Bubble-Sort Exchange Networks: Symmetry Selection Rules and Language-Model Distillation on Exchange-Only Spin Chains

AI summary

Overview

Research area: Quantum computing with exchange-only (EO) spin qubits in semiconductor quantum dots, combining SU(2) symmetry constraints, Born-machine generative modelling, and distillation of small language-model distributions.

Technical level: Advanced. The paper assumes familiarity with spin qubits, total-spin sectors, Pauli spin blockade, matrix-product states, and KL divergence.

Scope: The paper derives an exact symmetry selection rule for Pauli-spin-blockade readout of exchange-only spin chains and demonstrates, in noise-free simulation, that a variational bubble-sort exchange network can distil short GPT-2 next-token distributions at a level competitive with a bond-dimension-4 MPS.

What This Paper Is About

Linear chains of exchange-coupled quantum dots have only one fast two-body control, the Heisenberg exchange pulse, and a pulse of area π is a SWAP. The parallel bubble-sort SWAP network normally used only to route all-to-all interactions therefore becomes a native, parameterised circuit on such a chain, where a pulse of any area that is not a multiple of π both routes and entangles. The paper asks what such a network can actually produce at its Pauli-spin-blockade readout, given that exchange conserves total spin, and then uses the answer to distil short GPT-2 continuation distributions into the chain as a Born machine.

Key Contributions

1. A closed-form selection rule for PSB readout under exchange-only dynamics. For outcomes with k triplet pairs, the probability is bounded by the initial state's weight in total-spin sectors S ≤ k, with S = 0 excluded when k = 1; for any set of Hamming classes the bound uses the union of their sectors. A transportation/max-flow argument proves the rule is also sufficient: the closure of reachable readout distributions is exactly the polytope cut out by the rule, and the implied total-variation floor is attained.

2. Consequences for the two encodings. The three-spin EO logical space lies in a single total-spin sector and has no floor from a code state. Two-spin singlet–triplet encodings initialised in products of singlets and T₀ states have structural zeros, whereas the Néel state |↑↓↑↓⋯⟩ contains every sector, forbids no outcome class, and has a negligible floor for the two teachers studied.

3. End-to-end distillation of GPT-2 distributions. Six three-spin EO qubits at depth R = N = 18 reach KL = 0.011 nat, and 0.004 at depth 2N, the latter below the bond-dimension-4 MPS (0.005); six two-spin S–T qubits from the Néel state reach 0.014 at depth 24 and 0.004 at depth 36 (medians of five seeds), against 0.28–1.5 from products of singlets and T₀ states for the first teacher.

4. A depth bound and its numerical counterpart. No arrangement of disjoint pulses in any geometry brings every pair of spins together in fewer than N − 1 rounds, one fewer than the chain's N. At depth 2N the chain distils as well as or better than a two-dimensional array, a non-local round robin, and random matchings in the tests reported. A side result (Appendix A) compiles the orthogonal factor of the polar decomposition of a two-qubit weight block into the same network at depth 4N.

Main Findings

• Exchange is a native router and entangler: A pulse U(θ) = exp(−iθ Sᵢ·Sⱼ) is a SWAP at θ = π, the identity at θ = 0, and a partial SWAP otherwise, so one bubble-sort round has depth 1 on the chain versus depth 4 for the superconducting ZZ·SWAP round counted in the cited reference.

• Symmetry caps what readout can see: Because every exchange pulse commutes with total spin, the sector weights c_S² of the initial state are invariant. For the two-spin encoding at zero magnetisation, outcome classes with k triplet pairs can only draw on sectors S ∈ Λ(k), with Λ(k) = {0,…,k} for k ≠ 1 and Λ(1) = {1}.

• The bound is tight: A max-flow/min-cut argument shows the reachable distributions are the marginals of a transportation problem from sectors to outcomes, so the total-variation floor of Eq. (6) is the exact minimum over the closure. A direct numerical search in the 924-dimensional S_z = 0 sector of N = 12 (Adam, 1500 steps, 2–3 seeds agreeing to 10⁻⁴) matched the floor to four digits in all ten cases tested, seven of which had a non-zero floor.

