark uses a cubic polynomial in the sine signal, avoiding inverse-sine approximation in drift encoding. Recompiling this circuit at its original input and grid gives a fastest uncapped design outside the window’s qubit range at the reference physical error probability (Figure 4 c; Methods).
At the same 1% field-error target, refining 64 to 128 cells doubles steps from 33 to 66 and raises runtime from 1.62 to 5.09 h, with one extra logical qubit and less than 1% more physical qubits (Table 1 ). For resolved distributions, refinement may replace upwinding; strong contraction may require upwinding to suppress oscillations, as resolving them by centred refinement can exceed EFTQC budgets (Methods).
To assess sensitivity to factory design and the corresponding error modelling, we re-estimate the minimum physical-qubit count for the 33-step radial Walsh–KvN circuit using two additional factory models [37 , 46 ] and the same core model (Supplementary Figure S12 ). We also assess sensitivity to gate composition by varying the number of native T gates—those present before arbitrary-rotation synthesis—while holding the logical width and arbitrary-rotation count fixed. At each setting, we reoptimize the architecture and rotation-synthesis precision under each factory-supply policy. The qubit savings and runtime changes obtained by allowing waits for T states depend on both native T count and physical error probability; the runtime change can even reverse sign (Supplementary Figure S9 and Supplementary Figure S11 ). These results show why architecture, supply policy and synthesis precision must be optimized together for each circuit and resource budget.
Supplementary Figure S9. Physical-qubit changes under alternative factory-supply policies. Colours show the relative change between the independently minimized footprints, using the supply-enforced policy as the reference. The waiting-allowed control fixes one top-level unit, and the supply-enforced design meets mean core demand. The dashed line marks 7,392 native T gates, and outlined circles mark the largest sampled runtime increase at each marked physical error probability. Synthesis, distances and factory settings are reoptimized at each grid point. Note 10 and the sweep source data specify fixed inputs and grid controls.
Supplementary Figure S12. Error budget and resource-model sensitivity. a, Physical qubits and total task time for the same 1% relative RMS wake-field-error target under different per-attempt error budgets. The marked 10 − 4 10^{-4} 1 0 − 4 setting is the common comparison error budget, not the field tolerance. b, Minimum-qubit estimates for the same 33-step radial Walsh–KvN circuit under three factory models with a common core model. Bracketed labels use main-manuscript reference numbers.
Uncapped designs at physical error probability 5 × 10 − 4 5 \times10^{-4} 5 × 1 0 − 4 and per-attempt error budget 10 − 4 10^{-4} 1 0 − 4 . Wake rows target 1% relative RMS field error. Shots are accepted samples; times include failed attempts. T counts include rotation synthesis.
† Published LCHS–KvN task; total time assumes 815 accepted shots without a matched field-error target.
‡ Bounds exclude rotation synthesis and shots. n n n is the maximum coordinate-register width, N x = 2 n N_x = 2^n N x = 2 n , N t N_t N t the step count, T T T the normalized duration and A L C H S A_{\mathrm{LCHS}} A LCHS the LCHS generator (Supplementary Notes 3, 11 and 12).
Coupled nonlinear dynamics and wake growth
To test the accuracy of Walsh–KvN as polynomial degree and dimension increase, we compare circuit predictions with matched classical solutions (Figure 5 ). For N c N_c N c transported coordinates, at most N x N_x N x cells per coordinate and drift polynomials of total degree at most K K K , suppose each drift component depends on at most s c s_c s c coordinates. The transport then uses O ( N c s c K ( log 2 N x ) K ) O(N_cs_c^K(\log_2N_x)^K) O ( N c s c K ( log 2 N x ) K ) arbitrary-angle rotations per step, or O ( N c K + 1 ( log 2 N x ) K ) O(N_c^{K+1}(\log_2N_x)^K) O ( N c K + 1 ( log 2 N x ) K ) with full coupling. These bounds require the grid and boundary treatment to preserve the polynomial structure of the rotation angles (Supplementary Note 3).
Figure 5. Nonlinear benchmarks and three-coordinate wake simulation. a , Quadratic (Q2), cubic (C3), quintic (Q5) and two-dimensional coupled cubic (2D C3) drifts. Filled/open fixed points are stable/unstable. Blue, purple and red denote near-stable, interior and near-unstable inputs. Diamonds show shot-noise-free relative moment errors; squares/bars show medians and 2.5th–97.5th percentiles of 5,000 estimates, each with 10,000 shots. b , Initial (grey) and final quintic densities from circuit (blue) and characteristics (dashed). The Gaussian centre x 0 x_0 x 0 is defined before truncation to the spatial domain [ 0 , 1 ] [0,1] [ 0 , 1 ] . c , Qubits and task time for coupled cubic input, using 674 shots without a qubit cap. d , Three-coordinate wake means from growth (blue) and late (orange) circuits versus continuum ensembles (dashed). e , Physical resources without a qubit cap. Dashed/open and solid/filled designs minimize qubits and time. Panels c /e target 1% relative RMS error under a 10 − 4 10^{-4} 1 0 − 4 implementation-error budget. Grey marks qubits outside the study window.
We benchmark one-dimensional quadratic (Q2), cubic (C3) and quintic (Q5) drifts and a two-dimensional coupled cubic drift (2D C3) against matched characteristic solutions (Figure 5 a,Figure 5 b). At a common budget of 10,000 shots, all twelve displayed inputs have relative moment RMS errors below 1%, combining noiseless bias and sampling variance. Lower errors near stable points are consistent with contraction damping trajectory perturbations, whereas repelling dynamics amplify them near unstable points (Supplementary Note 5). Contraction also narrows the probability density and can increase the spatial resolution needed to represent it. In a separate Q5 near-boundary test, insufficient resolution produces parasitic grid-scale oscillations in the centred discretization and a moment bias that additional shots cannot remove (Supplementary Fig. S5). For the coupled cubic example, we estimate fault-tolerant resources using 674 shots chosen to meet the 1% relative RMS moment-error target (Figure 5 c; Methods).
To test wake growth without the shift closure, we retain the fitted three-coordinate quadratic dynamics through growth and saturation (Figure 5 d). The radius r r r and shift coordinate z z z use finite grids. The polar equations contain sine and cosine functions of angle; we use a Fourier basis to handle this angular dependence while retaining polynomial rotation angles for Walsh synthesis (Supplementary Note 12). Two flag ancillas suppress spurious wrap-around connections introduced by index shifts at the coordinate and Fourier-mode boundaries (Methods). The growth and late runs start at nondimensional times t ∗ = 29.8 t^*=29.8 t ∗ = 29.8 and t ∗ = 40 t^*=40 t ∗ = 40 , respectively, from Gaussian ensembles defined in Cartesian modal coordinates about the corresponding model states and then mapped to the polar grids; both end at t ∗ = 49.829 t^*=49.829 t ∗ = 49.829 . To select finite coordinate ranges, we classically integrate the fitted three-coordinate model from probability-weighted initial points representing each Gaussian ensemble. Based on these trajectories, we retain 0 ≤ r ≤ 1.45 0\le r\le1.45 0 ≤ r ≤ 1.45 and 0 ≤ z ≤ 1.50 0\le z\le1.50 0 ≤ z ≤ 1.50 for growth, and 0 ≤ r ≤ 1.65 0\le r\le1.65 0 ≤ r ≤ 1.65 and 0.30 ≤ z ≤ 1.70 0.30\le z\le1.70 0.30 ≤ z ≤ 1.70 for late. Numerical trajectory checks estimate probability below 10 − 6 10^{-6} 1 0 − 6 outside these ranges during propagation; grid resolution is then selected within the fixed ranges (Methods; Supplementary Note 12). With shots chosen for 1% relative RMS field error against each run’s continuum reference, the fastest fault-tolerant designs use fewer than one million physical qubits and are estimated to take about 9 hours for late and 26 days for growth (Table 1 ; Figure 5 e).
Discussion
In addition to a conditional exponential gain over deterministic Eulerian Liouville grids for sparse KvN Hamiltonians, amplitude estimation can provide a quadratic query gain over trajectory Monte Carlo[5 ] . Realizing these gains for a full PDE also requires efficient ensemble preparation and propagation. For a scalar field u ( x ) u(x) u ( x ) on M M M spatial points, N u N_u N u value bins per point give N u M N_u^M N u M joint field configurations[5 , 23 ] . Direct KvN embedding therefore needs O ( M log 2 N u ) O(M\log_2 N_u) O ( M log 2 N u ) position qubits. While this linear dependence on physical grid size can offset the potential quantum gain, model reduction replaces the grid-point variables with N c N_c N c modal coordinates, requiring O ( N c log 2 N x ) O(N_c\log_2 N_x) O ( N c log 2 N x ) position qubits for N x N_x N x cells per coordinate. An end-to-end quantum advantage over classical sampling is not established for this benchmark.
Combining SINDy-derived ROMs with Walsh synthesis[40 ] offers a further benefit as dimension grows. As in total-degree Carleman embeddings[47 ] , a library for N c N_c N c coordinates and total degree at most K K K contains ( N c + K K ) \binom{N_c+K}{K} ( K N c + K ) monomials, including the constant, compared with ( K + 1 ) N c (K+1)^{N_c} ( K + 1 ) N c combinations when degree 0 to K K K is allowed independently per coordinate. In low-dimensional models, increasing the polynomial degree gives O ( K N c ) O(K^{N_c}) O ( K N c ) terms. When more coordinates are retained but the degree stays low, the library grows only as O ( N c K ) O(N_c^K) O ( N c K ) , and SINDy can select a sparse active subset[1 ] . Bounded coordinate coupling also keeps the Walsh–KvN rotation bound linear in N c N_c N c at fixed degree and resolution. This is also the regime where KvN’s conditional gain over a classical joint grid of N x N c N_x^{N_c} N x N c configurations becomes relevant.
Giannakis et al.[48 ] demonstrated Koopman evolution on three qubits of IBM Quantum System One, with the hardware implementation limited to circle rotation at constant angular velocity. Using Walsh–KvN circuits, we demonstrate nonlinear radial drift proportional to r − r 3 r-r^3 r − r 3 towards an attracting state on IBM hardware with four qubits, 16 cells and four or six propagation steps (Extended Data Fig. 1). We mitigate noise using an effective depolarizing model fitted to reference circuits with the same initial state and zero-angle propagation rotations (Methods). One qubit selects cells within pairs; three address the pairs. Cubic drift permits 1 + 3 + 3 + 1 = 8 1+3+3+1=8 1 + 3 + 3 + 1 = 8 Walsh terms: one constant, three single-bit, three two-bit and one three-bit parity. All are nonzero here, versus up to eight pair rotations per layer. At 64 cells, five address bits limit cubic support to 1 + 5 + 10 + 10 = 26 1+5+10+10=26 1 + 5 + 10 + 10 = 26 of 32 terms. Several full-count deviations from ideal-circuit predictions exceed counting uncertainty (Supplementary Note 15). Mitigation permits negative probabilities and estimated means above the largest cell centre, 31/32. These limits motivate early fault-tolerant tests of rotation savings over longer evolutions with field reconstruction.
Methods
Physical model and computational task
We construct the reduced model from the antisymmetric component of finite-element wake snapshots, which separates the shedding oscillation from symmetric steady and mean-flow components. Proper orthogonal decomposition ranks these spatial patterns by their contribution to the sampled-field squared norm. We retain the leading oscillatory pair and obtain the shift mode by projecting the difference between the cycle-mean and independently computed steady fields onto the orthogonal complement of that pair and normalizing it. Least-squares projection of the original fields gives the modal coefficients. Coordinate normalization is paired with rescaling of the spatial modes so that the reconstructed field is preserved (Supplementary Note 1).[1 , 3 , 33 ]
We use six CFD trajectories with different initial perturbations, labelled cases A–F, for model fitting and validation. Case A supplies the spatial basis and 292 projected states over the nominal nondimensional interval 10–80. We fit the three coupled equations by sequentially thresholded ridge regression, using a quadratic library without a constant term. Reflection symmetry allows 13 coefficients; ridge parameter 10 − 4 10^{-4} 1 0 − 4 and threshold 0.03 in RMS-normalized coordinates retain 12. A separately fitted shift closure and uniform angular averaging give equation (1) . Cases B–F provide five other initial conditions for evaluating the frozen models without refitting. Selection uses case A alone; derivative stencils remain within their training or validation partitions, and forecast scores exclude the initialized state. Supplementary Note 1 and Source Data specify the normalization, coefficients, derivative estimation, fit intervals and tests on cases B–F.
For the accuracy–cost comparison in Figure 2 b, we evaluate polynomial degrees 1–5 over thirteen thresholds. Four contiguous 73-state blocks define the tests: each fit uses three blocks and predicts the fourth from its first projected CFD state. Errors against CFD projected onto the fixed POD basis are averaged over the four tests, so this diagnostic excludes projection error and does not validate learning of the basis. The plotted rotation bound sums the bounds for one interior layer in each modal coordinate on the same 64-cell Cartesian grid. It excludes boundaries, splitting multiplicities, steps, preparation, rotation synthesis, shots and error correction; it is not a count for the polar circuit. Candidate scores and additional forecast tests are in Supplementary Note 1 and Source Data.
The primary input is a Gaussian in normalized radius, truncated and normalized on [0,1]. Target cell masses are differences of its cumulative distribution function over 64 uniform cells. The pre-truncation standard deviation is 0.25; the Gaussian centre is adjusted so that the target cell-centre mean after truncation and discretization matches the initial projected CFD radius, 0.9017. The Gaussian is a prescribed algorithmic ensemble, with no uncertainty distribution fitted from CFD. A matrix product state (MPS) with bond dimension two approximates the square roots of these cell masses, the target amplitudes; its prepared masses define the piecewise-uniform continuum input shared by the matched circuits. Input compilation is offline, while every attempt executes the preparation circuit. Supplementary Note 2 gives the distribution parameters, compilation and fidelity. Quantum simulation errors are evaluated against the matched continuum reduced model; FEM comparisons assess physical model error. The 1% field target is chosen below the observed physical ROM errors.
For saved-window model audits, full-field error is the aggregate two-component velocity residual norm divided by the FEM reference norm. Projection error applies the same metric to the field reconstructed from the retained modes. Model validation instead compares against the projected CFD field, including the steady background in its denominator. Supplementary Note 1 separates these normalizations and the forecast windows.
Transport and readout
The radial generator is discretized on the uniform grid as a real skew-symmetric matrix. For neighbouring cells in increasing index order, the shared-face block A f a c e A_{face} A f a ce and its exact two-state evolution are
A f a c e = f f a c e 2 h ( 0 − 1 1 0 ) , e δ t A f a c e = R y ( f f a c e δ t h ) . (5) A_{face}=\frac{f_{face}}{2h}\begin{pmatrix}0 & -1 \\ 1 & 0\end{pmatrix},\ e^{\delta t A_{face}}=R_y\left(\frac{f_{face}\delta t}{h}\right). \tag*{(5)} A f a ce = 2 h f f a ce ( 0 1 − 1 0 ) , e δ t A f a ce = R y ( h f f a ce δ t ) . (5)
Here f f a c e f_{face} f f a ce is the drift at that face, h h h the cell width and δ t \delta t δ t the layer duration. The even–odd–even sequence described in the main text supplies the symmetric time step. We retain all internal face couplings and use the vanishing endpoint drift to eliminate the formal wrap-around coupling (Supplementary Note 3).
For N x = 2 n N_x=2^n N x = 2 n , the n − 1 n-1 n − 1 address bits identify a pair and the last bit distinguishes its two cells. Subtracting one from the cell index aligns odd pairs with the same target bit; the inverse shift restores the original indexing.
We implement these reversible shifts with exact multi-controlled flips using one shared clean work qubit, restored after each operation (Supplementary Note 4). A nonzero Walsh coefficient is implemented by one target rotation between CNOT parity networks. A degree-K K K term needs at most 2 K 2K 2 K CNOT gates. Equation (3) bounds the number of nonzero Walsh rotation terms; Supplementary Notes 3 and 4 give the coupled-coordinate construction, boundary conditions and gate inventories. The reported step counts are fixed rather than jointly optimized with shot count.
Two phase qubits, P X P_X P X and P Y P_Y P Y , start in equal superpositions prepared by Hadamard gates. Squared-radius-controlled phase updates surround each radial step. For a full step of duration δ \delta δ , we use
D ( δ ) = exp [ i γ δ r ^ 2 ( n ^ X + n ^ Y ) ] , U f ( δ ) = D ( δ / 2 ) [ U r ( δ ) ⊗ I p h ] D ( δ / 2 ) . (6) D(\delta)=\exp\left[i\gamma\delta\hat{r}^{2}\left(\hat{n}_X+\hat{n}_Y\right)\right],
\qquad
U_f(\delta)=D(\delta/2)\left[U_r(\delta)\otimes I_{ph}\right]D(\delta/2). \tag*{(6)} D ( δ ) = exp [ iγ δ r ^ 2 ( n ^ X + n ^ Y ) ] , U f ( δ ) = D ( δ /2 ) [ U r ( δ ) ⊗ I p h ] D ( δ /2 ) . (6)
Here U r U_r U r transports radius, r ^ \hat{r} r ^ is the radius operator, I p h I_{ph} I p h is the identity on the phase register and n ^ α = ( I − Z α ) / 2 \hat{n}_{\alpha}=(I-Z_{\alpha})/2 n ^ α = ( I − Z α ) /2 is the occupation operator of phase qubit P α P_{\alpha} P α for α = X , Y \alpha=X,Y α = X , Y , with Z α Z_{\alpha} Z α denoting Pauli Z on that qubit. Adjacent half updates combine between successive steps; the phase register remains coherent until readout. Repeating this joint step gives the propagation block in Fig. 1e. The moment estimators and reconstruction follow equation (4) and Supplementary Note 6.
We check continuum moments using DOP853 characteristic integration and 16- and 32-point Gauss–Legendre quadratures per original radial cell. The coupled three-mode field reference uses DOP853 at relative and absolute tolerances of 10 − 11 10^{-11} 1 0 − 11 and 10 − 13 10^{-13} 1 0 − 13 . Polynomial benchmark references and convergence checks are specified in Supplementary Notes 5–7.
For the oscillation checks, we compare the centred generator with an upwind generator that uses backward or forward differences according to the drift direction.[17 ] The wake comparison keeps the prepared input, radial spatial domain, evolution time window and joint radius–phase readout fixed. Both discretizations use classical finite-generator exponentials followed by normalization of the output. Supplementary Notes 4 and 5 give the distribution and moment errors. These checks compare discretizations; they do not add compiled LCHS resource estimates.
The three-coordinate calculation retains the fitted dynamics in radius, angle and shift. Its initial inputs are Cartesian Gaussians centred on the corresponding model characteristics, with widths 0.12, 0.12 and 0.09, conditioned on the retained radial and shift intervals. We compute polar cell masses with the radial Jacobian and approximate their square-root amplitudes with a bond-dimension-four MPS. Classical probability-weighted trajectories select the coordinate ranges before grid refinement. Candidate grids must keep deterministic field discrepancy below 1% at all reported times and admit a finite shot count for the final-time RMS target. At fixed steps and stencil, selection minimizes required shots times the per-attempt T count, including native T gates and rotations synthesized at a common precision (Supplementary Note 12). The domains, grids and propagation intervals are specified in Supplementary Note 12.
Finite radial and shift stencils omit exterior couplings, and the angular Fourier matrix is truncated at the retained mode boundaries. Two validity flags suppress coordinate-pair and angular-mode-pair connections that cyclic shifts would otherwise introduce. The same clean arithmetic ancilla is reused for shifts and checks; all work qubits are uncomputed after each matching layer. Each attempt includes MPS preparation, forward and inverse angular Fourier transforms, transport and coordinate readout. Preparation and endpoint transforms are verified by native-gate replay; propagation uses the resulting signed amplitudes and equivalent matching-block rotations. The deterministic bias includes the polar-grid MPS approximation and is measured against each input’s own continuum ensemble (Supplementary Note 12).
Field error and shot estimation
Each accepted radial shot measures radius in the computational basis and the two phase qubits in Pauli X X X and Y Y Y bases. The outcomes form Y = ( r s x , r s y , r 2 ) T \mathbf{Y}=(r s_x,r s_y,r^2)^{\mathsf{T}} Y = ( r s x , r s y , r 2 ) T , where r r r is the measured radius and s x , s y ∈ { − 1 , + 1 } s_x,s_y\in\{-1,+1\} s x , s y ∈ { − 1 , + 1 } are the phase outcomes. Using outcomes from the same shot preserves the correlations required by equation (4) .
Let b b b be deterministic relative field bias, Σ \boldsymbol{\Sigma} Σ the covariance of Y \mathbf{Y} Y , A \mathbf{A} A the linear map from moments to the field, W \mathbf{W} W the spatial norm weight, u ‾ r e f \overline{\mathbf{u}}_{\mathrm{ref}} u ref the continuum field reference and S S S the number of successful independent shots. The mean squared relative field error is
E [ ε f i e l d 2 ] = b 2 + v S , v = tr ( A T W A Σ ) ∥ u ‾ r e f ∥ W 2 . (7) \mathbb{E}\left[\varepsilon_{\mathrm{field}}^2\right]=b^2+\frac{v}{S},\qquad v=\frac{\operatorname{tr}\left(\mathbf{A}^{\mathsf{T}}\mathbf{W}\mathbf{A}\boldsymbol{\Sigma}\right)}{\left\lVert\overline{\mathbf{u}}_{\mathrm{ref}}\right\rVert_{\mathbf{W}}^2}. \tag*{(7)} E [ ε field 2 ] = b 2 + S v , v = ∥ u ref ∥ W 2 tr ( A T WA Σ ) . (7)
Here v v v is the single-shot field variance coefficient. The reported relative RMS error is the square root of the mean squared relative error in equation (7) , averaged over independent repetitions of the complete S S S -shot estimate. Bias is evaluated from exact circuit expectations against the matched continuum field, and covariance from the nominal joint output distribution. We use the uniform sampled-field norm, so W \mathbf{W} W is the identity and a common cell-area factor cancels. The channel-error allowance is converted to a total-variation bound for the measured distribution, accounting for conditioning on success where applicable. Observable and product ranges over the finite grid then bound changes in first and second moments, giving effective bias and variance bounds b e f f , v e f f b_{\mathrm{eff}},v_{\mathrm{eff}} b eff , v eff . For target ε \varepsilon ε , the required accepted-shot count is
S = ceil ( v e f f ε 2 − b e f f 2 ) , b e f f < ε . (8) S=\operatorname{ceil}\left(\frac{v_{\mathrm{eff}}}{\varepsilon^2-b_{\mathrm{eff}}^2}\right),\qquad b_{\mathrm{eff}}<\varepsilon. \tag*{(8)} S = ceil ( ε 2 − b eff 2 v eff ) , b eff < ε . (8)
We use the smallest integer satisfying this bound at the final time. If the effective bias bound meets or exceeds the target, this shot rule has no feasible solution. Supplementary Notes 7 and 8 derive the covariance and total-variation bounds and separate deterministic, sampling and implementation contributions. This criterion bounds RMS error rather than a fixed-confidence event. The implementation-error allowance is separate from the field tolerance.
Polynomial tests apply the same relative RMS criterion to the Euclidean moment vector. Scalar position samples supply the first and second moments; coupled samples supply x x x , y y y , x 2 x^2 x 2 , y 2 y^2 y 2 and x y xy x y . Figure 5 a uses a common shot count to compare inputs, whereas Figure 5 c chooses the smallest sufficient count under the 1% target and implementation-error allowance. Its interior coupled-cubic circuit uses a 32 × 32 32\times32 32 × 32 grid, 64 steps and 11 logical qubits. Input-specific references, moment covariances and sampling distributions are in Supplementary Note 5. Resource estimates use independent shots; coherent amplitude estimation requires additional preparation–propagation and inverse calls and reversible observable evaluation (Supplementary Note 7).[5 , 49 ]
Force observables are fitted polynomials of modal coordinates consistent with reflection symmetry.[32 ] Calibration uses FEM-projected coefficients and cylinder surface forces over 10 ≤ t ∗ ≤ 60 10\leq t^*\leq60 10 ≤ t ∗ ≤ 60 , including the plotted interval; Figure 2 f therefore assesses in-window reconstruction. At each of 34 saved states, 815 additional ideal shots are divided between XX, XY, YX and YY settings as 204, 204, 204 and 203, with the first letter labeling the higher-index phase qubit. Equal-weight setting-specific estimators recover the radius-weighted first and second phase harmonics, with rigid phase restored. These measurements use Clifford basis changes and are separate from the velocity-field shot budget. No 1% target is imposed on forces. Force RMS errors combine deterministic discrepancy against FEM and analytic sampling variance, normalized by the corresponding FEM RMS force. Supplementary Note 6 and Source Data give the maps, covariance, fixed-seed realization and standard errors.
Fault-tolerant resource estimation
Arbitrary rotations use the logarithmic Clifford+T precision model of Ross and Selinger.[36 ] Clifford basis changes map transport rotations to R Z R_Z R Z synthesis. We compare compact, intermediate and fast surface-code data blocks with one or two distillation levels.[37 , 38 ] Gidney cultivation calibrates factory inputs; preparation takes thirteen QEC rounds plus a reaction delay, rounded up to an integer number of first-level logical cycles.[50 ] Each round is 1 μ \mu μ s and reaction time is 10 μ \mu μ s. A distance-d d d patch occupies 2 ( d + 1 ) 2 2(d+1)^2 2 ( d + 1 ) 2 physical qubits. We use pitch d + 1 d+1 d + 1 for factory and connector areas; logical errors and timing use the actual code distance d d d .[51 ] Supplementary Notes 9 and 10 specify buffers, bus area, cultivation calibration, factory acceptance and stage-local storage.
The search varies four design choices. Code distances are chosen separately for the computation (core) and each factory stage; larger distances suppress logical errors but require more physical qubits. Each additional distillation stage further purifies its input T states. Parallel factory units increase the T-state supply rate. More T gates per arbitrary rotation improve its approximation. For fixed-length circuits, we first search odd core distances 13–49 and 45–200 T gates per rotation. Cultivation data give the error and preparation success probabilities of the T states entering the first distillation stage. Factory output rates follow from stage durations, acceptance probabilities and parallel unit counts. We test final-stage counts from one until mean supply meets core demand; preceding stages are sized to meet the next stage’s demand. We filter configurations by physical-qubit count, effective core error per consumed T state, factory-output error and T-state consumption interval. A configuration is discarded if another is no worse in all four and better in one. We then jointly select a retained configuration and rotation-synthesis precision under the per-attempt error budget (Figure 3 c). The fastest-design search has no physical-qubit cap and extends distance choices beyond the initial range. We skip choices whose attempt-time lower bound exceeds the fastest feasible time already found. Capacity calculations in Supplementary Fig. S10 instead cap physical qubits at 10 5 10^5 1 0 5 and 10 6 10^6 1 0 6 , and limit expected time per accepted shot to one day divided by 815. Supplementary Note 9 details the search and candidate-exclusion rules; Notes 10 and 13 specify comparison and capacity settings.
Before conditioning on circuit success, the search combines logical faults, factory-output faults and rotation-synthesis error under a modelled normalized-diamond-distance budget η \eta η of 10 − 4.52 10^{-4.52} 1 0 − 4.52 . A circuit consumes N T = N T n a t i v e + n T N R Z N_T=N_T^{native}+n_TN_{R_Z} N T = N T na t i v e + n T N R Z , where N T n a t i v e N_T^{native} N T na t i v e counts native T/T† gates, N R Z N_{R_Z} N R Z arbitrary rotations and n T n_T n T synthesized T gates per rotation. We evaluate
E m o d e l = 1 − ( 1 − E c o r e ) ( 1 − e M S F ) N T + N R Z 2 − n T / 3 (9) E_{\mathrm{model}}=1-\left(1-E_{\mathrm{core}}\right)\left(1-e_{\mathrm{MSF}}\right)^{N_T}+N_{R_Z}2^{-n_T/3} \tag*{(9)} E model = 1 − ( 1 − E core ) ( 1 − e MSF ) N T + N R Z 2 − n T /3 (9)
and require E m o d e l ≤ η E_{\mathrm{model}\leq\eta} E model ≤ η . Here E c o r e E_{\mathrm{core}} E core is the probability of at least one core logical fault per attempt and e M S F e_{\mathrm{MSF}} e MSF the error per delivered T state. The synthesis contribution uses the leading-order surrogate 2 − n T / 3 2^{-n_T/3} 2 − n T /3 ; a certified budget would require each rotation’s actual synthesis error (Supplementary Note 8). The per-round logical failure model is
p L ( p p h y s , d ) = 0.1 ( p p h y s p t h ) ( d + 1 ) / 2 (10) p_L(p_{\mathrm{phys}},d)=0.1\left(\frac{p_{\mathrm{phys}}}{p_{\mathrm{th}}}\right)^{(d+1)/2} \tag*{(10)} p L ( p phys , d ) = 0.1 ( p th p phys ) ( d + 1 ) /2 (10)
with p t h = 10 − 2 p_{\mathrm{th}}=10^{-2} p th = 1 0 − 2 .[37 , 38 ] The physical error probability p p h y s p_{\mathrm{phys}} p phys is treated as an effective stochastic rate, assuming coherent errors are rendered effectively stochastic by noise tailoring. The reference value is 5 × 10 − 4 5\times10^{-4} 5 × 1 0 − 4 and the sweep spans 2 × 10 − 4 2\times10^{-4} 2 × 1 0 − 4 –9 × 10 − 4 9\times10^{-4} 9 × 1 0 − 4 . Memory exposure uses the full T-state consumption interval, including waits (Supplementary Note 9).
The nominal core interval τ c o r e \tau_{\mathrm{core}} τ core is the larger of the reaction time and the scheduled period: fast, intermediate and compact blocks require one, five and nine code-distance-scaled QEC rounds per T state. The mean factory interval τ M S F \tau_{\mathrm{MSF}} τ MSF is the inverse final-level output rate, giving τ T = max ( τ c o r e , τ M S F ) \tau_T=\max(\tau_{\mathrm{core}},\tau_{\mathrm{MSF}}) τ T = max ( τ core , τ MSF ) . Each architecture reevaluates core-memory exposure when supply is slower than demand. For the supply-policy sensitivities, the waiting-allowed control fixes one final-level unit, while supply enforced optimizes the required count; each independently selects distances and synthesis precision. With Clifford gates commuted into the final measurement bases,[37 ] S S S accepted shots and nominal circuit success probability P s P_{\mathrm{s}} P s give
t a t t e m p t = N T τ T , t t a s k = S P s t a t t e m p t . (11) t_{\mathrm{attempt}} = N_{\mathrm{T}} \tau_{\mathrm{T}},\quad t_{\mathrm{task}} = \frac{S}{P_{\mathrm{s}}}t_{\mathrm{attempt}} . \tag*{(11)} t attempt = N T τ T , t task = P s S t attempt . (11)
Per-attempt time covers state preparation, coherent propagation and, for three-coordinate circuits, the Fourier transforms between the angular grid and its Fourier modes. Total task time includes failed attempts when collecting S S S accepted shots. Excluded are classical model fitting and circuit compilation before execution; the time before a factory’s first T-state output; additional qubit-reset or readout delays; and classical data-transfer latency. Times use the T-state consumption interval, including supply waits. Detailed chip routing and concurrent scheduling are outside the model. Alternative factory models and error-budget sensitivities are documented in Supplementary Notes 9, 10 and 14.
