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

arXiv:2610.22093 · quant-ph · Submitted 2026-10-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: Today's paper: "Bubble-Sort Exchange Networks".

Mira: The gist The bubble-sort exchange network is a one-parameter-per-pulse variational circuit whose depth equals its number of rounds, and at depth 2N, in the cases tested,

Kai: First, who's behind it and why it matters.

Title and authors: Kai: So we're looking at this paper now, "Bubble-Sort Exchange Networks: Symmetry Selection Rules and Language-Model Distillation on Exchange-Only Spin Chains." The paper starts by laying out the basic building blocks of these exchange networks, focusing on how the Heisenberg exchange pulse acts as a partial SWAP when the pulse area is not a multiple of pi.

Mira: Right, it explains that this specific structure is native to exchange-only spin qubits and shows that any interaction other than a full SWAP involves entanglement and routing simultaneously <ref:2610.22093#pg1>.

Lev: From an engineering standpoint, I’m interested in how they build this exchange network on the chain of N dots, because that’s what we have to actually implement on hardware.

Kai: Exactly. The core idea is using these bubble sort rounds of odd even transpositions with one parameter per pulse, and seeing how the Pauli spin blockade readout limits what you can actually learn >.

Mira: And they tie this all back to the readout using Pauli spin blockade, which conserves total spin and means the reachable readout is fixed by your initial state's total-spin content <ref:2610.22093#pg1>.

Lev: It’s interesting that they are looking at the performance for both two-spin and three-spin encodings in these different setups.

The paper's summary: Kai: So the paper summarizes that this bubble sort exchange network is native to exchange only spin qubits, which is a key thing because it uses the Heisenberg exchange pulse Uij theta on neighboring dots >.

Mira: Right, it’s not just any circuit; it uses the Heisenberg exchange pulse Uij theta on neighboring dots and shows that a pulse of area pi is a SWAP up to a phase, and anything else is a partial SWAP that both entangles and routes <ref:2610.22093#pg1>.

Lev: It’s like they are systematically mixing all the neighbors over R rounds of odd even transposition, which leads to every pair having interacted exactly once if you run R equals N rounds >.

Kai: And then they tie this all back to the readout using Pauli spin blockade, which conserves total spin and means the reachable readout is fixed by your initial state's total-spin content <ref:2610.22093#pg1>.

Mira: That constraint is what limits what you can possibly learn from the system >.

Lev: They also talk about selection rules based on the Hamming classes of outcomes, which they say are derived from a transportation problem between sectors and outcomes <ref:2610.22093#pg3>.

Kai: And they find that for the three-spin encoding, the logical space sits in a single sector with no floor at all >.

Mira: But for the two-spin S–T encoding, it has structural zeros because of parity restrictions <ref:2610.22093#pg3>.

The paper's improvements: Kai: So the paper outlines several ways they improve things. One big improvement is using a closed-form selection rule for Pauli spin blockade readout under exchange only dynamics >.

Mira: This rule, which depends on the initial state's total-spin sector weights, lets them precisely characterize the closure of the reachable set and find an implied total-variation floor <ref:2610.22093#pg3>.

Lev: That floor is something they were missing before because of those parity restrictions in the two-spin encoding >.

Kai: And then they show that for their three-spin encoding, it has no floor at all, which is a significant result compared to the two-spin version <ref:2610.22093#pg3>.

Mira: They also look at distillation methods using a smoothed divergence objective function called PKL epsilon, which they say achieves distillation accuracy comparable to or better than matrix product state Born machines at relevant bond dimensions >.

Lev: That objective function is what they use to train the network parameters theta via gradient descent.

Conclusion: Kai: So to wrap up on this paper, the authors show that by using these bubble sort exchange networks, you can achieve distillation of GPT-two distributions with three-spin qubits at a KL divergence of zero point zero one one to zero point zero one four nat for six qubits at depth R equals N eighteen >.

Mira: And they also show the two-spin chain from the Néel state reaches a value between zero point two eight and one point five for the first teacher, which is better than what was previously achievable with products of singlets and T0s <ref:2610.22093#pg3>.

