{"id":"cf791e09-f755-4acd-8a47-f04aa2640d44","arxiv_id":"2506.00528","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A top-magnitude sign quantisation, grounded in a family of convex polytopes, encodes embeddings as ternary vectors and delivers fast, accurate approximate similarity search.","lead":"Embedding vectors, long lists of numbers used to search by meaning, can be shrunk to tiny ternary codes of -1, 0 and 1 while keeping search rankings almost intact. Searching a million compressed embeddings can then run up to over 100 times faster on ordinary hardware, with rank correlations as high as 0.96.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The main claim is not yet established for arbitrary embeddings: the geometric argument presupposes high-dimensional uniformity, and the paper itself random-rotates the Laion data to avoid coordinate alignment, an implicit admission that the proxy can fail on real, axis-aligned, or…","rationale":"I read the paper as making a practical claim (ternary top-magnitude encoding plus bitwise scalar product gives fast, accurate kNN proxies) and a theoretical claim (this is guaranteed by EVP geometry). The experiments are real and impressive; the speedups are plausible and the method is simple. The reader's verdict of CONDITIONAL is appropriate. My stress-test found the weakest point exactly where the reader placed it: the theoretical bridge from EVP geometry to arbitrary embeddings. Section 3's argument concerns uniformly distributed points on a hypersphere; real embeddings are not uniform. The top-magnitude rule is coordinate-aligned, so data whose mass is concentrated in a few coordinates can break the concentration-of-measure premise. The paper's random rotation of Laion is strong evidence the authors know this; without a stated 'always rotate' step or an experiment without rotation, the unqualified claim is not established. I also note §2.5's claim of independent coordinates on the sphere is false, but the needed equiprobability follows from permutation symmetry, so I do not treat that as load-bearing. The missing code reference and deferred proof [9] are further reasons not to accept the theoretical claims as-is. A single ablation (unrotated Laion, plus a synthetic axis-aligned control) would settle the matter.","tokens_in":9476,"tokens_out":6634,"duration_ms":69591,"concrete_test":"Re-run the Section 5.1 Laion experiment exactly, but skip the random-rotation step: take the 500-dimensional PCA-reduced CLIP vectors, encode with the {333,500} EVP, and measure Spearman ρ (Table 2) and 30@100 recall (Figure 6) over the same 1000 queries. If ρ drops materially (say >0.05) or recall degrades, the headline result depends on a preprocessing step that is not part of the stated generic method. As a sanity check, repeat the same comparison on synthetic axis-aligned data with small intrinsic dimension (e.g., 500-D points supported on 25 coordinates plus noise) to confirm the failure mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the implicit transfer from EVP geometry to arbitrary real embeddings. Section 3 justifies the proxy by concentration of measure: for uniformly distributed points on a high-dimensional sphere, almost all the volume of a Voronoi cell lies near its boundary, so ℓ2(ui, vi) is nearly constant, and the angles ∠vi vj uj and ∠vj vi ui concentrate near π/2 (items 4–5). The mapping rule in §2.4 — keep the x = 2d/3 largest absolute coordinates with their signs — is coordinate-aligned, not rotation-invariant. If an embedding set has low effective dimensionality or is aligned with the coordinate axes, the same coordinates tend to be selected for many vectors; the ternary vectors then express mainly the dominant directions, and the residual directions that determine fine rank order are discarded. Consequently the concentration argument, which is a statement about the uniform measure on the sphere, need not apply. Section 5.1 concedes this: the Laion CLIP data is randomly rotated 'to avoid the principal components being tightly aligned with the EVP vertices', i.e., to force the data into the generic position the theory requires. No experiment reports performance without that rotation, and the method as stated does not tell practitioners to rotate. The reported Spearman values (e.g., 0.96 for Laion) and the 30@n recall therefore validate the pipeline that includes rotation, not the unqualified claim that arbitrary floating-point vectors can be replaced by ternary vectors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing ℓ2-normalised d-dimensional embedding vectors with ternary vectors in {-1,0,1}^d having exactly x non-zero entries, obtained by keeping the x coordinates of largest absolute value with their signs; this is shown to be the nearest vertex of the authors' 'Equi-Voronoi polytope' (Section 2.4). The manuscript claims that the Voronoi cells of these vertices have equal volume on the hypersphere (Section 2.5) and that, by concentration-of-measure arguments, vertex-to-vertex distances closely proxy the true distances between the points in the corresponding cells (Section 3); the full geometric argument is deferred to the in-preparation reference [9]. Encoding a ternary vector as a pair of bit masks v+ and v- allows the scalar