The Table 1 bounds fix drift degree, dimension, harmonic content and stencil order where applicable. The LCHS bound assumes efficient block encoding, controlled selection and weight preparation (Supplementary Notes 3, 11 and 12). For BE–KvN, the coordinate-amplitude block encoding of Guseynov et al. [35 ] is reconstructed for the same prepared 64-cell input, generator, interval and final observables. We compensate for the oracle’s coordinate scaling in the drift polynomial and block normalization. Degree-three and degree-two transformations encode radial transport and squared-radius phase coupling. We approximate the exponential of the joint generator over the full interval. Gate counts use the outer circuit template with every undetermined phase counted as an arbitrary rotation. Bias and shots use the certified polynomial on the finite generator spectrum. Supplementary Note 11 specifies the normalization, polynomial accuracy, acceptance and complete inventory. This reconstruction does not establish a lower bound for other block encodings of the same generator.
LCHS–KvN uses the published sine-based drift, a uniform superposition over 128 cells and ten steps of size 0.1.[17 ] Multi-controlled X gates use one shared clean work qubit, initialized to zero and restored after each gate, with the Khattar–Gidney construction.[53 ] Other controlled gates use the Zindorf–Bose decomposition. We retain the extra phase acquired only when the controls select the gate, which is needed to implement its full unitary.[54 ] For each hardware configuration, we choose the smallest integer number of T gates per arbitrary rotation that meets the implementation-error budget. The budget also includes the reported numerical error from converting the source rotation angles for synthesis. The published algorithmic success probability is approximately 0.88; this factor accounts for repeated attempts in the task time. Its task has no matched field-error target, so total time uses 815 accepted shots for comparison. General polynomial drift encoded through its sine signal requires inverse-sine approximations,[55 ] while the improved LCHS kernel is demonstrated on advection–diffusion.[56 ] Supplementary Note 10 specifies the reference and hardware compilation.
IBM hardware experiment
The hardware circuits combine bond-two MPS preparation with four- or six-step radial propagation on four qubits and 16 cells, omitting phase qubits (Extended Data Fig. 1). We analyse all 16,000 recorded shots per circuit for the four-step dataset and all 32,000 for the six-step dataset, over t ∗ = 40 t^*=40 t ∗ = 40 –49,8291. Reference circuits preserve logical preparation and propagation structure but set every Walsh rotation angle to zero. Supplementary Note 15 and Source Data give device, native-gate compilation, input verification, full-count calibration and the 32-cell control.
An effective global depolarizing model is fitted separately for each step count. Least squares fits the reference distribution to a mixture of the ideal MPS input and uniform probabilities, yielding a depolarizing contrast. The mitigated value of each observable is its ideal input value plus the transport–reference difference divided by this contrast (Supplementary equation (97) ), assuming equal effective noise in the two circuits. The correction preserves normalization but does not enforce nonnegative probabilities. Bootstrap resampling includes reference and transport counts; error bars quantify counting uncertainty and exclude residual model bias. These radial diagnostics are separate from the phase-readout and field-reconstruction task.
Use of generative AI
The authors used Claude (Anthropic) and ChatGPT (OpenAI) for assistance with language editing of the manuscript and with code development. The authors reviewed and verified all text, code and results, and take full responsibility for the content.
Data availability
The data supporting this study, including Supplementary Data 1–8 and the Source Data files cited in the text, are available from the corresponding author on request. These include flow snapshots, fitted model parameters, circuit inventories, IBM hardware counts and metadata, and numerical source data for the figures. Upon publication, the data will be deposited in a public repository with a permanent identifier.
Code availability
The custom code for model identification, circuit construction and verification, resource estimation and figure generation is available from the corresponding author on request, together with numerical inputs, dependencies and reproduction instructions. Upon publication, the code will be deposited in a public repository with a permanent identifier.
Acknowledgements
This work was supported by Korea National Research Foundation (NRF) grant No. NRF-RS-2026-25612745, ICT R&D program of IITP-2023-2021-0-01810, RS-2023-00225385, and AFOSR grant FA2386-22-1-4052. The work was partly supported by the Cleveland Clinic Quantum Start-up Catalyzer Program and Amazon Web Services. Authors acknowledge the use of IBM Quantum services for this work through the Cleveland Clinic’s Quantum Start-up Catalyzer Program. The views expressed are those of the authors, and do not reflect the official policy or position of IBM or the IBM Quantum team. This work was also supported by the National Quantum Laboratory at Maryland (QLab). We acknowledge QLab for providing quantum computing resources on IonQ devices for this work. The views expressed are those of the authors, and do not reflect the official policy or position of IonQ.
Author contributions
H.N. devised the overall framework for early fault-tolerant flow prediction, combining low-degree models identified with SINDy, Walsh synthesis of Koopman–von Neumann quantum circuits and velocity-field reconstruction from joint radius–phase readout. H.N. and U.O. worked on resource estimation and quantum circuit development. H.N. and S.B. worked on reduced-order modelling and computational fluid dynamics simulations. D.A. proposed investigating quantum algorithms for nonlinear dynamics and acquired funding. D.A. performed the IBM hardware experiments. H.N. wrote the manuscript. All authors discussed the results and commented on and revised the manuscript.
Competing interests
D. Ahn is a co-founder of Singularity Quantum Inc. and holds an equity interest in the company. The other authors declare no competing interests.
Figure 1 (PDF p. 23). Nonlinear radial KvN transport on IBM hardware. Experiments on ibm_cleveland use four qubits and 16 cells, with matrix product state preparation of bond dimension two. Four steps (a,b) use all 16,000 recorded shots per circuit; six (c,d) use all 32,000. a,c, Probability at normalized radius r ≥ 7 / 8 r \ge7/8 r ≥ 7/8 , an auxiliary diagnostic near the attracting state r = 1 r = 1 r = 1 . The last two cells start at 7 / 8 7/8 7/8 and have centres 29 / 32 29/32 29/32 and 31 / 32 31/32 31/32 . b,d, Mean radius. Dotted lines mark 31 / 32 31/32 31/32 , the maximum mean for nonnegative cell probabilities. Noise mitigation can yield estimates above this bound. Lines are ideal circuits, circles measured values, squares noise-mitigated estimates and dashed lines continuum means. Bars show bootstrap standard deviations (1,000 resamples) after mitigation and counting standard errors before mitigation; systematic bias is excluded. See Supplementary Note 15.
The Supplementary Information follows the computational pipeline: physical reduction and benchmark definition (Note 1), initial ensembles and state preparation (Note 2), transport circuits and validation (Notes 3–5), joint radius–phase readout (Note 6), sampling and complete prediction cost (Note 7), protected execution and resource optimization (Notes 8–10), the matched block-encoding circuit, BE–KvN (Note 11), the Fourier–Walsh circuit for the three-coordinate wake (Note 12), step capacity, factory supply and sensitivity analyses (Notes 13 and 14), and the IBM radial hardware experiment (Note 15). Native equations, circuit checks, continuum references and resource inputs accompany the final results. Bold symbols denote vectors and matrices; scalar components are italic. Numerical values are rounded for presentation; full-precision values are retained in the Source Data and calculation inputs.
Supplementary Note 1 Physical reduction and the frozen benchmark
Spatial representation and nonlinear model calibration
The Reynolds number is Re = U ∞ D / ν \operatorname{Re}=U_{\infty}D/\nu Re = U ∞ D / ν , where U ∞ U_{\infty} U ∞ is inflow speed, D D D cylinder diameter and ν \nu ν kinematic viscosity.
The sampled two-component velocity field uses an independently solved unstable steady field and three modes,
u ( t ∗ ) ≃ u s t + ϕ 1 a 1 ( t ∗ ) + ϕ 2 a 2 ( t ∗ ) + ϕ 3 a 3 ( t ∗ ) . (12) \mathbf{u}(t^{*}) \simeq\mathbf{u}_{\mathrm{st}}+\boldsymbol{\phi}_{1}a_{1}(t^{*})+\boldsymbol{\phi}_{2}a_{2}(t^{*})+\boldsymbol{\phi}_{3}a_{3}(t^{*}). \tag*{(12)} u ( t ∗ ) ≃ u st + ϕ 1 a 1 ( t ∗ ) + ϕ 2 a 2 ( t ∗ ) + ϕ 3 a 3 ( t ∗ ) . (12)
The first two coordinates form the oscillatory pair, and the third is a shift coordinate. POD and constrained polynomial regression provide a physically interpretable reduction[1 , 2 , 3 ] .
Let x x x , y y y be spatial coordinates, t ∗ t^{*} t ∗ nondimensional time and u = ( u , v ) T \mathbf{u}=(u,v)^{T} u = ( u , v ) T . Reflection about y = 0 y=0 y = 0 is ( R u ) ( x , y , t ∗ ) = ( u ( x , − y , t ∗ ) , − v ( x , − y , t ∗ ) ) T (\mathcal{R}\mathbf{u})(x,y,t^{*})=(u(x,-y,t^{*}),-v(x,-y,t^{*}))^{T} ( R u ) ( x , y , t ∗ ) = ( u ( x , − y , t ∗ ) , − v ( x , − y , t ∗ ) ) T . POD uses the antisymmetric field u a = ( u − R u ) / 2 \mathbf{u}_{a}=(\mathbf{u}-\mathcal{R}\mathbf{u})/2 u a = ( u − R u ) /2 . This isolates shedding oscillations from the symmetric steady and mean-flow components. Full-field reconstruction retains the complementary symmetric component.
For Re = 100 \operatorname{Re}=100 Re = 100 , the uncentred, unweighted two-component POD pair is covariance-normalized with positive phase orientation. Inverse-transpose spatial modes preserve its span and reconstruction but are generally not orthonormal. Snapshot selection, mesh, solid masking and frozen normalization are recorded in the Source Data methods record.
The shift mode uses the trapezoidal mean over two complete cycles, bounded by interpolated phase crossings. The steady reference uses the same mesh and boundary conditions; crossing times are recorded in the Source Data. Define
d = u ˉ − u s t , d ⊥ = ( I − Ψ 0 Ψ 0 T ) d , ϕ 3 = d ⊥ / ∥ d ⊥ ∥ 2 . (13) \begin{aligned}
\mathbf{d} &= \bar{\mathbf{u}}-\mathbf{u}_{\mathrm{st}},\\
\mathbf{d}_{\perp} &= \left(\mathbf{I}-\mathbf{\Psi}_{0}\mathbf{\Psi}_{0}^{T}\right)\mathbf{d},\\
\boldsymbol{\phi}_{3} &= \mathbf{d}_{\perp}/\lVert\mathbf{d}_{\perp}\rVert_{2}. \tag*{(13)}
\end{aligned} d d ⊥ ϕ 3 = u ˉ − u st , = ( I − Ψ 0 Ψ 0 T ) d , = d ⊥ / ∥ d ⊥ ∥ 2 . (13)
Here I \mathbf{I} I is the sampled-space identity and ∥ ⋅ ∥ 2 \lVert\cdot\rVert_{2} ∥ ⋅ ∥ 2 the Euclidean norm. Define Ψ = [ ϕ 1 , ϕ 2 ] \mathbf{\Psi}=[\boldsymbol{\phi}_{1},\boldsymbol{\phi}_{2}] Ψ = [ ϕ 1 , ϕ 2 ] , and let Ψ 0 \mathbf{\Psi}_{0} Ψ 0 have orthonormal columns spanning the same column space. Then Ψ 0 Ψ 0 T \mathbf{\Psi}_{0}\mathbf{\Psi}_{0}^{T} Ψ 0 Ψ 0 T is the orthogonal projector used in equation (S2); Ψ Ψ T \mathbf{\Psi}\mathbf{\Psi}^{T} Ψ Ψ T need not be a projector.
With Ψ a = ( ϕ 1 , ϕ 2 , ϕ 3 ) \mathbf{\Psi}_a=(\boldsymbol{\phi}_1,\boldsymbol{\phi}_2,\boldsymbol{\phi}_3) Ψ a = ( ϕ 1 , ϕ 2 , ϕ 3 ) and a = ( a 1 , a 2 , a 3 ) T \mathbf{a}=(a_1,a_2,a_3)^{T} a = ( a 1 , a 2 , a 3 ) T , project the original full field by solving the Gram system
G = Ψ a T Ψ a , \mathbf{G}=\mathbf{\Psi}_a^{T}\mathbf{\Psi}_a, G = Ψ a T Ψ a ,
G a ( t ∗ ) = Ψ a T [ u ( t ∗ ) − u s t ] , \mathbf{G}\mathbf{a}(t^*)=\mathbf{\Psi}_a^{T}\left[\mathbf{u}(t^*)-\mathbf{u}_{\mathrm{st}}\right], Ga ( t ∗ ) = Ψ a T [ u ( t ∗ ) − u st ] ,
u R O M ( t ∗ ) = u s t + Ψ a a ( t ∗ ) . (14) \mathbf{u}_{\mathrm{ROM}}(t^*)=\mathbf{u}_{\mathrm{st}}+\mathbf{\Psi}_a\mathbf{a}(t^*). \tag*{(14)} u ROM ( t ∗ ) = u st + Ψ a a ( t ∗ ) . (14)
Polynomial regression and shift closure
Sparse identification of nonlinear dynamics (SINDy) fits three linear and six quadratic monomials, excluding a constant. Reflection parity allows 13 coefficients; ridge parameter 10 − 4 10^{-4} 1 0 − 4 and sequential thresholding at 0.03 in RMS-normalized coordinate units retain 12. The threshold removes one allowed cross term from the shift equation; the other structural zeros are imposed by degree and symmetry. The regression selects coefficients within the prescribed quadratic library; the library degree remains a modelling and circuit-design choice.
A diagnostic using case A alone compares libraries of maximum degree 1–5 over thirteen thresholds and four contiguous 73-state validation blocks. Derivative stencils stay wholly within their training or validation partitions. Within the quadratic library, threshold 0.03 minimizes mean validation projected-field error among tested values: 5.62% for 12 coefficients, versus 6.36% for 13 and 8.85% for seven. The cubic fits with 18 coefficients at threshold 0.03 and 26 at threshold 0.01 give 3.27% and 2.85%, respectively. On a common illustrative Cartesian grid with 64 cells per coordinate, summed Walsh rotation-support upper bounds for one interior pair layer per coordinate are 226 for the adopted quadratic model, 671 for the 18-coefficient cubic fit and 1,297 for the 26-coefficient fit. These bounds exclude step count, preparation, synthesis, shots and error correction; they are not counts for the polar wake circuit or physical-resource estimates. The quadratic cap is an accuracy–synthesis-cost trade-off, not an accuracy optimum or automatically identified degree. The spatial basis remains fixed, so its learning is not validated; these errors differ from the 1% quantum-versus-ROM target. Full candidate scores and coefficients are supplied in Source Data. Figure 2 b of the main text displays the diagnostic.
The same three fixed fits are evaluated on all five non-training cases B–F at saved times 40.08–49.68 and 20.16–90, initialized at the first projected CFD state of each interval. The initialized state is excluded from scoring, as in the case-A validation. No candidate is selected or refitted from these scores. Short-window mean projected-field errors are 1.50%, 0.373% and 0.197% for 12, 18 and 26 coefficients; extended-window means are 11.4%, 7.70% and 0.651%. Corresponding mean full-field errors are 7.23%, 7.05%, 7.04% and 13.68%, 11.10%, 7.38%. These are arithmetic means over all five cases. Case A alone supplies training data. Cases B–F use five different initial perturbations of the CFD velocity field; their streamfunction amplitudes, centres and widths are listed in Source Data. The projected-field denominator is the reconstructed CFD projection including the steady background; full-field errors include the projection residual and use the full CFD denominator. The 5.62% case-A validation score excludes projection error, whereas the 6.42% FEM comparison at the final time and 6.77–9.50% radial audit include it. Source Data retain every rollout and error definition.
Modes and model selection use case A only, over the nominal 10–80 calibration interval with uniform growth weighting. Cases B–F provide additional initial conditions; the historical viewing roles remain in Source Data. The model-selection score windows are recorded in the Source Data methods record.
The fit uses 292 projected states. RMS-normalized coordinates, unit-norm feature columns and unsmoothed fourth-order centred derivatives define the regression; coefficients are restored to original units. Sampling intervals and retained derivative rows are recorded in the Source Data methods record.
For a ⊥ = ( a 1 , a 2 ) T \mathbf{a}_{\perp}=(a_1,a_2)^{T} a ⊥ = ( a 1 , a 2 ) T , the fitted quadratic model has the structure
d a ⊥ d t ∗ = ( L + a 3 B d y n ) a ⊥ , d a 3 d t ∗ = λ 3 a 3 + q 11 a 1 2 + q 12 a 1 a 2 + q 22 a 2 2 + q 33 a 3 2 . (15) \begin{aligned}
\frac{\mathrm{d}\mathbf{a}_{\perp}}{\mathrm{d}t^{*}}&=(\mathbf{L}+a_3\mathbf{B}_{dyn})\mathbf{a}_{\perp},\\
\frac{\mathrm{d}a_3}{\mathrm{d}t^{*}}&=\lambda_3a_3+q_{11}a_1^2+q_{12}a_1a_2+q_{22}a_2^2+q_{33}a_3^2. \tag*{(15)}
\end{aligned} d t ∗ d a ⊥ d t ∗ d a 3 = ( L + a 3 B d y n ) a ⊥ , = λ 3 a 3 + q 11 a 1 2 + q 12 a 1 a 2 + q 22 a 2 2 + q 33 a 3 2 . (15)
Here L \mathbf{L} L and B d y n \mathbf{B}_{dyn} B d y n are real 2 × 2 2\times2 2 × 2 matrices, and λ 3 , q 11 , q 12 , q 22 , q 33 \lambda_3,q_{11},q_{12},q_{22},q_{33} λ 3 , q 11 , q 12 , q 22 , q 33 are real scalars. Full-precision coefficients C C C and exponents p p p , the reduced parameters and the initial phase are given in the Source Data. The three-mode comparator evolves all three coordinates as a coupled dynamical system without the shift closure. The sparse fit sets the planar cross-term coefficient in the third equation to zero.
Let ϱ = ( a 1 2 + a 2 2 ) 1 / 2 \varrho=(a_1^2+a_2^2)^{1/2} ϱ = ( a 1 2 + a 2 2 ) 1/2 and θ = atan2 ( a 2 , a 1 ) \theta=\operatorname{atan2}(a_2,a_1) θ = atan2 ( a 2 , a 1 ) . The separate shift closure is a 3 ≃ c 0 + c 2 ϱ 2 a_3\simeq c_0+c_2\varrho^2 a 3 ≃ c 0 + c 2 ϱ 2 , with c 0 = 0 c_0=0 c 0 = 0 fixed. Its coefficient is fitted by least squares over all 292 training states:
c 2 = ∑ i ϱ ( t i ∗ ) 2 a 3 ( t i ∗ ) ∑ i ϱ ( t i ∗ ) 4 . (16) c_2=\frac{\sum_i\varrho(t_i^{*})^2a_3(t_i^{*})}{\sum_i\varrho(t_i^{*})^4}. \tag*{(16)} c 2 = ∑ i ϱ ( t i ∗ ) 4 ∑ i ϱ ( t i ∗ ) 2 a 3 ( t i ∗ ) . (16)
Brunton et al.[2 ] discuss this parabolic relation for rapidly relaxing shift dynamics. We assess closure error separately. It is not obtained from a ˙ 3 = 0 \dot{a}_3=0 a ˙ 3 = 0 and does not establish an invariant manifold.
Uniform angular averaging ⟨ g ⟩ θ = ( 2 π ) − 1 ∫ 0 2 π g ( θ ) d θ \langle g\rangle_{\theta}=(2\pi)^{-1}\int_{0}^{2\pi}g(\theta)\,\mathrm{d}\theta ⟨ g ⟩ θ = ( 2 π ) − 1 ∫ 0 2 π g ( θ ) d θ at fixed ϱ , a 3 \varrho,a_3 ϱ , a 3 removes second-angle harmonics. Applying it with a 3 ≃ c 0 + c 2 ϱ 2 a_3\simeq c_0+c_2\varrho^2 a 3 ≃ c 0 + c 2 ϱ 2 defines the linear radial growth rate μ \mu μ , cubic radial saturation coefficient β \beta β , radius-independent angular rate ω \omega ω and unscaled squared-radius angular coupling γ r a w \gamma_{\mathrm{raw}} γ raw by
μ = 1 2 tr ( L + c 0 B d y n ) , β = − 1 2 c 2 tr B d y n , ω = 1 2 [ ( L + c 0 B d y n ) 21 − ( L + c 0 B d y n ) 12 ] , γ r a w = 1 2 c 2 [ ( B d y n ) 21 − ( B d y n ) 12 ] . (17) \begin{aligned}
\mu&=\frac{1}{2}\operatorname{tr}(\mathbf{L}+c_0\mathbf{B}_{dyn}),\\
\beta&=-\frac{1}{2}c_2\operatorname{tr}\mathbf{B}_{dyn},\\
\omega&=\frac{1}{2}\left[(\mathbf{L}+c_0\mathbf{B}_{dyn})_{21}-(\mathbf{L}+c_0\mathbf{B}_{dyn})_{12}\right],\\
\gamma_{\mathrm{raw}}&=\frac{1}{2}c_2\left[(\mathbf{B}_{dyn})_{21}-(\mathbf{B}_{dyn})_{12}\right]. \tag*{(17)}
\end{aligned} μ β ω γ raw = 2 1 tr ( L + c 0 B d y n ) , = − 2 1 c 2 tr B d y n , = 2 1 [ ( L + c 0 B d y n ) 21 − ( L + c 0 B d y n ) 12 ] , = 2 1 c 2 [ ( B d y n ) 21 − ( B d y n ) 12 ] . (17)
Define R = μ / β R=\sqrt{\mu/\beta} R = μ / β and r = ϱ / R r=\varrho/R r = ϱ / R . Figure 2 uses modal coordinates x = a 1 / R x=a_1/R x = a 1 / R , y = a 2 / R y=a_2/R y = a 2 / R and z = ( a 3 − c 0 ) / ( c 2 R 2 ) z=(a_3-c_0)/(c_2R^2) z = ( a 3 − c 0 ) / ( c 2 R 2 ) , so r 2 = x 2 + y 2 r^2=x^2+y^2 r 2 = x 2 + y 2 and z ≃ r 2 z\simeq r^2 z ≃ r 2 . These are modal, not spatial, coordinates; a j a_j a j retain original units. To follow the late nonlinear approach to saturated shedding, take the Re = 100 \operatorname{Re}=100 Re = 100 benchmark start at 40 as the time origin and use τ = μ ( t ∗ − 40 ) \tau=\mu(t^{*}-40) τ = μ ( t ∗ − 40 ) .
The spatial modes labelled in Figure 1 are ϕ x = R ϕ 1 \boldsymbol{\phi}_x=R\boldsymbol{\phi}_1 ϕ x = R ϕ 1 , ϕ y = R ϕ 2 \boldsymbol{\phi}_y=R\boldsymbol{\phi}_2 ϕ y = R ϕ 2 and ϕ z = c 2 R 2 ϕ 3 \boldsymbol{\phi}_z=c_2R^2\boldsymbol{\phi}_3 ϕ z = c 2 R 2 ϕ 3 . Thus ϕ 3 \boldsymbol{\phi}_3 ϕ 3 denotes the unscaled shift mode; ϕ z \boldsymbol{\phi}_z ϕ z includes the coordinate normalization.
Figure 1. From wake dynamics to fault-tolerant circuit costs. a, Ensemble spreading near an unstable fixed point and concentration near a stable one. b, Proper orthogonal decomposition (POD) and sparse identification of nonlinear dynamics (SINDy) yield a three-mode polynomial model. c, Four Walsh coefficients represent eight rotation angles that vary linearly with the even-pair index on a 16-cell grid. d, Phase half steps flank the even–odd–even radial rotations. e, Radius and phase readouts yield ensemble-averaged modal coefficients for flow-field reconstruction with the steady field u s t u_{\mathrm{st}} u st and spatial modes φ x \varphi_x φ x , φ y \varphi_y φ y and φ z \varphi_z φ z . f, Circuit and shot counts, the error budget and hardware parameters determine the required number of physical qubits and the runtime.
d r d t ∗ = μ ( r − r 3 ) , d r d τ = r − r 3 , γ = γ r a w R 2 . (18) \frac{\mathrm{d}r}{\mathrm{d}t^{*}}=\mu(r-r^3),\qquad\frac{\mathrm{d}r}{\mathrm{d}\tau}=r-r^3,\qquad\gamma=\gamma_{\mathrm{raw}}R^2. \tag*{(18)} d t ∗ d r = μ ( r − r 3 ) , d τ d r = r − r 3 , γ = γ raw R 2 . (18)
d θ d t ∗ = ω + γ r 2 , d θ d τ = ω μ + γ μ r 2 . (19) \frac{\mathrm{d}\theta}{\mathrm{d}t^{*}}=\omega+\gamma r^2,\qquad\frac{\mathrm{d}\theta}{\mathrm{d}\tau}=\frac{\omega}{\mu}+\frac{\gamma}{\mu}r^2. \tag*{(19)} d t ∗ d θ = ω + γ r 2 , d τ d θ = μ ω + μ γ r 2 . (19)
Write θ = θ c + Φ \theta=\theta_c+\Phi θ = θ c + Φ , with known θ c = θ 0 + ω ( t ∗ − 40 ) \theta_c=\theta_0+\omega(t^*-40) θ c = θ 0 + ω ( t ∗ − 40 ) , Φ ( 40 ) = 0 \Phi(40)=0 Φ ( 40 ) = 0 and d Φ / d t ∗ = γ r 2 \mathrm{d}\Phi/\mathrm{d}t^*=\gamma r^2 d Φ/ d t ∗ = γ r 2 . We set θ 0 = atan2 ( a 2 , a 1 ) \theta_0=\operatorname{atan2}(a_2,a_1) θ 0 = atan2 ( a 2 , a 1 ) for all radii using the CFD-projected coefficients linearly interpolated to t ∗ = 40 t^*=40 t ∗ = 40 . Table S1 lists the frozen parameters.
Table S1. Frozen reduced-model parameters (rounded). Calculations use full-precision values.
Quantum reconstruction uses Ψ m T = ( φ 1 , φ 2 , c 2 φ 3 ) \mathbf{\Psi}_m^{T}=(\varphi_1,\varphi_2,c_2\varphi_3) Ψ m T = ( φ 1 , φ 2 , c 2 φ 3 ) , with the same steady reference and c 0 = 0 c_0=0 c 0 = 0 . The spatial-mode matrix Ψ m \mathbf{\Psi}_m Ψ m differs from scalar phase Φ \Phi Φ .
Let the radius–phase moment vector be m = ( m 1 , m 2 , m 3 ) T = ( E [ r cos Φ ] , E [ r sin Φ ] , E [ r 2 ] ) T \mathbf{m}=(m_1,m_2,m_3)^T=(\mathbb{E}[r\cos\Phi],\mathbb{E}[r\sin\Phi],\mathbb{E}[r^2])^T m = ( m 1 , m 2 , m 3 ) T = ( E [ r cos Φ ] , E [ r sin Φ ] , E [ r 2 ] ) T . The fixed map in equation (20) supplies the same coefficient rotation and scaling for field reconstruction.
In the rescaled spatial basis, the moment-to-coordinate map for the implemented zero-intercept closure is
M = ( R cos θ c − R sin θ c 0 R sin θ c R cos θ c 0 0 0 R 2 ) , u ‾ = u s t + Ψ m T M m . (20) \mathbf{M}=\begin{pmatrix}
R\cos\theta_c & -R\sin\theta_c & 0 \\
R\sin\theta_c & R\cos\theta_c & 0 \\
0 & 0 & R^2
\end{pmatrix},\qquad\overline{\mathbf{u}}=\mathbf{u}_{\mathrm{st}}+\mathbf{\Psi}_m^T\mathbf{M}\mathbf{m}. \tag*{(20)} M = R cos θ c R sin θ c 0 − R sin θ c R cos θ c 0 0 0 R 2 , u = u st + Ψ m T Mm . (20)
Independent flow calculation and physical errors
Six finite-element cases at Reynolds number 100 start from the same steady flow with separate initial perturbations. Taylor–Hood P2/P1 elements and a time step of 0.02 are fixed across cases; mesh, boundary and perturbation settings are supplied in Source Data. The frozen spatial modes and radial/phase model are not refitted.
Physical field error is the aggregate two-component velocity residual norm divided by the reference-field norm; projection error uses the same metric. The nominal 40–49.829 and 20–90 windows use saved states at 40.08–49.68 and 20.16–90, respectively. Each forecast starts from the first saved state in its window.
Supplementary Data 1 | Short window physical forecast audit (CSV).
The quantum 1% target concerns the continuum Radial ensemble, separately from physical forecast error. Source Data provide both window CSVs and the calculation code. Extended windows in Supplementary Supplementary Figure S1 delimit longer predictions.
Supplementary Figure S1. Physical forecast scope. a, Short window with nominal start 40 and final time 49.829, evaluated at 41 saved states from 40.08 to 49.68. b, Extended nominal window 20–90, evaluated at 292 saved states from 20.16 to 90. Each forecast starts from the first saved state in its window. The same frozen cubic radial/phase model and three-mode projection are compared for cases A–F. Larger extended errors delimit physical prediction scope and are distinct from quantum ensemble discretization error.
Figure 2 reproduction and Reynolds-number comparison
The Re=100 start is fixed at time 40 without start-time optimization. Interpolating modal coordinates before computing the normalized squared radius gives 0.81307 at this start and 0.95806 at the end. The interval therefore follows the late nonlinear approach to saturation. It overlaps temporal calibration and uses later spatial data, so it is a reproduction window rather than an independent forecast (Figure 2 ).
In the comparison regions marked at t ∗ = 40.30 t^*=40.30 t ∗ = 40.30 (Figure 2 a), the reconstructed positive and negative extrema have slightly greater magnitudes and lie farther upstream than the FEM extrema. In the projection onto squared radius and shift (Figure 2 d), FEM z z z starts below r 2 r^2 r 2 and approaches it as the wake grows, with a largest absolute difference of 0.047 over the circuit interval; the Walsh–KvN points lie on the diagonal because the radial model imposes z = r 2 z=r^2 z = r 2 .
Finite-shot moment estimates for Reynolds number 100 are given in Supplementary Note 7. Supplementary Data 2 compares Reynolds numbers 100, 200 and 300 using a separate steady reference, spatial basis and fitted model for each case. Each case uses 64 positive cells and 33 Walsh steps. Full Walsh gate checks cover all cases.
