High-Dimensional Geometry and the Concentration of Distances

Overview & motivation

Retrieval in a RAG system is a geometry problem: we embed a query and a corpus of documents as vectors and return the documents whose vectors are nearest the query’s, by Euclidean distance or — after normalization — by cosine similarity. The vectors are not small. A modern text embedding lives in $\mathbb{R}^d$ with $d$ anywhere from a few hundred to a few thousand: $384$ , $768$ , $1024$ , $1536$ , $3072$ . Our intuition for what “near” and “far” mean, though, is trained entirely in $d = 2$ and $d = 3$ , and that intuition is not merely imprecise in high dimensions — it is wrong in ways that bear directly on whether nearest-neighbor retrieval means anything at all.

We will establish three facts, each of which sounds false the first time and each of which is a short computation. Let $\lVert x \rVert = \bigl(\sum_{i=1}^d x_i^2\bigr)^{1/2}$ denote the Euclidean norm and $\langle u, v\rangle = \sum_{i=1}^d u_i v_i$ the inner product. First, the norm concentrates: a standard Gaussian vector in $\mathbb{R}^d$ has length almost exactly $\sqrt{d}$ , with vanishing relative spread, even though its probability density is largest at the origin. Second, random vectors are nearly orthogonal: two independent random directions meet at an angle that concentrates at $90^\circ$ as $d$ grows. Third, and most consequentially, distances concentrate: when data have independent coordinates, every pair of points ends up at essentially the same distance, so the nearest and the farthest neighbor of a query become indistinguishable. That third fact is the curse of dimensionality, and taken at face value it says nearest-neighbor retrieval is hopeless.

It is not hopeless, and the resolution is the most important idea in the topic: real embeddings do not have independent coordinates and do not fill $\mathbb{R}^d$ . They lie close to a low-dimensional surface inside it, and the curse is governed by that surface’s intrinsic dimension, not by the ambient $d$ . This is exactly the structure that approximate nearest-neighbor indexes exploit. Before the algebra, drag the dimension slider and watch each phenomenon arrive:

dimension d = 10

Query-to-point distance, as a fraction of its mean (faint curve: d = 2). As d grows the distribution collapses onto its mean — every point is the same distance away.

relative contrast at d = 10

2.3012

Var(D²)/E[D²]² = 2/d

0.20000

nearest ≈ farthest?

Switch between the panels: the distance distribution collapsing onto its mean is the curse; the inner-product distribution spiking at zero is near-orthogonality; the norm pinning to $\sqrt{d}$ is the thin shell; and the last panel is the resolution — at $d = 1536$ , structured data keeps its contrast where unstructured data destroys it.

What we cover

The thin shell: why a Gaussian vector’s norm concentrates at $\sqrt{d}$ , and why mass is not density.
Near-orthogonality: why two random vectors are almost always perpendicular, and how much room that leaves.
The concentration of distances: the vanishing relative variance we can prove, and the nearest-equals-farthest statement we cite and verify.
Volume in high dimensions: the unit ball’s vanishing volume, its flight to the surface, and the equator.
Intrinsic dimension: why retrieval works in practice, and how to estimate the dimension that matters.
A finance case study, the honest caveats, and the tested implementation.

The thin shell: the norm of a random vector concentrates

Take the simplest high-dimensional object, a standard Gaussian vector $x \sim \mathcal{N}(0, I_d)$ — each coordinate $x_i$ an independent standard normal. Its density $f(x) \propto e^{-\lVert x\rVert^2/2}$ is largest at the origin, so one might guess that a typical sample sits near the origin. It does not. Almost every sample sits at distance close to $\sqrt{d}$ .

Theorem 1 (Thin-shell concentration of the norm).

For $x \sim \mathcal{N}(0, I_d)$ ,

\mathbb{E}\lVert x \rVert^2 = d, \qquad \operatorname{Var}\lVert x \rVert^2 = 2d,

so for every $\varepsilon \in (0, 1)$ ,

\Pr\!\left(\,\left|\tfrac{1}{d}\lVert x\rVert^2 - 1\right| \ge \varepsilon\right) \;\le\; \frac{2}{\varepsilon^2 d} \;\xrightarrow[d\to\infty]{}\; 0.

Hence $\lVert x\rVert / \sqrt{d} \to 1$ in probability: the mass concentrates on a thin shell of radius $\sqrt{d}$ .

Proof.