Lev: The main implication here is that using this specific network structure, combined with these readout constraints, lets you get better results for distillation than some other methods have been able to achieve >.

Kai: It’s a lot of work to build those physical systems and get that fidelity on the three-spin chain at R equals 2N eighteen, but the result is solid >.

Mira: Lev, what’s your final thought on where this leads?

Lev: I think it shows that even with these constraints imposed by the physics of exchange only dynamics, we can still design a system that performs distillation tasks effectively for larger systems and different encoding types.

Kai: Yeah, it confirms the feasibility of using these specific network topologies for learning complex classical models in this quantum hardware environment >.

Mira: It really ties together the theory of entanglement routing with the practical goal of model compression in an accessible way <ref:2610.22093#pg1>.

Lev: It sets a benchmark for what kind of resource we can expect from these architectures before we try to scale them up further.

Yuichiro Minato

blueqat Inc.

quant-ph

Submitted: 2026-10-15

Updated: 2026-10-15

Comments: 19 pages, 9 figures, 7 tables. Code and data: https://github.com/minatoyuichiro/eoall-check

Code: https://github.com/minatoyuichiro/eoall-check

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: The gist The bubble-sort exchange network is a one-parameter-per-pulse variational circuit whose depth equals its number of rounds, and at depth 2N, in the cases tested, the linear chain distils as

Key concepts

Bubble-Sort Exchange Network: This is a variational quantum circuit built on a chain of spins where rounds of odd-even spin transpositions are performed. It uses one parameter per pulse and conserves total spin, allowing it to simulate complex operations like routing information between neighboring qubits.
Pauli Spin Blockade (PSB) Readout: This is the method used to measure the quantum state of the chain. It works by observing how spins interact; if they are in a specific configuration, they block further evolution, which allows researchers to read out the final distribution of outcomes.
Distillation via PKL Divergence: The training process involves using gradient descent on a smoothed divergence called PKL. This objective function guides the network parameters to learn and reproduce the probability distribution of a classical language model, effectively 'distilling' knowledge into the quantum circuit.

Terminology

Distillation — 1× in this paper · explained in 2 episodes

Summary

The gist The bubble-sort exchange network is a one-parameter-per-pulse variational circuit whose depth equals its number of rounds, and at depth 2N, in the cases tested, the linear chain distils as well as or better than two-dimensional or non-local layouts Bubble networks are native to exchange-only (EO) spin qubits and study it as a Born machine read out by Pauli spin blockade (PSB)

How it works

The core idea is to build a variational “bubble-sort exchange network” on a chain of N dots—R rounds of odd–even transposition, one real parameter per pulse—and study it as a Born machine read out by Pauli spin blockade (PSB) Exchange conserves total spin, so what the readout can reach is fixed by the total-spin content of the initial product state The network has depth R and ≈ R(N − 1)/2 pulses on N spins, and at R = N every pair of spins has interacted exactly once This family is called a bubble-sort exchange network, having depth R and ≈ R(N − 1)/2 pulses on N spins

The elementary operation is defined by the Heisenberg exchange pulse Uij (θ) = exp(−iθ Si ·Sj) on neighbouring dots This leads to the relationship θ = π being a SWAP up to a phase, and any intermediate θ is a partial SWAP that both entangles and routes The bubble-sort schedule with R rounds is BR = sum Y−1 r=0 Y i≡r (mod 2) Ui,i+1(θr,i), i.e. round r acts on the pairs (0, 1),(2, 3),... for even r and (1, 2),(3, 4),... for odd r

Selection Rules and Readout

The reachable readout distributions are constrained by a closed-form selection rule for PSB readout under exchange-only dynamics For any set K of Hamming classes of the outcome string (number of triplet pairs), P(b ∈ K) ≤ sum S∈Λ(K) c2S, where c2S is the weight of the initial state in total-spin sector S and Λ(k) = sum 0,..., k for k ≠ 1, Λ(1) = sum 1 The rule is also sufficient since the reachable distributions are marginals of a transportation problem from sectors to outcomes