product to be computed exactly by the bitwise function b2sp using four AND operations and additions (Section 5.1.2). Experiments on uniform, GloVe-100, PubMed-384, and PCA-reduced and randomly rotated Laion data (around one million vectors each) report Spearman correlations up to 0.96 (Table 2), 30@n recall values that improve with n (Figures 5-6), and end-to-end speedups of 33-158x for 100 exhaustive queries over 1M vectors (Table 5), with comparisons against 1-bit and b1.58 quantisation baselines.","tokens_in":9806,"tokens_out":16667,"duration_ms":157448,"significance":"If the empirical results are taken at face value, this is a practical and simple contribution: the mapping threshold is derived (x = 2d/3 from maximising the EVP vertex count), no parameters are fitted to the target data, the bitwise function b2sp computes the ternary scalar product exactly, and the representation is SIMD-friendly at about 2 bits per coordinate. The reported correlations beat the 1-bit and b1.58 baselines on all five data configurations, and speedups of one to two orders of magnitude for exhaustive search over one million vectors are substantial. The weakness is that the theoretical narrative is not commensurate with the claims: the geometric explanation is explicitly outsourced to an unpublished companion (Section 3), the only self-contained lemma rests on a false premise (Section 2.5), and the 'entirely generic' applicability claim is undermined by the need to randomly rotate the Laion data (Section 5.1). These are fixable within the scope of a revision, and the paper would then be a credible contribution to practical embedding search.","major_comments":[{"comment":"The equi-volume proof rests on a false premise. The text states that for uniformly distributed points on a hypersphere 'each vector element is necessarily drawn from an identical and independent distribution', but the coordinates of a uniform point on the sphere are identically distributed and not independent (for d=2, |y| is determined by |x|; the joint distribution is supported on a submanifold, and reference [15] itself derives a marginal Beta distribution rather than independence). Consequently the equiprobability of the nearest vertex does not follow as written, and the inference about equal cell volumes is not proven. The conclusion is in fact true, but only via a different argument: the signed permutation group acts transitively on the vertex set and preserves the uniform measure on the sphere, so the spherical Voronoi cells are congruent. Since equi-volume is stated in Section 2.2 as a defining property of EVPs, the paper should replace the given argument with the symmetry argument.","section":"2.5"},{"comment":"The theoretical core of the paper is an outline. Section 3 explicitly refers the reader to an unpublished companion paper [9] for the full geometric explanation, and items 4-5 assert the concentration of the angles ∠vi vj uj and ∠vj vi ui and of the tetrahedron rotation angle by citing [7,8] without giving a quantified statement. As written, the manuscript does not deliver the 'principled geometric analysis' promised in Section 6, and there is no derivation of, e.g., a bound on the variance of ℓ2(ui, uj) conditional on ℓ2(vi, vj). A theory should also be consistent with the ordering in Table 2, where the uniform data - the case the concentration reasoning addresses most directly - show the lowest correlations (0.79-0.80) and the real data show the highest (0.94-0.96); the paper should either supply the full argument in this manuscript or explicitly characterise the conditions under which the proxy error is small.","section":"3"},{"comment":"The claim of generic applicability (Abstract; Section 1.3: 'entirely generic') is not supported. The mapping in Section 2.4 selects coordinates by largest absolute magnitude and is therefore coordinate-aligned, while the geometric justification in Section 3 concerns the uniform measure on a high-dimensional sphere. Real embedding sets can have low effective rank or be aligned with the coordinate axes; in such cases the same coordinates tend to be selected for many vectors, and the residual directions that determine fine rank order may be discarded. The paper itself randomly rotates the Laion data 'to avoid the principal components being tightly aligned with the EVP vertices' (Section 5.1, item 4), which is an implicit admission that the proxy can fail in non-generic positions; yet no experiment is reported without this rotation and the method description in Sections 2 and 5 does not instruct practitioners to rotate. I request that the Laion experiments be repeated without rotation and with several rotation seeds, and that the paper state explicitly the data conditions under which the proxy is expected to be valid.","section":"5.1"},{"comment":"The experimental configuration for the Laion data is internally inconsistent. Section 5.1 item 4 and Table 2 state that the CLIP data was reduced by PCA to 500 dimensions and mapped with a {333,500} EVP, while Table 5 lists 'Laion' with 768 dimensions and a {512,768} EVP and attributes to that row the 158x speedup; the Table 5 header row also reads 'Glove Dino Laion', although no Dino dataset appears elsewhere in the paper. Because the 158x speedup is a