Figure 7.
Introduction
An entry of a covariance, coherence or spectral-density matrix does not vary freely. Its admissible values are bounded by the entries around it, and a displacement that looks modest in isolation can make the whole matrix singular. For a labelled connection the natural question is therefore not how far it can move in some ambient geometry, but how far it can move while every other entry is held fixed. That question fixes both a domain and a family of admissible paths, and the answer depends on the constraint as well as on the chosen matrix metric.
A fixed-complement model describes, for example, a controlled sensitivity analysis in which one connection is perturbed while the measured diagonal and all other connections are held at their reference values. The resulting distance quantifies admissible change along that slice. Statistical inference across observations with varying surrounding entries requires an additional model, as discussed in § 7 .
Other analyses of connectivity treat the matrix as a whole. For example, intrinsic local polynomial regression models how a positive-definite matrix varies with covariates [1 ] , while the Janashia–Lagvilava algorithm studied by Wang et al. factors a matrix-valued spectrum [42 ] . Here the surrounding entries instead serve as constraints on a single connection.
Geometry for correlation matrices already offers answers to adjacent questions. David and Gu formulated correlation geometry as a quotient by positive diagonal congruence, including the complex Hermitian case [12 ] . In the real two-variable case this yields a distance proportional to a difference of Fisher transforms [39 ] , and transformed means and dispersions have been used for anomaly analysis [13 ] . Later work established alternative correlation geometries and identified curvature restrictions on the full quotient-affine correlation manifold [38 , 40 ] . A quotient removes scale directions; requiring the diagonal and the complementary block to stay fixed instead restricts admissible motion. These are different constraints and they produce different metrics.
For covariance matrices the Schur-normalized entry is the classical partial correlation of the two variables after linear adjustment for the complement. After a fixed congruence our surface is exactly the unit-diagonal two-by-two Hermitian correlation family equipped with the restriction of the ambient metric. Results for that induced submanifold therefore apply verbatim. The quotient construction instead minimizes over diagonal rescaling velocities. On the real diameter the squared radial coefficients are respectively 2 ( 1 + r 2 ) / ( 1 − r 2 ) 2 2(1+r^2)/(1-r^2)^2 2 ( 1 + r 2 ) / ( 1 − r 2 ) 2 and 2 / ( 1 − r 2 ) 2 2/(1-r^2)^2 2/ ( 1 − r 2 ) 2 . The distinction also matters for comparisons with totally geodesic SPD submanifolds [41 ] , since the ambient geodesic leaves our slice (§ 4 ).
Metric choice remains consequential. Log-Euclidean and log-Cholesky constructions supply different coordinate geometries [2 , 3 , 30 ] , and statistical models on positive-definite manifolds depend on that choice [34 , 35 , 37 ] . Here the metric and path constraint are fixed before the distance is defined.
Fixed-entry positive-definite completion and partial-correlation parameterizations supply the algebraic background [14 , 21 , 25 , 29 ] . We use the Schur complement to isolate one conjugate pair and express its feasible values as a normalized disk. Schur feasibility, correlation geometry, Fréchet averaging [18 , 31 ] and covariance whitening [27 ] are established ingredients; the contribution developed here is the explicit fixed-complement reduction, the elementary geodesic-leg solution with its branch conditions, and the population and computational consequences on that induced surface.
Three features of the induced surface are worth stating at the outset. First, one congruence that is constant across the leaf removes the complement from the tangent energy, so the geometry is two-dimensional in the complex case and one-dimensional in the real case, independently of the ambient matrix dimension. Second, the complex surface is complete and nonpositively curved, with curvature varying between zero and − 1 / 4 -1/4 − 1/4 ; the Cartan–Hadamard theorem then gives unique minimizing geodesics and global logarithms, and the resulting convexity of the squared distance, together with a finite second moment, gives a unique empirical Fréchet centre. Third, the real restriction has an elementary signed arc coordinate, and that sign is exactly what a squared joint score cannot retain.
The geodesic problem on the disk reduces to a single scalar equation. Angular momentum is conserved, so every pair of points determines a Clairaut constant through either a monotone branch or a branch with one radial turning point. We show that both legs of that reduction are exactly integrable, which turns the numerical problem into one bounded root find followed by an evaluation, and we give the branch conditions and the limiting cases at the origin, on a diameter and at equal radii. The elementary legs avoid quadrature at a moving turning point, where a fixed-node rule can misjudge an objective difference. Stable limits and precision refinement are therefore part of the evaluation procedure.
