acceptodds
Under review as a conference paper at ICLR 2027

New Records for the Hadamard Maximal Determinant Problem

Abstract

The Hadamard Maximal Determinant problem seeks an matrix with entries in , which maximizes the determinant for a given order . We report matrices attaining new record determinants for orders between and , as well as . We provide a proof of optimality within a specific family of circulant–block constructions for . For these orders, we recast the Hadamard problem as a sequence reconstruction problem from a pair of periodic autocorrelation functions, subject to certain arithmetic conditions. We describe several learning–based strategies for reconstruction and provide a comparison with classical simulated annealing–style approaches as a benchmark.

Primary area:
optimization
Keywords:
Combinatorial optimization, Hadamard maximal determinant problem, simulated annealing, sequence reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6BVUOpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper consdiers the Hadamard problem (maximum determinant binary matrix) focused on matrix sizes congruent with 3 modulo 4. The problem is of interest in the combinatorics and coding theory communities. The paper describes a reformulation of the problem as a sequence reconstruction by focusing on circulant matrices and setting up a sequence admissibility condition. The search for suitable sequences is implemented via reinforcement learning.

Critical strengths

The use of RL makes the paper tangentially relevant to ICLR. This approach to matrix construction allows for improvements over the state of the art, although it is not clear if the problem is fully solved.

The proposed approach is feasible for exhaustive search for small values of nn, and shows improvements over the state of the art results in the literature.

Critical weaknesses

The narrative on residual algebraic obstruction is not clearly laid out in the manuscript (only detailed in an appendix), but appears crucial to the work.

Some notation can be simplified, e.g., the Kronecker product of two column vectors j⊗jj \otimes j can be written as a rank-one matrix jj⊤jj^\top.

Questions for authors

Can a definition of algebraically unobstructed residuals be given explicitly? Is it feasible to have it in the manuscript, or does it need to be relegated to an appendix? I was not able to find William Orrick's website or an archive of it. Can you be more explicit in your reference or provide an alternative source?

2 comments

We thank the reviewer for a thorough review of our submission. Below we address the questions raised in the review.


Q1: Can a definition of algebraically unobstructed residuals be given explicitly?

We have added the definition directly in Section 3.1, where algebraically unobstructed was first introduced. Specifically, we define a residual rr to be algebraically unobstructed if, for each divisor d>2d > 2 of mm, there exists bd∈Kdb_d \in K_d whose field norm bdb_d equals the projection of rr into cyclotomic field KdK_d. We have also clarified that this formulation allows us to detect algebraic obstructions through finitely many local tests at each dd, which are used for reward shaping later. The further details remain in Appendix E.

We thank the reviewer's suggestion of simplifying notation. We have replaced the Kronecker product j⊗jj \otimes j with jjTj j^T, and clarified several other notational conventions to improve clarity.

We also thank the reviewer for pointing out the access to William Orrick’s website. We have updated the citation to include an archived version of the website available through web.archive.org.

We hope that our reply addresses the questions you have highlighted in the review process.

Reviewer_6BVUOpenReview ↗

Thank you.

Thank you for your response. I was expecting these issues to be easily addressed and my original assessment had taken this into account.

Reviewer_KHBbOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the Hadamard maximal determinant problem for a certain type of matrix construction, where the matrix is constructed using two circulant matrices. Fourier analysis expresses the determinant through periodic autocorrelation targets, allowing promising targets to be enumerated before reconstructing binary sequence pairs. The authors explore reconstruction methods using arithmetic obstruction tests, exact partner-reconstruction oracles, learned search policies, and simulated annealing. The paper reports improved constructions for six matrix orders within specified construction, thereby improving the known lower bounds for the unrestricted maximal determinant problem at those orders.

Critical strengths

The paper reports new record determinants for {\pm1}-matrices at six different orders. These results provide concrete progress on the longstanding maximal determinant problem, and the explicit sequence constructions allow the reported matrices and their determinants to be independently verified.

Critical weaknesses

Insufficient motivation for the chosen construction. The bordered bicirculant construction in (1) is introduced largely as a starting assumption, making the transition from sec. 1 to sec. 2 feel abrupt. Since the original problem concerns general {\pm1}-matrices, I would appreciate an upfront explanation of why this particular family is a promising restriction. The subsequent Fourier analysis demonstrates its computational advantages, but its relevance to obtaining large determinants in the unrestricted problem remains insufficiently explained to me.