Every interval lasts 9.8291 nondimensional units (Supplementary Data 2). Re = 100 \operatorname{Re}=100 Re = 100 starts at t ∗ = 40 t^*=40 t ∗ = 40 ; Re = 200 \operatorname{Re}=200 Re = 200 and 300 start at the first saved upward crossing of r 2 r^2 r 2 through 0.8 after t ∗ = 10 t^*=10 t ∗ = 10 . Here r r r is normalized by each fitted saturated amplitude, so the windows compare similar stages of nonlinear growth.
DOP853 integrates all three coupled quadratic equations with tolerances 10 − 11 10^{-11} 1 0 − 11 relative and 10 − 13 10^{-13} 1 0 − 13 absolute. fig:2e compares the three-mode ROM field (3-mode) with FEM (grey), the reconstruction after shift closure and angular averaging (Radial) with 3-mode (red), and the circuit reconstruction (Circuit) with the matched continuum Radial ensemble field (blue). Grey/red single-flow errors use the FEM field norm; blue noiseless ensemble bias uses the matched continuum Radial field norm.
Closure RMS measures the dimensionless deviation of projected z z z from r 2 r^2 r 2 , over 34 states including initialization. Supplementary Data 2 also gives the signed mean and maximum absolute residual over those states. The field error uses 33 later states and is 100 ∑ ∥ u − u C F D ∥ 2 / ∑ ∥ u C F D ∥ 2 100\sqrt{\sum\lVert\mathbf{u}-\mathbf{u}_{\mathrm{CFD}}\rVert^2/\sum\lVert\mathbf{u}_{\mathrm{CFD}}\rVert^2} 100 ∑ ∥ u − u CFD ∥ 2 / ∑ ∥ u CFD ∥ 2 , including both components and background. Circuit-versus-CFD errors include physical-model error and ensemble-versus-single-state input differences.
Supplementary Data 2 | Reynolds-number comparison for separately fitted wake models (CSV).
Physical forecast scope
Physical dimension and polynomial transport
For s s s physical components at N s N_s N s spatial sites, let N f = s N s N_f=sN_s N f = s N s be the number of scalar field coordinates. A full configuration-space lift with N u N_u N u values per coordinate and a ROM lift with N c N_c N c retained coordinates and N r N_r N r bins per coordinate require
N f u l l = N u N f , q f u l l = N f ⌈ log 2 N u ⌉ , q R O M = N c ⌈ log 2 N r ⌉ . (21) N_{\mathrm{full}}=N_u^{N_f},\quad q_{\mathrm{full}}=N_f\lceil\log_2 N_u\rceil,\quad q_{\mathrm{ROM}}=N_c\lceil\log_2 N_r\rceil. \tag*{(21)} N full = N u N f , q full = N f ⌈ log 2 N u ⌉ , q ROM = N c ⌈ log 2 N r ⌉ . (21)
These data-register counts exclude the phase register and workspace. They differ from amplitude-encoded CFD: compression of N u N f N_u^{N_f} N u N f configurations does not establish exponential advantage in N s N_s N s , because even a poly( N f , log N u ) (N_f,\log N_u) ( N f , log N u ) update remains polynomial in the physical coordinate count. Joseph’s KVN construction obtains a quadratic query improvement over trajectory Monte Carlo through amplitude estimation (main-text ref. 5). Its coherent preparation, propagation and observable-evaluation costs must also be charged. Reducing the transported dimension before this lift can contain that workload if the reduced model preserves the relevant dynamics (main-text Methods). This complementarity does not establish an end-to-end advantage for the present shot-based calculations.
Initial ensembles and state preparation {#sec:Supplementary Note 2}
Both methods use the same compiled MPS input at cell centres ( j + 1 / 2 ) / 64 (j+1/2)/64 ( j + 1/2 ) /64 for j = 0 , … , 63 j=0,\ldots,63 j = 0 , … , 63 , without probability reweighting. Its cell masses define a piecewise-uniform continuum density propagated by characteristics. This is a prescribed algorithmic ensemble, not an uncertainty distribution inferred from CFD initial conditions. Input moments and the corresponding single-flow state are recorded in the Source Data; quantum points impose the shift closure z = E [ r 2 ] z=\mathrm{E}[r^2] z = E [ r 2 ] .
For the primary wake, the Gaussian centre before truncation is 1.4597 and its standard deviation is 0.25. The normalized target cell masses are differences of the normal cumulative distribution function over the 64 uniform cells on [ 0 , 1 ] [0,1] [ 0 , 1 ] . The construction chooses the centre so that the cell-centre mean of these target masses (0.90170 when rounded) matches the projected single-flow radius at t ∗ = 40 t^*=40 t ∗ = 40 to within 10 − 11 10^{-11} 1 0 − 11 before rounding. After bond-two MPS preparation, the mean is 0.90170 and the standard deviation of the piecewise-uniform input is 0.088203. The standard deviation evaluated only at cell centres is 0.088088. The actual prepared masses define the matched continuum input. Source Data include the construction parameters, target and prepared cell masses, and a script reproducing these values.
For the retained seven-model polynomial refinement suite (Supplementary Figure S4 ), all runs end at t = 1 t=1 t = 1 . There are N x = 2 n N_x=2^n N x = 2 n uniform cells and n n n data qubits per coordinate. The scalar Gaussian target has underlying mean 0.35 and standard deviation 0.12, truncated and normalized on [ 0 , 1 ] [0,1] [ 0 , 1 ] . The coupled target is the product of this distribution and a similarly truncated normal with underlying mean 0.65 and standard deviation 0.10 in y y y . Target cell masses are exact differences of the normal cumulative distribution function (CDF).
Supplementary Figure S4. Spatial and temporal verification for seven polynomial models. a, Exact finite-generator moment errors relative to continuum reference solutions started from the actual MPS-prepared input (filled circles) or the analytic Gaussian target (open squares). Each grid has an independently fitted MPS input. b, Compiled-state temporal error against exact finite-generator evolution from the same signed input, as defined by equation (33) . Spatial grids use 64 scalar cells or sixteen cells per coupled coordinate; dotted guides have slope minus two. Table S4 and equation (30) define the scalar and coupled fields. These curves do not establish fixed-input, fixed-accuracy protected-resource scaling.
We use a matrix product state (MPS) with maximum retained bond dimension χ \chi χ , the largest retained Schmidt rank across cuts of the n n n -qubit coordinate register; N x = 2 n N_x=2^n N x = 2 n is the cell count. A supplied MPS stores O ( n χ 2 ) O(n\chi^2) O ( n χ 2 ) coefficients and admits sequential preparation with O ( n χ 2 ) O(n\chi^2) O ( n χ 2 ) elementary gates at fixed rotation precision.1 The implemented real bond-two loader truncates successive singular-value decompositions to χ = 2 \chi=2 χ = 2 and compiles fixed-size two-qubit subspace rotations. One layer uses 4 n − 5 4n-5 4 n − 5 Ry rotations and 2 n − 3 2n-3 2 n − 3 CNOTs: 19 and nine at n = 6 n=6 n = 6 . Its actual gate-executed state supplies propagation and matched-input references.
Higher accuracy can require L M P S L_{\mathrm{MPS}} L MPS successive bond-two layers acting on residual states. Their counted circuit cost is O ( L M P S n ) O(L_{\mathrm{MPS}}\,n) O ( L MPS n ) ; the effective final bond dimension need not remain two. The refinement and sixteen-input suites require fidelity at least 0.9999 with L M P S ≤ 16 L_{\mathrm{MPS}}\leq16 L MPS ≤ 16 and record actual layer counts. For the one-layer loader, shrinking dense-array factorization costs O ( N x ) O(N_x) O ( N x ) at fixed precision and bounded local-solver work; general dense-to-MPS factorization has a conservative O ( N x χ 2 ) O(N_x\chi^2) O ( N x χ 2 ) construction bound. Neither fixed χ \chi χ nor fixed layer count guarantees resolution-independent accuracy. Our dense construction does not use the Irwin–Hall approximation of Ref. 4.
Gaussian construction, angle compilation and validation are offline preprocessing. They depend on the input distribution and grid, not the drift, so compiled inputs can be stored and reused across models and repeated measurements. Every attempt executes and counts the full preparation circuit. The MPS loading of Ref. 5 also has linear gate scaling at fixed layer count; its CNOT counts are specific to that construction.
The primary wake input has fidelity 1 − 4.38 × 10 − 7 1-4.38\times10^{-7} 1 − 4.38 × 1 0 − 7 , amplitude L2 error 0.00066178, and total variation 0.0002187 against the prescribed Gaussian cell masses. Its actual real amplitudes, including any signs from truncation, are retained. Native Ry gates are expanded as S S S , H H H , R z R_z R z , H H H , S S S , and their executed outputs agree with the native preparation. Supplementary Data 3 retains every coordinate/resolution input, including actual layer count, fidelity, total variation and negative weight. The three-coordinate inputs use the finite-bond MPS constructions and input-specific references detailed in Supplementary Note 12.
Supplementary Data 3 | Quality of the nine shared MPS preparation circuits (CSV).
Fidelity is the squared state overlap with the target amplitude vector. Total variation is half the absolute cell-mass difference. Negative weight sums the Born probabilities of negative real amplitudes after one global phase alignment. Inputs requiring additional MPS layers are reported with their actual circuit counts. The primary wake uses a single layer; generic polynomial inputs use the individually recorded layers.
For the sixteen additional tests, Table S5 defines four input families per selected model: an interior Gaussian, an input near a domain boundary, and inputs near stable or unstable fixed points. Centres and widths are specified before truncation to [ 0 , 1 ] [0,1] [ 0 , 1 ] . Each coordinate is compiled independently by the same MPS procedure, with at least 0.9999 target fidelity and recorded layer count. Actual gate-prepared cell probabilities define the matched continuum density. The coupled near-unstable case uses centres (0.02,0.98) and widths (0.04,0.04); the chosen width is the widest tested whose target cell-centre x x x mean is below 0.05. The narrower tests remain resolution diagnostics. These labels do not imply concentration at an exact equilibrium.
Table S5. Inputs and accuracy after propagation for all sixteen tests. Centres and widths specify the normal distributions before truncation. MPS fidelities compare with target cell masses and multiply across coordinates. Matched RMS uses characteristics of the actual prepared piecewise-uniform input; target RMS uses the analytic truncated normal. Equation (S20) averages two scalar or five coupled moments. All runs use 64 steps to time 1, with 64 scalar cells or sixteen cells per coupled coordinate. Unrounded values are supplied in CSV.
Continuity transport and polynomial complexity
For a deterministic vector field F \mathbf{F} F , the density equation is ∂ t ρ = − ∇ ⋅ ( F ρ ) \partial_t \rho=-\nabla\cdot(\mathbf{F}\rho) ∂ t ρ = − ∇ ⋅ ( F ρ ) . Set ρ = ψ ∗ ψ \rho=\psi^*\psi ρ = ψ ∗ ψ and evolve the wavefunction by
L ψ = − F ⋅ ∇ ψ − 1 2 ( ∇ ⋅ F ) ψ . (22) L\psi=-\mathbf{F}\cdot\nabla\psi-\frac{1}{2}(\nabla\cdot\mathbf{F})\psi. \tag*{(22)} L ψ = − F ⋅ ∇ ψ − 2 1 ( ∇ ⋅ F ) ψ . (22)
Equation (S11) gives the continuity equation ∂ t ∣ ψ ∣ 2 = − ∇ ⋅ ( F ∣ ψ ∣ 2 ) \partial_t|\psi|^2=-\nabla\cdot(\mathbf{F}|\psi|^2) ∂ t ∣ ψ ∣ 2 = − ∇ ⋅ ( F ∣ ψ ∣ 2 ) . The integration-by-parts term ⟨ f , L g ⟩ + ⟨ L f , g ⟩ \langle f,Lg\rangle+\langle Lf,g\rangle ⟨ f , Lg ⟩ + ⟨ L f , g ⟩ is a boundary flux that vanishes under the domain conditions used here, so H = i L H=iL H = i L is Hermitian on this domain.[6 , 7 ] Norm preservation is compatible with attraction because the probability density grows as its support contracts.
On uniform radial cells of width h h h , define the real skew matrix
Ω j + 1 , j = F j + 1 / 2 2 h , Ω j , j + 1 = − F j + 1 / 2 2 h . (23) \Omega_{j+1,j}=\frac{F_{j+1/2}}{2h},\quad\Omega_{j,j+1}=-\frac{F_{j+1/2}}{2h}. \tag*{(23)} Ω j + 1 , j = 2 h F j + 1/2 , Ω j , j + 1 = − 2 h F j + 1/2 . (23)
Every other entry is zero apart from these adjacent couplings. Splitting Ω \Omega Ω into even and odd faces gives Ω = Ω e + Ω o \Omega=\Omega_e+\Omega_o Ω = Ω e + Ω o . Each face exponential acts only on a two dimensional pair. On that pair it is the real rotation with R y R_y R y angle F j + 1 / 2 δ / h F_{j+1/2}\delta/h F j + 1/2 δ / h . The second order radial step is
U r ( δ ) = e δ Ω e / 2 e δ Ω o e δ Ω e / 2 + O ( δ 3 ) . (24) U_r(\delta)=e^{\delta\Omega_e/2}e^{\delta\Omega_o}e^{\delta\Omega_e/2}+O(\delta^3). \tag*{(24)} U r ( δ ) = e δ Ω e /2 e δ Ω o e δ Ω e /2 + O ( δ 3 ) . (24)
The local product-formula error gives second-order accumulated time error at fixed spatial discretization under the tested conditions. Note 4 describes the reversible odd-pair alignment. Its cyclic index wrap has zero physical coupling because the endpoint face velocities vanish; it does not identify nonzero density across the endpoints.
Walsh–KvN and the matched block-encoding circuit BE–KvN store the same 64 cell centres on [ 0 , 1 ] [0,1] [ 0 , 1 ] in a six-qubit position register (Notes 2 and 11).
For transport, write N x N_x N x for bins per reduced coordinate and n = ⌈ log 2 N x ⌉ n=\lceil\log_2N_x\rceil n = ⌈ log 2 N x ⌉ . Assume affine binary coordinates and local rotation angles of fixed total degree at most K K K . A scalar layer has at most m ≤ N c n m\le N_c n m ≤ N c n active control bits. Polynomial Walsh support follows[8 ] from equation (S16); the resulting costs are
C l a y e r r o t = O ( ( N c log N x ) K ) , C s t e p r o t = O ( N c ( N c log N x ) K ) . (25) \begin{aligned}
C_{\mathrm{layer}}^{\mathrm{rot}}&=O\left((N_c\log N_x)^K\right),\\
C_{\mathrm{step}}^{\mathrm{rot}}&=O\left(N_c(N_c\log N_x)^K\right). \tag*{(25)}
\end{aligned} C layer rot C step rot = O ( ( N c log N x ) K ) , = O ( N c ( N c log N x ) K ) . (25)
Symmetric splitting uses 4 N c − 1 4N_c-1 4 N c − 1 scalar angle layers to update all N c N_c N c components. Independent parity computation costs O ( K + 1 ) O(K+1) O ( K + 1 ) gates per term. If each component couples at most s c ≤ N c s_c\le N_c s c ≤ N c coordinates, the step uses O ( N c ( s c log N x ) K ) O\left(N_c\left(s_c\log N_x\right)^K\right) O ( N c ( s c log N x ) K ) rotations. A per-coordinate degree bound K K K instead permits total degree N c K N_cK N c K . Mixed terms and self-dependent drift are allowed only when discretization and boundary treatment preserve polynomial angles. Index arithmetic, coefficient compilation, synthesis and resolution-dependent step counts remain additional costs.
Evolution and complete output cost
Let N t N_t N t be the number of local steps. For the one-coordinate cubic wake, the radial structural costs are
C W a l s h = O ( N t n 3 ) , C B E = O ( T N x 5 / 2 n 2 ) , C L C H S , o p t = O ( ∥ A L C H S ∥ T n ) . (26) \begin{aligned}
C_{\mathrm{Walsh}} &= O\left(N_t n^3\right),\\
C_{\mathrm{BE}} &= O\left(TN_x^{5/2}n^2\right),\\
C_{\mathrm{LCHS,opt}} &= O\left(\lVert A_{\mathrm{LCHS}}\rVert Tn\right). \tag*{(26)}
\end{aligned} C Walsh C BE C LCHS , opt = O ( N t n 3 ) , = O ( T N x 5/2 n 2 ) , = O ( ∥ A LCHS ∥ T n ) . (26)
We select both block-encoding references for unitary building blocks whose polylogarithmic gate-count exponents do not grow with physical drift degree. Here a Clifford or arbitrary rotation has unit cost. The BE–KvN row is the direct coordinate-amplitude reconstruction of Supplementary Note 11 at degree three (equation (S84)), with T T T its normalized evolution duration; its grid-size power comes from restoring the physical coordinate scale. Quadratic phase updates add O ( N t n 2 ) O\left(N_tn^2\right) O ( N t n 2 ) to Walsh–KvN. Linear-size index shifts are available, although the primary circuit selects a cheaper quadratic decomposition at its register size (Note 4).[9 ] The optimized LCHS row uses the norm-scaled workload ∥ A L C H S ∥ T \lVert A_{\mathrm{LCHS}}\rVert T ∥ A LCHS ∥ T for generator A L C H S A_{\mathrm{LCHS}} A LCHS and normalized time T T T . This bound assumes linear-size block encoding and favorable controlled selector, weight-preparation, precision and success costs.
For the conditional gate-cost comparison, radial bounds before synthesis are O ( N t ( log 2 N x ) K ) O\left(N_t\left(\log_2N_x\right)^K\right) O ( N t ( log 2 N x ) K ) (Walsh–KvN), O ( K ( log 2 N x ) 2 [ T N x 1 + K / 2 + log ( 1 / ε e x p ) ] ) O\left(K\left(\log_2N_x\right)^2\left[TN_x^{1+K/2}+\log\left(1/\varepsilon_{\mathrm{exp}}\right)\right]\right) O ( K ( log 2 N x ) 2 [ T N x 1 + K /2 + log ( 1/ ε exp ) ] ) (BE–KvN, equation (S84)) and O ( ∥ A L C H S ∥ T log 2 N x ) O\left(\lVert A_{\mathrm{LCHS}}\rVert T\log_2N_x\right) O ( ∥ A LCHS ∥ T log 2 N x ) (optimized LCHS–KvN[10 ] ). Here A L C H S A_{\mathrm{LCHS}} A LCHS is the generator used in the LCHS–KvN discretization and T T T is the evolution duration in the normalized time of each construction. With N t = Θ ( N x a ) N_t=\Theta\left(N_x^a\right) N t = Θ ( N x a ) , a > 0 a>0 a > 0 , and ∥ A L C H S ∥ T = Θ ( N t ) \lVert A_{\mathrm{LCHS}}\rVert T=\Theta(N_t) ∥ A LCHS ∥ T = Θ ( N t ) , the BE–KvN bound has the larger power of grid size whenever a < 1 + K / 2 a<1+K/2 a < 1 + K /2 ; the LCHS–KvN bound carries one logarithmic factor instead of K K K . Compiled T T T counts increase from Walsh–KvN to BE–KvN to LCHS–KvN in the benchmark circuits (Supplementary Notes 10 and 11); the LCHS–KvN count is for the published task at its own settings.
Walsh synthesis and common-target alignment
Physical role of the retained degree
The cubic radial drift and quadratic phase coupling in Note 1 determine the angle degrees and thus the Walsh support bounded by equation (S16). Polynomial degree, rather than retained coefficient count alone, controls this support.
Supplementary Figure S2 shows MPS preparation of the position register, coherent initialization of both phase qubits and repeated half-phase–transport–half-phase steps. The radial transport uses symmetric even–odd–even splitting. This is the implemented Strang product, rather than the exact finite-generator exponential; adjacent phase half steps are merged before compilation.
Supplementary Figure S2. Gate-level construction of the radial Walsh–KvN circuit. a, MPS state preparation, coherent initialization of the two phase qubits and repeated symmetric half-phase–transport–half-phase steps, followed by measurement. Each radial block uses even–odd–even adjacent-pair layers; the displayed time increment is the step duration. b, Common-target transport: decrement aligns odd pairs, the aligned block uses odd-face angles, and increment restores the cell indices. c, A representative Walsh-character Pauli rotation, with a gate angle twice the coefficient in the displayed exponent. d, Phase update containing position-only and position–phase Z strings, the latter acting separately on the two phase qubits. Panels c and d show individual rotations; the compiled circuit reuses parity between compatible Gray-ordered masks (Table S2 ).
Support bound
The support restriction used here is an established property of polynomial Hadamard–Walsh transforms, stated in Theorem 4 of Guseynov et al.[8 ] . The following Boolean-bit derivation specializes it to the rotation-angle functions used in this circuit.
A grid coordinate on m m m control bits is affine: x ( b ) = x 0 + ∑ i w i b i x(b)=x_0+\sum_i w_i b_i x ( b ) = x 0 + ∑ i w i b i . Expanding a degree K K K polynomial in this coordinate produces products of at most K K K distinct bits, since every Boolean bit satisfies b i a = b i b_i^a=b_i b i a = b i for positive integer a a a . With z i = ( − 1 ) b i z_i=(-1)^{b_i} z i = ( − 1 ) b i and b i = ( 1 − z i ) / 2 b_i=(1-z_i)/2 b i = ( 1 − z i ) /2 , each such product expands into Walsh characters ∏ i ∈ S z i \prod_{i\in S}z_i ∏ i ∈ S z i of weight at most K K K . Thus the multiplexed angle has the representation
ϑ ( b ) = ∑ ∣ S ∣ ≤ K ϑ ^ S ∏ i ∈ S ( − 1 ) b i , N W ≤ ∑ j = 0 min ( K , m ) ( m j ) . (27) \vartheta(b)=\sum_{|S|\le K}\widehat{\vartheta}_S\prod_{i\in S}(-1)^{b_i},\qquad N_W\le\sum_{j=0}^{\min(K,m)}\binom{m}{j}. \tag*{(27)} ϑ ( b ) = ∣ S ∣ ≤ K ∑ ϑ S i ∈ S ∏ ( − 1 ) b i , N W ≤ j = 0 ∑ m i n ( K , m ) ( j m ) . (27)
Here ϑ ^ S \widehat{\vartheta}_S ϑ S are the Walsh coefficients of the angle and N W N_W N W is the number of nonzero coefficients. For the primary cubic drift, m m m equals five and the support bound is 1 + 5 + 10 + 10 = 26 1+5+10+10=26 1 + 5 + 10 + 10 = 26 . Three radial rotation layers give 78 arbitrary rotations per full radial step. Each Walsh character is implemented through a parity controlled rotation with appropriate fixed basis changes. The bound concerns one layer for fixed polynomial degree. It does not bound the number of time steps required as resolution increases and therefore is not a complete asymptotic speedup claim.
Common-target alignment and elementary gates
For the one-coordinate radial register, let N x = 2 n N_x=2^n N x = 2 n be the number of cells, n n n the coordinate-register width and q q q an integer pair index. The reversible decrement maps an odd pair ( 2 q + 1 , 2 q + 2 ) (2q+1,2q+2) ( 2 q + 1 , 2 q + 2 ) to ( 2 q , 2 q + 1 ) (2q,2q+1) ( 2 q , 2 q + 1 ) , so all aligned pairs share the least-significant-bit rotation target. The inverse increment restores the cell indices. The cyclic endpoint coupling is zero because the endpoint face velocities vanish (Note 3). The affine pair coordinate then leaves only the low-degree angle function to synthesize. In the implemented primary circuit, these index shifts use one-clean decompositions with exact Toffoli gates. The Walsh-angle network uses CNOT parity operations and single-qubit rotations. The index arithmetic is included in the 224 native T gates per radial step and 7,392 per complete attempt for method A.
Both index-shift methods use six radial qubits and one initially clean work qubit. Any borrowed data bits and the work qubit are restored before transport rotations. Method A uses a one-clean multi-controlled-X construction, whereas method B combines a clean-split increment construction with Cuccaro majority and unmajority carry primitives. Every Toffoli is expanded exactly into Clifford and seven T or T-dagger gates, retaining relative phases. Table S3 compares the complete preparation-to-readout circuits. The index shifts use established constructions[9 , 11 , 12 ] .
Table S3. Complete unsynthesized circuit counts.
In Supplementary Figure S2 , the aligned odd-layer block uses the odd-face rotation angles. It shares the even layer’s circuit template, rather than its angle values.
Akash et al. implement the same skew-symmetric generator form using Gray-coded adjacent-pair rotations[13 ] . Their pair-by-pair construction uses a number of selected rotations proportional to N x N_x N x . This comparison concerns the presented constructions and their rotation organization. A target-grouped angle table generally includes an activation mask. If that mask selects only one of the 2 m 2^m 2 m control indices, every coefficient of its exact m m m -bit Walsh transform is nonzero for a nonzero selected angle, even if the original velocity is constant. Thus an O ( n K + 1 ) O(n^{K+1}) O ( n K + 1 ) rotation bound does not follow from simply applying Walsh synthesis to each target table.
Alternative index organization
Keeping an explicit activation selector separate from a degree-K K K angle recovers sparse polynomial synthesis within each target group. Let W K ( m ) W_K(m) W K ( m ) denote the support bound in equation (S16) for m m m control bits, n n n the coordinate-register width, and R G r a y R_{\mathrm{Gray}} R Gray and R a l i g n e d R_{\mathrm{aligned}} R aligned the complete even–odd–even rotation counts. Then
R G r a y ≤ 2 W K ( n − 1 ) + 2 ∑ j = 1 n − 1 W K ( n − j − 1 ) = O ( n K + 1 ) , R a l i g n e d ≤ 3 W K ( n − 1 ) = O ( n K ) . (28) \begin{aligned}
R_{\mathrm{Gray}}&\le2W_K(n-1)+2\sum_{j=1}^{n-1}W_K(n-j-1)=O\left(n^{K+1}\right),\\
R_{\mathrm{aligned}}&\le3W_K(n-1)=O\left(n^K\right). \tag*{(28)}
\end{aligned} R Gray R aligned ≤ 2 W K ( n − 1 ) + 2 j = 1 ∑ n − 1 W K ( n − j − 1 ) = O ( n K + 1 ) , ≤ 3 W K ( n − 1 ) = O ( n K ) . (28)
At fixed degree, the selector construction has one target group per bit position; common-target alignment avoids that extra sum. Selector gates, index arithmetic and parity work must also be counted. These are analytical component bounds, not an additional transpiled wake benchmark; the implemented aligned circuit is counted below.
Parity reuse and verified gate reduction
Panels c and d of Supplementary Figure S2 show individual compute–rotate–uncompute Pauli-string rotations. The rotation angle is twice the coefficient in the displayed Pauli exponent. Independently doing this for each term repeats CX gates. A traversal through masks changes only the differing parity bits between consecutive terms and uncomputes at the end. This is an application of established Walsh and Gray order synthesis[14 ] , rather than a new diagonal unitary synthesis principle.
The method-A Walsh–KvN circuit retains the state, angles and native T count when parity is reused. Table S2 gives the component counts and the reduction from 23,217 to 12,577 CX gates per complete attempt.
Table S2. Components of the primary Walsh–KvN circuit.
The reduction is 45.828%. The one-step operator norm difference over all 256 joint input basis columns is 3.14 × 10 − 15 3.14 \times10^{-15} 3.14 × 1 0 − 15 . Synthesis of arbitrary rotation angles into discrete Clifford plus T strings is a separate resource step and is not included in this gate-equivalence statement.
State preparation and complete-attempt gate counts
For six radius bits, the radius squared phase has 22 Walsh coefficients. The identity term produces two rotations on the phase register after discarding a common global phase. Each of the other 21 coefficients produces three Z string rotations, giving 2 + 3 ( 21 ) = 65 2+3(21)=65 2 + 3 ( 21 ) = 65 arbitrary rotations per phase layer. Symmetric half phases merge across successive steps, leaving 34 phase layers over 33 radial steps. Table S2 reports the complete arbitrary-rotation and native T totals.
The phase update contains both position-only and position–phase Z strings, as shown in Supplementary Figure S2 . The latter act separately on the two phase qubits; the former act on the position register. The identity position-only contribution is a global phase and is discarded. Gray-code ordering of masks reuses parity within each compatible transport or phase-string family. This mask ordering is distinct from the Gray-coded pair construction compared in equation (28) .
For the selected method-A circuit, 19 preparation rotations encode the prescribed initial distribution. Phase-register basis changes before measurement are Clifford operations. The complete attempt counts below include preparation, evolution and readout.
Adjacent phase half steps merge. Counts include the actual preparation and every evolution operation. The unchanged basis changes before measurement enter the base Clifford count.
Counts precede approximate rotation synthesis. The Steps column counts transport steps.
Both one-clean methods were checked on all 256 basis columns of the joint input sector of the radius and phase registers, with the work qubit initially zero. The largest one-step operator norm errors against independently assembled pair rotations are 7.14 × 10 − 15 7.14 \times10^{-15} 7.14 × 1 0 − 15 for A and 1.24 × 10 − 14 1.24 \times10^{-14} 1.24 × 1 0 − 14 for B. A 24-angle preparation replay reproduced its saved verification input to amplitude error 4.65 × 10 − 15 4.65 \times10^{-15} 4.65 × 1 0 − 15 . Complete 33-step runs from that verification input and an arbitrary complex joint input retain all four branches of the phase register. The maximum joint-state Euclidean-norm errors are 1.89 × 10 − 13 1.89 \times10^{-13} 1.89 × 1 0 − 13 for A and 3.47 × 10 − 13 3.47 \times10^{-13} 3.47 × 1 0 − 13 for B, with work-register leakage below 3.39 × 10 − 30 3.39 \times10^{-30} 3.39 × 1 0 − 30 . The reused circuits contain 12,577 and 20,167 CX gates, respectively. These tests use exact arbitrary rotations and do not certify their approximate Clifford+T synthesis.
A separate 128-cell refinement tensors the original six-bit MPS with a low-order ∣ + ⟩ |+\rangle ∣ + ⟩ qubit, exactly preserving its piecewise-uniform continuum input. The same time interval uses 66 steps and ten logical qubits: seven coordinate, two phase and one clean work qubit. Its complete attempt contains 14,097 arbitrary rotations, 23,100 native T / T † T/T^\dagger T / T † gates and 38,947 CX gates. Gate simulation agrees with independent pair rotations to state-norm error below 7 × 10 − 13 7 \times10^{-13} 7 × 1 0 − 13 over the full evolution; the full 512-column clean-work input sector is checked for one step. The field bias is 0.048856%, and the 10 − 4 10^{-4} 1 0 − 4 implementation-error budget requires 809 accepted shots for 1% relative RMS field. At physical error probability 5 × 10 − 4 5 \times10^{-4} 5 × 1 0 − 4 and 1 μs QEC rounds, the fastest design uses 1,193,151 compiled T T T gates, core distance 19, 561,542 physical qubits and 5.0944 h. The minimum-qubit design uses 60,482 physical qubits and 59.957 h. Source Data, Radial128, contains the expanded gate program, state histories, shot-bound checks and resource-search results. This is a refinement of the radial task, not a matched reproduction of the published LCHS task.
