Hyperbolic Oracles · Explainer 11 · The 0.7 Claude Fence

The 0.7 Claude Fence: a Tracer

Tukey’s fence flags a value beyond Q3 + 1.5 IQR. In flat space that rule needs one choice: the coordinate it acts on. Hyperbolic space makes the choice for us. Fence the mass each point reaches, the volume of the ball about a robust centre out to it. Then a cloud spread as a hierarchy spreads has its fence on the doubling sphere of its upper quartile. In the hyperbolic plane that sphere lies within ln 2 = 0.693… of the quartile, and approaches exactly that as the cloud grows. This page follows one run stage by stage, with its exact numbers. It proves the nine lemmas and checks each one live.

Open the fence The live trace Lemmas CF1–CF9 Preregistered results Manual The fence in 3D
live checks: running… trace parity: running… ln 2 = 0.6931… the doubling gap of H² 3 / 3 preregistered claims supported
Contents
  1. Where this sits
  2. Tukey’s fence needs a coordinate
  3. Why 0.7
  4. The algorithm
  5. The live trace
  6. Lemmas
  7. The guard
  8. Results
  9. Live checks
  10. Manual
  11. The fence in 3D
  12. What is not claimed
01

Where this sits

The fence is one of the Hyperbolic Oracles, beside Hyperbolic PolyStep. Its points are exact, lifted by the ghost map like every state of the step. It runs on the browser core’s exact rationals. One of its datasets is version 2’s norm-5 tree in H³. Click any box to travel.

02

Tukey’s fence needs a coordinate

Tukey’s rule is about one number per observation. Take the quartiles Q1 and Q3 and the spread IQR = Q3 − Q1. A value past Q3 + 1.5 IQR is outside; past Q3 + 3 IQR it is far out.

Two facts decide everything that follows.

  • Shifting or stretching the coordinate changes nothing. Replace every value \(v\) by \(\alpha v + \beta\) with \(\alpha > 0\): the quartiles and the fences move with it, and every verdict stays the same.
  • Bending it changes the verdicts. Replace \(v\) by \(v^2\), \(e^v\) or \(\cosh v\): the quartiles still pick the same observations, because order is kept (CF1). The fences do not move with them, so the verdicts change.

A multivariate cloud gives no single number, so the usual move is to measure each point’s distance from a centre and fence the distances. Many monotone functions of that distance would serve as well: the distance itself, its square, the volume it encloses. In flat space the choice is a matter of taste. In hyperbolic space it is not, because volume grows exponentially with distance. A Tukey fence on distance ends up measuring the wrong thing (CF7).

The choice

Fence the mass: \(m(x) = \operatorname{vol} B(c, d(c,x))\), the volume of the ball about the centre that reaches \(x\). For a cloud spread evenly through a ball, mass is uniform in every curvature and every dimension (CF3), so Tukey’s rule treats the flat and hyperbolic cases alike. In the hyperbolic plane the mass is also a rational function of the data (CF4).

03

Why 0.7

The fence keeps Tukey’s 1.5 and 3. The 0.7 is not a dial. It is what those constants become in hyperbolic space.

On a uniform cloud the mass quartiles are \(V/4\) and \(3V/4\), where \(V\) is the volume of the data’s ball. Tukey’s outside fence is then \(3V/4 + 1.5 \cdot V/2 = 3V/2\): exactly twice the mass of the upper quartile. The fence is the upper quartile’s doubling sphere. How far beyond the quartile does that sphere lie? In flat space the gap is \((2^{1/n} - 1)\,Q3\), which grows with the cloud without bound. In the hyperbolic plane it never reaches ln 2 and tends to it (CF5):

$$\cosh(Q_3 + t) - 1 = 2(\cosh Q_3 - 1) \;\Longrightarrow\; t < \ln 2 = 0.6931\ldots,\qquad t \to \ln 2 \text{ as } Q_3 \to \infty.$$

In H³ the gap is below \(\tfrac12\ln 2\), and in \(H^n\) below \(\ln 2/(n-1)\). The far-out fence is the tripling sphere, within \(\ln 3/(n-1)\) of Q3. The same constant has a second life. ln 2 is the sharp four-point constant of hyperbolic space (CF9): no four points of \(H^n\) violate the tree condition by more than ln 2. That bound is the fence’s guard for data given as distances.