Questions for authors

My understanding towards the paper is that, within the adopted matrix construction, the paper reformulates the problem as reconstructing sequence pairs from a target PAF sum. Necessary conditions, e.g., the arithmetic norm conditions or the PSD bounds, then guide the search for a candidate sequence a, after which the oracle tells whether a matching sequence b exists. Figure 4 supports that arithmetic RL becomes better at finding candidates satisfying the necessary checks, and Table 1 shows higher reconstruction success rates than uniform random sampling. What I want to know is whether this improvement merely comes from producing a larger proportion of candidates passing the checks, from potentially selecting candidates with a higher probability of admitting a partner even among those passing the checks, or from both.

2 comments

We thank the reviewer for a detailed review of our submission. We address the questions/concerns raised in the review below.


Q1: Motivation behind the bordered bicirculant constructions.

This is a good question which we will address more fully here. Recall that our two main constructions for candidate ±1\pm 1 matrices are the, uncorrelated X(A,B)X(A,B) and correlated Xcorr.(A,B)X^{\mathrm{corr.}}(A,B) ansatze, defined as follows:

X(A,B):=(1j⊤−j⊤jAB−jB⊤−A⊤).Xcorr.(A,B):=(1j⊤−j⊤jABj−B−A⊤).X(A,B):=\begin{pmatrix} 1 & j^\top & -j^\top\\\\ j & A & B\\\\ -j & B^\top & -A^\top \end{pmatrix}. \quad X^{\mathrm{corr.}}(A,B):=\begin{pmatrix} 1 & j^\top & -j^\top\\\\ j & A & B\\\\ j & -B & -A^\top \end{pmatrix}.

where jj is an all-ones vector of size mm. For any fixed order n≡3(mod4)n\equiv 3\pmod 4, we focus on finding the circulant matrices AA and BB which maximize the determinant ∣det⁡X(A,B)∣|\det X(A,B)| (or ∣det⁡Xcorr.(A,B)∣|\det X^{\mathrm{corr}.}(A,B)|, whichever is larger).

The main reason for the bicirculant form is tractability. Computing the Gram matrix, the Gram matrices for each half are C=AAT+BBT+jjTC = AA^T + BB^T + jj^T. Between halves, this is just −jjT- jj^T. This allows us to change the search space from the space of (possibly multiple) ±1\pm 1 circulant matrices to the significantly smaller set of joint PAF targets.

Explaining how the specific structure of the ansatz arises may be divided into three parts.

  1. Explaining the role of the two circulants.
  2. Explaining the form of the border.
  3. Explaining the sign patterns for the bicirculant portion.

The role of two circulants

The bicirculant ansatz with our specific choice of border may be motivated by the fact that the discrete Fourier transform diagonalizes circulants. Taking the Fourier transform of XX, then permuting the modes accordingly to yield a zero-mode block and m−1m-1 independent 2×22\times2 blocks:

∣det⁡X∣=∣det⁡X^0∣⏟border⋅∏t>0∣det⁡X^t∣,X^t=(a^tb^tb^ˉt−a^ˉt).˜|\det X|=\underbrace{|\det\hat X_0|}_{\text{border}}\cdot\prod_{t>0}|\det\hat X_t|,\qquad \hat X_t=\begin{pmatrix}\hat a_t&\hat b_t\\\\ \bar{\hat b}_t&-\bar{\hat a}_t\end{pmatrix}\~.

This decomposition is possible because AA and BB are circulant and the border is constant on each half (more on this below) --- any other form would couple the different Fourier modes. We focus first on the minor formed from eliminating the border. Recall we are working with orders n≡3(mod4)n \equiv 3 \pmod 4, meaning the minor is the determinant of an (n−1)≡2(mod4)(n-1) \equiv 2 \pmod 4 submatrix. The block matrix ansatz X′=(AB\B⊤−A⊤)X' = \begin{pmatrix}A&B\\\B^\top&-A^\top\end{pmatrix} is natural here as it attains the Ehlich--Wojtas upper bound on determinants at this order (Ehlich, 1964a; Wojtas, 1964). This may be easily seen by forming the Gram matrix, which is block-diagonal and assumes the form G′=(2m−2)1m+2JmG' = (2m-2)\mathbf{1}_m + 2 J_m if one enforces p(s)=2p(s) = 2. Note that our situation is slightly different as we enforce p(s)∈±2p(s) \in \\{\pm 2\\}, this ensures the off-diagonal entries of the Gram matrix (excepting the border) are Gij∈−1,3G_{ij} \in \\{-1, 3\\}. This is empirically motivated - off diagonal entries of the Ehlich block form which yield an upper bound on the determinant for orders n≡3(mod4)n \equiv 3 \pmod 4 are in −1,3\\{-1, 3\\}.