To isolate the spatial accuracy cost of upwinding on the wake, we also apply the positive-drift first-order upwind KvN stencil of Novikau and Joseph (ref. 10). Its coordinates are the same cell centres in [0, 1], and couplings outside that finite grid are omitted. Each finer input tensors the same recorded six-qubit MPS with low-order ∣ + ⟩ |+\rangle ∣ + ⟩ qubits. Both phase qubits are retained, the coupled amplitude generator is propagated by classical matrix-exponential action over the same time interval, and the same readout at the final time is applied after normalization by the output norm. This acts on amplitudes, rather than upwinding the density equation. The squared output norm is recorded separately and is not identified with the success probability of a complete LCHS circuit.
For the same continuum field reference, upwind bias is 1.105%, 0.572%, 0.292%, 0.147% and 0.0735% at 64, 128, 256, 512 and 1,024 cells, respectively. Exact-time centred controls are also retained. The first tested upwind grid below the reported 64-cell circuit bias of 0.107% is 1,024 cells. This threshold matches deterministic bias, not the full 1% RMS field target. At 128 cells the ideal upwind distribution can instead meet that target with 1,187 accepted shots, excluding implementation error. Reference moments are independently regenerated by analytic radial and phase characteristics. Source Data, Wake_Upwind_Refinement, gives the operators, boundary rule, probabilities, variances, error and shot data, and reproduction code. No upwind LCHS gate circuit or matched physical-resource optimization is claimed.
Continuum references and nonlinear transport validation
If the prepared positive cell amplitudes are a j a_j a j and h 0 = 1 / 64 h_0=1/64 h 0 = 1/64 , define the continuum input density as ρ 0 = ∣ a j ∣ 2 / h 0 \rho_0=|a_j|^2/h_0 ρ 0 = ∣ a j ∣ 2 / h 0 within cell j. Refining each cell by a factor k k k splits its amplitude into k k k equal amplitudes a j / k a_j/\sqrt{k} a j / k . Thus refinement preserves a single continuum input. A newly fitted Gaussian on every grid would define a different comparison.
For each original cell, Gauss–Legendre nodes sample the uniform within-cell density. DOP853 integrates radius and accumulated phase at every quadrature node. Weighted sums then give the three continuum moments. A 16 versus 32 point per cell check changes the final moments by less than 10 − 15 10^{-15} 1 0 − 15 . A separate 32 point regeneration reproduces these moments within 2 × 10 − 13 2\times10^{-13} 2 × 1 0 − 13 . The reference uses numerical characteristic integration; the closed form radial solution is not inserted into the circuit or used to generate phase kick angles. At a fixed grid, exact finite-generator evolution isolates temporal splitting error; comparison with the continuum ensemble isolates spatial bias (Supplementary Figure S3 ). Total bias is evaluated directly because these vectors need not add in norm.
Supplementary Figure S3. Wake-transport refinement from the current MPS input. a, Initial, final circuit and matched characteristic densities for the primary 64-cell, 33-step calculation. b, Relative final field bias, including the 64-cell, 33-step primary point; the exact finite generator isolates spatial bias. c, Temporal product-formula error against the same finite generator on 64 cells, with an inverse-square guide. The 128-cell input subdivides the prepared 64-cell continuum density. The 128-cell, 66-step point is additionally compiled, gate-simulated and resource-estimated in Note 4; the remaining refinement points are finite-operator and independent pair-rotation checks.
Joint spatial and time refinement (CSV)
Supplementary Data 4 is supplied as CSV, with field and temporal errors expressed in percent. Finite-generator bias compares exact finite evolution with the fixed-input continuum reference; temporal error compares split propagation with the same finite generator. These error magnitudes need not add.
The 64-cell baseline supplies an explicit resource candidate. Discretization error depends on the input, interval, drift and observable; the piecewise-uniform input and endpoint concentration prevent inferring a universal spatial order from these finite data.
Independent of the wake-specific shift closure, six scalar polynomial fields and one coupled cubic field test spatial and temporal refinement. Four initial families for Q2, C3, Q5 and the coupled model give sixteen additional runs. These are analytic transport tests, not further SINDy calibration. Note 2 specifies their MPS inputs; Table S5 and Supplementary Figure S4 –Supplementary Figure S6 retain all declared cases, including the less accurate boundary and near-repeller cases.
Supplementary Figure S6. Coupled cubic mean trajectories for four inputs. Dashed black curves are matched-input characteristic means; blue squares sample complete-circuit means every eight steps. All runs use sixteen cells per coordinate and 64 steps to time 1. Full-domain panels and stable/unstable enlargements retain the local motion. Filled/open corners mark stable/unstable equilibria. Grey arrows are pointwise drift, not a closed equation for ensemble means. Table S5 gives pre-truncation input parameters and both reference errors; Note 5 diagnoses the near-repeller spatial limitation.
Test models and Gaussian targets
These tests do not refit the primary wake data or preparation and do not determine its protected-resource estimates. For each scalar case on [ 0 , 1 ] [0,1] [ 0 , 1 ] , the dynamics and speed normalization are
d x d t = a P ( x ) , a = 1 4 max x ∈ [ 0 , 1 ] ∣ P ( x ) ∣ . (29) \frac{\mathrm{d}x}{\mathrm{d}t}=aP(x),\quad a=\frac{1}{4\max_{x\in[0,1]}|P(x)|}. \tag*{(29)} d t d x = a P ( x ) , a = 4 max x ∈ [ 0 , 1 ] ∣ P ( x ) ∣ 1 . (29)
The positive constant a a a is determined once from the continuous polynomial’s critical points and endpoints, fixing peak absolute speed at 0.25 without changing equilibria or their stability signs. Table S4 gives the polynomials and normalizers.
Table S4. Scalar test polynomials and speed normalizers.
The coupled field on [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 is
d x d t = x ( 1 − x ) ( 1 + 2 y ) 3 , d y d t = y ( 1 − y ) ( − 2 + x ) 2 . (30) \frac{\mathrm{d}x}{\mathrm{d}t}=\frac{x(1-x)(1+2y)}{3},\quad\frac{\mathrm{d}y}{\mathrm{d}t}=\frac{y(1-y)(-2+x)}{2}. \tag*{(30)} d t d x = 3 x ( 1 − x ) ( 1 + 2 y ) , d t d y = 2 y ( 1 − y ) ( − 2 + x ) . (30)
Both components have total degree three, depend on both coordinates, and have peak absolute speed 0.25. Every tested component vanishes on its own boundary faces. Cyclic index arithmetic therefore creates zero wrap coupling without an additional boundary mask. The accompanying CSV gives the complete model definitions and continuous-domain normalization.
Circuit construction and complete counts
Notes 3–4 define the discretization and synthesis. The coupled circuit applies four directional even/odd generators in a symmetric seven-layer palindrome. Scalar tests use seven logical qubits (six data and one work qubit), and the 16-cell coupled test uses nine (eight data and one work qubit). Supplementary Data 5 is supplied as CSV and counts the actual MPS preparation from zero together with all 64 propagation steps. All index arithmetic uses the same one-clean decomposition as the primary radial circuit. The coupled resource estimate uses the corresponding complete circuit and the measurement workload of Figure 5 .
Supplementary Data 5 | Complete finite-circuit counts including MPS preparation (CSV).
The coupled baseline has sixteen cells per coordinate and 256 configurations. The CSV records total and per-step rotation, CX and native T counts. The seven-model refinement input and the sixteen initial-condition inputs have different preparation parameters; both inventories are given. The sixteen-cell coupled inputs serve as a spatial diagnostic; the common-accuracy comparisons and resource calculation in Figure 5 use 32 cells per coupled coordinate, with the same drift, Gaussian parameters, final time and 64 steps.
Output definitions and error components
The Cartesian output vector contains E [ x ] \mathrm{E}[x] E [ x ] and E [ x 2 ] \mathrm{E}[x^2] E [ x 2 ] in one dimension, and E [ x ] \mathrm{E}[x] E [ x ] , E [ y ] \mathrm{E}[y] E [ y ] , E [ x 2 ] \mathrm{E}[x^2] E [ x 2 ] , E [ y 2 ] \mathrm{E}[y^2] E [ y 2 ] and E [ x y ] \mathrm{E}[xy] E [ x y ] in two dimensions. For a predicted moment m j m_j m j , its declared reference m j r e f m_j^{\mathrm{ref}} m j ref , and M o M_o M o components, the absolute root mean square (RMS) error is
e R M S = [ 1 M o ∑ j = 1 M o ( m j − m j r e f ) 2 ] 1 / 2 . (31) e_{\mathrm{RMS}}=\left[\frac{1}{M_o}\sum_{j=1}^{M_o}\left(m_j-m_j^{\mathrm{ref}}\right)^2\right]^{1/2}. \tag*{(31)} e RMS = [ M o 1 j = 1 ∑ M o ( m j − m j ref ) 2 ] 1/2 . (31)
This absolute RMS is retained for the error-component diagnostics below. Figure 5 uses relative RMS, obtained by dividing the Euclidean moment error by the norm of its matched reference vector. Four reference output vectors are defined at each grid: m A \mathbf{m}_{A} m A is the characteristic output of the fixed analytic Gaussian input; m C \mathbf{m}_{C} m C is the characteristic output of the ideal Gaussian cell-mass input, uniform within each cell; m P \mathbf{m}_{P} m P is the characteristic output of the actual MPS-prepared cell masses, also uniform within each cell; and m H \mathbf{m}_{H} m H is the exact finite-generator output from the actual signed gate-prepared amplitude. The complete compiled output after N t N_t N t steps is m Q \mathbf{m}_{Q} m Q . Thus
m Q − m A = ( m Q − m H ) + ( m H − m P ) + ( m P − m C ) + ( m C − m A ) . (32) \mathbf{m}_{Q}-\mathbf{m}_{A}=\left(\mathbf{m}_{Q}-\mathbf{m}_{H}\right)+\left(\mathbf{m}_{H}-\mathbf{m}_{P}\right)+\left(\mathbf{m}_{P}-\mathbf{m}_{C}\right)+\left(\mathbf{m}_{C}-\mathbf{m}_{A}\right). \tag*{(32)} m Q − m A = ( m Q − m H ) + ( m H − m P ) + ( m P − m C ) + ( m C − m A ) . (32)
The four vectors in equation (32) isolate temporal splitting, matched-input transport, MPS fitting and ideal cell projection. Their vector sum is exact; their RMS norms are not additive. Matched transport uses the actual signed prepared amplitude, so agreement of Born probabilities alone does not remove finite-grid phase effects. Supplementary Data 6 is supplied as CSV, with RMS entries in units of 10 − 3 10^{-3} 1 0 − 3 and all underlying vectors retained. Supplementary Fig. S4a compares exact finite-generator output with the matched prepared-input and analytic-Gaussian references.
Supplementary Data 6 | Baseline moment-error components and totals (CSV; RMS in units of 10 − 3 10^{-3} 1 0 − 3 ).
Spatial accuracy and independent references
Supplementary Fig. S4a is a resolution study: the MPS input is fitted independently on each grid, so it is not a convergence sequence from one fixed prepared continuum input. Matched-input transport error decreases across the tested grids, while Gaussian-target error also includes changing preparation bias and need not decrease monotonically. Characteristics use DOP853 and Gauss–Legendre quadrature; independent tolerance and quadrature checks, full vectors and grid-specific results are supplied as CSV.
Gate agreement temporal order and support
Clean-work-register operators are checked against independent pair-rotation constructions, followed by complete elementary-gate replay including the recorded MPS preparation. The seven operator checks and 28 temporal runs retain their state discrepancies, norm defects and work-register leakage in CSV; leakage remains at floating-point roundoff. For temporal error, equation (33) compares the compiled state with exact evolution of the same finite generator and prepared amplitude.
For temporal product-formula (Trotter) error in Supplementary Fig. S4b, let N t N_t N t denote the propagation step count, T = 1 T=1 T = 1 the final time, and G N x G_{N_x} G N x the complete finite skew generator on N x N_x N x cells per coordinate, including both directions in the coupled model. Let ψ 0 \psi_0 ψ 0 be the actual signed gate-prepared data amplitude and ψ N t ( T ) \psi_{N_t}(T) ψ N t ( T ) the zero-work data component of the full elementary-gate state after N t N_t N t steps, retained without renormalization. With α ∈ R \alpha\in\mathbb{R} α ∈ R a global phase, define
ε s p l i t ( N t ) = min α ∈ R ∥ ψ N t ( T ) − e i α T G N x ψ 0 ∥ 2 . (33) \varepsilon_{\mathrm{split}}(N_t)=\min_{\alpha\in\mathbb{R}}\left\|\psi_{N_t}(T)-e^{i\alpha T G_{N_x}}\psi_0\right\|_2. \tag*{(33)} ε split ( N t ) = α ∈ R min ψ N t ( T ) − e i α T G N x ψ 0 2 . (33)
The exact exponential uses the same generator and initial amplitude as the compiled circuit. Work-register leakage is reported separately, without an acceptance check. Supplementary Fig. S4b measures the temporal product-formula error defined by equation (33) , which is distinct from the moment-error component in equation (32) . The original trajectory arrays and temporal-order checks are retained in the data.
Exact rational Walsh transforms test 64 directional angle tables: coefficients above the corresponding total polynomial degree vanish exactly. The floating-point residuals, zero threshold and reconstructed angles are retained with the support CSV. These checks validate the implemented low-degree constructions; preparation accuracy and boundary handling still require assessment for each new model.
Table S5 collects all sixteen initial distributions, preparation fidelities and matched/target RMS errors at the final time. Initial-condition and MPS settings are defined in Note 2; moment RMS is defined by equation (31) . These dimensionless moment errors are distinct from the relative wake-field error and from full-density norms.
For Figure 5 a, each multinomial position sample supplies all two scalar or five coupled moments jointly, preserving their covariance. Every displayed input uses the same fixed budget of 10,000 shots per estimate, with 5,000 independent nominal-circuit estimates. This common comparison condition reveals differences in sampling sensitivity; it is not chosen by solving the 1% accuracy criterion. The relative moment error uses the same matched-input characteristic reference for every repetition. Squares and ranges show medians and 2.5th–97.5th percentiles of individual errors, not confidence intervals for their medians. Seeds, moment samples and interval endpoints are given in the Source Data, which separately list accuracy-derived resource counts for all sixteen inputs, including infeasible boundary diagnostics. The 674-shot count in Figure 5 c is obtained from that target-based calculation.
For the twelve inputs displayed in Figure 5 a, combining squared noiseless relative bias with sampling variance at 10,000 shots gives relative moment RMS errors below 1%; the largest is 0.8341%. Classical tangent-flow spectral norms averaged over near-stable inputs range from 0.279 to 0.778, compared with 2.15 to 2.53 near unstable points (Source Data, Figure5_Bias_Analysis). These sensitivities support the qualitative contraction and repulsion interpretation of the error ordering. The RMS criterion concerns the average squared error over repeated estimates; individual estimates and percentile endpoints can exceed 1%.
Protected execution of the coupled example
The coupled resource estimate uses 32 cells per coordinate, ten position qubits and one shared clean arithmetic qubit. The drift vanishes at the coordinate boundaries, so no boundary flags are needed. Its complete 64-step circuit has 23,514 arbitrary rotations, 54,058 CX gates and 24,192 native T gates. All five moments are computed jointly from each position measurement. Their average over S S S independent shots defines the sample-mean vector m ^ \hat{\mathbf{m}} m ^ . For the nominal circuit mean vector m \mathbf{m} m , matched reference vector m r e f \mathbf{m}_{ref} m r e f and single-shot covariance Σ \boldsymbol{\Sigma} Σ , define the relative moment error by
ε m o m , r e l = ∥ m ^ − m r e f ∥ 2 ∥ m r e f ∥ 2 , b 2 = ∥ m − m r e f ∥ 2 2 ∥ m r e f ∥ 2 2 , V = tr Σ ∥ m r e f ∥ 2 2 . (34) \varepsilon_{\mathrm{mom},\mathrm{rel}}=\frac{\left\lVert\hat{\mathbf{m}}-\mathbf{m}_{ref}\right\rVert_{2}}{\left\lVert\mathbf{m}_{ref}\right\rVert_{2}},\quad b^{2}=\frac{\left\lVert\mathbf{m}-\mathbf{m}_{ref}\right\rVert_{2}^{2}}{\left\lVert\mathbf{m}_{ref}\right\rVert_{2}^{2}},\quad V=\frac{\operatorname{tr}\boldsymbol{\Sigma}}{\left\lVert\mathbf{m}_{ref}\right\rVert_{2}^{2}}. \tag*{(34)} ε mom , rel = ∥ m r e f ∥ 2 ∥ m ^ − m r e f ∥ 2 , b 2 = ∥ m r e f ∥ 2 2 ∥ m − m r e f ∥ 2 2 , V = ∥ m r e f ∥ 2 2 tr Σ . (34)
Then E [ ε m o m , r e l 2 ] = b 2 + V / S E\left[\varepsilon_{\mathrm{mom},\mathrm{rel}}^{2}\right]=b^{2}+V/S E [ ε mom , rel 2 ] = b 2 + V / S . For each one-shot moment value f j f_j f j , let w j w_j w j be its range over all outcomes and w j ( 2 ) w_j^{(2)} w j ( 2 ) the range of f j 2 f_j^{2} f j 2 . Under the measurement-total-variation error budget η = 10 − 4 \eta=10^{-4} η = 1 0 − 4 , sufficient bounds are
b η 2 = ∑ j ( ∣ m j − m j r e f ∣ + η w j ) 2 ∥ m r e f ∥ 2 2 , (35) b_{\eta}^{2}=\frac{\sum_j\left(\left|m_j-m_j^{ref}\right|+\eta w_j\right)^{2}}{\left\lVert\mathbf{m}_{ref}\right\rVert_{2}^{2}}, \tag*{(35)} b η 2 = ∥ m r e f ∥ 2 2 ∑ j ( m j − m j r e f + η w j ) 2 , (35)
V η = tr Σ + ∑ j [ η ( w j ( 2 ) + 2 ∣ m j ∣ w j ) + η 2 w j 2 ] ∥ m r e f ∥ 2 2 . (36) V_{\eta}=\frac{\operatorname{tr}\boldsymbol{\Sigma}+\sum_j\left[\eta\left(w_j^{(2)}+2\left|m_j\right|w_j\right)+\eta^{2}w_j^{2}\right]}{\left\lVert\mathbf{m}_{ref}\right\rVert_{2}^{2}}. \tag*{(36)} V η = ∥ m r e f ∥ 2 2 tr Σ + ∑ j [ η ( w j ( 2 ) + 2 ∣ m j ∣ w j ) + η 2 w j 2 ] . (36)
The common relative RMS target is ε = 0.01 \varepsilon=0.01 ε = 0.01 . When b η < ε b_{\eta}<\varepsilon b η < ε , the smallest integer satisfying this bound is
S = max ( 1 , ceil ( V η ε 2 − b η 2 ) ) . (37) S=\max\left(1,\operatorname{ceil}\left(\frac{V_{\eta}}{\varepsilon^{2}-b_{\eta}^{2}}\right)\right). \tag*{(37)} S = max ( 1 , ceil ( ε 2 − b η 2 V η ) ) . (37)
If b η ≥ ε b_{\eta}\ge\varepsilon b η ≥ ε , increasing the shot count cannot meet the target. This criterion is the same as the field-error target in Notes 7 and 8, applied here to the coordinate-moment vector rather than the reconstructed velocity field.
For the interior input, the relative bias is 0.0040131 and the variance coefficient is 0.054146. Their implementation-error upper bounds are 0.0042472 and 0.055209, respectively. The sufficient count is 674 shots. The ideal relative RMS is 0.98204% and the model-based bound is 0.999753%. At physical error probability 5 × 10 − 4 5 \times10^{-4} 5 × 1 0 − 4 , the fastest selected architecture uses 563,942 physical qubits and takes 7.4468 h; the minimum-qubit design uses 62,418 qubits and takes 89.537 h. The fastest search is uncapped; removing the former one-million-qubit cap leaves all eight selected designs unchanged. The 13-round preparation with cycle ceiling, padded patch area, block-dependent exposure and connection model are unchanged. The architecture frontier is Pareto-filtered using physical-qubit count, effective core error per consumed T state, factory-output error and T-state consumption interval; joint selection then optimizes synthesis precision. The Source Data supply the refined gate programs, preparation fits, relative-accuracy budgets, architecture frontiers and selected designs at eight physical error probabilities. These estimates use the same phenomenological logical-fault model and synthesis surrogate as the main estimates. The sixteen-cell calculations and fixed-10,000-shot estimates are diagnostics and do not define the Figure 5 c workload.
Variance ratios use variance at the final time divided by each method's own initial variance: Var(x) for scalar tests and Var(x)+Var(y) for the coupled test. Cell-centre and piecewise-uniform continuum inputs have different initial variances. A small initial denominator can amplify a ratio without implying large absolute spread. The separate time-0.30 checks are auxiliary; the coupled-case ratios below are supplied in a diagnostic CSV.
Two cases delimit accuracy at the displayed grids. For Q5 near the boundary, matched moment RMS is 0.016374, compared with exact finite-generator error 0.016415 and compiled-versus-finite discrepancy 4.17e-5. The density comparison retains total-variation distance 0.3688 (Supplementary Figure S5 ). This discrepancy is predominantly spatial rather than temporal. The Q5 circuit develops alternating-cell (grid-scale) oscillations during under-resolved contraction: matched characteristics put 93.3% of the mass in the first cell, compared with 56.7% for the circuit. With the reference-moment normalization used in Figure 5 a, the absolute RMS error of 0.016374 corresponds to 198.9% relative moment bias. This input cannot meet the 1% target by increasing the shot count; it is retained in Supplementary Figure S5 and the full sixteen-input data, rather than the twelve inputs displayed in Figure 5 a.
Supplementary Figure S5. Scalar densities across initial distributions. Rows show Q2, C3 and Q5; columns follow the input families in Table S5. Grey is the actual MPS-prepared initial cell density; blue and dashed black are circuit and matched-input characteristic densities after propagation. All runs use 64 cells and 64 steps to time 1. The Q5 near-boundary case remains visible despite its matched moment RMS of 0.0164 and density total-variation distance of 0.3688.
This is the spatial-discretization issue motivating the dissipative upwind formulation of Novikau and Joseph [10 ] ; LCHS implements the resulting nonunitary evolution. Walsh angle synthesis reproduces the chosen centred discretization and does not damp its grid-scale modes. Additional independent shots reduce sampling variance but cannot remove this deterministic bias.
A supplementary grid-refinement diagnostic keeps the Q5 boundary input and unit-time interval fixed. We tensor its recorded six-qubit bond-two MPS with low-order ∣ + ⟩ |+\rangle ∣ + ⟩ qubits, preserving the same piecewise-uniform continuum input, and use one symmetric even–odd–even step per cell. Both exact finite-generator action and the explicit pair product are retained in Source Data, Q5_Grid_Refinement. The pair-product result reproduces the archived 64-cell circuit state within 5 × 10 − 13 5 \times10^{-13} 5 × 1 0 − 13 in state-vector norm.
Using the continuous characteristic moments as the reference, relative moment bias falls from 198.9% at 64 cells to 6.78%, 1.90% and 0.609% at 512, 1,024 and 2,048 cells. Total variation on those refined grids is 8.42%, 8.18% and 6.27%, respectively; at 8,192 cells it is 1.86%. Aggregating the same distributions back into 64 bins gives much smaller total variations, including 0.292% at 512 cells. Such aggregation and cell-centre reference moments can hide unresolved structure and are not the continuum-moment accuracy criterion. Refinement reduces this error over the tested window; it does not establish that all oscillations have disappeared.
One-clean native reconstructions at 512 and 2,048 cells use ten and twelve logical qubits. With a rotation-synthesis error allocation of 5 × 10 − 5 5 \times10^{-5} 5 × 1 0 − 5 , their per-attempt synthesis-workload estimates are 2.30 × 10 7 2.30 \times10^{7} 2.30 × 1 0 7 and 3.12 × 10 8 3.12 \times10^{8} 3.12 × 1 0 8 T gates. These include MPS preparation and index arithmetic. Native one-step circuits are checked on three complex input states against independent pair rotations; the refined full evolutions use the pair product, and no complete refined gate-by-gate simulation or physical-resource optimization is claimed. At the common 10 − 4 10^{-4} 1 0 − 4 implementation-error budget, small reference moments amplify the range-based implementation-error term to a 1.20–1.21% worst-case floor on the tested grids, even at zero noiseless bias. Refinement alone therefore cannot pass the protected 1% check; that check does not diagnose residual oscillations. These diagnostics therefore do not establish an end-to-end EFTQC result or a matched cost advantage over an upwind LCHS circuit.
For the primary 64-cell, 33-step radial wake, all 34 saved states between nondimensional times 40 and 49.829 were checked against matched-characteristic cell masses, obtained by backtracking cell faces and integrating the actual prepared piecewise-uniform input. The maximum cell-mass total-variation distance is 0.004304 (0.4304%), and the value after 33 steps is 0.002418 (0.2418%). The circuit marginal cell-centre standard deviation decreases from 5.64 to 1.54 cell widths. No interior peak or trough with prominence at least 1% of the corresponding marginal maximum is detected. These diagnostics describe this radial input and time window; they do not establish the absence of smaller fluctuations or assess the three-coordinate runs. Source Data, Grid_Oscillation_Diagnostics, supplies the cell masses, time histories, definitions and reproduction code.
The mechanism combines spatial dispersion with the absence of numerical damping. For locally constant drift, centred differencing propagates short spatial wavelengths at incorrect speeds while retaining their amplitudes. Contraction can generate such modes as fine structures become poorly resolved; upwind dissipation attenuates them (ref. 10) and, being first-order accurate in space, also smooths resolved structure, as the wake control of Supplementary Note 4 and the Q5 control below show. In the radial wake, the matched-characteristic cell-centre standard deviation decreases from 5.64 to 1.53 cell widths during the reported window. The Q5 boundary reference instead places 93.3% of its mass in the first cell. These observations are consistent with a resolved-window wake calculation and stronger unresolved local concentration in Q5. They do not define a universal one-cell width threshold. Walsh synthesis compresses the implementation of the selected discrete transport and does not change this spatial-resolution requirement.
As a matched discretization control for this case, we apply the sign-dependent first-order upwind KvN stencil of Novikau and Joseph (ref. 10) to the same Q5 boundary input and unit-time interval. The sign of the cell-centre drift selects the upwind neighbour of each cell, and couplings outside the grid are omitted. Refined inputs append equal-superposition qubits as low-order index bits to the recorded six-qubit MPS-prepared state, preserving its amplitude signs and the same piecewise-uniform input density. Exact generator action is followed by normalization by the output norm, as in the wake control of Supplementary Note 4. The exact-time centred generator is evaluated alongside; its errors agree with the exact-generator rows of Source Data, Q5_Grid_Refinement, within 10 − 9 10^{-9} 1 0 − 9 . Of the 64 cells of the prepared input, 16 have negative amplitudes, the most negative being − 4.7 × 10 − 5 -4.7 \times10^{-5} − 4.7 × 1 0 − 5 ; together they carry probability 8.5 × 10 − 9 8.5 \times10^{-9} 8.5 × 1 0 − 9 . The centred solution carries 13.4% of its probability in negative-amplitude cells at 64 cells and 0.16% at 512 cells, against 2.4 × 10 − 8 2.4 \times10^{-8} 2.4 × 1 0 − 8 and 1.5 × 10 − 8 1.5 \times10^{-8} 1.5 × 1 0 − 8 for upwind. From 1,024 cells onward, this probability is of order 10 − 8 10^{-8} 1 0 − 8 for both schemes; it cannot by itself distinguish inherited input signs from transport-generated oscillations. Upwind native-cell total variation is 23.9% and 21.4% at 64 and 128 cells, below the centred 37.0% and 24.9%, and exceeds the centred value from 256 cells onward (18.2% against 8.37% at 512 cells; 6.67% against 1.84% at 8,192 cells). Upwind relative moment bias is 308.7%, 82.7%, 29.5% and 9.08% at 64, 512, 2,048 and 8,192 cells, against 199.4%, 6.66%, 0.490% and 0.046% for the centred control. Upwind mass in the first original cell is 0.694 at 64 cells and 0.927 at 8,192 cells; the reference value is 0.933. The squared upwind output norm is 0.252, 0.554, 0.774 and 0.916 on the same four grids. Distribution and moment errors are the primary diagnostics. An auxiliary positive-root control replaces the recorded amplitudes by the nonnegative square roots of their cell probabilities. Its total-variation and relative-moment errors and squared output norm differ by less than 0.00023 percentage points. For this auxiliary input, upwind amplitudes remain nonnegative and centred negative-amplitude probability falls below 10 − 9 10^{-9} 1 0 − 9 from 1,024 cells. These sign statements do not establish the absence of density ripples. The first-order upwind control retains a larger moment bias than centred refinement and a reduced output norm. These statements concern this input, boundary rule and moment metric; another wall closure or a higher-order dissipative scheme could differ. No LCHS circuit, success probability or resource estimate is claimed. Source Data, Q5_Upwind_Comparison, supplies the operators, distributions and reproduction code.
For the coupled near-unstable input, matched RMS is 0.015769, exact finite-generator error 0.015768 and compiled-versus-exact moment discrepancy 7.64e-7. Circuit/reference variance ratios are 5.8936/4.8105. The analytic-Gaussian-target RMS is 0.0065184 against a different reference; cancellation can make it smaller than matched RMS. Table S5 retains both reference errors. On the refined 32-cell grid used in Figure 5 a, this same near-unstable input has relative moment bias 0.82453%; the sixteen-cell result is retained here to document the resolved spatial limitation.
Joint radius–phase readout and field reconstruction
The phase Φ \Phi Φ obeys d Φ d t ∗ = γ r 2 \frac{d\Phi}{dt^*}=\gamma r^2 d t ∗ d Φ = γ r 2 along each radial trajectory. With both phase qubits initialized in equal superpositions, the phase kick D D D in main-text equation (6) produces four occupation branches. Each branch undergoes the same radial transport with its corresponding phase kick. The radius-dependent term remains in the joint evolution; only the constant full-register identity term is a global phase.
Pauli X X X on P x P_x P x and Pauli Y Y Y on P y P_y P y commute and can be measured in the same shot. Together with the measured radius r r r , the outcomes s x , s y ∈ { − 1 , 1 } s_x,s_y\in\{-1,1\} s x , s y ∈ { − 1 , 1 } give the estimators Y 1 = r s x Y_1=rs_x Y 1 = r s x , Y 2 = r s y Y_2=rs_y Y 2 = r s y and Y 3 = r 2 Y_3=r^2 Y 3 = r 2 . Their averages retain radius–phase correlations.
At finite resolution, phase kicks and radial transport do not commute, so the phase branches can have different radial marginals. We calculate all reported means, covariances and success probabilities from the complete discrete joint state. No conditional product-state approximation is imposed.