Drawn live by fence.js: the doubling gap \(t(Q_3)\) in H² (amber) and H³ (rose) against the quartile radius, with ln 2 and \(\tfrac12\ln 2\) dashed and the flat gap \((\sqrt2 - 1)Q_3\) in grey.
04

The algorithm

Seven stages, each resting on a lemma. The chart shows which lemma carries which stage, and where exact arithmetic ends and floats begin.

  1. Data. Exact points of \(H^2\) or \(H^3\): rational Poincaré points on the grid \(2^{-32}\), lifted by the ghost map \(G(p) = (1 + |p|^2,\, 2p)/(1 - |p|^2)\).
  2. Centre. The Fréchet median, the minimiser of \(\sum_i d(c, x_i)\), found by the hyperbolic Weiszfeld iteration \(c \leftarrow \mathrm{normalise}\big(\sum_i x_i / \sinh d(c, x_i)\big)\). The centre is carried as weights on the data, so the iteration reads only the Gram matrix \(G_{ij} = -\langle x_i, x_j\rangle\). The final weights are read as exact binary rationals, giving an exact integer vector \(S\) with \(c = S/\sqrt{\sigma}\), \(\sigma = -\langle S, S\rangle\) (CF2, CF8).
  3. Radii. \(b_i = -\langle S, x_i\rangle\), exact. So \(\cosh d_i = b_i/\sqrt\sigma\) and \(\sinh^2 d_i = (b_i^2 - \sigma)/\sigma\), both exact rationals (CF1).
  4. Quartiles. Order statistics: the \(\lceil N/4\rceil\)-th and \(\lceil 3N/4\rceil\)-th smallest radius. They are the same points in every monotone coordinate (CF1).
  5. Mass. \(m = \cosh d - 1\) in \(H^2\) (area over \(2\pi\)) and \(m = \sinh 2d - 2d\) in \(H^3\) (volume over \(\pi\)) (CF3, CF4).
  6. Fences and verdicts. Outside past \(m_3 + 1.5(m_3 - m_1)\), far out past \(m_3 + 3(m_3 - m_1)\). In \(H^2\) these are two rational comparisons, \(2b > 5b_3 - 3b_1\) and \(b > 4b_3 - 3b_1\). In \(H^3\), a point at or inside the upper quartile is decided by order alone. Any other point is decided in fixed point with a proved error bound, doubling the precision until the sign is certain (CF4, CF5).
  7. Guard. For data given as distances: any quadruple whose four-point defect exceeds ln 2 proves a wrong distance (CF9).
05

The live trace

A small cloud, fenced in this browser by the same code as the fence page. Step through the stages. Each one draws what it computes and prints its numbers, the exact ones as they are.

The tracer
06

Lemmas CF1–CF9

These are the lemmas the fence is derived from. Each one is checked live in section 9 against a control that must fail. CF8’s breakdown point and CF9’s upper bound are cited, not reproved here, and say so.

CF1 · Order statistics. Let \(\varphi\) be strictly increasing, and break ties by index. Then the \(k\)-th smallest of \(\varphi(v_1), \dots, \varphi(v_N)\) is \(\varphi\) of the \(k\)-th smallest \(v_i\). Quartiles taken as order statistics, \(v_{(\lceil N/4\rceil)}\) and \(v_{(\lceil 3N/4\rceil)}\), are therefore the same observations whether the coordinate is the distance \(d\), \(\cosh d\), \(\sinh^2 d\) or the mass.

A strictly increasing \(\varphi\) preserves every comparison \(v_i < v_j\), so it preserves the sorted order and each rank. An interpolating quantile, such as \(v_{(k)} + h(v_{(k+1)} - v_{(k)})\), averages two observations. \(\varphi\) of an average is not the average of the \(\varphi\)s unless \(\varphi\) is affine. So only order-statistic quantiles commute with the change of coordinate. That is also why the radii can be compared exactly: comparing \(b_i\) compares \(d_i\).