headline quantitative claim, the authors must correct these inconsistencies and state exactly which configuration produced Tables 2, 5, and Figure 6.","section":"Table 5"}],"minor_comments":[{"comment":"The line 'Since P_i xi = P_i yi = 1' should read 'since Σ x_i^2 = Σ y_i^2 = 1'; as printed, the displayed sums are not the squared norms and the derivation of the correlation between ℓ2 distance and negative scalar product is hard to follow.","section":"1.1"},{"comment":"The continuous Γ-function argument identifies a continuous maximiser at x = 2d/3, but the EVP requires an integer x; the paper should state the exact rounding rule used (e.g., floor, ceil, or nearest integer) for each d and clarify the sense of 'always' for small d.","section":"5.1.1"},{"comment":"The paper calls the encoding '1.58-bit', but for x = 2d/3 the vertex set has per-coordinate entropy log2(C(d, 2d/3) 2^{2d/3})/d ≈ 1.98 bits, and the implementation stores two bit masks (v+, v-) per coordinate, i.e., 2 bits per element before padding to 256-bit words; this terminology should be clarified relative to the unconstrained ternary set.","section":"Title / 1.2"},{"comment":"The Spearman ρ values are computed over pairs whose true-distance distributions differ strongly across data sets (uniform data concentrate near √2, compressing the dynamic range and mechanically lowering ρ); within-dataset method comparisons are valid, but cross-dataset comparison of ρ should be interpreted with care, and confidence intervals would strengthen the claim.","section":"Table 2"},{"comment":"The reported exhaustive-query times should state whether they include the cost of mapping the query vector to its EVP vertex (top-x selection), and whether the data index or query index is rebuilt during the measurement; also, the code URL and the sketch reference in Section 4.2 currently appear as '[?]' placeholders.","section":"Tables 4-5"},{"comment":"The caption asserts that the Voronoi partitions corresponding to the cuboctahedron vertices are congruent, while the figure shows non-uniform faces; the claim needs a brief justification or a more careful wording distinguishing the spherical Voronoi cells from the faces of the polytope.","section":"Figure 2"},{"comment":"The choice x = 2d/3 is justified only by maximising the number of vertices; no experiment in the paper varies x, so the claims in Section 6.3 that this configuration 'generally gives the best search results' and that smaller x decreases accuracy are asserted without the supporting ablation.","section":"5.1.1 / 6.3"},{"comment":"The b1.58 baseline needs one implementation sentence: how its variable number of non-zero entries is handled when the b2sp function is applied, and whether the 30@n recall for that baseline is computed in exactly the same way as for the EVP representation; without this, the baseline comparison is difficult to reproduce.","section":"4.3 / Table 2"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript leans heavily on the authors' own prior work ([7], [8]) and on an in-preparation companion [9] for the theoretical core, and the only self-contained proof in the paper (Section 2.5) is incorrect as written; I recommend that acceptance depend on the authors supplying the geometric argument in the paper or clearly scoping the validity conditions. The discrepancy between Table 2 (Laion 500, {333,500}) and Table 5 (Laion 768, {512,768}), together with the stray 'Dino' in the Table 5 header, suggests the paper was assembled from experiment logs with different configurations; the authors should reconcile these before the next round. The '1.58-bit' branding in the title is somewhat at odds with the 2-bit-per-coordinate implementation and the ~1.98-bit entropy of the constrained vertex set; the editorial team may want the terminology aligned with the actual encoding."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper has a genuinely simple quantisation scheme—keep the top 2d/3 coordinates by magnitude and encode them as ±1, the rest as 0—and it reports surprisingly strong rank correlations (Spearman up to 0.96) and large speedups (33x–158x on exhaustive search) over a million vectors. The bitwise scalar product b2sp is exact for the ternary vectors, and the comparison against 1-bit and b1.58 schemes is fair and informative. The x = 2d/3 choice, derived from maximising the number of Voronoi cells, appears to be new.\n\nThe soft spots are mostly in the theory, not the measurements. Section 3's explanation is explicitly an outline, with the full geometric argument deferred to an unpublished reference [9]. Section 2.5's equi-volume proof contains a false premise: coordinates of uniformly distributed points on a hypersphere are identically distributed but not independent, so the equiprobability of nearest vertices does not follow. More importantly, the proxy is justified by concentration of measure for uniform data on the sphere, but real embeddings can have low effective dimensionality or coordinate alignment. The paper's own decision to randomly rotate the Laion data to avoid 'principal components being tightly aligned with the EVP vertices' is an implicit admission that the method can fail on real, axis-aligned, or low-dimensional structure. No experiment reports the rotation-free case, so the unqualified claim about arbitrary float vectors is not established. Code is also not provided.