From amplitudes to joint samples
Let ψ k ( j ) \psi_k(j) ψ k ( j ) be the amplitudes of the four discrete phase branches at radius cell j j j . Basis changes before measurement implement Pauli X on P X P_X P X and Pauli Y on P Y P_Y P Y . With branch ordering fixed by the saved circuit, their unitary basis map U r e a d \mathbf{U}_{read} U r e a d gives
p i j = ∣ ∑ k = 0 3 U r e a d i k ψ k ( j ) ∣ 2 , p_{ij}=\left|\sum_{k=0}^{3}U_{read\,ik}\psi_k(j)\right|^2, p ij = k = 0 ∑ 3 U r e a d ik ψ k ( j ) 2 ,
m = ∑ i j p i j Y i j , \mathbf{m}=\sum_{ij}p_{ij}\mathbf{Y}_{ij}, m = ij ∑ p ij Y ij ,
Σ = ∑ i j p i j Y i j Y i j T − m m T . (38) \mathbf{\Sigma}=\sum_{ij}p_{ij}\mathbf{Y}_{ij}\mathbf{Y}_{ij}^{T}-\mathbf{m}\mathbf{m}^{T}. \tag*{(38)} Σ = ij ∑ p ij Y ij Y ij T − m m T . (38)
Here i i i labels the four joint phase outcomes and Y i j \mathbf{Y}_{ij} Y ij is the estimator vector for outcome i i i at radius cell j j j . The probabilities are conditioned on the actual joint success event when required. Walsh–KvN has unit norm and success probability. Replaying its saved branch history reproduces the measured means to numerical precision.
Supplementary Data 7 | Final Walsh–KvN joint outcome covariance (CSV).
These single-successful-shot covariances enter the field variance before division by the shot count. The Source Data records the continuum and circuit moments at the final time and the continuum Radial field norm. The moments retain nonlinear phase; equation (20) applies the known rigid phase and fixed field map separately.
Stored spatial modes and a small quantum output
Equation (20) maps the three moment estimates to modal coefficient expectations through the fixed matrix M \mathbf{M} M . The stored spatial modes then reconstruct the velocity field. These coefficients are physical observables, distinct from complex KvN amplitudes.
The quantum output is this small set of expectations. Classical reconstruction at N s N_s N s physical output sites costs O ( N s ) O(N_s) O ( N s ) arithmetic operations using the stored modes. Training and mode storage are offline costs, and the ROM representation error is assessed separately. Estimating the full ensemble probability vector would require a separate measurement task.2
Drag and lift observables
Following Deng et al. (main-text ref. 32), we fit force coefficients as polynomials of the normalized modes x x x , y y y and z z z . The finite-element method (FEM) reference integrates pressure and viscous stress on the cylinder, with force per unit span normalized by ρ U ∞ 2 D / 2 \rho U_{\infty}^{2}D/2 ρ U ∞ 2 D /2 , where ρ \rho ρ is density, U ∞ U_{\infty} U ∞ the upstream velocity and D D D the cylinder diameter. Drag uses 1 1 1 , z z z , x 2 x^2 x 2 , x y xy x y , y 2 y^2 y 2 , z 2 z^2 z 2 and lift uses x x x , y y y , x z xz x z , y z yz y z , respecting reflection symmetry. Calibration uses FEM-projected modes and surface forces. The calibration interval includes Figure 2 f, so the reported errors assess in-window force reconstruction with frozen maps. Coefficients, intervals and histories are supplied with Figure 2 f and in the numerical-detail CSVs.
For the radial circuit, x = r cos θ x=r\cos\theta x = r cos θ , y = r sin θ y=r\sin\theta y = r sin θ and z = r 2 z=r^2 z = r 2 . Lift uses radius-weighted first phase harmonics; drag additionally uses radial moments through r 4 r^4 r 4 and second phase harmonics. We evaluate their noiseless ensemble expectations from the saved phase coherences with the known rigid phase restored. We now estimate these expectations with 815 additional shots per time point, shared by drag and lift. These use the actual signed MPS-prepared state and the same coherent propagation as the velocity-field circuit.
The four measurement settings are XX, XY, YX and YY, with 204, 204, 204 and 203 shots, respectively. Each letter denotes a Pauli measurement on a phase qubit; the first letter labels the higher-index qubit. XX minus YY isolates the real part of four times the 00–11 coherence, and XY plus YX its imaginary part. These combinations provide the second phase harmonic for drag, while the single-qubit X and Y expectations provide the first harmonic for lift. Measured radius values supply the required radial powers and the deterministic rigid phase rotates the harmonics back to the physical angle. Each setting yields a two-component drag/lift estimator; their means are combined with equal weights of one quarter. The exact estimator covariance is the sum of the four single-shot covariances divided by sixteen times the corresponding setting shot counts. Thus both forces share 815 total shots, rather than 815 shots per basis. Source Data include the distributions, estimator values, seed, sampled counts and reproduction code.
At the 34 saved times, the relative RMS errors against FEM are 0.211% for drag and 10.47% for lift, combining deterministic in-window force reconstruction error and sampling variance. Their noiseless counterparts remain 0.193% and 8.81%; sampling alone contributes 0.0855% and 5.66% in the same RMS normalization. The FEM RMS normalizers are 1.33 for drag and 0.168 for lift. Drag includes a nonzero mean, whereas lift is oscillatory and uses first phase harmonics. Its smaller reference scale and sensitivity to phase mismatch can contribute to the unequal relative errors. The force-map, reduced-dynamics and ensemble-input contributions are not separately identified. Two thousand independent repetitions give measured-to-predicted mean-squared-error ratios of 1.003 and 0.996. The plotted open points are a single realization. Source Data retain theoretical standard errors, not confidence intervals. These additional force measurements are not included in Table 1 ’s velocity-field shot budget, and 815 force shots do not guarantee 1% force error. Only Clifford basis changes before measurement differ, adding neither qubits nor T gates; no new physical architecture is optimized for this diagnostic.
Sampling errors and complete prediction cost {#sec:Supplementary Note 7}
Deriving the mean square error
Let A = Ψ m T M \mathbf{A}=\boldsymbol{\Psi}_m^{T}\mathbf{M} A = Ψ m T M and m ^ = S − 1 ∑ s Y s \hat{\mathbf{m}}=S^{-1}\sum_s\mathbf{Y}_s m ^ = S − 1 ∑ s Y s . Define the field bias vector e = u s t + A m − u ˉ r e f \mathbf{e}=\mathbf{u}_{\mathrm{st}}+\mathbf{A}\mathbf{m}-\bar{\mathbf{u}}_{\mathrm{ref}} e = u st + Am − u ˉ ref . Independence gives Cov ( m ^ ) = Σ / S \operatorname{Cov}(\hat{\mathbf{m}})=\boldsymbol{\Sigma}/S Cov ( m ^ ) = Σ / S . For the field estimate u ^ = u s t + A m ^ \hat{\mathbf{u}}=\mathbf{u}_{\mathrm{st}}+\mathbf{A}\hat{\mathbf{m}} u ^ = u st + A m ^ , expanding the weighted field norm and using zero mean sample fluctuations yields
E ∥ u ^ − u ˉ r e f ∥ W 2 = ∥ e ∥ W 2 + tr ( A T W A Σ ) / S . (39) \mathbb{E}\left\lVert\hat{\mathbf{u}}-\bar{\mathbf{u}}_{\mathrm{ref}}\right\rVert_W^2
=
\left\lVert\mathbf{e}\right\rVert_W^2+
\operatorname{tr}\left(\mathbf{A}^{T}\mathbf{W}\mathbf{A}\boldsymbol{\Sigma}\right)/S. \tag*{(39)} E ∥ u ^ − u ˉ ref ∥ W 2 = ∥ e ∥ W 2 + tr ( A T WA Σ ) / S . (39)
For the uniform sampled-field norm, W = I \mathbf{W}=\mathbf{I} W = I and G = Ψ m Ψ m T \mathbf{G}=\boldsymbol{\Psi}_m\boldsymbol{\Psi}_m^{T} G = Ψ m Ψ m T , so equation (S28) uses M T G M \mathbf{M}^{T}\mathbf{G}\mathbf{M} M T GM with all off-diagonal covariances retained. A common cell-area factor cancels in the relative error.
Dividing equation (S28) by the squared reference-field norm gives b 2 + ν / S b^2+\nu/S b 2 + ν / S , with the normalized contributions defined by
b 2 = ∥ e ∥ W 2 / ∥ u ˉ r e f ∥ W 2 , ν = tr ( A T W A Σ ) ∥ u ˉ r e f ∥ W 2 . b^2=\left\lVert\mathbf{e}\right\rVert_W^2/\left\lVert\bar{\mathbf{u}}_{\mathrm{ref}}\right\rVert_W^2,\qquad
\nu=\frac{\operatorname{tr}\left(\mathbf{A}^{T}\mathbf{W}\mathbf{A}\boldsymbol{\Sigma}\right)}
{\left\lVert\bar{\mathbf{u}}_{\mathrm{ref}}\right\rVert_W^2}. b 2 = ∥ e ∥ W 2 / ∥ u ˉ ref ∥ W 2 , ν = ∥ u ˉ ref ∥ W 2 tr ( A T WA Σ ) .
Here b b b is the deterministic relative field bias, and ν \nu ν is the field sampling-variance coefficient before division by the accepted-shot count S S S . We report the square root of this mean-square expression as the relative RMS field error. The mean is over independent repetitions of an S S S -shot estimate. Solving the target inequality gives the ceiling rule in the main text. If b ≥ ε b\geq\varepsilon b ≥ ε , no finite shot count can meet the RMS target without reducing deterministic bias. This RMS criterion is not equivalent to requiring 95% of estimates to lie inside a fixed error ball.
Replacing radius-weighted phase expectations by products of marginal expectations increases the field bias at the final time. The comparison uses the same noiseless circuit distribution, field map and matched continuum Radial reference, leaving the squared-radius mean unchanged. The full and factorized moment vectors and their biases are supplied in the numerical-detail CSVs.
Radial concentration and the implemented field readout
Total covariance separates conditional mean variation from phase readout noise at fixed radius,
Σ = Cov r [ E ( Y ∣ r ) ] + E r [ Cov ( Y ∣ r ) ] . (40) \Sigma=\operatorname{Cov}_{r}\left[\mathbb{E}\left(\mathbf{Y}\mid r\right)\right]+\mathbb{E}_{r}\left[\operatorname{Cov}\left(\mathbf{Y}\mid r\right)\right]. \tag*{(40)} Σ = Cov r [ E ( Y ∣ r ) ] + E r [ Cov ( Y ∣ r ) ] . (40)
We evaluate equation (S29) using the actual finite-circuit conditional distribution. Radial and squared-radius variances decrease during the wake calculation, while total field variance increases because phase readout fluctuations dominate. The numerical-detail CSVs give the covariance decomposition and variance history. Radial concentration therefore does not by itself reduce the field shot budget.
Repeated sampling
Supplementary Data 8 verifies equation (S28) by repeated sampling of the saved circuit distributions. It supplies the tested steps and shot counts, fixed seed, moment and mean-square-error replicates, and replay code. The numerical-detail CSVs summarize the discrepancy check.
Supplementary Data 8 | Repeated finite-shot verification from the current MPS circuit (CSV).
These checks verify sampling error for the saved ideal distributions; they are not hardware-noise experiments or confidence guarantees for arbitrary inputs.
Generic exact preparation costs linearly in amplitude dimension.[15 , 16 ] The primary input uses the counted MPS circuit (Note 2); matched calculations use its actual cell masses. For S S S shots used for estimation, complete-attempt success probability P s P_s P s , preparation cost G p r e p G_{prep} G p r e p , step cost G s t e p G_{step} G s t e p and readout basis-change cost G r e a d G_{read} G r e a d , expected logical work is
C Q ≃ S P s [ G p r e p + N t G s t e p + G r e a d ] . (41) C_Q \simeq\frac{S}{P_s}\left[G_{prep}+N_tG_{step}+G_{read}\right]. \tag*{(41)} C Q ≃ P s S [ G p r e p + N t G s t e p + G r e a d ] . (41)
For the deterministic one-dimensional Walsh–KvN input family, let N x = 2 n N_x=2^n N x = 2 n be cells per coordinate, n n n the coordinate-register width, S S S the shot count, N t N_t N t the step count and K ≥ 2 K\ge2 K ≥ 2 the fixed aligned angle degree. With maximum MPS bond dimension χ \chi χ , generic sequential preparation gives the pre-synthesis logical-work bound
C t o t a l = C o f f l i n e + O ( S [ n χ 2 + N t n K ] ) (42) C_{total}=C_{offline}+O\left(S\left[n\chi^2+N_t n^K\right]\right) \tag*{(42)} C t o t a l = C o f f l in e + O ( S [ n χ 2 + N t n K ] ) (42)
Equation (S31) concerns one transported coordinate, as in the radial calculation. Supplementary Note 3 gives the dependence on the number of coordinates and their couplings.
C o f f l i n e C_{offline} C o f f l in e includes input construction, angle compilation and validation. For an input amplitude vector of length N N N , dense factorization has the O ( N χ 2 ) O\left(N\chi^2\right) O ( N χ 2 ) construction bound in Note 2; diagnostic checks are additional offline work. With L M P S L_{MPS} L M P S preparation layers, the implemented circuit has preparation term L M P S n L_{MPS}n L M P S n . The primary χ = 2 \chi=2 χ = 2 , L M P S = 1 L_{MPS}=1 L M P S = 1 circuit therefore has online work O ( S [ n + N t n K ] ) O\left(S\left[n+N_t n^K\right]\right) O ( S [ n + N t n K ] ) . Bond dimension, layers, synthesis precision, steps and shots can all depend on requested accuracy; this bound does not by itself give an end-to-end speedup.
For target relative RMS field error ε \varepsilon ε , suppose a declared input family has deterministic bias at most b m a x < ε b_{max}<\varepsilon b ma x < ε , field covariance contribution at most v m a x v_{max} v ma x , attempt success at least p m i n > 0 p_{min}>0 p min > 0 and complete-attempt gate cost at most G m a x G_{max} G ma x . Independent shots used for estimation then suffice with
S m a x = max ( 1 , ceil ( v m a x ε 2 − b m a x 2 ) ) , E [ C Q ] ≤ S m a x p m i n G m a x . (43) \begin{aligned}
S_{max}&=\max\left(1,\operatorname{ceil}\left(\frac{v_{max}}{\varepsilon^2-b_{max}^2}\right)\right),\\
\mathbb{E}\left[C_Q\right]&\le\frac{S_{max}}{p_{min}}G_{max}. \tag*{(43)}
\end{aligned} S ma x E [ C Q ] = max ( 1 , ceil ( ε 2 − b ma x 2 v ma x ) ) , ≤ p min S ma x G ma x . (43)
This follows from Note 7’s bias–variance relation and expected attempts per success. It bounds RMS error and expected work, not confidence or worst-case runtime. Uniform preparation, transport, readout and success bounds must be established for any claimed family; current results give finite instances. Offline learning and compilation must be added or amortized, with physical synthesis and timing charged through Notes 8, 9 and 10. The retained dimension and chosen polynomial library therefore link surrogate accuracy to transport-circuit cost, while the requested physical observable determines the sampling accuracy and resource budget.
Sampling crossover for a specified classical comparator
Let G c l G_{\mathrm{cl}} G cl be the work of a classical prediction task and g q g_q g q the complete quantum work per attempt, measured in a common declared cost model for the same input, dynamics, observables and accuracy. Let S S S denote independent shots used for estimation and P s P_s P s the complete-attempt success probability. The per-attempt cost includes preparation, all propagation steps, basis changes before measurement and any stipulated synthesis cost. Excluding one-time setup from both sides, the expected quantum work is smaller only if
E [ G q ] = S P s g q , S < P s G c l g q (44) \mathbb{E}[G_q] = \frac{S}{P_s}g_q,\qquad S < \frac{P_sG_{\mathrm{cl}}}{g_q} \tag*{(44)} E [ G q ] = P s S g q , S < g q P s G cl (44)
One-time fitting, compilation and other setup costs must be added or amortized consistently. For independent measurement settings j j j , replace the quantum work by ∑ j S j g q , j P s , j \sum_j S_j\frac{g_{q,j}}{P_{s,j}} ∑ j S j P s , j g q , j , where the sample count, per-attempt cost and success probability are specific to each setting. The physical runtime calculation in Note 10 uses the corresponding setting-dependent times under one shared architecture.
A task-level crossover depends on the complete preparation, transport and readout costs and on the estimator bias and variance in equation (S32). Preparation and physical-field reconstruction costs must be charged as specified in Notes 2 and 6. Reduced dimension can also reduce the matched classical cost, and trajectory sampling or moment closure may outperform the same-update comparator. A fixed shot count enters asymptotic work as a prefactor and can dominate at finite sizes. The finite comparisons in Note 10 quantify their declared tasks. A speedup claim requires a matched classical baseline and complete costs.
An Eulerian solver on the same N c N_c N c -coordinate ensemble grid has O ( N t N x N c ) O(N_tN_x^{N_c}) O ( N t N x N c ) work. If a quantum step costs poly ( N c , log N x ) \operatorname{poly}(N_c,\log N_x) poly ( N c , log N x ) and input/output costs remain controlled, then
C E u = O ( N t N x N c ) , C E u C Q ∼ N x N c ( S / P s ) poly ( N c , log N x ) . (45) C_{\mathrm{Eu}}=O\left(N_tN_x^{N_c}\right),\qquad\frac{C_{\mathrm{Eu}}}{C_Q}\sim\frac{N_x^{N_c}}{(S/P_s)\operatorname{poly}(N_c,\log N_x)}. \tag*{(45)} C Eu = O ( N t N x N c ) , C Q C Eu ∼ ( S / P s ) poly ( N c , log N x ) N x N c . (45)
This ratio compares the same ensemble grid, not all classical algorithms. Resolution-dependent N t N_t N t multiplies both costs and can erode absolute efficiency. Classical trajectory sampling has no obligatory ensemble-grid factor: for observable variance σ o 2 \sigma_o^2 σ o 2 , mean-square sampling tolerance ε o 2 \varepsilon_o^2 ε o 2 and trajectory cost C t r a j C_{\mathrm{traj}} C traj , its elementary cost is O ( C t r a j σ o 2 / ε o 2 ) O(C_{\mathrm{traj}}\sigma_o^2/\varepsilon_o^2) O ( C traj σ o 2 / ε o 2 ) . Independent quantum shots give no Monte Carlo speedup. For this wake, total cost relative to classical sampling is therefore not established.
Resolved-flow resource studies include the coupled costs of spatial and temporal resolution: a drag-extraction study finds no exponential advantage for its examined instances[17 ] , and an incompressible lattice-Boltzmann analysis retains bounded gains for selected outputs under its convergence, conditioning and readout assumptions[18 ] . The present reduced-ensemble task assesses physical ROM error separately from quantum output error; its candidate regime concerns reduced configuration grids rather than Reynolds-number scaling or fully resolved accuracy.
Implementation channels and protected output accuracy
The per-attempt error budget is η = 10 − 4 \eta=10^{-4} η = 1 0 − 4 . Physical-model error, continuum discretization and finite-representation bias are accounted for separately from the channel-error budget. A circuit attempt includes preparation, propagation and the success test. The notation below separates single-rotation synthesis errors from whole-attempt hardware faults, then connects their combined effect to field bias, variance and shots.
Rotation synthesis and implementation channels
Ross–Selinger synthesis model
An arbitrary R Z ( θ ) R_{\mathrm{Z}}(\theta) R Z ( θ ) rotation is approximated by a Clifford+T sequence. Let n T n_{\mathrm{T}} n T be the number of T/T† gates for this one rotation and ε R \varepsilon_{\mathrm{R}} ε R its operator-norm error tolerance. The typical-case Ross–Selinger scaling is
n T = 3 log 2 ( 1 / ε R ) + O ( log log ( 1 / ε R ) ) (46) n_{\mathrm{T}}=3\log_{2}\left(1/\varepsilon_{\mathrm{R}}\right)+O\left(\log\log\left(1/\varepsilon_{\mathrm{R}}\right)\right) \tag*{(46)} n T = 3 log 2 ( 1/ ε R ) + O ( log log ( 1/ ε R ) ) (46)
The resource search retains only the leading term, inverted as ε R ≈ 2 − n T / 3 \varepsilon_{\mathrm{R}}\approx2^{-n_{\mathrm{T}}/3} ε R ≈ 2 − n T /3 . [19 ] Here N T n a t i v e N_{\mathrm{T}}^{\mathrm{native}} N T native counts the exact T/T† gates already present before arbitrary-rotation synthesis.
The first logarithm measures the requested rotation precision in bits. The nested logarithm is an asymptotically smaller correction from the synthesis analysis; neither logarithm refers to grid size. This correction is omitted in the numerical leading-term model, where halving the per-rotation tolerance adds three T gates. The inverted relation estimates synthesis cost; it does not certify an actually synthesized gate sequence.
For a complete attempt, N R Z N_{\mathrm{RZ}} N RZ is the number of rotations requiring synthesis and N T N_{\mathrm{T}} N T the total T/T† count. Each T or T† consumes one magic state. With per-rotation cost n T n_{\mathrm{T}} n T , the compiled total is
N T = N T n a t i v e + n T N R Z . (47) N_{\mathrm{T}}=N_{\mathrm{T}}^{\mathrm{native}}+n_{\mathrm{T}}N_{\mathrm{RZ}}. \tag*{(47)} N T = N T native + n T N RZ . (47)
In the resource CSVs, the legacy column N_T stores the per-rotation quantity n T n_{\mathrm{T}} n T , whereas T_count stores the complete-attempt quantity N T N_{\mathrm{T}} N T . For the baseline fastest 64-cell radial design, 77 T gates for each of 4,803 rotations are added to 7,392 native T/T† gates, giving 377,223 consumed T states per attempt.
Measurement probabilities and normalized diamond distance
Let U i d ( ρ ) = U ρ U † U_{\mathrm{id}}(\rho)=U\rho U^{\dagger} U id ( ρ ) = U ρ U † be the ideal channel for one complete attempt, where U U U is the full ideal circuit and ρ \rho ρ is a quantum density operator. A channel maps an input density operator to its output. The trace-preserving implemented channel E i m p l E_{\mathrm{impl}} E impl includes synthesis and hardware imperfections before conditioning on circuit success. The normalized diamond distance is defined by[20 , 21 ]
D ⋄ ( E , U ) = 1 / 2 ∥ E − U ∥ ⋄ . (48) D_{\diamond}(E,U)=1/2\left\|E-U\right\|_{\diamond}. \tag*{(48)} D ⋄ ( E , U ) = 1/2 ∥ E − U ∥ ⋄ . (48)
The diamond norm optimizes over the quantum input and an entangled reference, so it controls every input and measurement. This reference is part of the distance definition, not an additional register in the resource estimate. Let P P P and Q Q Q be the ideal and implemented joint output distributions from the same input density operator ρ 0 \rho_0 ρ 0 , including all acceptance flags. Here y y y labels a complete classical readout record, and B B B is an arbitrary event: a set of such records. P ( B ) P(B) P ( B ) and Q ( B ) Q(B) Q ( B ) are the probabilities of that event. The supremum over events selects the largest probability difference over sets of outcomes, rather than only over individual cells. The total variation distance satisfies[20 ]
d T V ( P , Q ) = 1 / 2 ∑ y ∣ P ( y ) − Q ( y ) ∣ = sup B ∣ P ( B ) − Q ( B ) ∣ ≤ 1 / 2 ∥ E i m p l ( ρ 0 ) − U i d ( ρ 0 ) ∥ 1 ≤ D ⋄ ( E i m p l , U i d ) . (49) \begin{aligned}
d_{\mathrm{TV}}(P,Q)&=1/2\sum_y\left|P(y)-Q(y)\right|=\sup_B\left|P(B)-Q(B)\right|\\
&\leq1/2\left\|E_{\mathrm{impl}}(\rho_0)-U_{\mathrm{id}}(\rho_0)\right\|_1\\
&\leq D_{\diamond}(E_{\mathrm{impl}},U_{\mathrm{id}}). \tag*{(49)}
\end{aligned} d TV ( P , Q ) = 1/2 y ∑ ∣ P ( y ) − Q ( y ) ∣ = B sup ∣ P ( B ) − Q ( B ) ∣ ≤ 1/2 ∥ E impl ( ρ 0 ) − U id ( ρ 0 ) ∥ 1 ≤ D ⋄ ( E impl , U id ) . (49)
The factor one-half in equation (48) makes the channel distance lie between zero and one. Equation (49) concerns unconditioned outputs, including failed attempts; the following subsection conditions on circuit success.
Combining synthesis and stochastic faults
Let U s y n U_{\mathrm{syn}} U syn be the channel of the fault-free Clifford+T approximation to the complete ideal circuit. The index j j j labels a synthesized rotation: U j U_j U j is its ideal rotation matrix and U ~ j \widetilde{U}_j U j its Clifford+T approximation. The tilde marks coherent synthesis error, not a stochastic hardware fault. For N R z N_{R_z} N R z rotations with certified operator-norm errors ∥ U j − U ~ j ∥ ∞ ≤ ε R \lVert U_j-\widetilde{U}_j\rVert_{\infty}\leq\varepsilon_R ∥ U j − U j ∥ ∞ ≤ ε R , the unitary-channel bound and stability under composition give[20 , 21 ]
D ⋄ ( U s y n , U i d ) ≤ min { 1 , N R z ε R } . (50) D_{\diamond}(U_{\mathrm{syn}},U_{\mathrm{id}})\leq\min\{1,N_{R_z}\varepsilon_R\}. \tag*{(50)} D ⋄ ( U syn , U id ) ≤ min { 1 , N R z ε R } . (50)
For the error probabilities, uppercase E denotes a whole-attempt contribution and lowercase e a per-state contribution. Under the assumed independent effective stochastic Pauli model, E c o r e E_{\mathrm{core}} E core is the probability of at least one core logical fault during the attempt, whereas e M S F e_{\mathrm{MSF}} e MSF is the factory-output fault probability for each delivered T state. MSF stands for magic-state factory; this output error is distinct from the physical device error probability. With N T N_T N T magic states consumed, the probability of at least one core or magic-state fault is
p s t o c h = 1 − ( 1 − E c o r e ) ( 1 − e M S F ) N T . (51) p_{\mathrm{stoch}}=1-(1-E_{\mathrm{core}})(1-e_{\mathrm{MSF}})^{N_T}. \tag*{(51)} p stoch = 1 − ( 1 − E core ) ( 1 − e MSF ) N T . (51)
(51) multiplies the probability of a fault-free core by that of N T N_T N T fault-free magic states and takes the complement. In resource CSVs, p_T_final is the per-state quantity e M S F e_{\mathrm{MSF}} e MSF , whereas E_MSF is its accumulated contribution over a complete attempt.
Grouping attempts by whether a stochastic fault occurs gives E i m p l = ( 1 − p s t o c h ) U s y n + p s t o c h E f a u l t E_{\mathrm{impl}}=(1-p_{\mathrm{stoch}})U_{\mathrm{syn}}+p_{\mathrm{stoch}}E_{\mathrm{fault}} E impl = ( 1 − p stoch ) U syn + p stoch E fault , where E f a u l t E_{\mathrm{fault}} E fault is the average channel conditioned on at least one stochastic fault. This quantum channel is distinct from the classical event B in equation (S38). Since D ⋄ ( E i m p l , U s y n ) ≤ p s t o c h D_{\diamond}(E_{\mathrm{impl}},U_{\mathrm{syn}})\leq p_{\mathrm{stoch}} D ⋄ ( E impl , U syn ) ≤ p stoch , the triangle inequality gives
d T V ( P , Q ) ≤ D ⋄ ( E i m p l , U i d ) ≤ min { 1 , p s t o c h + N R z ε R } . (52) d_{\mathrm{TV}}(P,Q)\leq D_{\diamond}(E_{\mathrm{impl}},U_{\mathrm{id}})\leq\min\{1,p_{\mathrm{stoch}}+N_{R_z}\varepsilon_R\}. \tag*{(52)} d TV ( P , Q ) ≤ D ⋄ ( E impl , U id ) ≤ min { 1 , p stoch + N R z ε R } . (52)
The resource search combines the leading-term synthesis relation ε R ≃ 2 − n T / 319 \varepsilon_R\simeq2^{-n_T/319} ε R ≃ 2 − n T /319 with the assumed fault probabilities to estimate the whole-attempt error
E m o d e l = 1 − ( 1 − E c o r e ) ( 1 − e M S F ) N T + N R z 2 − n T / 3 . (53) E_{\mathrm{model}}=1-(1-E_{\mathrm{core}})(1-e_{\mathrm{MSF}})^{N_T}+N_{R_z}2^{-n_T/3}. \tag*{(53)} E model = 1 − ( 1 − E core ) ( 1 − e MSF ) N T + N R z 2 − n T /3 . (53)
Equation (S42) is a resource-model estimate; the bound in equation (S41) instead requires certified errors for each synthesized rotation. The tolerance ε R \varepsilon_R ε R belongs to one rotation, whereas η \eta η caps the complete attempt. Increasing n T n_T n T lowers the estimated synthesis error but increases N T N_T N T and hence the opportunities for core and factory faults at fixed architecture; the search balances these contributions. Preparation infidelity is distinct. For the published circuit of Ref. 10, an additional angle-normalization bound is included before conditioning on success, as recorded in the numerical-detail CSVs.
Circuit success and physical output errors
Let A A A denote success in the unconditioned distributions P , Q P,Q P , Q , with p A = P ( A ) = P s p_A=P(A)=P_s p A = P ( A ) = P s and q A = Q ( A ) q_A=Q(A) q A = Q ( A ) . The conditional distributions are P A ( B ) = P ( B ∩ A ) / P s P_A(B)=P(B\cap A)/P_s P A ( B ) = P ( B ∩ A ) / P s and Q A ( B ) = Q ( B ∩ A ) / q A Q_A(B)=Q(B\cap A)/q_A Q A ( B ) = Q ( B ∩ A ) / q A . (49) gives ∣ p A − q A ∣ ≤ η |p_A-q_A|\leq\eta ∣ p A − q A ∣ ≤ η , so q A ≥ P s − η > 0 q_A\geq P_s-\eta>0 q A ≥ P s − η > 0 . The success probability can change in either direction under implementation error.
First consider p A ≥ q A p_A\geq q_A p A ≥ q A , so ∣ p A − q A ∣ = p A − q A |p_A-q_A|=p_A-q_A ∣ p A − q A ∣ = p A − q A . For any accepted event, add and subtract its ideal unconditioned probability divided by q A q_A q A . The triangle inequality gives two contributions: at most η q A \frac{\eta}{q_A} q A η from the unconditioned probability difference, and at most ∣ p A − q A ∣ q A \frac{|p_A-q_A|}{q_A} q A ∣ p A − q A ∣ from renormalizing the ideal probability. The same bound holds for q A > p A q_A>p_A q A > p A by retaining the absolute success-probability difference. Each contribution is at most η q A \frac{\eta}{q_A} q A η ; using q A ≥ P s − η > 0 q_A\geq P_s-\eta>0 q A ≥ P s − η > 0 gives
d T V ( P A , Q A ) ≤ 2 η P s − η , P s > η . (54) d_{\mathrm{TV}}(P_A,Q_A) \le\frac{2\eta}{P_s-\eta},\quad P_s>\eta. \tag*{(54)} d TV ( P A , Q A ) ≤ P s − η 2 η , P s > η . (54)
The bound in equation (S43) is capped at one and conditions a trace-preserving attempt at the classical output. Deterministic Walsh–KvN uses δ T V = η \delta_{\mathrm{TV}}=\eta δ TV = η . For BE–KvN, we use its success lower bound in equation (S43). Resource rows charge complete failed attempts through S / P s S/P_s S / P s , where S S S is the number of accepted shots and P s P_s P s the nominal success probability. The success and conditional-error values are supplied in the numerical-detail CSVs. The error budget applies per attempt.