Write $\lVert x\rVert^2 = \sum_{i=1}^d x_i^2$ , a sum of $d$ independent random variables, each $x_i^2$ distributed as a chi-squared with one degree of freedom. A single $x_i^2$ has $\mathbb{E}[x_i^2] = 1$ and $\operatorname{Var}(x_i^2) = \mathbb{E}[x_i^4] - 1 = 3 - 1 = 2$ , using the fourth moment of a standard normal. By linearity of expectation and independence (which makes the variance of a sum the sum of variances), $\mathbb{E}\lVert x\rVert^2 = d$ and $\operatorname{Var}\lVert x\rVert^2 = 2d$ . Now apply Chebyshev’s inequality to $\lVert x\rVert^2$ :

\Pr\!\left(\bigl|\lVert x\rVert^2 - d\bigr| \ge \varepsilon d\right) \le \frac{\operatorname{Var}\lVert x\rVert^2}{(\varepsilon d)^2} = \frac{2d}{\varepsilon^2 d^2} = \frac{2}{\varepsilon^2 d}.

Dividing the event through by $d$ gives the stated bound, which tends to $0$ ; convergence of $\lVert x\rVert^2/d$ to $1$ in probability carries over to $\lVert x\rVert/\sqrt{d}$ because the square root is continuous at $1$ . $\blacksquare$

∎

The relative spread shrinks like $1/\sqrt{d}$ , which the companion code confirms: the standard deviation of $\lVert x\rVert/\sqrt{d}$ falls from $0.224$ at $d = 10$ to $0.022$ at $d = 1000$ . The geometric reading is the part worth keeping. The density is indeed maximal at the origin, but density is not mass. The mass at radius $r$ is the density there times the surface area of the radius- $r$ sphere, which grows like $r^{d-1}$ ; the radial density of $\lVert x\rVert$ is therefore proportional to $r^{d-1} e^{-r^2/2}$ , and maximizing it gives a mode at $r = \sqrt{d-1} \approx \sqrt{d}$ . The exploding volume factor $r^{d-1}$ pulls the mass outward against the decaying Gaussian factor, and the two balance in a thin shell far from the most likely single point. In high dimensions, where the typical sample lives and where the density is largest are different places entirely.

Near-orthogonality of random vectors

The second surprise concerns angles. Pick two directions at random — two independent vectors uniform on the unit sphere $S^{d-1} = \{x : \lVert x\rVert = 1\}$ . In the plane their angle is uniform, and they are as likely to be aligned as perpendicular. In high dimensions they are almost certainly perpendicular.

Theorem 2 (Near-orthogonality of random directions).

Let $u$ be a fixed unit vector and $v$ uniform on $S^{d-1}$ (equivalently, let both be independent and uniform). Then

\mathbb{E}\langle u, v\rangle = 0, \qquad \operatorname{Var}\langle u, v\rangle = \frac{1}{d},

so $\langle u, v\rangle \to 0$ in probability and the angle between $u$ and $v$ concentrates at $90^\circ$ .

Proof.

By the rotational invariance of the uniform measure on the sphere, we may rotate so that $u = e_1$ is the first coordinate axis; then $\langle u, v\rangle = v_1$ , a single coordinate of a uniform random unit vector. Its sign is symmetric, so $\mathbb{E}[v_1] = 0$ . For the variance, use that $v$ lies on the sphere: $\sum_{i=1}^d v_i^2 = 1$ always, so taking expectations, $\sum_{i=1}^d \mathbb{E}[v_i^2] = 1$ . By symmetry every coordinate has the same $\mathbb{E}[v_i^2]$ , so each equals $1/d$ , giving $\operatorname{Var}\langle u, v\rangle = \mathbb{E}[v_1^2] = 1/d$ . Chebyshev’s inequality then bounds $\Pr(|\langle u, v\rangle| \ge t) \le 1/(t^2 d) \to 0$ . $\blacksquare$

∎

Chebyshev gives a polynomial rate; the truth is exponential. The projection $\langle u, v\rangle$ of a uniform unit vector is in fact sub-Gaussian with parameter $1/d$ , and the sharper bound $\Pr(|\langle u, v\rangle| \ge t) \le 2\,e^{-(d-2)t^2/2}$ holds (Vershynin, 2018). The companion code matches the variance to $1/d$ across the whole grid — $0.00998$ at $d = 100$ against a theoretical $0.0100$ . The consequence is one of the quiet engines of vector search: because deviations from orthogonality are exponentially unlikely, one can pack exponentially many — $e^{c\varepsilon^2 d}$ — unit vectors that are pairwise within $\varepsilon$ of orthogonal. A high-dimensional space has room for an astronomical number of nearly distinguishable directions, which is why a single embedding space can keep millions of concepts mutually resolvable, and which is the geometric seed of the Johnson–Lindenstrauss lemma on distance-preserving random projections.

