Sequence-State Write in Neural & Quantum Systems

• Sequence-State Write (SSW) is a rank-1 update mechanism used in neural models and quantum circuits, enabling efficient memory control and state preparation.
• In neural models, SSW implements explicit memory updates via outer products and decay mechanisms, as seen in recent state-space and hybrid recurrent architectures.
• In quantum computing, SSW employs Fourier-analytic techniques to construct harmonic amplitude states, facilitating scalable block-encoding with precise error bounds.

Sequence-State Write (SSW) encompasses two distinct, technically rigorous frameworks that arise in advanced neural sequence modeling and quantum state preparation. In neural models, SSW refers to the primitive operation of updating a memory matrix via rank-1 outer products at each step in state-space and hybrid recurrent architectures. In quantum computing, SSW appears as a circuit design for preparing states with amplitudes following specific numerical sequences, notably the harmonic sequence, primarily through Fourier-analytic techniques. Both frameworks leverage efficient, low-rank, and interpretable updates as fundamental organizing principles for memory, state, and dynamics.

1. Sequence-State Write in State-Space and Hybrid Neural Models

Modern state-space and hybrid recurrent LLMs, including Gated DeltaNet, Mamba-2, and RWKV-7, structurally depart from conventional transformer residual streams by maintaining an explicit memory matrix $S_t \in\mathbb{R}^{d_k\times d_v}$. On each token $t$, the model computes key and value vectors, $\mathbf{k}_t\in\mathbb{R}^{d_k}$ and $\mathbf{v}_t\in\mathbb{R}^{d_v}$, and propagates state by a rank-1 update coupled with a decay of previous contents: $$S_t = \alpha_t\left(I - \beta_t\,\mathbf{k}_t\mathbf{k}_t^\top\right) S_{t-1} + \beta_t\,\mathbf{k}_t \mathbf{v}_t^\top,$$ where $\alpha_t, \beta_t$ are model-dependent scalar gates. The $\beta_t\,\mathbf{k}_t\mathbf{v}_t^\top$ term, designated as the Sequence-State Write (SSW), serves as a rank-1 insertion to a shared state cache. Later layers read from $S_t$ by contracting with query vectors, thus directly influencing output predictions (Young, 12 May 2026).

2. Sparse Autoencoder Decomposition: WriteSAE

WriteSAE is the first sparse autoencoder tailored for SSW in state-space and hybrid recurrent models. Unlike conventional SAEs operating on dense residual streams, WriteSAE restricts its atoms to the native rank-1 write shape, i.e., each atom $i$ is parameterized as

$$\mathrm{atom}_i = \mathbf{v}_i \mathbf{w}_i^\top, \quad \mathbf{v}_i \in \mathbb{R}^{d_k}, \; \mathbf{w}_i \in \mathbb{R}^{d_v}.$$

This explicit structural constraint permits literal “swap-in” of atom-derived updates in place of the native SSW at a single cache slot, matched by Frobenius norm to localize the causal effect. The primary WriteSAE training objective is a Top-$t$0 sparse autoencoder loss on vectorized $t$1, with auxiliary dead-feature revival and an explicit norm constraint $t$2, ensuring overcompleteness and direct interpretability (Young, 12 May 2026).

3. Mechanistic Analysis and Causal Interventions

The mechanics of SSW enable closed-form causal analysis. Perturbing $t$3 by $t$4 and propagating to downstream output at position $t$5, the first-order logit shift $t$6 for output token “tok” in prompt $t$7 is

$t$8

where $t$9 accumulates the decay coefficients along the path. Empirically, this formula predicts measured logit shifts with median $\mathbf{k}_t\in\mathbb{R}^{d_k}$0. SSW’s strict rank-1 structure underlies this strong causal localization: interventions (swap, ablate, random) at a single write site translate linearly and predictably to output distributional changes (Young, 12 May 2026).

The distinctive consequences for model interpretability and controllability are demonstrated by cache-slot substitution tests and behavioral install experiments. For example, at Qwen3.5-0.8B L9 H4, atom substitution outperforms matched-norm ablation (KL divergence to baseline) in 92.4% of $\mathbf{k}_t\in\mathbb{R}^{d_k}$1 firings across layers, generalizes to diagonal-write architectures (Mamba-2-370M at 88.08% over 2,500 firings), and supports precise behavioral control: sustained three-position atom installs at $\mathbf{k}_t\in\mathbb{R}^{d_k}$2 amplitude boost midrank continuation rates from 33.3% to 100% under greedy decoding (Young, 12 May 2026). This establishes SSW as both interpretable and steerable, with interventions at the primitive state update site.

4. Sequence-State Write for Quantum Harmonic State Preparation

In quantum computing, Sequence-State Write refers to a procedure for preparing a quantum state with amplitudes $\mathbf{k}_t\in\mathbb{R}^{d_k}$3 on an $\mathbf{k}_t\in\mathbb{R}^{d_k}$4-qubit register (target state $\mathbf{k}_t\in\mathbb{R}^{d_k}$5, $\mathbf{k}_t\in\mathbb{R}^{d_k}$6). The SSW method first constructs a “sawtooth” (linear) superposition $\mathbf{k}_t\in\mathbb{R}^{d_k}$7 via modular, efficient quantum primitives:

• Exponential superposition $\mathbf{k}_t\in\mathbb{R}^{d_k}$8 created with a gadget of T-depth $\mathbf{k}_t\in\mathbb{R}^{d_k}$9,
• Ancilla-using SELECT oracles via controlled-SWAPs to realize the linear amplitude profile,
• T-depth for $\mathbf{v}_t\in\mathbb{R}^{d_v}$0 preparation: $\mathbf{v}_t\in\mathbb{R}^{d_v}$1 (Rempfer et al., 27 Feb 2026).