Write P ⋆ , Q ⋆ P_{\star},Q_{\star} P ⋆ , Q ⋆ for the ideal and implemented estimator distributions, conditioned when necessary, and δ T V \delta_{\mathrm{TV}} δ TV for their total variation bound. For a real observable Y i ∈ [ a i , b i ) Y_i\in[a_i,b_i) Y i ∈ [ a i , b i ) , define R i = b i − a i R_i=b_i-a_i R i = b i − a i and m i = E P ⋆ Y i m_i=\mathbb{E}_{P_{\star}}Y_i m i = E P ⋆ Y i . Integrating event probabilities for the rescaled observable in [ 0 , 1 ) [0,1) [ 0 , 1 ) bounds its mean change; applying the same argument to Y i Y j Y_iY_j Y i Y j , of range R i j R_{ij} R ij , gives
e i = E Q ⋆ Y i − m i , ∣ e i ∣ ≤ R i δ T V , d i j = E Q ⋆ ( Y i Y j ) − E P ⋆ ( Y i Y j ) , ∣ d i j ∣ ≤ R i j δ T V . (55) \begin{aligned}
e_i&=\mathbb{E}_{Q_{\star}}Y_i-m_i,\quad|e_i|\le R_i\delta_{\mathrm{TV}},\\
d_{ij}&=\mathbb{E}_{Q_{\star}}(Y_iY_j)-\mathbb{E}_{P_{\star}}(Y_iY_j),\quad|d_{ij}|\le R_{ij}\delta_{\mathrm{TV}}. \tag*{(55)}
\end{aligned} e i d ij = E Q ⋆ Y i − m i , ∣ e i ∣ ≤ R i δ TV , = E Q ⋆ ( Y i Y j ) − E P ⋆ ( Y i Y j ) , ∣ d ij ∣ ≤ R ij δ TV . (55)
For covariance matrices Σ P , Σ Q \boldsymbol{\Sigma}_P,\boldsymbol{\Sigma}_Q Σ P , Σ Q , subtracting the products of their means yields
( Σ Q − Σ P ) i j = d i j − m i e j − m j e i − e i e j , (\boldsymbol{\Sigma}_Q-\boldsymbol{\Sigma}_P)_{ij}=d_{ij}-m_ie_j-m_je_i-e_ie_j, ( Σ Q − Σ P ) ij = d ij − m i e j − m j e i − e i e j ,
∣ ( Σ Q − Σ P ) i j ∣ ≤ ( R i j + ∣ m i ∣ R j + ∣ m j ∣ R i ) δ T V + R i R j δ T V 2 . (56) |(\boldsymbol{\Sigma}_Q-\boldsymbol{\Sigma}_P)_{ij}|\le(R_{ij}+|m_i|R_j+|m_j|R_i)\delta_{\mathrm{TV}}+R_iR_j\delta_{\mathrm{TV}}^2. \tag*{(56)} ∣ ( Σ Q − Σ P ) ij ∣ ≤ ( R ij + ∣ m i ∣ R j + ∣ m j ∣ R i ) δ TV + R i R j δ TV 2 . (56)
Apply the bias–variance identity in equation (S28) to the implemented distribution Q ⋆ Q_{\star} Q ⋆ , using the same field map A A A and spatial weight W W W as in Note 7. Bounds on its mean and covariance give conservative field-bias and variance contributions b e f f b_{\mathrm{eff}} b eff and v e f f v_{\mathrm{eff}} v eff , defined below.
Let m = ( m 1 , m 2 , m 3 ) T \mathbf{m}=(m_1,m_2,m_3)^T m = ( m 1 , m 2 , m 3 ) T be the ideal finite-circuit mean vector, whose components are defined in equation (S44), and let m r e f \mathbf{m}_{\mathrm{ref}} m ref be the corresponding moment vector of the matched continuum Radial ensemble. The field metric H \mathbf{H} H maps a moment discrepancy to squared relative field bias. Define
H = A T W A ∥ u r e f ∥ W 2 , Δ m 0 = m − m r e f (57) \mathbf{H}=\frac{\mathbf{A}^{T}\mathbf{W}\mathbf{A}}{\lVert\mathbf{u}_{\mathrm{ref}}\rVert_{\mathbf{W}}^2},\quad\Delta\mathbf{m}_0=\mathbf{m}-\mathbf{m}_{\mathrm{ref}} \tag*{(57)} H = ∥ u ref ∥ W 2 A T WA , Δ m 0 = m − m ref (57)
Equation (S44) bounds the change of each mean by its observable range times δ T V \delta_{\mathrm{TV}} δ TV . Allowing these shifts independently gives a rectangular box around the ideal mean; this is a conservative enlargement of the changes that an actual channel can produce. The squared field bias is the quadratic form with metric H \mathbf{H} H . Since this metric is positive semidefinite, the quadratic form is convex, so its maximum over the box is attained at a corner.
There are three measured moments and therefore eight corners. Let R = ( R 1 , R 2 , R 3 ) T \mathbf{R}=(R_1,R_2,R_3)^T R = ( R 1 , R 2 , R 3 ) T collect the observable ranges, and let the sign vector s ∈ { − 1 , 1 } 3 \mathbf{s}\in\{-1,1\}^3 s ∈ { − 1 , 1 } 3 select the lower or upper shift of each component. The symbol ⊙ \odot ⊙ denotes componentwise multiplication. Adding each corner shift to the nominal discrepancy Δ m 0 \Delta\mathbf{m}_0 Δ m 0 and evaluating the field metric gives equation (S47). Let Γ i j ( δ T V ) \Gamma_{ij}(\delta_{\mathrm{TV}}) Γ ij ( δ TV ) denote the covariance bound on the right-hand side of equation (S45).
b e f f 2 ( δ T V ) = max s ∈ { − 1 , 1 } 3 ( Δ m 0 + δ T V R ⊙ s ) T H ( Δ m 0 + δ T V R ⊙ s ) (58) b_{\mathrm{eff}}^2(\delta_{\mathrm{TV}})=\max_{\mathbf{s}\in\{-1,1\}^3}(\Delta\mathbf{m}_0+\delta_{\mathrm{TV}}\mathbf{R}\odot\mathbf{s})^T\mathbf{H}(\Delta\mathbf{m}_0+\delta_{\mathrm{TV}}\mathbf{R}\odot\mathbf{s}) \tag*{(58)} b eff 2 ( δ TV ) = s ∈ { − 1 , 1 } 3 max ( Δ m 0 + δ TV R ⊙ s ) T H ( Δ m 0 + δ TV R ⊙ s ) (58)
v e f f ( δ T V ) = tr ( H Σ P ) + ∑ i , j ∣ H i j ∣ Γ i j ( δ T V ) (59) v_{\mathrm{eff}}(\delta_{\mathrm{TV}})=\operatorname{tr}(\mathbf{H}\boldsymbol{\Sigma}_P)+\sum_{i,j}|H_{ij}|\Gamma_{ij}(\delta_{\mathrm{TV}}) \tag*{(59)} v eff ( δ TV ) = tr ( H Σ P ) + i , j ∑ ∣ H ij ∣ Γ ij ( δ TV ) (59)
Applying equation (S45) entrywise in the trace, including off-diagonal terms, gives the variance bound. Substitution in equation (S28) yields the smallest positive integer shot count below, where ε \varepsilon ε is the target relative RMS field error. For positive variance, a nonpositive denominator makes the bound infeasible.
S min = max { 1 , ⌈ v e f f ( δ T V ) ε 2 − b e f f 2 ( δ T V ) ⌉ } (60) S_{\min}=\max\left\{1,\left\lceil\frac{v_{\mathrm{eff}}\left(\delta_{\mathrm{TV}}\right)}{\varepsilon^{2}-b_{\mathrm{eff}}^{2}\left(\delta_{\mathrm{TV}}\right)}\right\rceil\right\} \tag*{(60)} S m i n = max { 1 , ⌈ ε 2 − b eff 2 ( δ TV ) v eff ( δ TV ) ⌉ } (60)
For deterministic Walsh–KvN, δ T V \delta_{\mathrm{TV}} δ TV equals the per-attempt error budget. The mean and covariance bounds hold for coherent or stochastic deviations within that total variation distance; the stochastic fault model enters through equation (S42). Supplementary fig:S12a fixes the wake task and physical error probability, recomputes the observable bounds and integer shot count at each error budget, and reoptimizes synthesis precision and factory counts. The complete sensitivity values are supplied as CSV. These are resource-model bounds, without additional noisy-flow simulation or per-angle synthesis certification.
This is the RMS criterion behind main-text equation (8), not a fixed-confidence guarantee. It connects implementation error to the field audit without replacing physical ROM error, transport convergence or validation of the prepared input family.
Physical resource assumptions and timing policies
Hardware parameters and provenance
Architecture and notation
The architecture comprises a core processor that runs the algorithm and a multilevel magic-state factory that supplies high-fidelity states. The core contains a data block for logical data and Pauli operations, and a buffer for magic-state growth, storage, injection, correction and scheduling. The data block includes data patches and lattice-surgery buses. The buffer includes growth and magic-state storage regions, correction and correction-storage regions, and their connecting buses.
The factory contains preparation units, distillation data patches, output-port patches, magic-state storage, correction and correction storage, buses and growth regions. Distillation units operate in parallel within each level; states expand and transfer between levels with different code distances. Because distillation is probabilistic, the relation between supply and consumption determines buffer occupancy and additional waiting.[23 ]
Let L L L be the number of distillation levels, d l d_l d l the code distance at level l l l , d c o r e d_{\mathrm{core}} d core the core code distance and u l u_l u l the number of parallel distillation units at that level. The physical error probability is p p h y s p_{\mathrm{phys}} p phys , the duration of one QEC round is t r o u n d t_{\mathrm{round}} t round , and the classical reaction time is t r e a c t t_{\mathrm{react}} t react . The compiled circuit consumes N T N_{\mathrm{T}} N T T and T† gates. The count N R z N_{\mathrm{R_z}} N R z includes arbitrary Rz _z z rotations requiring synthesis, and n T n_{\mathrm{T}} n T is the T count assigned to each Rz _z z rotation. At each code distance, one logical cycle is approximated by d d d QEC rounds. Preparation takes t p r e p = 13 t r o u n d + t r e a c t = 23 t_{\mathrm{prep}}=13\,t_{\mathrm{round}}+t_{\mathrm{react}}=23 t prep = 13 t round + t react = 23 μs, including the final cultivation decoder delay; its cycle count is rounded up at every first-level code distance. Full timing provenance and search settings are supplied as CSV.
The logical width is n L n_L n L ; it is distinct from the core tile count N c o r e N_{\mathrm{core}} N core and the physical-qubit count Q c o r e Q_{\mathrm{core}} Q core .
Pauli-based computation and magic-state injection
Following Litinski[26 ] , we write the Clifford+T circuit as a sequence of multiqubit Pauli π/8 rotations and absorb the residual Clifford into the final measurement bases; a T† gate reverses the sign of the Pauli axis. Each rotation consumes one magic state through lattice-surgery joint measurements, and the resource model approximates time and error using these T-state consumption slots. This rewriting does not set measurement, reaction or storage costs to zero. We assume auto- or post-corrected injection: a correction ancilla formed and stored in advance is measured in the appropriate basis to supply the Pauli-frame information, so later noncommuting operations need not wait for a classical correction. Its ancilla, storage, bus and classical reaction-time costs are included in the core buffer and factory correction-storage regions.[23 , 26 ] The 15-to-1 protocol detects all single and double input errors, which gives the cubic term in equation (61) ; logical lattice-surgery faults are accounted for separately.
Multilevel magic-state factories
Preparation and distillation errors
In the primary Silva-based model, first-level input magic states are prepared by Gidney cultivation. The recorded preparation circuit-qubit count, accepted-state error and acceptance probability are taken from the supplied cultivation simulation records at the corresponding physical error probability and code distance. For physical-area accounting, we use the effective preparation-unit footprint defined in equation (81) . The duration and integer logical-cycle ceiling are defined in Table S6 and the architecture notation. The recorded errors include expansion and escape; no additional entry-growth error is charged.[25 ]
Table S6. Hardware parameters. QEC, quantum error correction.
At distillation level l l l , factory-output error and acceptance probability are approximated by[23 , 29 ]
e o u t , l ≃ 35 e i n , l 3 + 7.1 e c l i f f , l ≃ 35 e i n , l 3 + 7.1 e m e m , l , (61) e_{\mathrm{out},l} \simeq35e_{\mathrm{in},l}^{3}+7.1e_{\mathrm{cliff},l}\simeq35e_{\mathrm{in},l}^{3}+7.1e_{\mathrm{mem},l}, \tag*{(61)} e out , l ≃ 35 e in , l 3 + 7.1 e cliff , l ≃ 35 e in , l 3 + 7.1 e mem , l , (61)
P l ≃ 1 − 15 e i n , l (62) P_l \simeq1-15e_{\mathrm{in},l} \tag*{(62)} P l ≃ 1 − 15 e in , l (62)
Equation (61) retains cubic input-error suppression and a Clifford/lattice-surgery contribution, approximating e c l i f f , l e_{\mathrm{cliff},l} e cliff , l by e m e m ( d l ) e_{\mathrm{mem}}(d_l) e mem ( d l ) as in Silva et al.'s numerical simplification. For acceptance, Silva et al. include a − 356 e c l i f f , l -356e_{\mathrm{cliff},l} − 356 e cliff , l correction; our equation (62) retains only 1 − 15 e i n , l 1-15e_{\mathrm{in},l} 1 − 15 e in , l . The resulting omission is quantified in the numerical-detail CSVs, while logical faults remain in output infidelity and the complete error budget. The level-indexed memory-error term is equation (64) evaluated at that level’s code distance.
Modelling assumption for coherent hardware errors
We model hardware-induced faults using effective stochastic physical and logical error probabilities after noise tailoring.[30 , 31 , 32 , 33 , 34 ] The Clifford+T approximation error of arbitrary R Z R_Z R Z rotations enters the error budget separately in equation (53) .
The logical failure probability per QEC round and the memory error over one logical cycle are
p L ( p p h y s , d ) = 0.1 ( p p h y s p t h ) ( d + 1 ) / 2 , (63) p_L(p_{\mathrm{phys}},d)=0.1\left(\frac{p_{\mathrm{phys}}}{p_{\mathrm{th}}}\right)^{(d+1)/2}, \tag*{(63)} p L ( p phys , d ) = 0.1 ( p th p phys ) ( d + 1 ) /2 , (63)
e m e m ( d ) = 1 − { 1 − p L ( p p h y s , d ) } d (64) e_{\mathrm{mem}}(d)=1-\left\{1-p_L(p_{\mathrm{phys}},d)\right\}^{d} \tag*{(64)} e mem ( d ) = 1 − { 1 − p L ( p phys , d ) } d (64)
Numerical calculations evaluate this accumulated probability directly.
Growth and connection errors between levels
Approximating distillation-output, routing and growth faults as independent gives the next-level input error in equation (65) . At each level, the distillation output in equation (61) uses the propagated e i n , l e_{\mathrm{in},l} e in , l , consistently with the general recursion of Silva et al. [equation (B13)]; we do not use the fixed e p r e p e_{\mathrm{prep}} e prep appearing in their equation (B16) for every level.
e i n , l + 1 = 1 − ( 1 − e o u t , l ) ( 1 − e p a t h , l ) ( 1 − e g r o w , l ) . (65) e_{\mathrm{in},l+1}=1-\left(1-e_{\mathrm{out},l}\right)\left(1-e_{\mathrm{path},l}\right)\left(1-e_{\mathrm{grow},l}\right). \tag*{(65)} e in , l + 1 = 1 − ( 1 − e out , l ) ( 1 − e path , l ) ( 1 − e grow , l ) . (65)
For the final transfer to the core, we set d l + 1 = d c o r e d_{l+1}=d_{\mathrm{core}} d l + 1 = d core .
Growth error is charged only when the code distance increases. For candidates with d l + 1 ≥ d l d_{l+1}\ge d_l d l + 1 ≥ d l ,
e g r o w , l = { e m e m ( d l ) , d l + 1 > d l , 0 , d l + 1 = d l . (66) e_{\mathrm{grow},l}=\begin{cases} e_{\mathrm{mem}}(d_l), & d_{l+1}>d_l, \\ 0, & d_{l+1}=d_l. \end{cases} \tag*{(66)} e grow , l = { e mem ( d l ) , 0 , d l + 1 > d l , d l + 1 = d l . (66)
Silva et al. leave growth error as a protocol-dependent function of the initial and final code distances. We instead approximate growth by one logical cycle of memory exposure at the source distance when the distance increases, and separately model routing faults through the following mean-path approximation based on the Silva bus geometry:
N p a t h , l = 1 + 11 u l ∑ j = 0 u l − 1 ∣ j − u l − 1 2 ∣ (67) N_{\mathrm{path},l}=1+\frac{11}{u_l}\sum_{j=0}^{u_l-1}\left|j-\frac{u_l-1}{2}\right| \tag*{(67)} N path , l = 1 + u l 11 j = 0 ∑ u l − 1 j − 2 u l − 1 (67)
e p a t h , l = 1 − { 1 − p L ( p p h y s , d l ) } N p a t h , l d l = 1 − { 1 − e m e m ( d l ) } N p a t h , l . (68) e_{\mathrm{path},l}=1-\left\{1-p_L\left(p_{\mathrm{phys}},d_l\right)\right\}^{N_{\mathrm{path},l}d_l}=1-\left\{1-e_{\mathrm{mem}}(d_l)\right\}^{N_{\mathrm{path},l}}. \tag*{(68)} e path , l = 1 − { 1 − p L ( p phys , d l ) } N path , l d l = 1 − { 1 − e mem ( d l ) } N path , l . (68)
Here N p a t h , l N_{\mathrm{path},l} N path , l is the path length used for error accounting, in units of a patch of code distance d l d_l d l . It is an explicit inter-unit bus approximation rather than a geometric simulation of every route, and is distinct from physical-area accounting.
For u u u equally spaced factory units with a central connection, the average path includes the final one-patch segment traversed by every output. Equation (S56) equals 1 for one unit, 1 + 11 u / 4 1+11u/4 1 + 11 u /4 for even u u u , and 1 + 11 ( u 2 − 1 ) / ( 4 u ) 1+11(u^2-1)/(4u) 1 + 11 ( u 2 − 1 ) / ( 4 u ) for odd u u u . Its mean exposure, multiplied by the actual code distance, enters equation (S57). Full bus area remains in the factory footprint. This is a geometry assumption, not a verified routing schedule.
The first-level input satisfies e i n , 1 = e p r e p e_{\mathrm{in},1}=e_{\mathrm{prep}} e in , 1 = e prep . The error of a magic state delivered from the last factory level to the core is
e M S F = 1 − ( 1 − e o u t , L ) ( 1 − e p a t h , L ) ( 1 − e g r o w , L ) (69) e_{\mathrm{MSF}}=1-\left(1-e_{\mathrm{out},L}\right)\left(1-e_{\mathrm{path},L}\right)\left(1-e_{\mathrm{grow},L}\right) \tag*{(69)} e MSF = 1 − ( 1 − e out , L ) ( 1 − e path , L ) ( 1 − e grow , L ) (69)
For N T N_T N T supplied states with the same error probability, independent injection faults give the total factory-related fault probability
E M S F = 1 − ( 1 − e M S F ) N T . (70) E_{\mathrm{MSF}}=1-\left(1-e_{\mathrm{MSF}}\right)^{N_T}. \tag*{(70)} E MSF = 1 − ( 1 − e MSF ) N T . (70)
Supply rates and waiting
The first distillation level takes O 1 = 13 O_1=13 O 1 = 13 logical cycles and subsequent levels take O l = 15 O_l=15 O l = 15 logical cycles. For a 15-to-1 protocol producing one state per unit, the output and input-consumption rates D l D_l D l and C l C_l C l are [23 ]
τ l = O l d l t r o u n d , (71) \tau_l=O_l d_l t_{\mathrm{round}}, \tag*{(71)} τ l = O l d l t round , (71)
D l = u l P l τ l , (72) D_l=\frac{u_l P_l}{\tau_l}, \tag*{(72)} D l = τ l u l P l , (72)
C l = 15 u l τ l (73) C_l=\frac{15u_l}{\tau_l} \tag*{(73)} C l = τ l 15 u l (73)
The nominal T-state consumption interval is set by the data block and the classical controller. The block uses b ∈ { 1 , 5 , 9 } b\in\{1,5,9\} b ∈ { 1 , 5 , 9 } logical cycles per state for fast, intermediate and compact layouts, respectively:
τ c o r e = max ( b d c o r e t r o u n d , t r e a c t ) , C c o r e = τ c o r e − 1 (74) \tau_{\mathrm{core}}=\max\left(b\,d_{\mathrm{core}}t_{\mathrm{round}},t_{\mathrm{react}}\right),\quad C_{\mathrm{core}}=\tau_{\mathrm{core}}^{-1} \tag*{(74)} τ core = max ( b d core t round , t react ) , C core = τ core − 1 (74)
The supply-enforced policy requires D L ≥ C c o r e D_L\ge C_{\mathrm{core}} D L ≥ C core and D l − 1 ≥ C l D_{l-1}\ge C_l D l − 1 ≥ C l , which determine the unit counts u l u_l u l . The waiting-allowed policy permits D L < C c o r e D_L<C_{\mathrm{core}} D L < C core , increasing the T-state consumption interval when final-level supply falls short. Preceding levels still meet downstream demand.
Let τ M S F = 1 / D L \tau_{\mathrm{MSF}}=1/D_L τ MSF = 1/ D L denote the mean MSF supply interval. The T-state consumption interval is τ T = max ( τ c o r e , τ M S F ) \tau_T=\max\left(\tau_{\mathrm{core}},\tau_{\mathrm{MSF}}\right) τ T = max ( τ core , τ MSF ) . For S S S accepted shots and nominal success probability p s p_s p s , the attempt time and total task time, in expectation over repeated attempts, are
t a t t e m p t = N T τ T , t t a s k = S P s t a t t e m p t . (75) t_{\mathit{attempt}}=N_T\tau_T,\quad t_{\mathit{task}}=\frac{S}{P_s}t_{\mathit{attempt}}. \tag*{(75)} t attempt = N T τ T , t task = P s S t attempt . (75)
Effective factory tile counts
The factory footprint includes preparation, distillation, output ports, correction storage, growth regions and connecting buses in the Silva-based layout used here. At the first level, we use ( 11 ( u 1 − 1 ) + 1 ) (11(u_1-1)+1) ( 11 ( u 1 − 1 ) + 1 ) inter-unit bus patches, consistent with equation (15) of Silva et al. and the ( − 10 ) (-10) ( − 10 ) constant in equation (79) , rather than ( 11 u 1 + 1 ) (11u_1+1) ( 11 u 1 + 1 ) as written in the accompanying text of Silva et al. For higher levels, the corresponding count is ( 10 ( u l − 1 ) + 1 ) (10(u_l-1)+1) ( 10 ( u l − 1 ) + 1 ) . The throughput-dependent preparation count and effective areas are specified in Equations (S65)–(S70). Correction storage at each factory level uses that level’s code distance; the core uses its own code distance:
N c o r r + b u s , l = 1.5 ⌈ t r e a c t t r o u n d d l ⌉ (76) N_{\mathrm{corr+bus},l}=1.5\left\lceil\frac{t_{\mathrm{react}}}{t_{\mathrm{round}}d_l}\right\rceil \tag*{(76)} N corr + bus , l = 1.5 ⌈ t round d l t react ⌉ (76)
Here ceil ( x ) \operatorname{ceil}(x) ceil ( x ) denotes the least integer not smaller than x x x . Following Silva et al., we allocate one effective bus patch per two correction-storage patches and permit extension on both sides of the bus (Supplementary Fig. S8). Under the ten-storage-position reference layout shown, exceeding that width would require a code distance below one with the adopted timing. We therefore add no correction-storage-induced path-length increment and do not assume a narrower reference unit. The total factory footprint is3
Q M S F = n 0 Q p r e p , u n i t ( d 1 ) + ∑ l = 1 L n l Q s u r f a c e ( d l ) , (77) Q_{MSF}=n_0Q_{\mathrm{prep,unit}}(d_1)+\sum_{l=1}^{L}n_lQ_{\mathrm{surface}}(d_l), \tag*{(77)} Q M S F = n 0 Q prep , unit ( d 1 ) + l = 1 ∑ L n l Q surface ( d l ) , (77)
n 0 = u 1 15 13 ( c p r e p P p r e p + 1 ) , (78) n_0=u_1\frac{15}{13}\left(\frac{c_{\mathrm{prep}}}{P_{\mathrm{prep}}}+1\right), \tag*{(78)} n 0 = u 1 13 15 ( P prep c prep + 1 ) , (78)
n 1 = u 1 [ 40 + 1.5 ⌈ t r e a c t t r o u n d d 1 ⌉ ] − 10 , (79) n_1=u_1\left[40+1.5\left\lceil\frac{t_{\mathrm{react}}}{t_{\mathrm{round}}d_1}\right\rceil\right]-10, \tag*{(79)} n 1 = u 1 [ 40 + 1.5 ⌈ t round d 1 t react ⌉ ] − 10 , (79)
n l = u l [ 38 + 1.5 ⌈ t r e a c t t r o u n d d l ⌉ ] − 9 , l ≥ 2. (80) n_l=u_l\left[38+1.5\left\lceil\frac{t_{\mathrm{react}}}{t_{\mathrm{round}}d_l}\right\rceil\right]-9,\quad l\ge2. \tag*{(80)} n l = u l [ 38 + 1.5 ⌈ t round d l t react ⌉ ] − 9 , l ≥ 2. (80)
Here P p r e p P_{\mathrm{prep}} P prep is preparation acceptance. The count n 0 n_0 n 0 and the correction-storage term describe steady-state effective areas rather than a detailed integer placement. For each complete Gidney preparation unit, we allocate the larger of the padded surface-code patch footprint and the recorded physical-qubit count, including auxiliary qubits:
Q p r e p , u n i t ( d 1 ) = max { Q s u r f a c e ( d 1 ) , Q G i d n e y ( d 1 ) } (81) Q_{\mathrm{prep,unit}}(d_1)=\max\{Q_{\mathrm{surface}}(d_1),Q_{\mathrm{Gidney}}(d_1)\} \tag*{(81)} Q prep , unit ( d 1 ) = max { Q surface ( d 1 ) , Q Gidney ( d 1 )} (81)
Unlike the ( 2 d 2 − 1 ) (2d^2-1) ( 2 d 2 − 1 ) physical-qubit allocation per logical patch used by Silva et al., we use the padded convention Q s u r f a c e ( d ) = 2 ( d + 1 ) 2 Q_{\mathrm{surface}}(d)=2(d+1)^2 Q surface ( d ) = 2 ( d + 1 ) 2 . The same pitch increment is applied to other factory-area components and allocated connector dimensions. The Gidney preparation-unit footprint is the maximum of this padded allocation and the recorded physical-qubit count, which includes auxiliary qubits. For the supplied circuits, Q G i d n e y ( d ) = 2 d 2 + 4 Q_{\mathrm{Gidney}}(d)=2d^2+4 Q Gidney ( d ) = 2 d 2 + 4 , so the padded allocation determines the preparation-unit footprint. Actual code distances, without padding, are used for logical-error probabilities, timing, factory-output errors and connection exposure. This effective-area allocation does not certify a fully routed concurrent layout.4
Patch-area conventions change physical allocation while leaving actual code distances, logical-error laws and service timing unchanged for a fixed architecture. Reoptimization can nevertheless select a different factory count, synthesis precision or code distance, so workload-level time changes need not follow area scaling.
Core resources and stochastic errors
Data-block and buffer area
For n L n_L n L logical data qubits, the tile counts of the three Litinski block models are listed in Table S7 .26 ^{26} 26 Each data-block formula is rounded up to an integer tile allocation.
Table S7. Core data-block models.
Following the core-buffer layout illustrated in Fig. 1 of Silva et al., we allocate 18 fixed effective tiles: four growth patches, four magic-state storage patches, two two-tile correction patches and six bus patches. Additional correction storage and associated bus area are accounted for separately using the same two-storage-per-bus allocation (Supplementary Figure S8 ). The total core tile count and physical footprint are
Supplementary Figure S8. Correction-storage and bus allocation. a–c, Schematic extensions of the Silva et al. layouts for higher-level distillation units, the core buffer and level-1 distillation units, respectively. Magenta and green tiles denote correction-storage and bus patches; storage and buses extend outward from the fixed region as needed. The area model allocates one effective bus patch per two correction-storage patches; the schematic tile counts do not depict this allocation ratio. Panel a illustrates ten storage positions per higher-level unit. All 74 selected two-level architectures use one higher-level correction-storage patch per unit, so the illustrated width is sufficient under the adopted timing and distance choices.
N c o r e = N b l k ( n L ) + 18 + 1.5 ⌈ t r e a c t t r o u n d d c o r e ⌉ , (82) N_{\mathrm{core}}=N_{\mathrm{blk}}\left(n_L\right)+18+1.5\left\lceil\frac{t_{\mathrm{react}}}{t_{\mathrm{round}}d_{\mathrm{core}}}\right\rceil, \tag*{(82)} N core = N blk ( n L ) + 18 + 1.5 ⌈ t round d core t react ⌉ , (82)
Q c o r e = N c o r e Q s u r f a c e ( d c o r e ) . (83) Q_{\mathrm{core}}=N_{\mathrm{core}}Q_{\mathrm{surface}}\left(d_{\mathrm{core}}\right). \tag*{(83)} Q core = N core Q surface ( d core ) . (83)
The tile count N c o r e N_{\mathrm{core}} N core must be distinguished from the physical-qubit count Q c o r e Q_{\mathrm{core}} Q core when constructing error volumes.
Active and idle core exposure
Core errors are evaluated for sequential consumption of one magic state per non-Clifford Pauli rotation. We retain an active volume and separately charge memory exposure over the entire core, including their overlap. This is more conservative bookkeeping than Silva's original volume model:
V a c t = b N T [ N b l k ( n L ) + 18 ] , V i d l e , b a s e = b N T N c o r e . (84) V_{\mathrm{act}}=bN_T\left[N_{\mathrm{blk}}\left(n_L\right)+18\right],\quad V_{\mathrm{idle,base}}=bN_TN_{\mathrm{core}}. \tag*{(84)} V act = b N T [ N blk ( n L ) + 18 ] , V idle , base = b N T N core . (84)
Both volumes are measured in tile–logical cycles. The active term covers the data block and fixed buffer. Correction storage contributes to the idle term and physical footprint.