The form of the border

Given AA and BB, the border is essentially forced. To maintain the clean factorization of the Fourier modes, the border must commute with shifting both halves cyclically at the same time. This forces the border to be constant on each half (±j)(\pm j) and the only freedom lies in the choice of sign. Recalling ∣det⁡X^t∣=p^t\vert \det \hat{X}_t \vert = \hat p_t, the contribution of the zero-mode block decouples. In other words, for a given AA and BB, one may maximize over 16 possible choices of border signs (the corner may always be fixed to +1),

max⁡border,∣a∣,∣b∣∣det⁡X∣={(∣a∣2+∣b∣2+2mmax⁡(∣a∣,∣b∣))∏t>0p^t,˜uncorr.(∣∣a∣2−∣b∣2∣+2m(∣a∣+∣b∣))∏t>0p^t,˜corr..˜\max_{\text{border}, \vert a \vert, \vert b \vert}|\det X|= \begin{cases} \big(|a|^2+|b|^2+2m\max(|a|,|b|)\big)\prod_{t>0}\hat p_t \~, \quad\text{uncorr.}\\\\ \big(||a|^2-|b|^2\big|+2m(|a|+|b|)\big) \prod_{t>0} \hat p_t\~, \quad \text{corr.}\~. \end{cases}

The optimal choice of the border then depends on whether the construction is correlated or uncorrelated and the choice of segment sums ∣a∣\vert a \vert and ∣b∣\vert b \vert. Equivalently, one may treat the ansatz as the bicirculant core, whose determinant is defined as the maximum over borders. The borders we use are the optimal ones for the segment sums of the sequences which attain the highest determinant targets, so our results are unaffected.

Sign patterns for bicirculant core

For every choice of the 32 possible sign patterns together with the transposes (up to negations and reversals), each Fourier mode block has two rows rir_i with ∣ri∣2=p^t\vert r_i\vert^2 = \hat p_t, so

∣det⁡X^t∣=(p^t2−∣⟨r1,r2⟩∣2)1/2≤p^t,˜\vert \det \hat X_t \vert = \left(\hat p_t^2 - \vert \langle r_1, r_2 \rangle \vert^2 \right)^{1/2} \leq \hat p_t\~,

with equality iff the rows are orthogonal. The uncorrelated form is the sign pattern (BT,−AT)(B^T, -A^T) for the second row of the block ansatz. Together with its negation, the inequality is saturated for every ∣a∣,∣b∣\vert a\vert, \vert b \vert for these choices. The rows of the correlated ansatz are no longer automatically orthogonal unless we additionally impose the conditions in Equation (36) of the paper, but may yield a better zero mode for certain orders, though we have only observed this for n=51n= 51.

Q2: Does the observed improvement over the baseline come from producing a larger proportion of candidates passing the checks, from potentially selecting candidates with a higher probability of admitting a partner even among those passing the checks, or from both?

The arithmetic-free baseline only enforces the positivity condition (Eq. 28) on the Fourier spectrum a^t\hat{a}_t. This is a necessary but not sufficient condition for the existence of a pair (a,b)(a,b) which realize the target p(s)p(s). In terms of the sequence elements, we have an inclusion of problems Z[Cm]⊂Q[Cm]⊂R[Cm]\mathbb{Z}[C_m] \subset \mathbb{Q}[C_m] \subset \mathbb{R}[C_m]. The arithmetic-free approach is strictly weaker as it works over R[Cm]\mathbb{R}[C_m], which is the least restricted ring as it ignores the fact that the entries of (a,b)(a,b) must be ±1\pm 1. The arithmetic approach works at the stronger level of the rationals Q[Cm]\mathbb{Q}[C_m] and encourages the vanishing of certain auxiliary conditions (the local field norm obstructions in Section 3.3 and Appendix E of the draft) for all d∣md \vert m. For a∼πφa \sim \pi_{\varphi} sampled from the policy, the presence of an obstruction indicates that no partner bb in Q[Cm]\mathbb{Q}[C_m] and hence Z[Cm]\mathbb{Z}[C_m] exists, and vanishing of all obstructions is still necessary but not sufficient.