CF2 · Exact invariance. For any rational Lorentz isometry \(g \in O^+(1,n)\), the fence of \(g x_1, \dots, g x_N\) returns the same verdicts, the same certificates and the same floats as the fence of \(x_1, \dots, x_N\), bit for bit, and its centre is \(g c\).

The Gram matrix \(G_{ij} = -\langle x_i, x_j\rangle\) is invariant under \(g\), and it is an exact rational, correctly rounded to the same double. The medoid’s sums and every Weiszfeld step read only \(G\), in a fixed order of operations, so they produce the same weights \(\lambda\) bit for bit. The exact centre vector is \(S = \sum \lambda_i x_i\) up to a positive scale, and it becomes \(gS\). Then \(b_i = -\langle gS, g x_i\rangle\) is unchanged, and so are \(\sigma\), the order, both rational comparisons and every fixed-point evaluation.

CF3 · The doubling sphere. Let the data be uniform on a ball \(B(c, R)\) of any space form, flat or hyperbolic, in any dimension. Then the mass \(m(X)\) is uniform on \([0, V(R)]\). Tukey’s outside fence on mass is \(\tfrac32 V(R) = 2 m_3\), the doubling sphere of the upper quartile, and the far-out fence is \(\tfrac94 V(R) = 3 m_3\).

By definition \(m(x) = V(d(c,x))\) with \(V\) increasing, so \(P(m(X) \le t) = P(d \le V^{-1}(t)) = V(V^{-1}(t))/V(R) = t / V(R)\). The quartiles are \(m_1 = V/4\) and \(m_3 = 3V/4\), with IQR \(V/2\). So \(m_3 + 1.5\,\mathrm{IQR} = 3V/2 = 2m_3\) and \(m_3 + 3\,\mathrm{IQR} = 9V/4 = 3m_3\). No inlier has mass above \(V\), so asymptotically the fence flags none of them. Sample quartiles approach these values at rate \(N^{-1/2}\).

CF4 · Exact in the plane, certified in space. In \(H^2\) the area of the disc about \(c\) through \(x\) is \(2\pi(\cosh d - 1)\). The verdicts are exactly \(2b > 5b_3 - 3b_1\) (outside) and \(b > 4b_3 - 3b_1\) (far out), comparisons of rationals. In \(H^3\) the volume is \(\pi(\sinh 2d - 2d)\). Its sign tests are decided by fixed-point evaluation with a proved error bound, or by order when the point is at or inside \(Q_3\).

Integrate the circumference \(2\pi\sinh r\) of the circle of radius \(r\) to get the area \(2\pi(\cosh d - 1)\). With \(\cosh d = b/\sqrt\sigma\), the rule \(m > m_3 + \tfrac32(m_3 - m_1)\) becomes \(2(\cosh d - 1) > 5(\cosh d_3 - 1) - 3(\cosh d_1 - 1)\). The constants cancel, leaving \(2\cosh d > 5\cosh d_3 - 3\cosh d_1\). Multiplying by \(\sqrt\sigma > 0\) gives \(2b > 5b_3 - 3b_1\). The far-out rule follows in the same way. In \(H^3\), integrating \(4\pi\sinh^2 r\) gives \(\pi(\sinh 2d - 2d)\). With \(a = \cosh d\) this is \(2a\sqrt{a^2-1} - 2\ln(a + \sqrt{a^2 - 1})\), computed from the exact \(a^2\) by integer square roots and an integer series for ln, each with its error bounded (mass_fixed). A sign is accepted only when the value exceeds the summed bounds. Otherwise the precision doubles, from 64 bits up to 8,192, and an undecided sign would be reported as unresolved. None was, in 1,800 runs. Mass is increasing in \(a\) and the fence is at least \(m_3\), so \(a \le a_3\) already means the point is inside.

CF5 · The 0.7. Let a ball of radius \(Q\) be grown to \(f \ge 1\) times its mass, reaching radius \(Q + t\). In \(H^2\), \(t \le \ln f\), and \(t \to \ln f\) as \(Q \to \infty\). In \(H^3\), \(t \le \tfrac12 \ln f\), again with equality in the limit. For Tukey’s outside fence on a uniform cloud, \(f = 2\), so the fence lies within \(\ln 2 = 0.6931\ldots\) of Q3 in the plane. For any data, \(f = 1 + 1.5(1 - m_1/m_3) \le 2.5\).