The extra idle interval per consumed state is the time beyond the block’s gate interval b d c o r e t r o u n d bd_{\mathrm{core}}t_{\mathrm{round}} b d core t round :
Δ τ = max ( 0 , τ T − b d c o r e t r o u n d ) (85) \Delta\tau=\max\left(0,\tau_T-bd_{\mathrm{core}}t_{\mathrm{round}}\right) \tag*{(85)} Δ τ = max ( 0 , τ T − b d core t round ) (85)
This includes controller delay beyond the gate interval and any supply delay beyond the nominal core interval. We charge the extra idle exposure in QEC rounds:
E c o r e = 1 − { 1 − e m e m ( d c o r e ) } V a c t + V i d l e , b a s e { 1 − p L ( p p h y s , d c o r e ) } N T N c o r e Δ τ / t r o u n d (86) E_{\mathrm{core}}=1-\left\{1-e_{\mathrm{mem}}\left(d_{\mathrm{core}}\right)\right\}^{V_{\mathrm{act}}+V_{\mathrm{idle,base}}\left\{1-p_L\left(p_{\mathrm{phys}},d_{\mathrm{core}}\right)\right\}^{N_TN_{\mathrm{core}}\Delta\tau/t_{\mathrm{round}}}} \tag*{(86)} E core = 1 − { 1 − e mem ( d core ) } V act + V idle , base { 1 − p L ( p phys , d core ) } N T N core Δ τ / t round (86)
Equivalently, equation (S53) gives the core survival probability per consumed T state as
s c o r e = ( 1 − e m e m ( d c o r e ) ) b [ N b l k ( n L ) + 18 ] + ξ N c o r e , s_{\mathrm{core}}=\left(1-e_{\mathrm{mem}}\left(d_{\mathrm{core}}\right)\right)^{b\left[N_{\mathrm{blk}}\left(n_L\right)+18\right]+\xi N_{\mathrm{core}}}, s core = ( 1 − e mem ( d core ) ) b [ N blk ( n L ) + 18 ] + ξ N core ,
where ξ = τ T d c o r e t r o u n d \xi=\frac{\tau_T}{d_{\mathrm{core}}t_{\mathrm{round}}} ξ = d core t round τ T is the number of core logical cycles required to consume one T state. The overlapping charges provide conservative exposure accounting for the modeled core.
Standalone distillation protocols
We also evaluate the analytical blocks in A Game of Surface Codes and the numerical protocols in Magic State Distillation: Not as Costly as You Think. Their factory timing and error calibrations are protocol specific. All comparisons share the core data-block and buffer model, active-volume coefficient, idle-memory charge and rotation-synthesis-error budget. The full analytical catalogue is supplied as CSV; Table S8 lists the numerical catalogue. The Litinski benchmarks use their source preparation models, while the primary Silva-based estimates use Gidney cultivation.[25 , 26 , 28 ]
Table S8. Numerically evaluated standalone protocols at physical error probability 0.0005.
Analytical standalone protocols
For example, the eleven-tile 15-to-1 analytical block has factory-output error 35 e 3 35e^{3} 35 e 3 , eleven logical cycles per batch and acceptance ( 1 − e ) 15 (1-e)^{15} ( 1 − e ) 15 , where e e e is its input magic-state error.[26 ] The CSV records all six analytical candidates, including output multiplicities and acceptance approximations.
These analytical output terms assume ideal Clifford operations inside the distillation block. Factory lattice-surgery faults are charged separately through logical patch-cycle exposure.[26 , 30 ]
E l o g i c a l = 1 − [ 1 − p L ( p p h y s , d ) ] V , V = N p a t c h N c y c l e s (87) E_{\mathrm{logical}}=1-\left[1-p_{\mathrm{L}}\left(p_{\mathrm{phys}},d\right)\right]^{V},\ V=N_{\mathrm{patch}}N_{\mathrm{cycles}} \tag*{(87)} E logical = 1 − [ 1 − p L ( p phys , d ) ] V , V = N patch N cycles (87)
Here p L ( p p h y s , d ) p_{\mathrm{L}}(p_{\mathrm{phys}},d) p L ( p phys , d ) is the logical failure probability per QEC round, N p a t c h N_{\mathrm{patch}} N patch the exposed factory patches, N c y c l e s N_{\mathrm{cycles}} N cycles the QEC rounds, and V V V their product. Global A Game of Surface Codes accounting includes core and factory exposure; when a common Silva core is used in the backend comparison, its factory exposure is charged separately without duplicating the core term.
Numerically evaluated protocol catalogue
Table S8 gives the supplied density-matrix catalogue at p p h y s = 5 × 10 − 4 p_{\mathrm{phys}}=5\times10^{-4} p phys = 5 × 1 0 − 4 . The tuple ( d X , d Z , d m ) (d_X,d_Z,d_m) ( d X , d Z , d m ) specifies the protocol code distances. Mean batch times are expressed in QEC rounds, rather than logical cycles.[28 ]
For the first-level protocols, acceptance can be inferred from the mean production time. Failure and retry overheads of the two-level protocols are already included in their tabulated mean batch times, so acceptance is not applied a second time in the supply rate. This is not an assumption of 100% acceptance; a separate overall two-level success probability is not supplied. Protocols ending in 20-to-4 produce four states per accepted batch, which is included in their output rate.
The Not as Costly factory-output-error catalogue already includes its noisy logical measurements, storage and error propagation at the specified d X d_X d X , d Z d_Z d Z and d m d_m d m distances. Their tabulated mean production times already include failures and retries. A second internal factory logical-error term or a second two-level acceptance factor would therefore double count these contributions. External connection and patch-growth errors remain separate in equations (88) –(91) .[28 ]
External connections for standalone factories
Standalone distillation blocks connect to the common core through an output face of length L o u t L_{\mathrm{out}} L out and a bridge of length L b r i d g e L_{\mathrm{bridge}} L bridge (Supplementary Figure S7 ). With u u u parallel factory units, the complete path span is
Supplementary Figure S7. Connections between standalone factories and the core. a, Example first- and second-level block geometry defining the output-face length L o u t L_{\mathrm{out}} L out and bridge length L b r i d g e L_{\mathrm{bridge}} L bridge . b, A shared route connects u u u factory units; its total span includes the width of the output patches. c, Even and odd unit counts give the centre-referenced distances used by the mean-path approximation. Lengths are measured in physical lattice-site spacings, and n o u t n_{\mathrm{out}} n out denotes the output multiplicity of one unit. Added path and bridge rectangles are charged separately from the published standalone block geometry. This effective placement model does not certify a fully routed concurrent schedule.
L p a t h = L o u t ( u − 1 ) + n o u t d o u t (88) L_{\mathrm{path}} = L_{\mathrm{out}}(u-1) + n_{\mathrm{out}}d_{\mathrm{out}} \tag*{(88)} L path = L out ( u − 1 ) + n out d out (88)
Here n o u t n_{\mathrm{out}} n out is the number of output patches per one factory unit and d o u t d_{\mathrm{out}} d out is their code distance. The span L p a t h L_{\mathrm{path}} L path includes the width of the output patches (Supplementary Figure S7 ). For uniformly placed units and a central core connection, the mean route length is approximated by
L ‾ p a t h = { u L p a t h 4 ( u − 1 ) , u ≥ 2 , even , ( u + 1 ) L p a t h 4 u , u ≥ 3 , odd , 0 , u = 1. (89) \overline{L}_{\mathrm{path}} =
\begin{cases}
\frac{uL_{\mathrm{path}}}{4(u-1)}, & u \ge2,\ \text{even},\\
\frac{(u+1)L_{\mathrm{path}}}{4u}, & u \ge3,\ \text{odd},\\
0, & u=1.
\end{cases} \tag*{(89)} L path = ⎩ ⎨ ⎧ 4 ( u − 1 ) u L path , 4 u ( u + 1 ) L path , 0 , u ≥ 2 , even , u ≥ 3 , odd , u = 1. (89)
Both lengths L ‾ p a t h \overline{L}_{\mathrm{path}} L path , L b r i d g e L_{\mathrm{bridge}} L bridge are measured in physical lattice-site spacings. Dividing by the output distance d o u t d_{\mathrm{out}} d out converts them to effective tile lengths for the error model:
N p a t h = L ‾ p a t h + L b r i d g e d o u t , e p a t h = 1 − { 1 − e m e m ( d o u t ) } N p a t h . (90) N_{\mathrm{path}} = \frac{\overline{L}_{\mathrm{path}}+L_{\mathrm{bridge}}}{d_{\mathrm{out}}}, \qquad e_{\mathrm{path}} = 1-\left\{1-e_{\mathrm{mem}}(d_{\mathrm{out}})\right\}^{N_{\mathrm{path}}}. \tag*{(90)} N path = d out L path + L bridge , e path = 1 − { 1 − e mem ( d out ) } N path . (90)
A state that must expand to the core code distance also incurs the growth error e g r o w e_{grow} e g r o w , giving
e M S F = 1 − ( 1 − e o u t ) ( 1 − e p a t h ) ( 1 − e g r o w ) (91) e_{MSF}=1-\left(1-e_{out}\right)\left(1-e_{path}\right)\left(1-e_{grow}\right) \tag*{(91)} e M S F = 1 − ( 1 − e o u t ) ( 1 − e p a t h ) ( 1 − e g r o w ) (91)
The standalone factory and Silva layouts use their own declared mean output-to-core connection geometry. Only the one-output mean distance enters connection error; the full bus and bridge allocation remains in area. These connections are not added twice to the internal factory footprint. Growth, core exposure and synthesis policies remain common across comparisons.
Published component dimensions supply the standalone factory area; their allocated extents use the same replacement of actual code distance d d d by pitch d + 1 d+1 d + 1 . Define L p a t h , a r e a L_{path,area} L p a t h , a r e a and L b r i d g e , a r e a L_{bridge,area} L b r i d g e , a r e a as the horizontal span and bridge length obtained by applying that replacement to their component dimensions. The added allocated rectangles have area L p a t h , a r e a ( d o u t + 1 ) + u n o u t ( d o u t + 1 ) L b r i d g e , a r e a L_{path,area}(d_{out}+1)+u\,n_{out}(d_{out}+1)L_{bridge,area} L p a t h , a r e a ( d o u t + 1 ) + u n o u t ( d o u t + 1 ) L b r i d g e , a r e a . Output multiplicity n o u t n_{out} n o u t is per factory unit, and u u u counts all units. These physical rectangles receive no additional factor of two. The actual lengths in equations (S77)–(S80), and hence connection errors and timing, remain unchanged. This effective allocation is not a verification of a fully routed concurrent schedule.
Search scope and execution timing
Architecture frontier. We enumerate core blocks and distances, factory distances and levels, and final-level unit counts from one through mean-supply saturation. Preceding levels meet downstream demand. Each architecture reevaluates preparation, distillation, growth, mean-path and core errors. For Figure 3 c we retain nondominated architectures without a physical-qubit cap in four quantities, physical qubits Q Q Q , effective core error per consumed T state e c o r e ( T ) = 1 − s c o r e e_{core}^{(T)}=1-s_{core} e cor e ( T ) = 1 − s cor e , final factory-output error e M S F e_{MSF} e M S F , and T-state consumption interval τ T \tau_T τ T . The filter compares the error quantities through their equivalent negative log-survival exponents. An architecture dominates another if it is no worse in all four quantities and is better in at least one. Including τ T \tau_T τ T retains faster geometries even when their area or error is larger.
Uncapped fixed-length search. The initial pass uses odd core distances 13–49, 45–200 T gates per arbitrary rotation and one or two distillation levels. First-level distances are restricted to the cultivation calibration table; higher-level distances range from the first-level distance to the core distance. A feasible initial design supplies a runtime bound. The subsequent fastest-design search removes the physical-qubit cap and extends code distances beyond 49, retaining the lower distance limit, the synthesis range and the two-level limit.
Runtime exclusion. Let n m i n n_{min} n min be the smallest integer T count per arbitrary rotation allowed by the synthesis-only error bound, including source-angle conversion error, and the lower search limit of 45. The compiled T-count lower bound is N T , m i n = N T n a t i v e + n m i n N R z N_{T,min}=N_{T}^{native}+n_{min}N_{R_z} N T , min = N T na t i v e + n min N R z , where N T n a t i v e N_{T}^{native} N T na t i v e counts native T/T† gates and N R z N_{R_z} N R z counts arbitrary rotations. Let t b e s t t_{best} t b es t be the fastest feasible attempt time already found. Since the attempt time is at least N T , m i n τ c o r e N_{T,min}\tau_{core} N T , min τ cor e , a core-distance branch is excluded when N T , m i n τ c o r e > t b e s t N_{T,min}\tau_{core}>t_{best} N T , min τ cor e > t b es t . For a fixed factory geometry, the same test uses the supply-aware interval τ T \tau_T τ T instead of τ c o r e \tau_{core} τ cor e . The weaker bound obtained by using one code-distance-scaled QEC round per T state sets a finite upper core distance, rounded upward to an odd integer. These are optimization bounds, separate from the external time cap used for capacity calculations.
Factory enumeration. At fixed distances, final-stage unit counts increase from one until mean supply meets core demand; lower stages are sized to supply the next stage. The search can stop earlier if the geometry is invalid or its optimistic factory-error contribution already exceeds the entire per-attempt budget. Beyond supply saturation, more units cannot shorten the consumption interval and increase area and connection exposure. The supplied Source/RE/uncapped.py implements these checks; its saved search audits record the distance bound and checked geometry counts.
Joint selection. For Supplementary Figure S10 , we compare the retained architectures at 45–125 synthesis T gates per arbitrary rotation and allow final-level supply delays. The error model includes the resulting memory exposure. For each architecture and synthesis cost, the expected time per accepted shot bounds the candidate number of steps. Integer bisection finds the largest step count within that time bound and the per-attempt error budget. The physical-qubit cap is 10 5 10^5 1 0 5 or 10 6 10^6 1 0 6 and the time cap is 86400 / 815 s per accepted shot. We break step-count ties by shorter expected time, then fewer qubits and then fewer synthesis T gates per rotation. The saved boundary checks test N t m a x N_t^{max} N t ma x and N t m a x + 1 N_t^{max}+1 N t ma x + 1 against the enumerated geometries. Fixed-length estimates minimize footprint or runtime at the prescribed propagation length and supply policy under the same implementation-error bound. Minimum-qubit breaks ties by runtime, and minimum runtime breaks ties by footprint. Supplementary fig:S10b reports τ c o r e / τ T \tau_{\mathrm{core}} / \tau_T τ core / τ T for the maximum-step selections. Unity denotes core-limited execution; values below unity denote factory-limited execution.
Execution timing. Following Litinski26, we represent the Clifford+T circuit by Pauli-product π / 8 \pi/8 π /8 rotations and absorb the residual Clifford into the final measurement bases or classical interpretation. Equation (75) estimates online time from T-state consumption intervals, including input preparation on every attempt. Separate Clifford, measurement, reset and input–output latencies, factory startup, stochastic arrivals and offline fitting lie outside this model. A fully routed schedule would resolve these operations and latencies explicitly.
Common inputs are the matrix-product-state-prepared wake distribution, error budget 10 − 4 10^{-4} 1 0 − 4 per attempt, 1 1 1 μ \mu μ s QEC rounds, a 10 10 10 μ \mu μ s reaction time, padded patch areas 2 ( d + 1 ) 2 2(d+1)^2 2 ( d + 1 ) 2 , thirteen preparation rounds with a cycle ceiling, and input-only distillation rejection. Supplementary Fig. S10 compares both qubit caps using 815 accepted Walsh–KvN shots and a one-day total time budget. The 10 6 10^6 1 0 6 -qubit cap also applies to Supplementary Fig. S12a. Supplementary Figs. S11 and S9 and both Fig. 5c and Fig. 5e use uncapped estimates. The waiting control in Supplementary Fig. S11 and Supplementary Fig. S9 fixes one final-level factory unit. The supply-enforced reference optimizes the required count.
Resource comparisons and circuit-specific supply optimization
The classical reference uses the same MPS-prepared continuum input. A midpoint characteristic quadrature and its refinement checks are supplied in the Source Data. The protected estimates quantify the declared quantum task under the specified resource model.
Walsh–KvN and BE–KvN (Supplementary Note 11) share the wake input, prediction interval and reconstructed-field-error target. The published LCHS–KvN reference below has a different drift, input and time, so its runtime is not a matched wake speedup ratio.[10 ]
The published nonlinear circuit uses 128 cells, ten steps (τ = 0.1 \tau=0.1 τ = 0.1 , t 1 = 1 t_1=1 t 1 = 1 ) and 25 algorithm qubits; seven initial Hadamards are already in its 108,018,555 single-target multi-controlled gates. One shared clean work qubit gives 26 logical qubits. Controlled X uses the same Khattar–Gidney construction as Walsh–KvN; controlled H follows by target-basis conjugation. Zindorf–Bose decomposition retains the conditional determinant phase for phase and general controlled unitaries.[11 , 35 ] The work qubit is restored after each primitive. The resulting inventory is 307,305,994 arbitrary Rz, 5,914,351,326 native T/T† and 8,781,048,741 Clifford gates, without cancellation between primitives. A primitive-level audit checks 1,566 distinct primitives, with maximum state error 1.25 × 10 − 13 1.25 \times10^{-13} 1.25 × 1 0 − 13 and work leakage 8.11 × 10 − 16 8.11 \times10^{-16} 8.11 × 1 0 − 16 ; this is not a new full-circuit state simulation. The published example reports a success probability of approximately 0.88 (ref. 10). The resource totals use the archived nominal input 0.87628 in Data/RE/workloads_v38.json; its additional digits have no independently traceable derivation in this package. Its fig:4c total applies the Walsh–KvN count of 815 accepted shots to that published task.
Factory supply as a separate design variable
Mean waiting reduces the required supply hardware while adding core-memory exposure (Supplementary Fig. S11a). Its control fixes one final-level unit; supply enforced optimizes the required count. Each policy independently selects factory/core distances and synthesis precision. Changing native T count can therefore change both area and runtime; the controlled rotation inventory remains fixed (Supplementary Fig. S11b–d).
Allowing mean waiting reduces the minimum-qubit design at the physical error probabilities illustrated in Supplementary Figure S11 , while the change in per-attempt time depends on the selected architecture. Table S11 gives representative designs; the CSVs retain all physical error probabilities, synthesis precisions and factory counts.
Table S11. Selected circuit-workload sensitivity controls. Changes compare waiting-allowed and supply-enforced minimum-qubit optima. Core T counts include the reoptimized synthesis workload.
Matched field prediction and the plotted selection objectives
Table S9 gives four representative designs for the primary Walsh–KvN index-shift method and matched BE–KvN. The full CSV also includes the alternative Walsh index-shift decomposition at the same field-error target, together with architecture choices, per-attempt times and synthesis precisions.
Table S9. Matched-field designs at physical error probability 5 × 10 − 4 5 \times10^{-4} 5 × 1 0 − 4 .
Table S10 separates minimum-qubit, capped-fastest and uncapped-fastest designs for the published LCHS–KvN reference. Its totals use the Walsh–KvN accepted-shot count and its own nominal success probability for a different physical task. Uncapped-fastest BE–KvN designs switch to two-level distillation at physical error probability 6 × 10 − 4 6 \times10^{-4} 6 × 1 0 − 4 ; the published LCHS–KvN reference switches at 5 × 10 − 4 5 \times10^{-4} 5 × 1 0 − 4 . These discrete architecture changes move both above the plotted qubit range. The CSVs retain all selected designs and transitions.
Table S10. Published LCHS–KvN reference at physical error probability 5 × 10 − 4 5 \times10^{-4} 5 × 1 0 − 4 for 815 accepted shots.
Circuit-specific supply sensitivity
At fixed architecture, area changes leave protection and time unchanged; reoptimization changes distances, supply and synthesis. Table S11 samples the inventory sweep with input, width, arbitrary rotations, shots and error budget fixed. These are resource controls, not new flow simulations. The associated source CSVs retain the complete sweep and wake gate, shot, error and timing data.
Source data retain gate programs, covariance, shot allocation, architectures and workload sweeps. Offline costs follow Note 9.
Reference provenance and comparison scope
Figure 4 a and Supplementary Figure S12 use main-manuscript reference numbers, independently of the Supplementary list. The map below distinguishes published hardware proxies and protected estimates from this study’s backend calculations; complete citations and definitions are given in the Source Data. Hardware error metrics have different calibration meanings, and STAR/RSA coordinates are illustrative. The Not-as-Costly minimum at physical error probability 0.0009 is conditional on nine available factories out of thirteen candidates. Backend geometry, factory-output errors and timing remain protocol-specific.
Table 2.
BE–KvN circuit
We reconstruct the polynomial generator block encoding of Guseynov et al.[8 ] for the 64-cell R e = 100 \mathrm{Re}=100 Re = 100 wake using their Corollary 1 coordinate-amplitude oracle and call the resulting circuit BE–KvN. BE–KvN and Walsh–KvN use the same prepared initial state, three observables at the final time and continuum Radial field reference. Walsh–KvN advances in 33 transport steps; BE–KvN evolves both the radius and phase registers in one exponential over the full window.
For N x N_x N x radial cells on the unit interval, let h = 1 / N x h=1/N_x h = 1/ N x and r j = ( j + 1 / 2 ) h r_j=(j+1/2)h r j = ( j + 1/2 ) h for j = 0 , … , N x − 1 j=0,\ldots,N_x-1 j = 0 , … , N x − 1 . The physical coordinate operator is X = diag ( r j ) X=\operatorname{diag}(r_j) X = diag ( r j ) . The projected block of the coordinate oracle is X / β X X/\beta_X X / β X , where β X \beta_X β X is 2 \sqrt{2} 2 times the Euclidean norm of the vector of cell centres. The radial branch uses a degree-three transformation of this signal. Averaging g ( r ) = ( 1 + 3 h 2 / 4 ) r − r 3 g(r)=(1+3h^2/4)r-r^3 g ( r ) = ( 1 + 3 h 2 /4 ) r − r 3 over adjacent cell centres gives the drift r − r 3 r-r^3 r − r 3 at their shared face. This yields H r = i Ω H_r=i\Omega H r = i Ω with the cubic face stencil Ω \Omega Ω in equation (S12) and zero endpoint coupling.
The joint generator is H = H r ⊗ I − κ X 2 ⊗ ( n X + n Y ) H=H_r\otimes I-\kappa X^2\otimes(n_X+n_Y) H = H r ⊗ I − κ X 2 ⊗ ( n X + n Y ) , where I I I is the identity, n X n_X n X and n Y n_Y n Y project the two phase qubits onto their occupied states and κ = 3.2626 \kappa=3.2626 κ = 3.2626 is the phase rate in normalized time. The phase branch uses a degree-two transformation of the same coordinate oracle. The joint block normalization is α = 17,695 \alpha=17{,}695 α = 17 , 695 . The normalized duration of the prediction window is T = 0.90901 T=0.90901 T = 0.90901 . A degree-16,197 outer exponential approximation and one round of oblivious amplitude amplification[36 ] give 48,591 block calls. We verified all 256 projected generator columns. The exponential error is bounded by 10 − 4 10^{-4} 1 0 − 4 on the normalized signal interval.
Gate counts are per attempt, including preparation and readout. Calls include controlled uses. T equivalents add native T/T† and 109 T gates per arbitrary RZ. Bias is noiseless relative field error at the final time. Shot counts target 1% relative RMS field error with a separate 10 − 4 10^{-4} 1 0 − 4 error budget. For circuits with acceptance checks, shot counts refer to accepted shots. The outer inventory uses a circuit template and counts every undetermined phase as an arbitrary rotation; individual outer phase angles were not synthesized. Bias and shots use the certified polynomial evaluated on the finite generator spectrum. For comparison, the Walsh–KvN circuit uses 4,803 arbitrary rotations and 7,392 native T/T† gates on nine logical qubits (Table S3), or 377,223 T gates per attempt at the 77 T gates per rotation of its fastest design.
Direct diagonalization of the 64-cell joint generator in normalized time gives a spectral norm of 26.372. The block normalization α = 17,695 \alpha=17{,}695 α = 17 , 695 is 671.00 times this norm. This ratio quantifies normalization overhead within the reconstruction; oracle gates, synthesis, sampling and hardware scheduling also contribute to the runtime ratio. For a normalization sensitivity, replace the normalization by this spectral norm while keeping the effective gate inventory per call, encoded width, synthesis precision, shots, success probability and selected architecture fixed. A Jacobi–Anger expansion at the archived degree 16,197 has an absolute Chebyshev-tail bound of 2.11 × 10 − 5 2.11\times10^{-5} 2.11 × 1 0 − 5 . At the reduced normalization, degree 37 meets the same bound and requires 111 calls, compared with 48,591 calls originally. Scaling the call-linear workload gives about 2.55 × 10 6 2.55\times10^6 2.55 × 1 0 6 T gates and 14.4 h at the fixed 986,060-qubit architecture. This is a conditional workload sensitivity, not a compiled alternative or lower bound. Fixed preparation and readout are included in the original effective per-call average. The calculation and all assumptions are supplied in Source Data.
At fixed polynomial degree, derivative order and term count, a generator-block call admits an O ( n 2 ) O(n^2) O ( n 2 ) elementary-gate upper bound.[8 ] For the joint evolution, normalized duration T T T and exponential tolerance ε exp \varepsilon_{\exp} ε e x p give a sufficient outer degree d exp = O ( α T + log ( 1 / ε exp ) ) 36 d_{\exp}=O(\alpha T+\log(1/\varepsilon_{\exp}))^{36} d e x p = O ( α T + log ( 1/ ε e x p ) ) 36 . The full evolution uses O ( d exp n 2 ) O(d_{\exp}n^2) O ( d e x p n 2 ) elementary gates before rotation synthesis and repetitions, with n = log 2 N x n=\log_2N_x n = log 2 N x index qubits. The normalization depends on grid resolution. We use K K K below for the physical drift degree, separately from the exponential approximation degree d exp d_{\exp} d e x p .
In the Fig. 4 surface-code model at physical error probability 5 × 10 − 4 5\times10^{-4} 5 × 1 0 − 4 and error budget 10 − 4 10^{-4} 1 0 − 4 per attempt, the minimum-qubit BE–KvN design uses 146,436 physical qubits and takes 3,198 days, and the fastest uncapped design uses 986,060 qubits and takes 262.9 days (Table S9). The corresponding Walsh–KvN designs use 53,382 and 559,942 qubits and take 20.07 h and 1.62 h. These estimates include repetitions and use the T-paced timing model. Both matched circuits include memory exposure during classical reaction and factory waits.
Polynomial degree and coordinate normalization
For normalized grid coordinates x i x_i x i satisfying ∑ i x i 2 = 1 \sum_i x_i^2=1 ∑ i x i 2 = 1 and the stated boundedness and parity conditions for transformed polynomials, Guseynov et al.[8 ] give O ( N t e r m K n 2 + N t e r m m n ) O\left(N_{\mathrm{term}}Kn^2+N_{\mathrm{term}}mn\right) O ( N term K n 2 + N term mn ) elementary gates per generator-block call. Here n = log 2 N x n=\log_2N_x n = log 2 N x is the coordinate-register width, K K K the maximum polynomial degree, m m m the maximum derivative order and N t e r m N_{\mathrm{term}} N term the number of operator terms. At fixed m m m and N t e r m N_{\mathrm{term}} N term , polynomial degree enters multiplicatively without changing the quadratic exponent of log 2 N x \log_2N_x log 2 N x . The detailed time-evolution bound also retains the generator normalization and evolution duration T T T . Our fixed physical grid violates this normalization; coordinate rescaling introduces the coefficient dependence quantified below.
For the unit interval used here, h = 1 / N x h=1/N_x h = 1/ N x and r j = ( j + 1 / 2 ) h r_j=(j+1/2)h r j = ( j + 1/2 ) h are the cell width and centre coordinate, with j = 0 , … , N x − 1 j=0,\ldots,N_x-1 j = 0 , … , N x − 1 . These physical coordinates do not satisfy the normalization: ∑ j r j 2 = N x / 3 − 1 / ( 12 N x ) \sum_j r_j^2=N_x/3-1/(12N_x) ∑ j r j 2 = N x /3 − 1/ ( 12 N x ) , which equals 21.33 on 64 cells and grows with the grid. In the coordinate-oracle convention used in this reconstruction, the signal is z = r / β X z=r/\beta_X z = r / β X , where
β X = ( 2 ∑ j = 0 N x − 1 r j 2 ) 1 / 2 = ( 2 N x 3 − 1 6 N x ) 1 / 2 = Θ ( N x 1 / 2 ) . (92) \beta_X=\left(2\sum_{j=0}^{N_x-1}r_j^2\right)^{1/2}=\left(\frac{2N_x}{3}-\frac{1}{6N_x}\right)^{1/2}=\Theta\left(N_x^{1/2}\right). \tag*{(92)} β X = ( 2 j = 0 ∑ N x − 1 r j 2 ) 1/2 = ( 3 2 N x − 6 N x 1 ) 1/2 = Θ ( N x 1/2 ) . (92)
Let f K ( r ) = ∑ k = 0 K a k r k f_K(r)=\sum_{k=0}^{K}a_kr^k f K ( r ) = ∑ k = 0 K a k r k be the physical drift. Hold K ≥ 1 K\ge1 K ≥ 1 fixed as the grid is refined, with bounded coefficients a k a_k a k and a highest-degree coefficient that tends to a nonzero constant. Direct degree-K K K transformation of the coordinate signal requires the normalized polynomial
P K ( z ) = f K ( β X z ) C K , C K = max ∣ z ∣ ≤ 1 ∣ f K ( β X z ) ∣ . (93) P_K(z)=\frac{f_K(\beta_Xz)}{C_K},\qquad C_K=\max_{|z|\le1}\left|f_K(\beta_Xz)\right|. \tag*{(93)} P K ( z ) = C K f K ( β X z ) , C K = ∣ z ∣ ≤ 1 max ∣ f K ( β X z ) ∣ . (93)
The normalization C K C_K C K must bound the polynomial on the full signal interval [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , including values outside the occupied physical signal range. For this direct polynomial construction, C K = Θ ( β X K ) = Θ ( N x K / 2 ) C_K=\Theta\left(\beta_X^K\right)=\Theta\left(N_x^{K/2}\right) C K = Θ ( β X K ) = Θ ( N x K /2 ) . Odd and even parts can be transformed separately when required. Combining the coefficient block with the physical first-derivative block, whose normalization is 1 / h = N x 1/h=N_x 1/ h = N x , gives
λ K = C K h = Θ ( N x 1 + K / 2 ) . (94) \lambda_K=\frac{C_K}{h}=\Theta\left(N_x^{1+K/2}\right). \tag*{(94)} λ K = h C K = Θ ( N x 1 + K /2 ) . (94)
The cubic drift therefore gives Θ ( N x 5 / 2 ) \Theta\left(N_x^{5/2}\right) Θ ( N x 5/2 ) , whereas a fourth-degree drift with the same coefficient assumptions gives Θ ( N x 3 ) \Theta\left(N_x^3\right) Θ ( N x 3 ) . The grid-corrected cubic used above has a bounded linear coefficient and a fixed cubic coefficient, so it has the same leading scaling. The degree-dependent factor is a normalization cost of this encoding route; it is not growth of the physical drift on the fixed interval.