To answer your question directly, the improvement arises as our pretraining procedure combined with finetuning via potential-based reward shaping encourages the policy to produce sequences a∈Q[Cm]a \in \mathbb{Q}[C_m] which satisfy stronger necessary conditions than the baseline approach. Empirically, this translates into a given sampled sequence from the trained policy being more likely to admit a rational partner. Informally, the pass@1 (the probability a sequence sampled from the policy passes the oracle check) factorizes as ρ1=p(pass)⋅p(oracle∣pass)\rho_1 = p(\textsf{pass}) \cdot p(\textsf{oracle} \vert \textsf{pass}). Both training procedures do not incorporate feedback from the oracle and can only improve the prior factor. Lastly, whether or not a given sequence a′a' admits a partner b′b' is deterministic, so policy degeneracy is actively harmful and the novelty reward in the arithmetic agent helps to incentivize sample diversity. For a more detailed discussion on the quantitative results on p(oracle∣pass)p(\textsf{oracle} \vert \textsf{pass}), please see our response to the Q2 of the reviewer 26QY.

Reviewer_26QYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper restricts the Hadamard maximal determinant problem for n≡3(mod4)n \equiv 3 \pmod{4} to a bordered bicirculant family. Using Fourier block-diagonalization, it shows that the determinant depends only on the segment sums (∣a∣,∣b∣)(|a|, |b|) and a PAF-sum target p(s)∈{±2}p(s) \in \{\pm 2\}. This enables exhaustive target enumeration, so realizing the top target certifies optimality within this PAF-restricted family, and reduces the problem to reconstructing a sequence pair. Reconstruction combines a compression/lift CUDA oracle with two RL samplers: a GNN agent whose reward is shaped by local Hilbert-symbol obstructions, and an arithmetic-free transformer. The authors validate this with new records at n=103,111,115n = 103, 111, 115 (via RL) and n=51,107,119n = 51, 107, 119 (via simulated annealing), near-complete reconstruction of held-out top-5% targets, and 10310^3–104×10^4\times pass@1 gains over random sampling.

Critical strengths

(1) The explicit record constructions are independently checkable, and exhaustive target enumeration supports optimality within the stated family, namely X(A,B)X(A, B) with p(s)∈{±2}p(s) \in \{\pm 2\} for s≠0s \neq 0. Objective outcomes like these are rare and valuable in AI-for-math work.

(2) Using local Hilbert-symbol obstructions as a potential-based reward is an original way to densify an extremely sparse existence signal, and Figure 4 shows high per-episode success at reaching obstruction-free sequences. This is the paper's main methodological idea and a promising direction for related sequence-construction problems such as Legendre pairs, although its contribution to the records is not yet established.

Critical weaknesses
  1. The paper does not yet establish the computational advantage of its learned reconstruction pipeline over competitive classical search. The abstract promises a comparison with simulated annealing–style approaches, yet the only non-learning comparator for the lift-based pipeline is uniform random sampling; Figure 6 concerns a different, full-state-space setting. Annealing alone produced the n=107,119n = 107, 119 records without the lift oracle, which motivates a direct comparison at n=103n = 103–115115, although the different arithmetic structure (mm prime vs. composite) means that state-space sizes alone do not establish comparable difficulty. Without such a comparison, the 10310^3–104×10^4\times gains over random sampling do not show an improvement over existing methods.

  2. The validation does not isolate the key idea. The numerical agreement between Table 4 and Table 1 indicates that Figure 5 and Table 4 report the arithmetic-free transformer, which does not use the obstruction-based reward, and the manuscript does not state which model produced the Table 2 records. The evaluation protocol behind the "Arithmetic / rand." column is also not explicit, and there are no controlled ablations of VdV_d, novelty, and RND, nor variance across independently trained seeds. The results thus support gains from learned sampling over random sampling, but the specific contribution of obstruction-based shaping remains unisolated; more broadly, the correspondence between methods, evaluation protocols, and result tables should be made explicit.