\n\nThese flaws do not sink the practical contribution: the method is simple, hardware-friendly, and easy to test on any embedding set, and the measured correlations are strong. They do mean the theoretical claims should be read as heuristic until the missing analysis is supplied or the claims are scaled back. The paper is worth a serious referee: it would be a good venue for a revision that either includes the full proof or explicitly labels the geometric argument as empirical motivation, and that reports performance without the random rotation on at least one real-world set. For my own work I wouldn't cite it yet, but I'd bring it to a reading group to discuss where the concentration argument breaks down.","headline":"A practical and fast 1.58-bit embedding quantisation with strong measured correlations, but the geometric proof is deferred and partly flawed, and generality beyond uniform high-dimensional data is untested.","tokens_in":686,"tokens_out":678,"would_cite":false,"duration_ms":23495,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Shrink embeddings to 1.58 bits per value and keep search rank at 0.96","keywords":["ternary quantisation","embedding search","Equi-Voronoi polytope","1.58-bit encoding","bitwise scalar product","k-nearest neighbour","high-dimensional geometry","vector compression"],"falsifier":"Sample pairs from a real embedding set such as GloVe-100, compute the angles $\\angle(v_i, v_j, u_j)$ and $\\angle(v_j, v_i, u_i)$ for each pair's nearest EVP vertices, and check whether the angles are tightly concentrated near $\\pi/2$; a wide spread would falsify the geometric bound that the proxy depends on.","tokens_in":9297,"feed_emoji":"⚡","tokens_out":14784,"duration_ms":109848,"temperature":0.7,"pith_summary":"The paper claims that any set of $\\ell^2$-normalised high-dimensional embedding vectors can be compressed to ternary vectors with exactly $2d/3$ non-zero entries drawn from $\\{-1,0,1\\}$, and that the bitwise scalar product of these compressed vectors gives a proxy for Euclidean distance that preserves rank order with Spearman correlation as high as 0.96. This would mean $k$-nearest-neighbour search over millions of embeddings can run tens to hundreds of times faster on commodity hardware with little loss of result quality. The mapping is generic: it does not depend on the source of the embeddings, only on the geometry of high-dimensional spheres. The authors ground the mapping in a family of convex polytopes they call Equi-Voronoi Polytopes, whose vertices define equal-volume Voronoi cells on the hypersphere.","feed_headline":"Shrink embeddings to 1.58 bits per value and keep search rank at 0.96","feed_subtitle":"A top-magnitude sign rule and a bitwise scalar product deliver 33x to 158x speedups on million-scale data.","key_machinery":"The $\\{x,d\\}$ Equi-Voronoi Polytope (EVP): the convex polytope whose vertices are all $d$-vectors with exactly $x$ non-zero entries each drawn from $\\{-1,0,1\\}$. The nearest-vertex mapping selects the $x$ largest absolute coordinates, assigns $1$ or $-1$ by sign, and zeroes the rest; choosing $x = 2d/3$ maximises the vertex count. Distance between two ternary vectors is computed as the bitwise scalar product $\\mathrm{b2sp}(v,w) = (\\mathrm{bsp}(v^+,w^+) + \\mathrm{bsp}(v^-,w^-)) - (\\mathrm{bsp}(v^+,w^-) + \\mathrm{bsp}(v^-,w^+))$ over two bit vectors per ternary vector, using only SIMD-friendly bit-count operations.","core_discovery":"The central discovery is that the nearest vertex of an Equi-Voronoi Polytope, computed by keeping the signs of the largest-magnitude coordinates and zeroing the rest, yields a distance-preserving proxy in high dimensions. Because a hypersphere in high dimensions concentrates almost all of its volume near its surface, a random point is nearly equidistant from its Voronoi centre, and angles subtended at the two Voronoi vertices by a point and the other vertex are tightly concentrated near right angles. These constraints leave little variance in the length of the unknown side between two points, so the vertex-to-vertex distance is a strong predictor of the true distance. The practical consequence is that the scalar product of two such ternary vectors, evaluated as a bitwise operation over two-bit encodings, reproduces the rank order of Euclidean distance on real datasets such as GloVe, PubMed, and CLIP.","pith_inferences":["The random rotation the authors apply to the Laion data before mapping hints that real embeddings with axis-aligned principal components may need a similar pre-rotation for the concentration assumption to hold; this preprocessing would be a natural extension for other datasets.","The masked-addition variant, in which only one operand is quantised while the other stays in floating point, could offer better accuracy for mixed-precision workloads and is a direct, untested extension of the same geometry.","The same vertex-space mapping could support approximate, multiplication-free matrix multiplication by applying the top-magnitude sign rule to one factor of each scalar product; the paper notes this possibility without evaluating it."],"forward_implications":["On the