In \(H^2\) set \(h(Q) = \cosh(Q + \ln f) - 1 - f(\cosh Q - 1)\). Expanding the exponentials, \(h(Q) = (f-1)\big[1 - e^{-Q}\tfrac{f+1}{2f}\big]\). This is at least \((f-1)^2/2f \ge 0\), because \(e^{-Q} \le 1\) and \(\tfrac{f+1}{2f} \le 1\). So the ball of radius \(Q + \ln f\) already holds at least \(f\) times the mass, and since mass increases with radius, \(t \le \ln f\). As \(Q \to \infty\), \((\cosh(Q+t) - 1)/(\cosh Q - 1) \to e^t\), so \(t \to \ln f\). In \(H^3\) set \(h(Q) = \sinh(2Q + \ln f) - 2Q - \ln f - f(\sinh 2Q - 2Q)\). Then \(h(0) = \sinh(\ln f) - \ln f \ge 0\) and \(h'(Q) = (f - 1)\big[2 - e^{-2Q}(1 + 1/f)\big] \ge 0\), so \(h \ge 0\) and \(t \le \tfrac12\ln f\). The ratio of volumes tends to \(e^{2t}\). The bound on \(f\) follows from \(m_1 \ge 0\).

CF6 · The flat limit. Let \(x_i = \exp_o(\varepsilon v_i)\) for flat vectors \(v_i\) with no ties among \(|v_i - \mathrm{med}|\). For all small enough \(\varepsilon\), the Claude fence returns the Euclidean volume fence: Tukey’s rule on \(|v_i - \mathrm{med}|^n\) about the Euclidean geometric median.

As \(\varepsilon \to 0\), hyperbolic distances are \(\varepsilon\) times flat distances up to \(O(\varepsilon^3)\). The hyperbolic volume of a ball of radius \(r\) is \(\omega_n r^n (1 + O(r^2))\). The Fréchet median minimises \(\sum d\), which converges after division by \(\varepsilon\) to the flat sum of distances, whose minimiser is the geometric median. Tukey’s rule ignores the positive factor \(\omega_n \varepsilon^n\). Each verdict is a strict inequality between continuous functions of \(\varepsilon\), so away from ties it holds for all small \(\varepsilon\).

CF7 · Distance-Tukey saturates. For the uniform ball of radius \(R\) in \(H^n\), as \(R \to \infty\) the radius quartiles satisfy \(d_p = R + \ln p/(n-1) + o(1)\). The IQR tends to \(\ln 3/(n-1)\), whatever the size. Tukey’s fence on distance sits \((1.5\ln 3 - \ln\tfrac43)/(n-1) \approx 1.36/(n-1)\) past the data’s edge, enclosing about 3.9 times its volume. The Claude fence sits \(\ln 1.5/(n-1)\) past the edge, enclosing 1.5 times.

\(V(r) = C_n e^{(n-1)r}(1 + o(1))\). So \(V(d_p) = pV(R)\) gives \(e^{(n-1)(d_p - R)} \to p\). Substituting \(p = \tfrac14, \tfrac34\) gives the IQR \(\ln 3/(n-1)\), and \(d_3 + 1.5\,\mathrm{IQR} - R = (\ln\tfrac34 + 1.5\ln 3)/(n-1)\). The volume ratio is \(e^{1.5\ln 3 + \ln(3/4)} = 3^{1.5}\cdot\tfrac34 \approx 3.90\). For the Claude fence, CF3 gives \(V(r_F) = 1.5\,V(R)\). A point planted at twice the data’s volume is therefore invisible to distance-Tukey and plain to the Claude fence, which is what CF-B measured.

CF8 · The centre. (a) The empty centre. In a uniform ball of \(H^n\), a point lies within \(r\) of the centre with probability \(V(r)/V(R) \approx e^{-(n-1)(R-r)}\). With \(N\) points, none is expected nearer than about \(R - \ln N/(n-1)\), so the medoid, the best data point, sits off-centre. (b) The Fréchet median is unique unless all points lie on one geodesic. The fixed points of the Weiszfeld map are exactly the critical points of \(\sum_i d(c, x_i)\). (c) Its breakdown point is \(\tfrac12\) (Fletcher, Venkatasubramanian and Joshi, 2009, for manifolds of non-positive curvature).