• Initial state determines capacity: Products of singlets and T₀ states have fixed spin parity, so whole outcome classes are forbidden. For the cat teacher, |S⟩^⊗6 gives a floor of 0.0528 and |T₀⟩^⊗6 a floor of 0.162; for the begin teacher, |S⟩^⊗6 and |T₀⟩^⊗6 give 0.248. The Néel state gives a floor of 0 for both teachers.

• Three-spin encoding has no symmetry floor: Starting from a code state, all weight sits in S = n/2 and every outcome string remains reachable, so the reachable polytope is the full simplex.

• State preparation is efficient in rounds: For n = 2 (N = 6), R = N + 2 = 8 rounds (20 pulses) prepare every Haar-random or MPS logical target tested to F = 1 − 10⁻⁸ with leakage 10⁻⁸, while R = 6 and R = 7 do not; the staircase schedule needs L = 4 (20 pulses). For n = 3 (N = 9), R = 9, 14, 18, 22, 27, 32 (36, 56, 72, 88, 108, 128 pulses) give F = 0.923, 0.995, 0.9999, 0.9999, 1 − 10⁻⁵, 1 − 10⁻¹⁰.

• Leakage is used as a resource: Final distillation states are 55–99% outside the code space, and leakage is never penalised; only the PSB readout distribution is compared with the teacher.

• Scaling behaviour differs by encoding: Sector-restricted simulation with an adjoint-method gradient reaches ten to fourteen logical qubits (1024–16384 outcomes, up to 30 dots). At fixed depth in units of N the two-spin chain loses ground to the MPS as size grows (for the first teacher, from the χ = 4 level at six qubits to the χ = 3 level at twelve), while the three-spin chain again reaches the χ = 4 level at depth 2N for the first teacher.

• MPS control values: The MPS Born machine reaches KL of 0.351/0.136/0.042/0.005 nat for cat and 0.729/0.168/0.010/0.001 for begin at χ = 1/2/3/4, and zero at χ = 8; the χ = 4 level is used as the operational criterion for successful distillation.

• Teacher definitions: GPT-2 small with a fixed prompt and restricted top-V vocabulary: cat ("The cat sat on", V = 8, L = 2, entropy 3.78 bit) and begin ("In the beginning", V = 4, L = 3, 3.45 bit), giving 64 outcomes encoded as six bits.

Methodology in Plain English

The author starts from an existing hardware trick: on a line of qubits, a bubble sort (odd–even transposition) swaps neighbouring pairs in alternating rounds and, after N rounds, has brought every pair together once. On a superconducting chip each of those blocks is a separate interaction plus a separate SWAP. On an exchange-coupled quantum-dot chain, the same block is a single exchange pulse whose area can be tuned, so the routing step and the entangling step merge into one trainable operation.

The circuit family is therefore simple: R rounds of alternating pulses, one real angle per pulse, applied to a chain of N spin-1/2 dots. The output is not read as a full quantum state but as a bit string from Pauli spin blockade on the first pair of each encoding block, which distinguishes singlet from triplet. That readout distribution is the model's probability distribution.

Because exchange conserves total spin, the author works out bookkeeping rules for which total-spin sectors can feed which readout patterns. The sectors are labelled by total spin S, and a readout with k triplet pairs can only be fed by certain sectors; the initial state's weight in those sectors therefore upper-bounds that outcome class. A max-flow argument then shows the bound is achievable in the limit of arbitrary depth, so it is not merely a heuristic.

For the machine-learning part, two short GPT-2 continuation distributions are compressed into matrix-product states of increasing bond dimension as a classical yardstick, and the same distributions are then learned directly by the exchange network, training only the pulse angles on the exact KL divergence and comparing readout distributions. Larger systems are handled by restricting the simulation to the relevant total-spin sectors and using adjoint-method gradients, which allows chains of up to 30 dots.

Why This Matters

The paper turns a hardware constraint into a design principle: instead of treating the SWAP network as routing overhead, it makes the swap angle the variational parameter, and it replaces vague statements about symmetry with an exact, provable ceiling on what the readout can produce. That gives a clean way to tell whether a poor result comes from the ansatz, the optimisation, or the initial state—an important distinction for anyone using spin chains as generative models.