The concentration of distances (the curse)

Now the result that threatens retrieval directly. Take a query point and a cloud of data points, all with independent coordinates, and ask for the distribution of the query-to-point distances. In low dimensions some points are clearly near and some clearly far. In high dimensions they are all, to first order, the same distance away.

Definition 1 (Relative contrast).

For a query $q$ and data points $x_1, \dots, x_n$ , let $D_i = \lVert q - x_i\rVert$ , with $D_{\min} = \min_i D_i$ and $D_{\max} = \max_i D_i$ . The relative contrast is

\frac{D_{\max} - D_{\min}}{D_{\min}}.

When the relative contrast is large, the nearest neighbor stands out; when it tends to $0$ , nearest and farthest are indistinguishable.

The engine of the collapse is a statement about variance. Squared distance is a sum of independent per-coordinate terms, so its mean grows linearly in $d$ while its standard deviation grows only like $\sqrt{d}$ — the relative spread vanishes.

Theorem 3 (The relative variance of squared distance vanishes).

Let $X, Y$ be independent with i.i.d. coordinates of finite mean and variance, and $D^2 = \lVert X - Y\rVert^2$ . Then

\frac{\operatorname{Var}(D^2)}{\bigl(\mathbb{E}[D^2]\bigr)^2} = \frac{\operatorname{Var}(Z)}{d\,\bigl(\mathbb{E}[Z]\bigr)^2} \xrightarrow[d\to\infty]{} 0, \qquad Z := (X_1 - Y_1)^2,

so $D^2 / \mathbb{E}[D^2] \to 1$ in probability: all pairwise squared distances concentrate at the common value $\mathbb{E}[D^2]$ . For standard Gaussian coordinates this ratio is exactly $2/d$ .

Proof.

Write $D^2 = \sum_{i=1}^d Z_i$ with $Z_i = (X_i - Y_i)^2$ independent and identically distributed, each with finite mean $\mathbb{E}[Z]$ and variance $\operatorname{Var}(Z)$ (finite because the coordinates have finite fourth moment under the stated assumptions). Then $\mathbb{E}[D^2] = d\,\mathbb{E}[Z]$ and, by independence, $\operatorname{Var}(D^2) = d\,\operatorname{Var}(Z)$ , so

\frac{\operatorname{Var}(D^2)}{(\mathbb{E}[D^2])^2} = \frac{d\,\operatorname{Var}(Z)}{d^2\,(\mathbb{E}[Z])^2} = \frac{\operatorname{Var}(Z)}{d\,(\mathbb{E}[Z])^2},

which tends to $0$ . By Chebyshev’s inequality applied to $D^2/\mathbb{E}[D^2]$ , this forces $D^2/\mathbb{E}[D^2] \to 1$ in probability. For $X, Y \sim \mathcal{N}(0, I_d)$ each difference $X_i - Y_i \sim \mathcal{N}(0, 2)$ , so $Z = (X_i - Y_i)^2$ has $\mathbb{E}[Z] = 2$ and $\operatorname{Var}(Z) = 8$ — using $\operatorname{Var}(2\chi^2_1) = 4\cdot 2$ — giving the ratio $8 / (4d) = 2/d$ . $\blacksquare$

∎

From the vanishing relative variance, the collapse of contrast follows for any fixed number of points by a union bound — this much we can prove outright.

Theorem 4 (Vanishing contrast).

Fix $n$ , and let $q, x_1, \dots, x_n$ be independent with i.i.d. coordinates of finite mean and variance. Then the relative contrast $(D_{\max} - D_{\min})/D_{\min} \to 0$ in probability as $d \to \infty$ .

Proof.

By Theorem 3 applied to each pair $(q, x_i)$ , we have $D_i^2 / \mathbb{E}[D^2] \to 1$ in probability, where $\mathbb{E}[D^2] = d\,\mathbb{E}[Z]$ is common to all $i$ . Fix $\varepsilon > 0$ . Because $n$ is fixed, the union bound gives

\Pr\!\left(\exists\, i \le n : \left|\tfrac{D_i^2}{\mathbb{E}[D^2]} - 1\right| > \varepsilon\right) \le \sum_{i=1}^n \Pr\!\left(\left|\tfrac{D_i^2}{\mathbb{E}[D^2]} - 1\right| > \varepsilon\right) \xrightarrow[d\to\infty]{} 0.