(a) follows from CF7’s asymptotics for \(V\). (b) A distance function is geodesically convex on a CAT(0) space and strictly convex along any geodesic not through its point, so the sum is strictly convex unless every point lies on one geodesic. At \(c\) the gradient of \(d(c, x)\) is \(-(x + \langle c, x\rangle c)/\sinh d\), a unit tangent vector pointing away from \(x\). So the gradient of the sum vanishes exactly when \(\sum_i x_i/\sinh d_i\) is parallel to \(c\), which is the Weiszfeld fixed-point condition. (c) is cited and checked live: a far cluster of fewer than half the points leaves the centre near the inliers, and one of more than half carries it away. A data point that coincides with the centre gets weight 0 for that step, as in Weiszfeld’s original rule.

CF9 · The four-point bound. Take any four points of \(H^n\). Let \(S_1 \ge S_2 \ge S_3\) be the three sums \(d(x,y) + d(z,w)\), \(d(x,z) + d(y,w)\) and \(d(x,w) + d(y,z)\). Then the defect \((S_1 - S_2)/2\) is at most \(\ln 2\). The bound is sharp: four points at radius \(R\) on the corners of a square reach \(\ln 2\) as \(R \to \infty\).

Sharpness: two points at radius \(R\) with angle \(\theta\) between them have \(\cosh d = \cosh^2 R - \sinh^2 R\cos\theta\), so \(d = 2R + 2\ln\sin(\theta/2) + o(1)\). For a square, adjacent pairs have \(\theta = \pi/2\), giving \(-\ln 2\) each, and diagonals have \(\theta = \pi\), giving 0. The sums are \(4R\), \(4R - 2\ln 2\) and \(4R - 2\ln 2\), so the defect tends to \(\ln 2\). The upper bound is the Gromov-product form of the inequality, \((x|z)_w \ge \min\{(x|y)_w, (y|z)_w\} - \ln 2\). It follows from strong hyperbolicity of CAT(−1) spaces, \(e^{-(x|z)_w} \le e^{-(x|y)_w} + e^{-(y|z)_w}\) (Nica and Špakula, 2016): the right side is at most \(2e^{-\min}\). We cite this and do not reprove it. The live check tests both the inequality and the bound on random quadruples at three scales in \(H^2\) and \(H^3\).

07

The guard

Sometimes data arrive as distances, not points: a graph metric, or an embedding’s output. Then CF9 becomes a test. A quadruple with defect above ln 2 cannot come from hyperbolic space, so one of its six distances is wrong.

The guard checks every quadruple, or a sample of them when there are too many. For each pair of points it counts how many violating quadruples contain that pair. A single corrupted distance \(d_{ij}\) appears in every quadruple that holds both \(i\) and \(j\), so it rises to the top of the list. In the reference trace, two distances among 16 points are increased by 2.5. The guard finds 125 violating quadruples, and its top two suspects are exactly the two corrupted pairs. On the clean distances it finds none: the largest defect is below ln 2, as CF9 requires.

Limits

The guard catches distances that break hyperbolicity. A corrupted distance that happens to stay consistent with some hyperbolic configuration goes unnoticed, and so does a corruption too small to push any quadruple past ln 2. Its float comparisons use a tolerance of 10⁻⁹, and it is not exact.

08

Preregistered results

The claims, cells, seeds and decision rule were written into PREREGISTRATION.md, Part F and committed before any measured run. The predictions came from CF3 and CF7, not from runs. Then 1,800 runs were made: 60 seeds per cell, 200 inliers each, with exact verdicts in H² and certified verdicts in H³.