Statistical standardization requires a probability law in addition to a metric. Intrinsic means and covariance-weighted logarithms are established tools [31 ] ; whitening fixes the covariance while retaining rotational freedom in the interpretation of components [27 ] . We apply these tools to a declared fixed-complement law, separate individual dispersion from uncertainty in the estimated centre [10 ] , and give a conservative rank for a score frozen after fitting under exchangeability [28 ] . The limit theory we need is the Fréchet-mean central limit theorem and its refinements on manifolds with and without cut loci [8 , 16 , 17 , 24 , 26 ] , together with convergence rates for empirical barycenters in metric spaces [1 ] ; regression and analysis of variance for manifold-valued responses provide the natural population frameworks [15 , 32 ] . Common normalized geometry neither identifies a common distribution nor supplies conditional calibration for arbitrary pooled complements [5 ] , and the conformal literature makes the finite-sample guarantee and its exchangeability requirement precise [4 , 6 , 7 , 11 , 33 ] .
§ 2 constructs the surface and proves completeness and the curvature bound. § 3 develops the reference population, the covariance standardization and the rank cases. § 4 compares the constrained distance with ambient and quotient constructions, and § 5 records the invariances of the declared slice. § 6 solves the geodesic problem and defines logarithms and parallel transport. § 7 treats estimation, calibration and independent-replicate inference. § 8 reports the numerical validation and § 9 discusses scope and limits. The appendices collect the ambient support reduction, the varying-complement derivative, the stable evaluation of the arc coordinate and the moment and derivative conditions used by the statistical limit theory.
An isolated connection as a Riemannian surface
The declared slice and its reduction
A connection is defined in a declared matrix frame. Throughout, n ≥ 2 n \ge2 n ≥ 2 , A A A is Hermitian positive definite (HPD), and only the entries A 12 = z A_{12}=z A 12 = z and A 21 = z ˉ A_{21}=\bar{z} A 21 = z ˉ are permitted to vary. After a simultaneous row and column permutation that places the selected pair first, write
A ( z ) = ( B ( z ) E E ∗ C ) , B ( z ) = ( a z z ˉ b ) , (1) A(z)=\begin{pmatrix} B(z) & E \\ E^* & C \end{pmatrix}, \qquad B(z)=\begin{pmatrix} a & z \\ \bar{z} & b \end{pmatrix}, \tag*{(1)} A ( z ) = ( B ( z ) E ∗ E C ) , B ( z ) = ( a z ˉ z b ) , (1)
with C ∈ C ( n − 2 ) × ( n − 2 ) C \in\mathbb{C}^{(n-2)\times(n-2)} C ∈ C ( n − 2 ) × ( n − 2 ) and E ∈ C 2 × ( n − 2 ) E \in\mathbb{C}^{2\times(n-2)} E ∈ C 2 × ( n − 2 ) . For n = 2 n=2 n = 2 the blocks E E E and C C C are absent and no empty inverse is used.
The metric convention throughout is the affine-invariant Riemannian metric
g A ( U , V ) = Re tr ( A − 1 U A − 1 V ) , (2) g_A(U,V)=\operatorname{Re}\operatorname{tr}(A^{-1}UA^{-1}V), \tag*{(2)} g A ( U , V ) = Re tr ( A − 1 U A − 1 V ) , (2)
on Hermitian tangent vectors with real scalar coefficients. A complex connection therefore carries two real degrees of freedom. The adjoint ∗ * ∗ denotes conjugate transpose, and the norm associated with Eq. (2) is its positive square root. Entries may carry covariance units; ρ \rho ρ , lengths under Eq. (2) and standardized scores are dimensionless.
Definition 2.1 (Fixed-complement slice). For n > 2 n>2 n > 2 , fix a , b ∈ R a,b\in\mathbb{R} a , b ∈ R , E E E and C ≻ 0 C\succ0 C ≻ 0 , and let H = E C − 1 E ∗ H=EC^{-1}E^* H = E C − 1 E ∗ , with α = a − H 11 \alpha=a-H_{11} α = a − H 11 , β = b − H 22 \beta=b-H_{22} β = b − H 22 and γ = H 12 \gamma=H_{12} γ = H 12 . Assume α , β > 0 \alpha,\beta>0 α , β > 0 . The slice is the set S = { A ( z ) : z ∈ C , A ( z ) ≻ 0 } \mathcal{S}=\{A(z):z\in\mathbb{C}, A(z)\succ0\} S = { A ( z ) : z ∈ C , A ( z ) ≻ 0 } together with the restriction of Eq. (2) . The normalized connection is ρ = ( z − γ ) / α β \rho=(z-\gamma)/\sqrt{\alpha\beta} ρ = ( z − γ ) / α β . For n = 2 n=2 n = 2 , set H = 0 H=0 H = 0 and γ = 0 \gamma=0 γ = 0 , so α = a \alpha=a α = a and β = b \beta=b β = b .