Potential applications:

• Semiconductor spin-qubit hardware: native compilation of swap-network circuits onto exchange-coupled dot arrays, where exchange is the only fast all-electrical two-body control.
• Quantum generative modelling: Born machines whose samples are token distributions, with PSB as the sampling mechanism and no classical readout head.
• Model compression and distillation: re-expressing small classical or neural sequence distributions in a parametrised quantum circuit, using MPS bond dimension as a resource yardstick.
• Benchmarking and verification: the selection rule provides a cheap, depth-independent check on whether a device or simulator result is physically plausible.

Industry relevance: the work is directly aimed at exchange-only and singlet–triplet spin-qubit platforms, the two encodings most pursued for scalable semiconductor quantum processors, and at the pipeline connecting tensor-network compression to quantum circuit training. The author is affiliated with blueqat Inc., and the results are framed explicitly as noise-free simulation with no claim of quantum advantage.

Future Directions

• Hardware demonstration: the distillation and selection-rule experiments are all noise-free simulations; testing PSB-readout distillation on an actual dot array, including readout imperfections and pulse-area errors, is the obvious next step.
• Scaling beyond 30 dots: the adjoint-method sector-restricted simulator reaches ten to fourteen logical qubits (1024–16384 outcomes, up to 30 dots); whether the three-spin chain's advantage at depth 2N persists at larger sizes is left open for the second teacher in particular.
• Multi-qubit unitary compilation: Appendix A compiles the four two-qubit logical unitaries tested at R = 4N with F_proc = 1 − 10⁻⁶, but three-qubit unitaries were not reached at that depth by the optimisation, leaving the compilation question unresolved.
• Initial-state engineering: the two-spin encoding's limitation is entirely in the initial state, since free sector weights give KL = TV = 0; finding easily prepared states with good sector coverage for arbitrary teachers, beyond the Néel state, is a natural extension, especially where a teacher's forbidden class carries non-negligible mass.

Target Audience

Researchers and graduate students in quantum information working on spin-qubit architectures, symmetry-constrained variational circuits, or quantum generative models; tensor-network and Born-machine practitioners interested in distillation benchmarks; and hardware-oriented engineers who need to know which readout distributions an exchange-only device can and cannot produce.

Authors’ abstract

On a chain of exchange-coupled quantum dots a Heisenberg exchange pulse of area pi is a SWAP, so the parallel bubble-sort SWAP network of superconducting chains is native to exchange-only (EO) spin qubits, and any other pulse area fuses routing with an entangling interaction. We build a variational "bubble-sort exchange network" (odd-even transposition, one parameter per pulse) and study it as a Born machine read out by Pauli spin blockade. Exchange conserves total spin, so the reachable readout distributions are fixed by the total-spin content of the initial product state. For the two-spin singlet-triplet encoding we derive a selection rule -- outcomes with k triplet pairs are bounded by the initial weight in sectors S <= k (S = 0 excluded for k = 1) -- and prove by a transportation argument that it is also sufficient: the reachable distributions form the polytope cut out by the rule, and its total-variation floor is attained. Products of singlets and T0's forbid whole outcome classes; the Neel state forbids none; the three-spin EO code has no floor. We then distil, in noise-free simulation, next-token distributions of GPT-2 over restricted vocabularies into the network, with matrix-product-state (MPS) Born machines as classical control. At six logical qubits the three-spin chain reaches KL = 0.011 nat at depth N and 0.004 at 2N, below the bond-dimension-4 MPS; the two-spin chain from the Neel state reaches 0.014 at 2N and 0.004 at 3N, against 0.28-1.5 from singlet/triplet products for the first teacher. Final states lie largely outside the code space: leakage is a resource, not an error. No arrangement of disjoint pulses meets every pair in fewer than N-1 rounds, and at depth 2N the chain distils as well as a 2D array or non-local couplings. Sector-restricted simulation extends this to 10-14 logical qubits (up to 30 dots).