ClaimStatementPredictionMeasuredStatus
CF-AOn uniform balls the Claude fence flags at most 0.70% of inliers in every cellabout 0%0–0.025% in all six cellssupported
CF-BAt twice the data’s mass, in the four broad cells, the Claude fence catches ≥ 90% and distance-Tukey ≤ 10%about 100% vs 0%97.2–100% vs 0% in all four cellssupported
CF-CA preregistered weakness: on wrapped normals the Claude fence flags more than 2% of inliers and more than distance-Tukey, in every cell5–17% vs about 1%6.1–16.8% vs 0.9–1.2%supported

Every cell, with the shaded rows deciding CF-B. The last column is the mean gap between the outside fence and Q3, in distance. On uniform balls it climbs towards ln 2 = 0.6931 in H² and \(\tfrac12\ln 2\) = 0.3466 in H³, as CF5 says.

CloudSpaceSizeClaude falseDist-Tukey falseClaude caughtDist-Tukey caughtMean gap
uniform ballH²R = 10.0%0.0%––0.3242
uniform ballH²R = 40.0%0.0%––0.6651
uniform ballH²R = 80.0%0.0%––0.6886
uniform ballH³R = 10.017%0.0%––0.2047
uniform ballH³R = 30.025%0.0%––0.3352
uniform ballH³R = 50.025%0.0%––0.3434
planted ×1.75H²R = 10.0%0.0%90.8%1.3%0.3304
planted ×2H²R = 10.0%0.0%100.0%15.2%0.3304
planted ×3H²R = 10.0%0.0%100.0%100.0%0.3304
planted ×1.75H²R = 40.0%0.0%86.8%0.0%0.6648
planted ×2H²R = 40.0%0.0%100.0%0.0%0.6648
planted ×3H²R = 40.0%0.0%100.0%12.5%0.6648
planted ×1.75H²R = 80.0%0.0%86.7%0.0%0.6875
planted ×2H²R = 80.0%0.0%100.0%0.0%0.6875
planted ×3H²R = 80.0%0.0%100.0%6.5%0.6875
planted ×1.75H³R = 10.008%0.0%82.2%0.2%0.2063
planted ×2H³R = 10.008%0.0%99.3%2.0%0.2063
planted ×3H³R = 10.008%0.0%100.0%75.0%0.2063
planted ×1.75H³R = 30.017%0.0%80.3%0.0%0.3346
planted ×2H³R = 30.017%0.0%97.3%0.0%0.3346
planted ×3H³R = 30.017%0.0%100.0%12.8%0.3346
planted ×1.75H³R = 50.017%0.0%80.0%0.0%0.3424
planted ×2H³R = 50.017%0.0%97.2%0.0%0.3424
planted ×3H³R = 50.017%0.0%100.0%8.3%0.3424
wrapped normalH²σ = 0.56.0%1.1%––0.3629
wrapped normalH²σ = 17.9%1.1%––0.6003
wrapped normalH²σ = 211.9%1.1%––0.8051
wrapped normalH³σ = 0.58.9%0.9%––0.2687
wrapped normalH³σ = 112.1%0.9%––0.3928
wrapped normalH³σ = 216.8%0.9%––0.4496

Not decided, but reported. At 1.75 times the mass, just past the fence’s expected 1.58 with 10 planted points among 210, the Claude fence catches 80–91%. At three times, distance-Tukey catches all of them in the flat-ish H² ball (R = 1) but at most 13% in the broad ones. The float track agreed with the exact and certified verdicts in all 1,800 runs, no sign was left unresolved, and 64 bits always sufficed. Raw records: reference/results/fence.jsonl; cell means and decisions: reference/results/fence_decisions.json.

What CF-C means

The Claude fence assumes the shape of a hierarchy, whose mass is spread evenly out to an edge. A Gaussian-like cloud has no edge, and its mass has a long upper tail, so the fence flags 6–17% of such a cloud. On Gaussian-like data, Tukey on distance is the better rule. The preregistration said so before the runs, and they agree.

09

Live checks and parity

CF1–CF9 in this browser
LemmaTrials passedControl rejectedWhat is measured
Running…
 
Parity with the Python reference

The reference’s trace reference/traces/fence.json holds the portable float kernels, the certified mass at 64 and 128 bits with its error bounds, certified signs, exact ties, and eight datasets fenced end to end. It also holds one dataset moved by a rational isometry (CF2) and the guard on clean and corrupted distances. Every float is compared bit for bit; the exact centre and the radii by SHA-256.