With dense inputs one Cholesky factorization of C C C costs O ( ( n − 2 ) 3 ) O((n-2)^3) O (( n − 2 ) 3 ) operations. Two triangular solves and their Gram matrix give H H H in a further O ( ( n − 2 ) 2 ) O((n-2)^2) O (( n − 2 ) 2 ) operations. The same H H H serves all endpoints on the slice. Each subsequent normalization has constant cost in n n n , while the scalar geodesic solver also depends on endpoint conditioning and precision. Small Schur pivots amplify relative errors in the normalization. Singular positive-semidefinite estimates lie outside this model. Regularization changes the matrix and thereby the connection being scored.
Theorem 2.2 (Feasibility and dimension-independent reduction). Under Definition 2.1 , A ( z ) ≻ 0 A(z)\succ0 A ( z ) ≻ 0 if and only if ∣ ρ ∣ < 1 |\rho|<1 ∣ ρ ∣ < 1 . *Moreover a single congruence, independent of ρ \rho ρ , maps A ( z ) A(z) A ( z ) to diag ( R ( ρ ) , I n − 2 ) \operatorname{diag}(R(\rho),I_{n-2}) diag ( R ( ρ ) , I n − 2 ) with
R ( ρ ) = ( 1 ρ ρ ˉ 1 ) , (3) R(\rho)=\begin{pmatrix}1&\rho\\\bar{\rho}&1\end{pmatrix}, \tag*{(3)} R ( ρ ) = ( 1 ρ ˉ ρ 1 ) , (3) and carries the metric Eq. (2) on the slice to that of R ( ρ ) R(\rho) R ( ρ ) alone. The complementary block contributes exactly zero to the induced metric.*
Proof. For n > 2 n>2 n > 2 multiply on the left by L = ( I 2 − E C − 1 0 I n − 2 ) L=\begin{pmatrix}I_2&-EC^{-1}\\0&I_{n-2}\end{pmatrix} L = ( I 2 0 − E C − 1 I n − 2 ) and on the right by L ∗ L^* L ∗ . Block multiplication gives L A ( z ) L ∗ = diag ( S ( z ) , C ) LA(z)L^*=\operatorname{diag}(S(z),C) L A ( z ) L ∗ = diag ( S ( z ) , C ) with
S ( z ) = B ( z ) − E C − 1 E ∗ = ( α z − γ z ˉ − γ ˉ β ) . (4) S(z)=B(z)-EC^{-1}E^*=\begin{pmatrix}\alpha&z-\gamma\\\bar{z}-\bar{\gamma}&\beta\end{pmatrix}. \tag*{(4)} S ( z ) = B ( z ) − E C − 1 E ∗ = ( α z ˉ − γ ˉ z − γ β ) . (4)
Since L L L is invertible, Sylvester’s law of inertia gives A ( z ) ≻ 0 ⟺ C ≻ 0 A(z)\succ0\Longleftrightarrow C\succ0 A ( z ) ≻ 0 ⟺ C ≻ 0 and S ( z ) ≻ 0 S(z)\succ0 S ( z ) ≻ 0 . For the 2 × 2 2\times2 2 × 2 matrix S ( z ) S(z) S ( z ) , positive definiteness is equivalent to α > 0 \alpha>0 α > 0 and det S = α β − ∣ z − γ ∣ 2 > 0 \det S=\alpha\beta-|z-\gamma|^2>0 det S = α β − ∣ z − γ ∣ 2 > 0 ; together these imply β > 0 \beta>0 β > 0 . Writing z − γ = ρ α β z-\gamma=\rho\sqrt{\alpha\beta} z − γ = ρ α β turns the determinant condition into ∣ ρ ∣ < 1 |\rho|<1 ∣ ρ ∣ < 1 , proving the first claim.