The consequence for the two encodings is that the threespin EO logical space lies in a single sector and has no floor; the two-spin S–T encoding with any product of singlets and T0’s has a parity-restricted sector content and hence structural zeros, but the Néel product state ↑↓↑↓ · · ·⟩ = N(S⟩ + T0⟩)/√2 contains every sector, leaves no outcome class forbidden, and has a negligible floor for the teachers studied

Distillation of GPT-2 Distributions

The paper focuses on distillation: taking a probability distribution produced by a classical model—here the distribution of two- or three-token continuations of GPT-2 over a restricted vocabulary—and re-learning it in a Born machine whose samples are produced by projective readout of the spin chain This connects to the tensor-network route into quantum circuits, where a classical model or its output distribution is first compressed into a matrix-product state (MPS)

The training involves gradient descent on the smoothed divergence PKLϵ(pt∥pθ) = sum b pt(b) [ln(pt(b) +ϵ) −ln(pθ(b) +ϵ)] with ϵ = 10−12 The network parameters θ are trained by gradient descent on the smoothed divergence PKLϵ(pt∥pθ) = sum b pt(b) [ln(pt(b) +ϵ) −ln(pθ(b) +ϵ)] with ϵ = 10−12 and no renormalisation

Key Findings on Performance

The three-spin encoding comes within a factor of three of the χ = 4 level; the two-spin encoding from products of singlets and T0’s stalls between the χ = 1 and χ = 2 levels The Néel state removes the floor for both teachers but only wins for the teacher whose floor was non-zero

The distillation results show that six three-spin EO qubits at depth R = N = 18 reach KL = 0.011–0.014 nat (within a factor of three of the bond-dimension-4 MPS, 0.005) The two-spin chain from the Néel state reaches 0.28–1.5 from products of singlets and T0’s for the first teacher, against 0.28–1.5 from products of singlets and T0’s for the second (whose forbidden class carries negligible mass, the alternating product state is competitive)

Hardware and Scaling

The chain is not what limits the network, since the N − 1-round bound is geometry-independent and at depth 2N nothing we tried optimises better The Sz conservation of the network means that simulation can be restricted to a single sector, = 18564 instead of 2 18 amplitudes at n = 6

The three-spin chain reaches the χ = 4 level at R = 2N at six and, for the first teacher, at ten qubits The two-spin chain from products of singlets and T0’s stalls between the χ = 4 level at n = 12 (0.49 against 0.43)

The final states lie largely outside the code space; for a Born machine the leakage space is a resource, not an error The leakage is not by itself predictive of the KL

The three-spin chain at R = 2N reaches 0.0038, so at equal depth the two encodings end within a factor 1.

Improvements for AI systems

The improved system can perform distillation of GPT-2 next-token distributions into an exchange-only circuit with high fidelity, specifically reaching KL = 0.011–0.014 nat at depth N for six three-spin EO qubits and 0.014 at depth 2N for the two-spin chain when initialized in the Néel state, outperforming products of singlets and T0’s (which stall at 0.28–1.5).
The system can leverage a closed-form selection rule for PSB readout under exchange-only dynamics, which is based on the initial state's total-spin sector weights, to precisely characterize the closure of the reachable set and attain an implied total-variation floor.
The system can be optimized using a training objective of smoothed divergence PKLϵ(pt∥pθ) = b pt(b) [ln(pt(b) +ϵ) −ln(pθ(b) +ϵ)] to train the exchange network parameters, achieving distillation accuracy that is comparable to or better than MPS Born machines at relevant bond dimensions.
The system can utilize a leakage shortcut, where the final state has 35–99% of its weight outside the code space, treating leakage as a resource for a Born machine instead of an error, enabling distillation even when the fidelity is not directly penalized by leakage-free constraints.
The system can utilize sector-restricted simulation methods to extend distillation capabilities to larger systems, reaching 24–30-dot chains and teachers with up to 16384 outcomes, where the three-spin chain still reaches the χ = 4 level at R = 2N.