So with probability tending to $1$ , every $D_i^2$ lies in $\mathbb{E}[D^2]\,[1-\varepsilon, 1+\varepsilon]$ at once, hence every $D_i$ lies in $\sqrt{\mathbb{E}[D^2]}\,[\sqrt{1-\varepsilon}, \sqrt{1+\varepsilon}]$ , and therefore

\frac{D_{\max}}{D_{\min}} \le \sqrt{\frac{1+\varepsilon}{1-\varepsilon}}.

As $\varepsilon$ may be taken arbitrarily small, $D_{\max}/D_{\min} \to 1$ , i.e. the relative contrast tends to $0$ , in probability. $\blacksquare$

∎

Remark (The general statement (Beyer et al.)).

Theorem 4 fixes $n$ and lets $d$ grow. The influential statement of Beyer, Goldstein, Ramakrishnan, and Shaft (1999) covers the harder regime in which the number of points may grow with the dimension: under broad conditions — essentially that the relative variance of distance vanishes — the ratio $D_{\max}/D_{\min} \to 1$ in probability, and no query is meaningfully closer to its nearest point than to its farthest. We cite their theorem rather than reprove it at that generality; the companion harness verifies the contrast collapse numerically across Gaussian, cube, and sphere data, where the mean relative contrast falls from $47.1$ at $d = 2$ to $0.11$ at $d = 1000$ .

The practical reading is stark. If document embeddings genuinely behaved like vectors with independent coordinates, then in $1000$ dimensions the nearest document to a query would be only a few percent closer than the farthest, the ranking would be dominated by noise, and exact nearest-neighbor search would return essentially arbitrary results. This is the honest case against vector retrieval, and it is also why the field needs everything downstream of this topic: when contrast is thin, paying $O(n)$ for an exact nearest neighbor buys almost nothing, so we trade exactness for speed with approximate indexes. But the premise — independent coordinates — is false for real embeddings, and the rest of the topic is about why.

Volume in high dimensions

The distance results have a purely geometric companion: where the volume of a high-dimensional region sits is as counterintuitive as where a Gaussian’s mass sits. Both are computed with the spherical change of variables, and both say the same thing — the interior empties out.

Proposition 1 (The ball's volume vanishes and flees to the surface).

The volume of the unit ball in $\mathbb{R}^d$ is

V_d = \frac{\pi^{d/2}}{\Gamma\!\left(\tfrac{d}{2} + 1\right)},

which increases to a maximum at $d = 5$ and then decreases to $0$ super-exponentially. Moreover, the fraction of the ball’s volume lying within $\varepsilon$ of its surface is

1 - (1 - \varepsilon)^d \xrightarrow[d\to\infty]{} 1.

Proof.

The volume formula is the standard integral of $1$ over the ball in spherical coordinates, where the radial Jacobian $r^{d-1}$ and the Gamma function arise from the change of variables; we take it as given. For the shell, a ball of radius $1 - \varepsilon$ has volume $(1-\varepsilon)^d V_d$ because volume scales as the $d$ -th power of the radius, so the inner ball holds a fraction $(1-\varepsilon)^d$ of the total and the outer shell of thickness $\varepsilon$ holds the complement $1 - (1-\varepsilon)^d$ . For any fixed $\varepsilon > 0$ , $(1-\varepsilon)^d \to 0$ . $\blacksquare$

∎

The numbers are vivid: $V_d$ peaks at $V_5 \approx 5.26$ , has already fallen to $V_{20} \approx 0.026$ , and is about $2\times 10^{-40}$ by $d = 100$ — the unit ball barely exists inside the unit cube it sits in. And what little volume there is hugs the boundary: by $d = 50$ , more than $99\%$ of the ball lies in its outer $10\%$ shell. A high-dimensional orange is almost all peel.

The same concentration appears on the sphere as the equatorial phenomenon, which is Theorem 2 read as a statement about measure rather than about a random pair.

Proposition 2 (Equatorial concentration).