Running…
10

The fence page, manual

The fence page builds a cloud, fences it, and shows every intermediate. Changing a control refences at once.

ControlWhat it does
Fence it · New cloudRefence the current settings, or draw a new random seed (keys f/Enter and n).
spaceH² (exact verdicts) or H³ (certified verdicts); keys 2 and 3.
clouduniform ball (mass uniform out to radius R, the shape of a hierarchy); wrapped normal (a Gaussian in the tangent space at the origin, spread σ); norm-5 tree (version 2’s exact tree in H³, every node to the chosen depth).
points · R · σ · depthThe cloud’s size and spread.
planted · at mass × · asOutliers added on top, marked with white rings. A scattered shell puts them in random directions at the given multiple of the data ball’s mass. One far cluster puts them all near one point, the test of the centre’s robustness. Deeper nodes adds tree nodes one to three levels below the cloud.
centreThe Fréchet median, or the medoid as a control: in a broad cloud no data point is central (CF8).
viewdistance: each point drawn at its hyperbolic distance from the centre, so the fence spheres stand apart. Poincaré: the true model about the centre, where broad clouds crowd the rim. Key v. In H³ drag to turn; click a point to read it.
distance-TukeyDraws the baseline’s fence, a dashed cyan sphere.
GuardAdds the chosen amount to k distances of the first 120 points, then runs the four-point guard: the defect histogram against ln 2, the count above it, and the suspected pairs.

The panels: Tukey’s box, twice draws the box and fences in mass (top) and distance (bottom), with the same quartile points in both. The 0.7 shows the fence’s gap beyond Q3 against ln 2/(n − 1). Verdicts compares both fences with the truth. Certificates counts how each verdict was decided. Self-test replays the reference trace on load and runs CF1–CF9 on request. URL parameters: dim, kind, N, R, sigma, L, deep, k, f, mode, seed, centre, view, dt, ck, cd, checks.

1 · Watch the 0.7 appear open

H², uniform ball, no planted points. Move R from 1 to 9 and read the 0.7 panel. The gap starts near 0.3 and climbs to 0.68–0.69, while the mass factor stays near 2. In H³ it climbs to about 0.34.

2 · Twice the mass open

Ten points at twice the data ball’s mass. The Claude fence catches them and flags no inliers, while distance-Tukey’s dashed sphere lies beyond them all. Raise the factor to 4 and distance-Tukey starts to catch some.

3 · The weakness open

A wrapped normal with no planted points. The Claude fence flags around a tenth of the cloud, because a Gaussian’s mass has a long tail. Distance-Tukey flags about 1%. This is CF-C, preregistered as expected.

4 · The empty centre open

H³, R = 4, centred on the medoid. The best data point sits far from the true centre, every mass is distorted, and inliers get flagged. Switch the centre to the Fréchet median and they come back.

5 · Blame a distance open

Two corrupted distances. Every quadruple is checked, dozens beat ln 2, and the corrupted pairs head the suspect list. Set the corruption to 0 and the largest defect drops below ln 2.

6 · A hierarchy open

Version 2’s norm-5 tree to depth 3, plus nodes two levels deeper. They are caught, mostly far out. With one level deeper, some are inside the fence: in this tree, nodes at one depth already differ widely in radius.

11

The fence in 3D

The 3D page runs the same algorithm, the same fence.js, in H³ only, and draws what the flat page can only project: the cloud inside its nested fence shells, seen from the centre.

What changes in H³. The mass is the volume \(\pi(\sinh 2d - 2d)\). Volume grows like \(e^{2d}\), so every hyperbolic length in the lemmas halves. The doubling gap is below \(\tfrac12\ln 2 = 0.3466\) (CF5), the distance IQR of a broad ball saturates at \(\tfrac12\ln 3\), and distance-Tukey sits 0.68 past the edge (CF7). The verdicts are certified rather than exact. A point at or inside the upper quartile is settled by order. Any other point is settled in fixed point at 64 bits or more, and the precision doubles until the sign is certain (CF4). The four-point bound ln 2 is the same in every dimension (CF9).