For the second, let D = diag ( α , β ) D=\operatorname{diag}(\alpha,\beta) D = diag ( α , β ) and set T = diag ( D − 1 / 2 , C − 1 / 2 ) L T=\operatorname{diag}(D^{-1/2},C^{-1/2})L T = diag ( D − 1/2 , C − 1/2 ) L . Then T A ( z ) T ∗ = diag ( R ( ρ ) , I n − 2 ) TA(z)T^*=\operatorname{diag}(R(\rho),I_{n-2}) T A ( z ) T ∗ = diag ( R ( ρ ) , I n − 2 ) , because D − 1 / 2 S ( z ) D − 1 / 2 = R ( ρ ) D^{-1/2}S(z)D^{-1/2}=R(\rho) D − 1/2 S ( z ) D − 1/2 = R ( ρ ) and C − 1 / 2 C C − 1 / 2 = I n − 2 C^{-1/2}CC^{-1/2}=I_{n-2} C − 1/2 C C − 1/2 = I n − 2 .
It remains to show that T T T is a tangent isometry for Eq. (2) . For any invertible T T T , ( T A T ∗ ) − 1 = T − ∗ A − 1 T − 1 (TAT^*)^{-1}=T^{-*}A^{-1}T^{-1} ( T A T ∗ ) − 1 = T −∗ A − 1 T − 1 . Substituting U ↦ T U T ∗ U\mapsto TUT^* U ↦ T U T ∗ and V ↦ T V T ∗ V\mapsto TVT^* V ↦ T V T ∗ gives
tr ( T − ∗ A − 1 T − 1 T U T ∗ T − ∗ A − 1 T − 1 T V T ∗ ) = tr ( A − 1 U A − 1 V ) \begin{aligned}
\operatorname{tr}(T^{-*}A^{-1}T^{-1}TUT^*T^{-*}A^{-1}T^{-1}TVT^*) \\
&=\operatorname{tr}(A^{-1}UA^{-1}V)
\end{aligned} tr ( T −∗ A − 1 T − 1 T U T ∗ T −∗ A − 1 T − 1 T V T ∗ ) = tr ( A − 1 U A − 1 V )
by cyclicity of the trace, so Eq. (2) is invariant under congruence. Applying this to T T T and restricting to variations d A dA d A supported on the selected pair leaves exactly Re tr ( R − 1 d R R − 1 d R ) \operatorname{Re}\operatorname{tr}(R^{-1}dR R^{-1}dR) Re tr ( R − 1 d R R − 1 d R ) . For n = 2 n=2 n = 2 the same argument applies with T = D − 1 / 2 T=D^{-1/2} T = D − 1/2 only. □
The set S \mathcal{S} S is a submanifold of the HPD manifold, and the question of which submanifolds of the positive-definite cone are totally geodesic has been studied in its own right [41 ] ; the slice here is not one of them in general, which is exactly why its induced geodesics differ from ambient ones. If no positive complement or no positive α , β \alpha,\beta α , β exists, the slice is empty; a disk written formally over a non-positive block is not the geometry of the original matrices. On a non-empty slice the normalization is invertible, so distinct normalized connections are identifiable once the frame and the complementary entries have been declared. Figure 1 draws the admissible disk schematically, and evaluates the induced stretch and an ambient path that leaves the slice.
Figure 1. An isolated connection has its own path geometry. (A) The admissible disk of normalized connections; the heat contours show intrinsic distance from the origin. Only the selected conjugate pair varies. (B) Radial and tangential stretch agree at the origin and diverge toward the boundary. (C) The ambient geodesic changes the diagonal entries, whereas a slice path keeps them fixed.