Questions for authors
  1. Please compare on n=103,111,115n = 103, 111, 115, under matched hardware and search budgets, against: (a) the simulated annealing used for n=107,119n = 107, 119; (b) a classical generator for AA (e.g., annealing over AA) combined with the same feasibility filters and lift oracle, or another applicable structured search such as compression-and-lifting or an adaptation of the orbits method (Kotsireas et al., 2019). Please report both end-to-end cost including training and cost amortized over a fixed held-out target set. A reproducible advantage over a reasonably tuned baseline would materially strengthen the case for the learned approach.

  2. Do both learned-policy columns in Table 1 measure oracle-certified recovery of a partner BB, on the same target set and with the same notion of a sample (a complete AA vs. a multi-step episode)? In Figure 4, the policy is scored per episode but the random baseline per sampled sequence; how do they compare under matched candidate evaluations? What fraction of obstruction-free sequences AA admit a partner B∈{±1}mB \in \{\pm 1\}^m?

  3. Which model (Section 3.3 or 3.4), reward terms, and oracle settings produced each Table 2 record? Could you ablate VdV_d, novelty, and RND over multiple seeds to test whether obstruction-based shaping drives the gains?

5 comments

We thank the reviewer for providing constructive feedback and a thorough review of our submission. Below we address the questions raised in the review.


Firstly, regarding the points outlined in Weakness 2. We thank the reviewer for suggesting additional comparisons and we have updated the Fig. 5 to also include arithmetic RL pass@k curves. We have also clarified that the records in Table 2 were produced using arithmetic RL model. We address the remaining points in our answers to the reviewer's questions below.

Q1(a)

Our simulated annealing-experiments were aimed towards reconstructing a single PAF-sum target for each order attaining the maximal determinant. The timing of the reconstruction, over the same hardware, including the number of samples explored, is shown in the table below.

Order nSimulated annealing (hrs)Estimated number of samples (billions)
10329.141310.4
11172.143906.2
11541.131800.9

We have also included this table in the paper in appendix A.2

Q1(b)

The reviewer points out that another fair comparison would've been to use a similar simulated-annealing approach on the restricted state-space for constructing obstruction-free candidate sequences aa. We agree. Therefore, we have also conducted tests using simulated-annealing using the energy function −∑d∣mVd-\sum_{d\mid m}V_d, where VdV_d is the same potential used for reward shaping in arithmetic RL. We would like to emphasize that the key metric is the throughput of obstruction-free sequences aa, as the oracle reconstructing bb is identical in both cases. Therefore, provided that the distribution of the samples is sufficiently uniform, the likelihood of an obstruction-free aa admitting bb will be same as that of RL (which we report in our answer to Q2 below). Estimated times for generating 10001000 obstruction-free sequences aa on the same hardware are shown in the table below:

Order nArithmetic RL, train + 1k obstruction-free samples (sec)Arithmetic SA, 1k samples (sec)
1032311.9(train) + 43.6(infer.)3602
1114926.3(train) + 80.4(infer.)91600
1153125.6(train) + 52.6(infer.)64145

Once arithmetic RL model has been trained, it generalized across different targets (which is how we were able to apply it to previously unseen PAF-sum targets, especially the target corresponding to the maximal determinant within the family considered). Below, we also include a table comparing the estimated sampling budgets required for 95%95\% probability of reconstructing at least one pair (a,b)(a,b) on the maximal-determinant targets.

Order nArithmetic RL (million samples)Arithmetic SA (million samples)
1032.355(train) + 0.122(infer.)18.936
1112.355(train) + 0.761(infer.)244.550
1152.355(train) + 0.742(infer.)500.157

Q2

Yes, both learned-policy columns in Table 1 measure oracle-certified recovery of a partner. The table of empirical probabilities pb∣a:=p(b exists∣[t−aa∗]=0)p_{b\mid a}:= p(b~\text{exists}\mid [t-aa^*]=0) for different orders is shown in the table below.