In plain terms, the oracle stores the list of cell coordinates as a normalized quantum state, so every coordinate is divided by the factor in equation (S81) before loading. On 64 cells this factor is 6.53, and the cubic term of the drift is then represented at about 1/279 of its physical size. Restoring that scale raises the block normalization, and hence the number of block calls, by a factor of this order, which grows with the cell count and with the drift degree. This dependence is not specific to uniform grids: it holds for fixed-interval grid families retaining a nonvanishing fraction of cells at order-one distances from the chosen origin. This reconstruction keeps the origin at zero radius, where the cubic drift and the quadratic phase coupling have definite parity.
At fixed term count and first-derivative stencil, the direct construction requires O ( K n 2 ) O(Kn^2) O ( K n 2 ) elementary gates per generator-block call. A sufficient outer simulation degree is d exp = O ( λ K T + log ( 1 ε exp ) ) d_{\exp}=O\left(\lambda_KT+\log\left(\frac{1}{\varepsilon_{\exp}}\right)\right) d e x p = O ( λ K T + log ( ε e x p 1 ) ) , giving the upper bound
G e x p = O ( K n 2 [ T N x 1 + K 2 + log ( 1 ε e x p ) ] ) . (95) G_{\mathrm{exp}}=\mathcal{O}\left(K n^2\left[T N_x^{1+\frac{K}{2}}+\log\left(\frac{1}{\varepsilon_{\mathrm{exp}}}\right)\right]\right). \tag*{(95)} G exp = O ( K n 2 [ T N x 1 + 2 K + log ( ε exp 1 ) ] ) . (95)
Here G e x p G_{\mathrm{exp}} G exp counts elementary gates before rotation synthesis and measurement repetitions. At fixed degree, duration and tolerance it is O ( N x 1 + K / 2 ( log 2 N x ) 2 ) \mathcal{O}\left(N_x^{1+K/2}\left(\log_2 N_x\right)^2\right) O ( N x 1 + K /2 ( log 2 N x ) 2 ) . The leading dependence remains linear in the evolution duration T T T ; it is the power of the grid size that increases with K K K . This bound describes the direct coordinate-amplitude reconstruction used in Table S12 and is not a lower bound for other block encodings or polynomial implementations. For comparison, the Walsh-support bound in Supplementary Note 3 gives O ( ( log 2 N x ) K ) \mathcal{O}\left((\log_2 N_x)^K\right) O ( ( log 2 N x ) K ) transport rotations per step at fixed model dimension and drift degree. These comparisons concern different components of the propagation cost; full resource estimates also include evolution length, approximation accuracy, rotation synthesis and repetitions. They provide complementary motivations for retaining physically accurate low-degree models.
Table S12. Logical inventory of the matched 64-cell BE–KvN circuit.
Three-coordinate Fourier–Walsh wake circuit
Polar representation and transport
Set x = r cos θ x=r\cos\theta x = r cos θ and y = r sin θ y=r\sin\theta y = r sin θ . Write the planar drift as a matrix acting on ( x , y ) (x,y) ( x , y ) , with entries a , b a,b a , b in the first row and c , d c,d c , d in the second. Each entry is affine in z z z . Define A p = ( a + d ) / 2 A_p=(a+d)/2 A p = ( a + d ) /2 , B p = ( a − d ) / 2 B_p=(a-d)/2 B p = ( a − d ) /2 , C p = ( b + c ) / 2 C_p=(b+c)/2 C p = ( b + c ) /2 and D p = ( c − b ) / 2 D_p=(c-b)/2 D p = ( c − b ) /2 . The polar radial drift is r ( A p + B p cos 2 θ + C p sin 2 θ ) r(A_p+B_p\cos2\theta+C_p\sin2\theta) r ( A p + B p cos 2 θ + C p sin 2 θ ) and the angular drift is D p + C p cos 2 θ − B p sin 2 θ D_p+C_p\cos2\theta-B_p\sin2\theta D p + C p cos 2 θ − B p sin 2 θ . The third drift is q ( z ) + r 2 ( G + H cos 2 θ + I sin 2 θ ) q(z)+r^2(G+H\cos2\theta+I\sin2\theta) q ( z ) + r 2 ( G + H cos 2 θ + I sin 2 θ ) , where q q q is the quadratic self term and G , H , I G,H,I G , H , I are the constant and second-harmonic coefficients of the original planar quadratic term. This is the exact change of coordinates for r > 0 r>0 r > 0 . The radial drift vanishes at r = 0 r=0 r = 0 and the cell centres are at positive radius.
For polar cell amplitudes we use the square root of the probability density relative to d r d θ d z dr\,d\theta\,dz d r d θ d z , so the Cartesian density contributes with Jacobian r r r . With f \mathbf{f} f the polar drift, the Koopman–von Neumann generator is − i ( f ⋅ ∇ + ∇ ⋅ f 2 ) -i\left(\mathbf{f}\cdot\nabla+\frac{\nabla\cdot\mathbf{f}}{2}\right) − i ( f ⋅ ∇ + 2 ∇ ⋅ f ) . In the Fourier basis, multiplication by exp ( ± 2 i θ ) \exp(\pm2i\theta) exp ( ± 2 i θ ) shifts the signed mode index by ± 2 \pm2 ± 2 . In the angular derivative, the coupling between modes k k k and k ′ k' k ′ has the mean momentum ( k + k ′ ) / 2 (k+k')/2 ( k + k ′ ) /2 multiplying the corresponding angular drift coefficient. We truncate the Fourier matrix at the ends of the signed mode range and suppress connections across the truncation boundary.
Disjoint matchings split the radial and third-coordinate differences and the angular mode shifts into two-state or four-state blocks. The coefficients for a layer are at most quadratic in the index values, and a fixed Pauli product specifies each block. We use Walsh rotations for these coefficients and a symmetric product of the matching layers. The growth and late runs use order-4 and order-2 radial and third-coordinate stencils, with 89 and 49 matching applications per time step, respectively. The angle register is transformed before the first step and transformed back for the final joint position measurement.
Boundary flags prevent cyclic address arithmetic from adding couplings absent from the finite generator. For example, an eight-cell odd layer retains pairs ( 1 , 2 ) (1,2) ( 1 , 2 ) , ( 3 , 4 ) (3,4) ( 3 , 4 ) and ( 5 , 6 ) (5,6) ( 5 , 6 ) , but excludes the formal wrap pair ( 7 , 0 ) (7,0) ( 7 , 0 ) . The radial benchmark automatically gives this wrap pair zero angle because the drift vanishes at the endpoint. In the three-coordinate calculation, the drift generally does not vanish at the artificial radial and shift truncation limits, so invalid wrap pairs are explicitly suppressed. The angle is periodic, but the retained Fourier modes are truncated rather than cyclically connected: a mode shift must not connect the two ends of the retained range. A matching layer can combine a radial or shift-coordinate pair with an angular-mode pair, requiring up to two independent validity flags. The flags are computed from the pair-address bits before the Walsh rotations, which cancel the angles of invalid pairs. The pair-address bits are unchanged by the rotations, so reversing the flag computation restores both flags to zero after the layer. The shared one-clean arithmetic ancilla is separately restored after each multi-controlled X operation and reused for both checks and index shifts. No measurement or reset is required.
For Table 1 , let n n n be the maximum coordinate-register width, measured in qubits. At fixed stencil order and MPS bond dimension, the quadratic drift above gives a grid-independent number of matching layers. Their at-most-quadratic angle coefficients require at most O ( N t n 2 ) O(N_t n^2) O ( N t n 2 ) gates before rotation synthesis over N t N_t N t steps, using the bounded-degree construction of Note 3 and the selected quadratic-size index arithmetic. Endpoint Fourier transforms add O ( n 2 ) O(n^2) O ( n 2 ) , and fixed-bond preparation adds O ( n ) O(n) O ( n ) . This conditional circuit-structure bound does not bound the grid-dependent number of steps needed for fixed field accuracy.
Both inputs are Cartesian Gaussians, with widths 0.12, 0.12 and 0.09, integrated over polar cells with the radial Jacobian. The growth centre at time 29.8 is ( − 0.74116 , − 0.088441 , 0.47820 ) (-0.74116,-0.088441,0.47820) ( − 0.74116 , − 0.088441 , 0.47820 ) . The late centre at time 40 is ( 0.83119 , 0.51072 , 0.90498 ) (0.83119,0.51072,0.90498) ( 0.83119 , 0.51072 , 0.90498 ) , obtained by propagating only that central ODE point. The runs use independently prescribed ensembles. Each Gaussian is conditioned on its retained radial and shift intervals. Both polar-grid inputs are real MPS states with bond dimension 4. Their fidelities to the target polar cell amplitudes are 1 − 1.65 × 10 − 5 1-1.65\times10^{-5} 1 − 1.65 × 1 0 − 5 for growth and 1 − 9.76 × 10 − 7 1-9.76\times10^{-7} 1 − 9.76 × 1 0 − 7 for late. The compiled growth and late circuits use 24 and 20 logical qubits, 3,670,409 and 151,117 arbitrary Rz rotations, and 13,540,166 and 530,890 native T/T† gates per attempt, respectively, including preparation and both transforms.
Sequential Schmidt truncations construct the finite-bond MPS, choosing the first tested bond dimension in 2, 4, 8, 16, 32 and 64 with fidelity at least 0.9999. Real Givens rotations in Gray order compile each local isometry; each controlled rotation uses two single-qubit rotations and two exact clean-work multi-controlled X operations. All rotations and native gates are counted. The growth and late preparation circuits contain 773 and 601 arbitrary rotations and 5,390 and 4,186 native T/T† gates. These bond-four preparations require no additional work qubits. The position registers have 21 and 17 qubits. Both circuits use three transport work qubits: one shared clean arithmetic ancilla and two matching-validity flags. The arithmetic uses the radial circuit’s one-clean decomposition (Note 4). The complete growth and late preparation circuits contain 15,975 and 12,411 elementary gates and are replayed on the full 21- and 17-qubit state vectors. Their amplitude errors against the MPS contraction, without replacing signs or aligning away relative phases, are 2.97 × 10 − 13 2.97 \times10^{-13} 2.97 × 1 0 − 13 and 2.31 × 10 − 13 2.31 \times10^{-13} 2.31 × 1 0 − 13 . Both native endpoint Fourier circuits are also replayed; their forward-transform errors against the normalized classical Fourier transform, up to a global phase, are below 9 × 10 − 16 9 \times10^{-16} 9 × 1 0 − 16 . Propagating the actual gate-prepared amplitudes through all 528 and 66 steps of the equivalent matching-block implementation reproduces the reported biases at the final time of 0.244% and 0.221% and accepted-shot counts of 234 and 104. The changes in relative bias are below 3 × 10 − 13 3 \times10^{-13} 3 × 1 0 − 13 . Figure 5 d is therefore a double-precision state-vector calculation with verified native MPS preparation and endpoint transforms. The full transport circuit is still evaluated through matching blocks, not replayed elementary gate by elementary gate. Source Data and the supplied generator retain the complete preparation gate lists and the verification results.
The field budget uses the full covariance of the joint Cartesian modal observables r cos θ r \cos\theta r cos θ , r sin θ r \sin\theta r sin θ and z z z . Each input has its own continuum ODE reference. The growth and late references integrate their respective prescribed Gaussian inputs. The nominal deterministic field bias includes polar-grid MPS approximation. The per-attempt implementation-error budget is 10 − 4 10^{-4} 1 0 − 4 in normalized diamond distance, which bounds the total variation distance of the measured distribution (equation (49) ). We bound the change in the first and second moments by the observable and product ranges over the finite polar grid, and choose the smallest shot count whose mean squared relative field error at the final time is at most 10 − 4 10^{-4} 1 0 − 4 . Figure 5e uses the uncapped surface-code search of Supplementary Note 9. The growth and late inputs require 234 and 104 accepted shots, respectively. Both continuum quadratures use 64 radial, 128 angular and 64 shift nodes; reducing these to 48, 96 and 48 changes the reconstructed reference fields by less than 10 − 6 10^{-6} 1 0 − 6 relatively.
The finite domains are fixed before grid resolution. We retain r ∈ [ 0 , 1.45 ] r \in[0, 1.45] r ∈ [ 0 , 1.45 ] and z ∈ [ 0 , 1.50 ] z \in[0, 1.50] z ∈ [ 0 , 1.50 ] for growth, and r ∈ [ 0 , 1.65 ] r \in[0, 1.65] r ∈ [ 0 , 1.65 ] and z ∈ [ 0.30 , 1.70 ] z \in[0.30, 1.70] z ∈ [ 0.30 , 1.70 ] for late. Continuum Gaussian quadrature follows the adopted ODE characteristics; the sampled trajectory weight outside these domains is below 10 − 6 10^{-6} 1 0 − 6 for both runs. This is a numerical quadrature support check, not a rigorous trajectory enclosure. The radial lower bound is the polar axis r = 0 r = 0 r = 0 , where radial drift vanishes, to include the full nonnegative radial interval without an added inner radial truncation boundary. The angular domain is a full period for both runs.
Growth uses 256 × 32 × 256 cells in ( r , θ , z ) (r, \theta, z) ( r , θ , z ) , giving spacings 0.0056641, 0.19635 radians and 0.0058594. Halving r r r , θ \theta θ and z z z separately gives peak relative field discrepancies of 1.220%, 0.944% and 0.825%. The radial half grid fails the noiseless 1% check. The angular and shift half grids pass that check but require 2,900 and 752 shots, respectively, increasing the total synthesis workload relative to the selected grid.
Late uses 64 × 32 × 64 cells in ( r , θ , z ) (r, \theta, z) ( r , θ , z ) , giving spacings 0.025781, 0.19635 radians and 0.021875. Halving r r r , θ \theta θ and z z z separately gives peak relative field discrepancies of 0.399%, 0.494% and 0.469%. All three pass the noiseless 1% check, but their required counts of 135, 116 and 129 shots increase shots × T count per attempt relative to the selected 104-shot grid.
Grid selection uses a bounded coordinatewise search at the stated time-step count and stencil. Among candidates with noiseless relative field discrepancy below 1% at all 67 reported times and a finite shot count for 1% relative RMS field error at the final time including implementation error, we select the grid minimizing shots × T count per attempt. For this grid comparison, the synthesis T count per rotation is rounded up from three times the base-two logarithm of twice the rotation count divided by η \eta η , with η = 10 − 4 \eta= 10^{-4} η = 1 0 − 4 . This comparison assigns half the implementation-error budget to rotation synthesis. It does not establish a global minimum over grids or physical runtime. The selected circuits then undergo the full surface-code and synthesis-precision search without fixing that division of the error budget. The complete candidate list, axis-halving results, domains and preparation fidelities are supplied in Source Data.
Coherent step capacity and circuit-specific factory supply
Coherent step capacity
These analyses retain the architecture and synthesis searches of Notes 8–10. Capacity is constrained by physical qubits, implementation error and expected time per accepted shot. The fixed shot budget does not impose the 33-step benchmark’s field-error target on longer predictions.
Search settings follow Methods: a 10 − 4 10^{-4} 1 0 − 4 per-attempt implementation-error budget, the baseline physical error probability and timings, physical-qubit caps of 10 5 10^{5} 1 0 5 and 10 6 10^{6} 1 0 6 , and expected time per accepted shot limited to the 24-hour budget divided by 815 accepted shots.
When runtime sets the step limit, the limit scales approximately as the inverse T-state consumption interval for a fixed T-state count per step and fixed time available after preparation. Faster core processing raises this limit when the factory can meet the higher demand.
Supplementary Figure S10 compares the maximum step count N t max N_t^{\max} N t m a x and τ T \tau_T τ T at the baseline. The interval ratio in Supplementary Figure S10 shows that the fast block is supply-limited throughout the sweep at 10 5 10^{5} 1 0 5 qubits and core-limited at 10 6 10^{6} 1 0 6 . At 10 5 10^{5} 1 0 5 qubits, compact becomes supply-limited at 6 × 10 − 4 6 \times10^{-4} 6 × 1 0 − 4 , as higher selected code distances leave room for one factory unit, down from two at the baseline. Raising the cap at baseline reduces the fast block’s τ T \tau_T τ T from 110.50 to 21 μs and raises N t max N_t^{\max} N t m a x from 64 to 393. These changes show why data-block choice and factory allocation must be optimized together within the qubit budget. The step count also reflects synthesis and error constraints. Longer predictions require a separate assessment of field error. Supplementary Figure S11 fixes the evolution length to examine how a smaller code distance and permitted waits change qubit count, synthesis cost and runtime.
Factory supply and native T count
At a fixed evolution length, we jointly select the architecture and synthesis precision to minimize qubit count under each supply policy in Supplementary Figure S11 . At the baseline error probability, both minimum-qubit designs use one final-level factory unit. Allowing waits reduces the selected core code distance from 25 to 19. The nominal core interval falls from 225.00 to 171.00 μs, below the 221.00 μs mean MSF supply interval, leaving a 50.00 μs wait per T state. The waiting design uses 27.40% fewer physical qubits and takes 3.16% longer per attempt (Supplementary Figure S11 ). Its higher synthesis precision raises the compiled T-state count from 382,026 to 401,238, outweighing the slight reduction in τ T \tau_T τ T . The balance varies with physical error probability and native T inventory (Supplementary Figure S11 ,Supplementary Figure S11 ). For the 200,000-native-T control, waiting allowed saves 19.69% of qubits and increases runtime by 18.95% (Supplementary Figure S11 ).
To assess sensitivity to the native T count, we vary only the native T/T† inventory, fixing the number of arbitrary R z R_z R z rotations at 4,803 (Supplementary Figure S9 ). At each setting, we reevaluate the feasible hardware and rotation-synthesis choices for each supply policy. Synthesis error therefore need not scale proportionally with native T count. For a given circuit, the runtime change must therefore be evaluated against the available time budget when selecting a qubit-saving waiting policy. The different runtime changes at fixed width and arbitrary-rotation count show why architecture, supply policy and synthesis precision must be optimized together for each circuit and budget.
Error budget and resource model comparisons
Error-budget sensitivity
Relaxing the error budget reduces protection costs but increases the field bias bound and required measurement repetitions (Supplementary Figure S12 ). At error budget 0.01, Walsh–KvN uses 358,370 qubits yet requires 7,266 shots and 8.66 h for the same 1% field-error target. Error budget 0.0001 is a conservative comparison setting, rather than the field tolerance itself (Supplementary Notes 8–10).
A target-accuracy sensitivity keeps the selected circuit, physical design and implementation-error allowance fixed. Relaxing the field-error target from 1% to 5% requires 33 accepted shots and takes 3.94 min on 559,942 physical qubits for the fastest baseline design. This calculation does not reoptimize the architecture (Source Data).
Resource-model comparison
To assess sensitivity to factory design and the corresponding error modelling, Supplementary Figure S12 compares minimum-qubit estimates for the same 33-step radial Walsh–KvN circuit using two additional factory models[26 , 28 ] and the same core model. All three baseline estimates lie within the physical-qubit range of the study window. Notes 9–10 specify the shared core model and the factory-specific geometry, timing and error calibration.
The primary hardware experiment combines MPS preparation with the nonlinear dissipative drift
d r d t ∗ = μ ( r − r 3 ) \frac{d r}{d t^*}=\mu\left(r-r^3\right) d t ∗ d r = μ ( r − r 3 )
, with μ = 0.0924819 \mu=0.0924819 μ = 0.0924819 , on the unit interval. Four qubits encode 16 cells; a five-qubit control uses 32 cells. The target cell masses are Gaussian cumulative-distribution differences, normalized after truncation to [ 0 , 1 ] [0,1] [ 0 , 1 ] . The pre-truncation width is 0.25, and the centre is chosen to give cell-centre mean 0.9017043. Successive singular-value decompositions truncate the positive target amplitudes to MPS bond dimension two. A staircase of nearest-neighbour two-qubit unitaries prepares the compressed state.
Here r r r is the normalized radius on [ 0 , 1 ] [0,1] [ 0 , 1 ] , with an attracting state at r = 1 r=1 r = 1 . On the 16-cell grid labelled from zero, cells 14 and 15 cover [ 7 / 8 , 15 / 16 ] [7/8,15/16] [ 7/8 , 15/16 ] and [ 15 / 16 , 1 ] [15/16,1] [ 15/16 , 1 ] , with centres 29 / 32 29/32 29/32 and 31 / 32 31/32 31/32 . Thus P ( r ≥ 7 / 8 ) P(r \ge7/8) P ( r ≥ 7/8 ) sums their probabilities and diagnoses movement towards saturation. The threshold 7 / 8 7/8 7/8 is a convenient grid boundary, not a separate physical criterion. With nonnegative probabilities summing to one, the cell-centre mean is at most 31 / 32 31/32 31/32 , attained when all probability lies in cell 15; the continuous coordinate can still reach 1. This probability alone does not establish distribution contraction or field accuracy. The sampling checks below concern the radial probability and mean separately and do not include phase readout or establish the 1% field-reconstruction target.
This is the same input-approximation family used in Supplementary Note 2. Noiseless preparation fidelities are 0.9999998152 at 16 cells, 0.9999996254 at 32 cells and 0.9999995620 at 64 cells. The 64-cell compressed state has overlap fidelity above 1 − 10 − 13 1-10^{-13} 1 − 1 0 − 13 with the recorded manuscript input; the target mean differs by 4.8 × 10 − 9 4.8 \times10^{-9} 4.8 × 1 0 − 9 because of rounding. These are classically computed state overlaps. Hardware initialization is assessed separately from measured populations. The executed preparation circuits use seven CZ gates for 16 cells and ten for 32 cells. No 64-cell hardware transport is reported.
Table S13 | IBM MPS execution and full-count analysis. Device times refer to full acquisitions. The 32-cell control executes one step of length T / 4 T/4 T /4 , where T = 9.8291 T=9.8291 T = 9.8291 .
Table S13.
Experiments were performed on ibm_cleveland (Heron r2) on 16 November 2026. Both MPS jobs use gate and measurement twirling with 32 randomizations and XpXm dynamical decoupling. Each circuit is transpiled with six candidate seeds, selecting the smallest summed negative log gate survival from the backend target. The four-step job contains nine 16-cell circuits and three 32-cell circuits; the six-step job contains thirteen 16-cell circuits. Job identifiers and all counts are retained in Source Data.
Extended Data Figure 1 uses all acquired shots: 16,000 per circuit in the four-step job and 32,000 in the six-step job, including each paired reference. The 32-cell control also uses all 16,000 shots. Original counts, metadata and device times are preserved in Source Data.
For 16 cells, the interval t ∗ = 40 t^*=40 t ∗ = 40 –49.8291 49.8291 49.8291 is divided into four or six equal steps. For each step count k k k , a fresh circuit prepares the MPS input, executes k k k even–odd–even transport steps and performs computational-basis measurement. There is no intermediate readout or re-preparation. Three pair-address bits give eight nonzero Walsh coefficients for the cubic drift, so this grid tests transport rather than rotation compression. The 32-cell control executes initialization and one T / 4 T/4 T /4 step, ending at t ∗ = 42.457275 t^*=42.457275 t ∗ = 42.457275 . Its layers use fifteen nonzero Walsh coefficients out of sixteen possible, a small reduction that does not establish the large-grid scaling. The circuits contain no phase register or velocity-field reconstruction.
The four-step transport circuits have 7, 91, 172, 259 and 342 device CZ gates for k = 0 , … , 4 k=0,\ldots,4 k = 0 , … , 4 . The six-step circuits have 7, 91, 172, 256, 342, 427 and 510 for k = 0 , … , 6 k=0,\ldots,6 k = 0 , … , 6 . The 32-cell initialization and one-step circuits have 10 and 281 CZ gates. References share the logical transport/address skeleton; independent compilation can give slightly different native counts. For example, six-step references at k = 5 k=5 k = 5 and 6 have 429 and 511 CZ gates.
Noise mitigation and uncertainty
At each nonzero k k k , the reference retains the MPS input and logical transport/address skeleton but sets every Walsh rotation angle to zero. Its ideal output is the compressed input distribution. This noise control is separate from the continuum Radial model defining the 1% field-error target. Let p 0 \mathbf{p}_0 p 0 be the ideal MPS-prepared probability vector, p r e f , k \mathbf{p}_{\mathrm{ref},k} p ref , k the measured reference probability vector, and u \mathbf{u} u the uniform probability vector on the chosen grid. The fitted least-squares contrast is
F k = ( p 0 − u ) T ( p r e f , k − u ) ∥ p 0 − u ∥ 2 . (96) F_k=\frac{(\mathbf{p}_0-\mathbf{u})^{T}(\mathbf{p}_{\mathrm{ref},k}-\mathbf{u})}{\lVert\mathbf{p}_0-\mathbf{u}\rVert^2}. \tag*{(96)} F k = ∥ p 0 − u ∥ 2 ( p 0 − u ) T ( p ref , k − u ) . (96)
For a diagonal radial observable, O 0 O_0 O 0 is its expectation in the ideal MPS input, while O k O_k O k and O r e f , k O_{\mathrm{ref},k} O ref , k are the measured transport and reference means. The noise-mitigated estimator is
O c o r r , k = O 0 + O k − O r e f , k F k . (97) O_{\mathrm{corr},k}=O_0+\frac{O_k-O_{\mathrm{ref},k}}{F_k}. \tag*{(97)} O corr , k = O 0 + F k O k − O ref , k . (97)
Mitigation assumes the same effective global depolarizing contrast for the paired transport and reference circuits. F k F_k F k is a probability contrast, not a measured quantum-state fidelity. In the 16-cell four-step job it decreases from 0.755 to 0.390; in the six-step job it decreases from 0.750 to 0.275. For the 32-cell one-step control it is 0.444. No positivity constraint or clipping is applied.
Extended Data Fig. 1 shows the complete 16-cell acquisitions. Measured bars are counting standard errors at the full sample size. Mitigated bars are standard deviations of 1,000 independent multinomial bootstrap resamples of paired histograms, with F k F_k F k refitted in every resample and seed zero for each job. These bars exclude systematic noise-model mismatch and temporal correlations within a job. The known input is not shown as a mitigated hardware point at k = 0 k=0 k = 0 . The continuum means use the ideal MPS cell probabilities as a piecewise-uniform input.
To express the hardware analysis in the bias–variance form of main-text equation (7) , let S S S be the independent shot count in each transport and reference circuit. For each observable and step k k k , b k b_k b k is the absolute relative bias of the full-count corrected estimate against its matched continuum radial expectation, and v k v_k v k is its relative single-shot variance coefficient. First-order propagation through equation (97) includes uncertainty in F k F_k F k and its covariance with the reference mean. The relative mean-squared error and the estimated 5% shot requirement are
r M S E k ≃ b k 2 + v k S , S r e q , k = ⌈ v k 0.05 2 − b k 2 ⌉ ( b k < 0.05 ) . (98) rMSE_k \simeq b_k^2+\frac{v_k}{S},\quad S_{\mathrm{req},k}=\left\lceil\frac{v_k}{0.05^2-b_k^2}\right\rceil\ (b_k<0.05). \tag*{(98)} r M S E k ≃ b k 2 + S v k , S req , k = ⌈ 0.0 5 2 − b k 2 v k ⌉ ( b k < 0.05 ) . (98)
If F k F_k F k is held fixed, the leading counting variance scales as 1 F k 2 \frac{1}{F_k^2} F k 2 1 ; the full propagated variance contains additional contrast-uncertainty and covariance terms. At the four-step terminal point, fixed-contrast propagation gives a standard deviation of 0.0139 for 16,000 shots, compared with 0.0165 from full linear propagation and 0.0166 from bootstrap resampling.
Across both radial observables and all propagated times, equation (98) gives maximum empirical requirements of 2,507 shots per circuit for four steps and 8,613 for six steps. These include the full-count bias; a 5% counting-only criterion gives 1,763 and 3,190, respectively. The acquired 16,000 and 32,000 shots are 6.4 and 3.7 times the bias-inclusive requirements, and all are used in the figure. These budgets are retrospective plug-in estimates from the full counts.
Observed transport and mitigation limits
Using all shots, the outer-cell probability after four steps is 0.96571 ± 0.01661 and after six steps is 0.96234 ± 0.01601, compared with ideal-circuit values 0.99419 and 0.99521. The 32-cell one-step control gives 0.81774 ± 0.01383 against ideal 0.82889. Across the six-step outer-cell probabilities, all corrected values lie below the ideal-circuit curve by 1.9–3.7 bootstrap standard deviations. The maximum discrepancy is 3.7 standard deviations at step five. These deviations exceed counting uncertainty at several times and are consistent with residual noise-model mismatch. Additional shots reduce counting variance but do not remove systematic calibration bias. The 1% field target is assessed separately.
The fitted depolarizing model leaves reference total-variation residuals of 0.042–0.088 for 16 cells and 0.091 for the 32-cell control. Normalized nonnegative probabilities on the 16-cell grid give a mean between 1 / 32 1/32 1/32 and 31 / 32 = 0.96875 31/32 = 0.96875 31/32 = 0.96875 . The correction preserves normalization but does not enforce nonnegative probabilities, so sampling noise and model mismatch can produce estimates outside this interval. The full-count means after four and six steps are 0.98453 and 0.99792. Some corrected cell weights and formal variance estimates are negative. Such estimates do not indicate motion outside the physical domain. Full-distribution recovery and variance contraction remain unvalidated.
Earlier runs using general exact StatePreparation are retained as controls: 8,000 shots per circuit for four steps and 32,000 for six. In the original full acquisitions, the MPS four-step initialization has population total-variation distance 0.042 from its ideal input, compared with 0.100 in the earlier exact-loader four-step job; the MPS six-step initialization has distance 0.030. The jobs and compiled circuits differ, so these observations do not isolate the effect of changing the loader. The exact-loader probability after six steps, 0.893 ± 0.017, falls below the full-acquisition MPS estimate; this does not define a universal depth threshold.
The archive also retains the earlier zero-noise extrapolation controls. Gate folding uses noise factors 1,3,5 with twirled readout-error extinction (TREX); probabilistic error amplification uses factors 1,1.5,2,3 with TREX. Amplification gives the exact-loader four-step k = 2 k = 2 k = 2 probability 0.6866 ± 0.0306 against ideal 0.9364 and probability after six steps 0.2740 ± 0.1074 against ideal 0.9952. These controls do not establish reliable generic mitigation for the transport circuits.
Independent NumPy replay checks all 25 MPS logical–device circuit pairs. Their probability distributions agree within 8.8 × 10 − 15 8.8 \times10^{-15} 8.8 × 1 0 − 15 ; direct adjacent-pair propagation from an independently compressed input agrees with logical QASM within 5.4 × 10 − 15 5.4 \times10^{-15} 5.4 × 1 0 − 15 . Continuum means use analytical characteristics and 64-point Gaussian quadrature per initial cell. Raw counts, logical and native QASM/QPY circuits, full-precision calibration data and reproduction code accompany these results.
Supplementary references