PartWhat it shows
The cloud inside its fenceEvery point at its distance (or its Poincaré radius) in its direction from the centre. The shells are the Q1 and Q3 spheres, the outside fence (amber) and the far-out fence (rose), with distance-Tukey’s sphere on request. Each shell is translucent, with three great circles. Points behind the centre are dimmer and drawn before the front. Drag to turn, wheel to zoom, click a point to read it; space stops the turning.
cutslab keeps only a thin slice through the centre, which shows the radial layering as a cross-section. half removes every point in front of the centre, so the inside of the shells is visible.
Sky mapEach point’s direction from the centre, in longitude and latitude on the sphere at infinity, with dot size growing with mass. A uniform ball fills the sky evenly, the norm-5 tree shows its octahedral branches, and a far cluster is a single spot.
The mass onionPoints per mass shell, from the centre out, against a uniform ball’s expectation: N/4 in each of the four shells inside the data’s edge, none past the outside fence (CF3). Planted points are outlined in white.
The 0.7 in H³The gap between the outside fence and Q3, against \(\tfrac12\ln(\text{factor})\) and \(\tfrac12\ln 2\).
Verdicts and certificatesHow many points were decided by order and how many in fixed point, with the most bits used. Any unresolved sign would be listed. The float track is compared too.
PolyStep 3D particlesA live dataset: two structures of the PolyStep 3D twin system (B1 in H³, n = 30, three particles per channel) run for the chosen frames with the chosen coupler mode. Every factor point of structure 1 is then fenced: 270 points.

URL parameters: kind, N, R, sigma, L, deep, F, cp, k, f, mode, seed, centre, view, cut, spin, dt, ck, cd, checks.

1 · The half-ln-2 shell open

A broad uniform ball, no planted points. The onion shows four nearly equal shells and empty ones beyond. The 0.7 panel reads a gap close to 0.34, just under \(\tfrac12\ln 2\), with the mass factor near 2.

2 · Twice the mass, in space open

Ten points at twice the data’s volume sit between the outside fence and distance-Tukey’s cyan shell. The Claude fence flags them all, and distance-Tukey flags none (CF-B, in its H³ cells). Cut half to see them against the inner shells.

3 · The tree’s branches open

The norm-5 tree to depth 3, with six nodes two levels deeper. The sky map shows the six octahedral directions and their children. The deeper nodes are caught, and so are some depth-3 nodes, because in this tree nodes of one depth already differ widely in radius.

4 · Fence a search open

The particle cloud of PolyStep 3D after 240 frames, steered. It is small, with Q3 near 0.4, so the fence works in its nearly flat regime and the gap is far below \(\tfrac12\ln 2\). It flags 4–13% of the factor points, the ones that wandered from the bulk. A search cloud is shaped more like a Gaussian than a hierarchy, so this is the over-flagging that CF-C predicts. Read the flags as a description of the cloud, not as faults in the search. Switch the coupler to carry and then off: the two clouds are identical, point for point, as K2 says they must be. Only steer changes the positions.

5 · The guard in space open

Three corrupted distances among the first hundred points. Thirty thousand sampled quadruples are checked against ln 2, and the corrupted pairs lead the suspect list. With no corruption the largest defect stays below ln 2.

12

What is not claimed

  • Not a universal outlier rule. The fence is derived for clouds spread as hierarchies spread. On Gaussian-like clouds it over-flags, measured at 6–17% (CF-C). The name does not make 0.7 a tuning constant: Tukey’s 1.5 and 3 are kept, and ln 2 is what they become (CF3, CF5).
  • Cited, not reproved. The breakdown point ½ of the Fréchet median (CF8c) and strong hyperbolicity of CAT(−1) spaces behind CF9’s upper bound are taken from the literature and checked numerically here.
  • The centre is a float iteration. It is exact in what follows from it: its weights are read as exact rationals, so every verdict is exact or certified for that centre, and the centre is exactly equivariant (CF2). A Weiszfeld iteration stopped at a tolerance of 10⁻¹³ is not the true minimiser to the last bit.
  • Finite samples. CF3 and CF5 are statements about populations. On 200 points, the gap and the mass factor scatter around them; the runs report how much.
  • The guard is a float test with a tolerance, and it sees only corruptions that break hyperbolicity.