OrderProbability pb∣ap_{b\mid a}
1030.081%
1110.014%
1150.015%

We did not observe any substantial difference between probabilities pb∣ap_{b\mid a} sampled using our arithmetic RL method, versus arithmetic SA.

Finally, regarding the baselines in Fig. 4, since RL has a fixed episode length budget for each evaluation, we have treated the whole trajectory as one sample, hence the single-draw comparison with the baseline. We acknowledge that matching the methods based on the number of samples encountered is a more fair comparison, therefore, we have now updated the figure to match the number of draws with the episode length budget. Furthermore, we have also included simulated-annealing as an additional baseline in Fig. 4. Our conclusions regarding the efficiency and effectiveness of our arithmetic RL method (especially when compared to baselines) remain unchanged.

Q3

The records were produced using arithmetic RL model. We would also like to clarify that our pretraining procedure, which involves imitation learning (behavior cloning) on a synthetically sampled PAF-sum data (we outline the methodology in Appendix D.4), is a significant contributing factor to the stability of the training of arithmetic RL. The reward structure helps to sustain the performance and allows the agent to increase generalizability (as indicated by the training curves in Fig. 4). This is partly due to the fact that the reward, while shaped through Hilbert symbols, is still extremely sparse, making it difficult to train the agent directly through the signals coming from VdV_d without any pretraining.

Regarding the novelty reward term rnovelr_{\text{novel}} and RND, the key motivation for the two was to ensure the agent doesn't collapse to constructing only one specific sequence aa for a given target tt. As we have mentioned in our answer to Q1, we would like to ensure that our sampler is not only efficient, but also produces a diverse set of obstruction-free sequences aa. Below we outline a brief ablation study indicating effect of scaling rnovelr_{\text{novel}} by λ≥0\lambda\geq 0 on the diversity of obstruction free samples aa. We limited the RL training to 20 PPO updates over 10 different seeds [1]. In order to focus on the effect of the novelty reward, we use the same BC starting point for all values of λ\lambda (since BC never uses rnovelr_{\mathrm{novel}}, it is unaffected by λ\lambda).

Novelty weightNum duplicatesReduction of duplicates
03725.6—
0.53419.37.81%
13476.66.34%
23280.711.89%
42941.420.73%
82313.837.80%
161130.069.60%

Where the reduction of duplicates is measured relative to λ=0\lambda = 0. We would also like to clarify that we don't necessarily claim rnovelr_{\text{novel}} as instrumental to the success of our RL approach, but rather it encourages regularization of the policy beyond RND and entropy regularizations.

We hope that our reply addresses the weaknesses and the questions highlighted in the review.


[1]: Each run uses the same 10610^6 inference attempts. We use the first 470456 obstruction-free sequences per run (which is the minimum of the success counts across 10 seeds). The duplicate counts are averaged over multiple seeds (modulo cyclic shifts and reflections).

Dear reviewer 26QY,

Thanks again for your valuable and constructive feedback. Our updated tables now provide additional evidence supporting the end-to-end computational advantage of arithmetic RL over classical search, accounting for both training and inference costs. We would like to follow up on our rebuttal and check whether our responses have addressed your concerns, or if there are any additional questions that would need further clarification.

Comments on the paper

1 comment

Dear reviewers and Area Chairs,

We thank the reviewers for their constructive feedback, which has helped us to strengthen our work and improve the quality of the paper. Below, we summarize the main outcomes of the discussion.

  • End-to-end efficiency advantage of arithmetic RL (26QY): Our additional experiments and timing estimates provide supporting evidence that arithmetic RL outperforms classical search methods, even when accounting for all end-to-end computational costs, including dataset generation, BC pretraining, RL training, inference, and oracle calls (See our answer to 26QY, follow-up Q2). We have also corrected the normalization of the sampling budgets for the reported random-search baselines (26QY, follow-up Q1) and performed multi-seed ablations of the arithmetic potential-based shaping (26QY follow-up Q3).
  • Motivation for bicirculant matrices (KHBb): We provide further motivation for the bicirculant ansätze used in the paper, and explain in more detail the mechanism through which our main approach outperforms a natural baseline presented in our paper.
  • Notation and clarity (6BVU): As suggested, we provide a more accessible definition of an “algebraically unobstructed sequence” in Section 3.1 of the paper and simplify the notational conventions used in the description of our principal method.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.