Does the curse of dimensionality break nearest-neighbour search over 1536-dimension embeddings?

## What the curse of dimensionality actually claims The result people are reaching for comes from work in the late 1990s on when nearest-neighbour queries are meaningful. Take n points drawn with independent, identically distributed coordinates in d dimensions and a query point drawn the same way. As d grows, the distance from the query to its nearest point and to its farthest point converge in relative terms: the ratio of the gap to the smallest distance goes to zero. Every point becomes about equally far away, so the notion of "the nearest one" carries almost no information, and any index that prunes by distance loses its ability to prune. The intuition is concentration of measure. A distance is a sum of d per-coordinate contributions. The sum's mean grows roughly like d while its standard deviation grows only like the square root of d, so the *relative* spread — the contrast — shrinks like one over the square root of d. Independent coordinates mean no shared structure to keep some points genuinely close. ## Why real embeddings largely escape it The theorem's premise is the part that fails. A learned text embedding's coordinates are not independent draws: - **Low intrinsic dimension.** The vectors live on a curved, much thinner manifold inside the ambient space. Estimates of intrinsic dimension for sentence embeddings usually land far below the ambient width. Concentration behaviour tracks the intrinsic dimension, not the number of coordinates you happen to store. - **Correlated coordinates.** Training ties dimensions together; the model spends capacity encoding shared semantic structure rather than filling the space uniformly. - **Genuine cluster structure.** Documents about the same subject really are close together, which is the entire reason retrieval works. That structure is signal that concentration arguments assume away. Empirically this shows up as embedding search working perfectly well at 768, 1536, or 3072 dimensions, which is exactly what every production RAG system demonstrates daily. A candidate who says "cosine similarity is meaningless above a few hundred dimensions" is quoting a theorem outside its hypotheses. ## What high dimension does cost you Being right that retrieval still works is only half the answer. The costs that are real: **Memory and bandwidth.** Bytes scale linearly in d, and the searchable working set usually has to be resident to hit latency targets. This is normally the binding constraint. **Compute per comparison.** A dot product or L2 distance is one pass over d values, so every candidate evaluation scales linearly. **Index efficiency.** Approximate indexes rely on geometry to avoid looking at everything. As dimension grows, partitions become less selective and neighbourhood structure less informative, so you must examine more candidates to hit the same recall. The practical symptom is not wrong answers but a worse recall-versus-latency curve: you spend more work per query for the same quality. **Data hunger for anything fitted on top.** Classifiers or reducers fitted over embeddings need more examples as width grows, since the parameter count scales with d. ## How to check your own data instead of arguing If you genuinely suspect concentration, it is measurable in an afternoon: 1. Sample a few thousand query vectors and a large sample of corpus vectors. 2. For each query, compute the distance to its nearest corpus vector and the mean distance to the sample. 3. Report **relative contrast**: mean distance divided by nearest distance. A value comfortably above one means neighbours are meaningfully distinguished; a value hovering near one across your queries is the warning sign. 4. Repeat at several truncated or reduced widths. If contrast is stable while recall@10 is stable, dimension is not the thing hurting you. Pair that with the only metric that decides anything: recall@10 on labeled queries at each candidate width. ## The correct conclusion Dimension is a capacity-versus-cost dial, not a cliff you fall off at some magic width. More dimensions give the model more room to encode distinctions, with diminishing returns; fewer dimensions save memory, comparison work, and index effort. Between those, the quality curve on a real corpus is usually flat over a wide range and then drops. Choose by sweeping that curve. Where concentration reasoning genuinely earns its keep is as a caution about *padding*: if you inflate width without giving the model more real information — say by concatenating redundant features or projecting up — you add all of the cost and none of the capacity, and you do push the geometry toward the uninformative regime. High dimension is only worth paying for when the extra coordinates carry extra signal.