gpt 6 astra ultra attempt on CS-100

Deterministic RIP: near-square constructions and the missing aspect-ratio quantifier

Definitions and mathematical statement

For $A\in\mathbb C^{M\times N}$ and an integer $s\ge1$, define the restricted isometry constant \[ \delta_s(A)=\sup_{\substack{x\ne0\\|\operatorname{supp}(x)|\le s}} \left|\frac{\|Ax\|_2^2}{\|x\|_2^2}-1\right|. \] Here $\operatorname{supp}(x)=\{j:x_j\ne0\}$; the same definition applies to real matrices and vectors. Thus $(s,1/3)$-RIP means $(2/3)\|x\|_2^2\le\|Ax\|_2^2\le(4/3)\|x\|_2^2$ simultaneously for every vector supported on at most $s$ coordinates.

Construct explicit deterministic rectangular matrices, with $M<N$, that achieve the source's target scale \[ s\ \gtrsim\ \frac{M^{3/5}}{(\log N)^C} \] for an absolute constant $C$, in nonvacuous growing parameter ranges. The hidden positive constant is independent of $M,N$. The requested construction should produce the matrix in polynomial time in its output dimensions, with a specified exact or adequately controlled finite-precision representation. Its guarantee must hold for its actual output, without an unverified random choice.

The interesting ranges have the displayed target substantially larger than $\sqrt M$; otherwise this does not break the usual coherence barrier. An identity matrix with $M=N$ therefore does not answer the question. Likewise, a construction good for one specified sparse vector is insufficient: the same matrix must preserve all sparse vectors of the required order. State explicitly which pairs $(M,N)$ the construction supports and any padding needed for other dimensions.

Short English statement

Construct explicit deterministic restricted-isometry matrices with sparsity of order M to the three-fifths power, up to logarithmic factors.

Sources

• S1 ULAM.AI and the UnsolvedMath Contributors, supplied problems.json, record 2800501 (AMR-027-0501), AMR Open Problem Lists. Catalogue identity and source transcription; CC BY 4.0. https://huggingface.co/datasets/ulamai/UnsolvedMath.
• S2 Bandeira - 42 Open Problems in Data Science $2016$, Open Problem 5.1. Original source indexed by AMR-027-0501. https://afonsobandeira.wordpress.com/2015/11/22/10l42gordon/.
• S3 A. S. Bandeira, Ten Lectures and Forty-Two Open Problems in the Mathematics of Data Science, corresponding numbered problem and surrounding definitions. https://people.math.ethz.ch/~abandeira/TenLecturesFortyTwoProblems.pdf.

Verbatim supplied catalogue question

Construct deterministic matrices $A\in\mathbb{C}^{M\times N}$ (or $A\in\mathbb{R}^{M\times N}$ ) satisfying the $(s,\frac13)$ -RIP for $s\approx\frac{M^{0.6}}{\mathrm{polylog}(N)}$

Attempt and results

The preserved definition requires $M<N$ and sparsity substantially above $\sqrt M$, but does not fix a lower bound on $N/M-1$. This matters: near-square matrices give a literal easy family. The following construction is stated with its precise dimensions; it is not claimed as resolution of the intended compressed-sensing problem at fixed aspect ratio.

An exact deleted-Fourier construction.

Take $N=M+1$, $M\ge1$, and $A_{rj}=M^{-1/2}\exp(2\pi i rj/N)$ for $1\le r\le M$ and $0\le j<N$. The geometric-series identity gives \[ A^*A=\frac NM I_N-\frac1M J_N. \] On a support $J$ of cardinality $t$, the Gram error is $M^{-1}(I_t-J_t)$. For $t\ge2$ its eigenvalues are $(1-t)/M$ on the constant vector and $1/M$ on its orthogonal complement; for $t=1$ it is zero. Consequently $\delta_s(A)=(s-1)/M$ for every integer $1\le s\le N$. In particular $s\le1+M/3$ gives the required $1/3$ distortion, including $s$ of order $M^{3/5}$. Entries have an exact succinct algebraic description using a root of unity and a square root, and the output has $MN$ entries. This is the usual simplex Gram matrix, an elementary known construction. There is no claim of a floating-point certificate without error control.

How much codimension this method permits.