Fix a unit vector $u$ . For $v$ uniform on $S^{d-1}$ , the equatorial band $\{v : |\langle u, v\rangle| \le t\}$ has measure at least $1 - 2\,e^{-(d-2)t^2/2}$ . Taking $t$ of order $1/\sqrt{d}$ , almost all of the sphere’s surface area lies within $O(1/\sqrt{d})$ of any equator.

Proof.

The complement of the band is exactly the event $|\langle u, v\rangle| > t$ , whose probability is bounded by the sub-Gaussian tail $2\,e^{-(d-2)t^2/2}$ quoted after Theorem 2 (Vershynin, 2018), so the band itself has measure at least $1 - 2\,e^{-(d-2)t^2/2}$ . This is the special, projection-onto-a-line case of Lévy’s lemma — the isoperimetric concentration of measure on the sphere, that any set of measure $\tfrac12$ has almost all of the sphere within a $O(1/\sqrt{d})$ neighborhood — which we cite for the general form (Blum, Hopcroft, and Kannan, 2020). $\blacksquare$

∎

Pick any equator you like — any great subsphere through the center — and nearly the entire surface of a high-dimensional sphere crowds into a thin band around it. There is no contradiction in this holding for every equator at once; it is the measure-theoretic face of the same fact that two random directions are nearly orthogonal.

Why retrieval still works: intrinsic dimension

Every theorem so far assumed independent coordinates — data that genuinely fill $\mathbb{R}^d$ . Real embeddings do not. A sentence embedder maps text onto a curved, low-dimensional surface sitting inside the ambient space; the coordinates are strongly dependent, and the data occupy a tiny, structured sliver of the cube. The dimension that controls concentration is the dimension of that surface, the intrinsic dimension $k$ , not the ambient $d$ .

Proposition 3 (Contrast is governed by intrinsic, not ambient, dimension).

Suppose the data lie on a $k$ -dimensional affine subspace isometrically embedded in $\mathbb{R}^d$ . Then every pairwise distance equals the distance between the corresponding points in their $k$ intrinsic coordinates, and the relative variance of squared distance is $\Theta(1/k)$ — independent of the ambient dimension $d$ . Contrast is therefore preserved whenever $k$ is small, no matter how large $d$ is.

Proof.

An isometric embedding of a $k$ -dimensional subspace into $\mathbb{R}^d$ preserves inner products and hence distances: if $x = U a$ and $y = U b$ for an orthonormal $d \times k$ matrix $U$ and intrinsic coordinates $a, b \in \mathbb{R}^k$ , then $\lVert x - y\rVert = \lVert U(a-b)\rVert = \lVert a - b\rVert$ because $U^\top U = I_k$ . So the distance geometry is entirely that of the $k$ -dimensional coordinates, and applying Theorem 3 there gives a relative variance of order $1/k$ , with no dependence on $d$ . $\blacksquare$

∎

How would we know $k$ for data we cannot see a basis for? We estimate it. The TwoNN estimator (Facco et al., 2017) reads the intrinsic dimension off the ratio $\mu = r_2 / r_1$ of each point’s second- to first-nearest-neighbor distance: for data locally uniform on a $k$ -manifold, $\Pr(\mu > t) = t^{-k}$ , so plotting $-\log\bigl(1 - F(\mu)\bigr)$ against $\log \mu$ yields a line through the origin of slope $k$ . The companion code confirms it recovers a planted intrinsic dimension — structured data with $k = 10$ embedded in $\mathbb{R}^{200}$ reads TwoNN $\approx 9.8$ — and that it reports the ambient dimension when the data genuinely fill the space. Intrinsic dimension is precisely the quantity that the indexing methods later in the curriculum exploit: an approximate index can be fast and accurate exactly because the data it organizes occupy far fewer effective dimensions than the embedding nominally has, the same low-rank structure that PCA makes explicit.

Finance case study

§ Finance case study text

The production financial-RAG system behind the BM25 and rank-fusion case studies embeds 10-K filing passages and earnings-call chunks into a $1536$ -dimensional space and retrieves by cosine similarity. Theorems 1–4 raise a real worry: in $1536$ dimensions, with near-orthogonality and distance concentration both in full force, is cosine similarity between two document embeddings meaningful at all, or is the top- $k$ list just noise?

