Why do cosine, dot product and L2 rank identically once vectors are unit-normalized?

## The algebra, in full Start from the definition of the dot product in terms of the angle: a·b = ‖a‖ · ‖b‖ · cos(θ) Unit-normalizing means replacing each vector v with v/‖v‖, so that ‖a‖ = ‖b‖ = 1. Substituting gives a·b = cos(θ). The dot product and the cosine are then literally the same number, not merely correlated. Now expand squared Euclidean distance: ‖a - b‖² = (a - b)·(a - b) = ‖a‖² + ‖b‖² - 2(a·b) With unit norms this becomes 2 - 2cos(θ), so the distance itself is √(2 - 2cos θ). Both are strictly decreasing functions of the cosine: as similarity rises from -1 to 1, squared distance falls from 4 to 0. A strictly monotone transform never reorders anything, so the top-k under any of the three functions contains the same items in the same relative order — provided you sort each in its correct direction. ## What this buys you operationally Because the three coincide, normalizing on write turns the metric choice into an implementation detail. Systems typically normalize once at indexing time and once per query, then compare with the dot product, since it is the cheapest of the three: no square roots, no per-comparison division, and it is the operation hardware and libraries optimize hardest. The chosen metric still has to be declared consistently at index build and query time — the equivalence tells you the *ranking* is the same, not that a mismatched configuration is harmless. ## What normalization throws away Normalization is a projection onto the unit sphere and it is not reversible: the magnitude is gone. Whether that matters depends entirely on the model. For most text-embedding models trained with a cosine objective, the magnitude was never a supervised quantity and discarding it is free. For models trained with an inner-product objective, magnitude can encode a learned prior — a per-item bias, a confidence, a popularity term — and normalizing silently deletes a signal the training put there on purpose. Deciding to normalize is therefore a modelling decision, not a hygiene step. ## The asymmetric-normalization case A nuance worth having ready: suppose documents are normalized but queries are not. For the *dot product*, the ranking is unaffected. Scoring is q·d over a fixed q, and scaling q by a constant scales every candidate score by the same constant, which cannot reorder them. For *Euclidean distance*, the ranking absolutely does change — ‖q - d‖² = ‖q‖² + 1 - 2(q·d), and although ‖q‖² is constant, the cross term is now weighted by the query's length relative to a fixed 1, so the geometry shifts. This is why "we normalize documents, that's enough" is correct under one metric and wrong under another. The reverse asymmetry — normalized queries against unnormalized documents — breaks dot-product ranking too, because each candidate carries its own different length into the score. Only the query side is free. ## Scores are not portable, only order is The most common practical trap here is treating the equivalence as if it made the values interchangeable. It does not. A minimum-similarity cut of 0.8 on cosine corresponds to a squared distance of 2 - 2(0.8) = 0.4, that is, a distance of about 0.632. Pasting 0.8 into a distance comparison keeps items out to a far larger radius, which quietly loosens the filter to near-uselessness. When you migrate a system from one metric to the other, convert every constant through the identity rather than reusing it. ## Operations that break unit norm Normalization is a property of the vectors as stored, and ordinary post-processing destroys it. Averaging several chunk vectors into one document vector produces something with length below 1 unless the chunks were identical. Summing vectors, applying a weighted combination, adding a small perturbation, or dimensionality-truncating a vector all leave the unit sphere. Any pipeline that manipulates vectors after embedding needs an explicit renormalization step at the end, otherwise the system drifts back into the unnormalized regime where the three metrics diverge again — and it does so gradually, without an error.