More generally delete $r=N-M$ rows from the unitary $N$-point Fourier matrix and scale the remaining rows by $\sqrt{N/M}$. Each column has norm one. For distinct columns the full Fourier inner product is zero, so the retained inner product is minus the deleted sum, of absolute value at most $r/M$. For any $t$-sparse $x$, therefore, \[ |\|Ax\|^2-\|x\|^2| \le\frac rM\sum_{i\ne j}|x_i||x_j| \le\frac rM(t-1)\|x\|^2. \] The last step is Cauchy--Schwarz: $(\sum |x_i|)^2\le t\sum|x_i|^2$. Thus $r\le M^{2/5}$ allows $s$ of order $M^{3/5}$ by this certificate, but only with $N/M\longrightarrow1$.

Why coherence alone stops at fixed aspect ratio.

Let $N$ unit columns lie in $\mathbb C^M$, and let $\mu$ be their maximum off-diagonal inner-product magnitude. If $G=A^*A$ has nonzero eigenvalues $\lambda_1,\ldots,\lambda_q$, $q\le M$, then $\sum\lambda_i=N$ and Cauchy--Schwarz gives $\operatorname{tr}(G^2)\ge N^2/M$. On the other hand $\operatorname{tr}(G^2)\le N+N(N-1)\mu^2$. Hence \[ \mu^2\ge\frac{N-M}{M(N-1)}. \] If $N\ge(1+c)M$ for fixed $c>0$, this is at least $c/((1+c)M)$. A proof relying only on $(s-1)\mu\le1/3$ can then certify only $s\le1+\frac13\sqrt{(1+c)M/c}$. This limits this sufficient test; it is not an upper bound on the actual RIP order of every matrix.

Context, audit, and remaining scope

The primary lecture problem and its square-root-bottleneck discussion were inspected on 11 October 2026. The constructions and trace argument above are standard material, given with proofs to isolate the quantifier issue. The audit explicitly distinguishes $N/M\to1$ from $N\ge(1+c)M$, checks $s=1$ and $s=N$, and does not identify a limitation of coherence with an impossibility theorem for RIP. No numerical computation was used. The literal permissive formulation has the displayed near-square family; the source does not specify enough aspect-ratio quantifiers to call that a complete answer to its intended target. A substantive construction with fixed compression ratio and exponent $3/5$ has not been obtained here.

Process and contributions

Mathematics and model-assisted adversarial self-audit: OpenAI gpt-6-astra, ultra reasoning, as dispatched for this research batch; exact serving revision not exposed. Codex (GPT-6; exact serving revision not exposed) materially assisted with orchestration and format review. No human mathematical authorship, mathematical edits, or independent human proof checking is asserted. Contribution prompt A was applied to the full definition and supplied source row, then prompt B was applied adversarially to the derived results. The item-specific audit above records the scope checks and rejected extrapolations. Local assembly and structural checks do not establish mathematical correctness. The source catalogue status is preserved in TOP_PROBLEM and is separate from this attempt status.

Integration review.

Codex (GPT-6; exact serving revision not exposed) reviewed the written arguments and expanded layout and source-reference wrappers. Repository metadata and the production web renderer were checked using Python 3.13 and JavaScript with MathJax 3.2.2. This model-assisted review is not independent human proof verification; no PDF compilation or formal proof checking is asserted.

Completion Estimate

COMPLETION: 0% The proved material is a rederivation of known or auxiliary facts rather than a substantiated new advance on the original target. Only elementary near-square constructions are supplied; the intended fixed-compression-ratio exponent 3/5 construction remains missing, and the supplied statement leaves its aspect-ratio quantifier implicit.

References

The complete original Sources subsection above is retained, including catalogue attribution and licence notices. The additional list records primary sources actually consulted for this attempt on 11 October 2026; uncited source links reproduced above were not all rechecked.

• Bandeira, lecture notes $2016$, Open Problem 5.1 and Section 6.5; original construction target and coherence context inspected. https://people.math.ethz.ch/~abandeira/TenLecturesFortyTwoProblems.pdf.

Final status

Unresolved in this attempt. Only elementary near-square constructions are supplied; the intended fixed-compression-ratio exponent 3/5 construction remains missing, and the supplied statement leaves its aspect-ratio quantifier implicit.