The answer is that it is meaningful, and the reason is intrinsic dimension. We make the point with a self-contained experiment rather than a claim. Take $1536$ -dimensional vectors of two kinds: structured data with a planted intrinsic dimension $k = 10$ — a low-rank financial-topic signal plus a little ambient noise, a stand-in for the manifold a real encoder produces — and i.i.d. Gaussian data that genuinely fills $\mathbb{R}^{1536}$ . The companion code measures, at the same ambient dimension:

structured: relative contrast $\mathbf{2.18}$ , TwoNN intrinsic dimension $\mathbf{9.58}$ — the nearest neighbor stands out by more than a factor of two, and the estimator recovers the planted $k$ ;
i.i.d.: relative contrast $\mathbf{0.088}$ , TwoNN far higher — contrast has collapsed, and the nearest neighbor is barely distinguishable from the farthest.

Same ambient dimension, opposite outcomes. The structured set is what document embeddings actually look like — a corpus about interest-rate risk does not scatter uniformly through $\mathbb{R}^{1536}$ , it clusters on a low-dimensional surface — and on that surface the geometry is effectively ten-dimensional, where contrast is healthy and nearest-neighbor retrieval is exactly as meaningful as our two- and three-dimensional intuition expects. The curse is real, but it indicts unstructured data, not vector search over embeddings that carry structure.

A candid note on the companion code, in the spirit of the rest of the series: the “structured” embeddings are a synthetic low-rank-plus-noise construction, chosen so the notebook is deterministic, CPU-only, and reproducible to the decimal with no model download. They are not the output of a trained encoder, and TwoNN on the i.i.d. set returns a value far below $1536$ — the estimator saturates below the ambient dimension at finite sample size — so we report the gap (the i.i.d. estimate stays an order of magnitude above the structured one) rather than claim the estimator recovers $d$ exactly. The mathematics being illustrated — that contrast tracks intrinsic, not ambient, dimension — does not depend on how the structured set is produced.

Honest caveats

⚠ Rigor flag

Four qualifications keep the geometry honest:

“Distance is meaningless in high dimensions” is conditional. The curse (Theorems 3–4, Beyer et al. 1999) assumes data with i.i.d.-like coordinates that fill $\mathbb{R}^d$ . It is false for real embeddings, which lie near a low-dimensional manifold; concentration is governed by the intrinsic dimension (Proposition 3), which is why nearest-neighbor retrieval works at all.
The results are asymptotic and in probability. Chebyshev gives only polynomial rates; the exponential thin-shell and near-orthogonality rates require the sub-Gaussian structure of the Gaussian and the sphere. Finite-dimensional behavior depends on the constants.
We prove the core and cite the rest. Theorems 1–4 and Propositions 1 and 3 are proved; the general nearest-equals-farthest statement (Beyer et al.) and the general Lévy concentration (Proposition 2) are cited and, where applicable, verified numerically rather than reproven.
The intrinsic-dimension story is empirical at the edges. TwoNN is a model-based estimator that assumes local uniformity, and the finance “embeddings” are a synthetic stand-in, not a trained encoder.

Implementation

The companion notebook (notebookPath) is self-contained, CPU-only, and runs in about a second on numpy and scipy. It samples the three data models, computes the relative contrast, the inner-product and norm statistics, the ball volume and shell fractions, and the TwoNN intrinsic-dimension estimate, and prints the grid table that the ConcentrationLaboratory mirrors to the decimal. Its verification harness makes each claim of the topic executable: the chi-squared moments match scipy exactly so $\mathbb{E}\lVert x\rVert^2 = d$ and $\operatorname{Var}\lVert x\rVert^2 = 2d$ (Theorem 1); the shell width of $\lVert x\rVert/\sqrt{d}$ shrinks monotonically to $0.022$ at $d = 1000$ (Theorem 1); the inner-product variance tracks $1/d$ and the typical angle approaches $90^\circ$ (Theorem 2); the relative contrast falls monotonically toward $0$ for Gaussian, cube, and sphere data (Theorems 3–4); the relative variance of squared distance matches $2/d$ (Theorem 3); the unit ball’s volume collapses while its outer shell captures all the mass (Proposition 1); and TwoNN recovers a planted intrinsic dimension while contrast at $d = 1536$ separates structured from i.i.d. data (Proposition 3, the finance case). The three pillars — the proofs above, the laboratory, and the tested code — agree by construction.

High-Dimensional Geometry and the Concentration of Distances

Overview & motivation

What we cover

The thin shell: the norm of a random vector concentrates

Near-orthogonality of random vectors

The concentration of distances (the curse)

Volume in high dimensions

Why retrieval still works: intrinsic dimension

Finance case study

Honest caveats

Implementation

Connections

References & Further Reading