Application of the Quantum Fourier Transform (QFT) to $\mathbf{v}_t\in\mathbb{R}^{d_v}$2 maps the state to one with amplitudes proportional to $\mathbf{v}_t\in\mathbb{R}^{d_v}$3 for $\mathbf{v}_t\in\mathbb{R}^{d_v}$4. A post-processing circuit with extra ancillas then combines cotangent asymptotes, yielding the final $\mathbf{v}_t\in\mathbb{R}^{d_v}$5 amplitudes, with systematized error analysis ($\mathbf{v}_t\in\mathbb{R}^{d_v}$6 for rotation-approximation error $\mathbf{v}_t\in\mathbb{R}^{d_v}$7, and $\mathbf{v}_t\in\mathbb{R}^{d_v}$8 ancillas for truncation precision).

Block-encoding extends SSW for diagonalization tasks—preparing block-encodings of matrices with harmonic sequence diagonals, via a circulant matrix built from $\mathbf{v}_t\in\mathbb{R}^{d_v}$9, two QFT layers, and an LCU (Linear Combination of Unitaries) construction, with T-depth dominated by the QFT layers (Rempfer et al., 27 Feb 2026).

5. Complexity, Generalization, and Practical Aspects

For state-space models, SSW’s core characteristic is its rank-1 nature, which allows for highly granular, targeted, norm-matched interventions not possible with residual stream architectures. WriteSAE’s autoencoder atoms, matched in both shape and norm to the primitive update, enable overcomplete decomposition and efficient causal control.

In quantum SSW, all dominant resource costs trace to the QFT: for linear-state preparation and QFT of an $$S_t = \alpha_t\left(I - \beta_t\,\mathbf{k}_t\mathbf{k}_t^\top\right) S_{t-1} + \beta_t\,\mathbf{k}_t \mathbf{v}_t^\top,$$0-qubit register, total T-depth scales as $$S_t = \alpha_t\left(I - \beta_t\,\mathbf{k}_t\mathbf{k}_t^\top\right) S_{t-1} + \beta_t\,\mathbf{k}_t \mathbf{v}_t^\top,$$1 for error $$S_t = \alpha_t\left(I - \beta_t\,\mathbf{k}_t\mathbf{k}_t^\top\right) S_{t-1} + \beta_t\,\mathbf{k}_t \mathbf{v}_t^\top,$$2, qubit overhead as $$S_t = \alpha_t\left(I - \beta_t\,\mathbf{k}_t\mathbf{k}_t^\top\right) S_{t-1} + \beta_t\,\mathbf{k}_t \mathbf{v}_t^\top,$$3, and block-encoding complexity includes additional LCU register costs. The methods generalize to arbitrary sequences whose discrete Fourier coefficients are efficiently computable, accommodating $$S_t = \alpha_t\left(I - \beta_t\,\mathbf{k}_t\mathbf{k}_t^\top\right) S_{t-1} + \beta_t\,\mathbf{k}_t \mathbf{v}_t^\top,$$4, sinc-to-rect, and exponential-to-Lorentzian mappings (Rempfer et al., 27 Feb 2026).

Amplitude normalization and error correction in the quantum context require classical repeat-until-success protocols, with subnormalization constants factoring into expected run-times and success probabilities. These considerations do not impact the interpretability and closure of the classical neural SSW mechanism.

6. Significance, Unification, and Outlook

Sequence-State Write is foundational for both modern memory architectures in neural sequence modeling and for efficient, analytically-based state preparation in quantum computation. In neural models, SSW’s rank-1 update provides a “causal wedge” for state control, interpretability, and mechanistic probing—properties unattainable by dense, higher-rank, or residual-stream alternatives. In quantum systems, SSW transforms Fourier characteristics of simple states into structured target distributions, handling singularities such as $$S_t = \alpha_t\left(I - \beta_t\,\mathbf{k}_t\mathbf{k}_t^\top\right) S_{t-1} + \beta_t\,\mathbf{k}_t \mathbf{v}_t^\top,$$5 efficiently and scalably.

A plausible implication is that SSW-type primitives will continue to serve as model- and hardware-aligned intervention points for both behavioral control (in the neural context) and fast state syntehsis or block-encoding (in the quantum context), given their compatibility with the underlying dynamical and memory structures. The shared low-rank, efficiently analyzable update motif suggests structural convergence between mechanistic LLM interpretability and quantum state engineering.

Paper/Method	Domain	arXiv ID
WriteSAE, SSW in LLMs	Neural Sequence Models	(Young, 12 May 2026)
Harmonic sequence SSW (state prep, block enc.)	Quantum Computation	(Rempfer et al., 27 Feb 2026)