three million-element real-world datasets tested, exhaustive 100-query searches are 33x, 103x, and 158x faster with the bitwise form than with floating-point Euclidean distance.","The Spearman correlation between the proxy and true distances exceeds both 1-bit and 1.58-bit baselines on every dataset, reaching 0.96 on the Laion CLIP embeddings.","An $\\{x,d\\}$ EVP representation of a 384-dimensional vector occupies 1024 bits, about 8% of the space of a 32-bit float vector, with the gap widening as dimensionality grows.","Because the mapping is defined by a simple coordinate-wise rule and uses only bit-count operations for comparison, it can be dropped into existing approximate-neighbour search pipelines as a fast prefilter."],"supporting_citations":[{"why":"Supplies the high-dimensional facts that volume concentrates near the hypersphere surface and that random vectors are nearly orthogonal, anchoring the geometric constraints.","marker":"[2]"},{"why":"Sources the PubMed and CLIP embedding data used in the correlation and search experiments.","marker":"[5]"},{"why":"Gives the angular concentration result for angles formed by triples of selected points, used to bound the angles at the Voronoi vertices.","marker":"[7]"},{"why":"Extends the angular constraints to the dihedral angle between the two triangles sharing the common edge, closing the tetrahedral bound.","marker":"[8]"},{"why":"Provides the fuller geometric exposition the paper refers to for the complete explanation of why vertex distances approximate true distances.","marker":"[9]"},{"why":"Defines the 1.58-bit ternary quantisation baseline from LLM research that the EVP approach is compared against.","marker":"[12]"},{"why":"Supplies the GloVe-100 word-embedding data set used in the experiments.","marker":"[13]"},{"why":"Supplies the CLIP/Laion embedding data set used in the experiments.","marker":"[14]"},{"why":"Establishes that uniformly distributed points on a hypersphere have independent coordinate distributions, supporting the equal-volume Voronoi cell argument.","marker":"[15]"}],"fun_headline_variants":["Ultra-quantisation: 1.58-bit encodings for fast embedding search","Ternary vectors: bitwise scalar product mimics Euclidean order","Equi-Voronoi trick: extreme compression with high accuracy","From floats to trits: embedding search at 1.58 bits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that high-dimensional geometry guarantees a sampled vector is nearly equidistant from its nearest vertex and that the angles at the two nearest vertices are near right angles; this fails for data with low effective dimensionality or strong axis alignment.","fun_headline_variants_meta":{"raw":{"variants":["Ultra-quantisation: 1.58-bit encodings for fast embedding search","Ternary vectors: bitwise scalar product mimics Euclidean order","Equi-Voronoi trick: extreme compression with high accuracy","From floats to trits: embedding search at 1.58 bits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000941,"raw_usage":{"total_tokens":4017,"prompt_tokens":934,"completion_tokens":3083,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":3006}},"tokens_in":550,"tokens_out":3083,"duration_ms":23644,"temperature":1.0,"reasoning_tokens":3006,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:03:33.168631+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample pairs from a real embedding set such as GloVe-100, compute the angles $\\angle(v_i, v_j, u_j)$ and $\\angle(v_j, v_i, u_i)$ for each pair's nearest EVP vertices, and check whether the angles are tightly concentrated near $\\pi/2$; a wide spread would falsify the geometric bound that the proxy depends on.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the high-dimensional facts that volume concentrates near the hypersphere surface and that random vectors are nearly orthogonal, anchoring the geometric constraints."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sources the PubMed and CLIP embedding data used in the correlation and search experiments."},{"cited_title":"In: Similarity Search and Applications","cited_arxiv_id":null,"evidence_quote":"Gives the angular concentration result for angles formed by triples of selected points, used to bound the angles at the Voronoi vertices."},{"cited_title":"ACM Transactions on Knowledge Discovery from Data18(6), 1–44 (2024)","cited_arxiv_id":null,"evidence_quote":"Extends the angular constraints to the dihedral angle between the two triangles sharing the common edge, closing the tetrahedral bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fuller geometric exposition the paper refers to for the complete explanation of why vertex distances approximate true distances."},{"cited_title":"In: Empirical Methods in Natural Language Processing (EMNLP)","cited_arxiv_id":null,"evidence_quote":"Supplies the GloVe-100 word-embedding data set used in the experiments."},{"cited_title":"github.io/2024/datasets/#768d_clip_embeddings_clip768 16","cited_arxiv_id":null,"evidence_quote":"Supplies the CLIP/Laion embedding data set used in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that uniformly distributed points on a hypersphere have independent coordinate distributions, supporting the equal-volume Voronoi cell argument."}],"review_version":1}