The induced metric
Theorem 2.3 (The induced complex metric). Identify ρ = x + i y \rho=x+iy ρ = x + i y with q = ( x , y ) T q=(x,y)^{\mathsf{T}} q = ( x , y ) T in the open unit disk D \mathbb{D} D and put r = ∣ ρ ∣ r=|\rho| r = ∣ ρ ∣ . The metric induced by Eq. (2) is
G ( q ) = 2 1 − r 2 I 2 + 4 ( 1 − r 2 ) 2 q q T . (5) G(q)=\frac{2}{1-r^2}I_2+\frac{4}{(1-r^2)^2}qq^{\mathsf{T}}. \tag*{(5)} G ( q ) = 1 − r 2 2 I 2 + ( 1 − r 2 ) 2 4 q q T . (5)
In polar coordinates ρ = r e i θ \rho=re^{i\theta} ρ = r e i θ , away from the origin,
d s 2 = a ( r ) d r 2 + b ( r ) d θ 2 , (6) ds^2=a(r)\,dr^2+b(r)\,d\theta^2, \tag*{(6)} d s 2 = a ( r ) d r 2 + b ( r ) d θ 2 , (6)
where
a ( r ) = 2 ( 1 + r 2 ) ( 1 − r 2 ) 2 , b ( r ) = 2 r 2 1 − r 2 . (7) a(r)=\frac{2(1+r^2)}{(1-r^2)^2},\qquad b(r)=\frac{2r^2}{1-r^2}. \tag*{(7)} a ( r ) = ( 1 − r 2 ) 2 2 ( 1 + r 2 ) , b ( r ) = 1 − r 2 2 r 2 . (7)
Proof. Write Δ = 1 − ∣ ρ ∣ 2 \Delta=1-|\rho|^2 Δ = 1 − ∣ ρ ∣ 2 and w = d ρ w=d\rho w = d ρ . Inverting R ( ρ ) R(\rho) R ( ρ ) gives
R − 1 = Δ − 1 ( 1 − ρ − ρ ˉ 1 ) R^{-1}=\Delta^{-1}\begin{pmatrix}1&-\rho\\-\bar{\rho}&1\end{pmatrix} R − 1 = Δ − 1 ( 1 − ρ ˉ − ρ 1 )
and
d R = ( 0 w w ˉ 0 ) , dR=\begin{pmatrix}0&w\\\bar{w}&0\end{pmatrix}, d R = ( 0 w ˉ w 0 ) ,
hence
R − 1 d R = Δ − 1 ( − ρ w ˉ w w ˉ − ρ ˉ w ) . R^{-1}dR=\Delta^{-1}\begin{pmatrix}-\rho\bar{w}&w\\\bar{w}&-\bar{\rho}w\end{pmatrix}. R − 1 d R = Δ − 1 ( − ρ w ˉ w ˉ w − ρ ˉ w ) .
Taking the trace of the square,
g ( w , w ) = Δ − 2 ( ρ 2 w ˉ 2 + ρ ˉ 2 w 2 + 2 ∣ w ∣ 2 ) = Δ − 2 ( 2 ∣ w ∣ 2 + 2 Re ( ρ ˉ 2 w 2 ) ) . \begin{aligned}
g(w,w)=\Delta^{-2}\left(\rho^2\bar{w}^2+\bar{\rho}^2w^2+2|w|^2\right) \\
&=\Delta^{-2}\left(2|w|^2+2\operatorname{Re}(\bar{\rho}^2w^2)\right).
\end{aligned} g ( w , w ) = Δ − 2 ( ρ 2 w ˉ 2 + ρ ˉ 2 w 2 + 2∣ w ∣ 2 ) = Δ − 2 ( 2∣ w ∣ 2 + 2 Re ( ρ ˉ 2 w 2 ) ) .
Since ρ ˉ 2 w 2 = ( ρ ˉ w ) 2 \bar{\rho}^2w^2=(\bar{\rho}w)^2 ρ ˉ 2 w 2 = ( ρ ˉ w ) 2 and Re [ ( ρ ˉ w ) 2 ] = 2 [ Re ( ρ ˉ w ) ] 2 − r 2 ∣ w ∣ 2 \operatorname{Re}[(\bar{\rho}w)^2]=2[\operatorname{Re}(\bar{\rho}w)]^2-r^2|w|^2 Re [( ρ ˉ w ) 2 ] = 2 [ Re ( ρ ˉ w ) ] 2 − r 2 ∣ w ∣ 2 ,
g ( w , w ) = 2 ∣ w ∣ 2 Δ + 4 [ Re ( ρ ˉ w ) ] 2 Δ 2 . g(w,w)=\frac{2|w|^2}{\Delta}+\frac{4[\operatorname{Re}(\bar{\rho}w)]^2}{\Delta^2}. g ( w , w ) = Δ 2∣ w ∣ 2 + Δ 2 4 [ Re ( ρ ˉ w ) ] 2 .
Now ∣ w ∣ 2 = d x 2 + d y 2 |w|^2=dx^2+dy^2 ∣ w ∣ 2 = d x 2 + d y 2 and Re ( ρ ˉ w ) = x d x + y d y = q ⋅ d q \operatorname{Re}(\bar{\rho}w)=x\,dx+y\,dy=q\cdot dq Re ( ρ ˉ w ) = x d x + y d y = q ⋅ d q , which gives Eq. (5) . Polarizing recovers the bilinear form, including its mixed terms. In polar coordinates ∣ w ∣ 2 = d r 2 + r 2 d θ 2 |w|^2=dr^2+r^2d\theta^2 ∣ w ∣ 2 = d r 2 + r 2 d θ 2 and Re ( ρ ˉ w ) = r d r \operatorname{Re}(\bar{\rho}w)=r\,dr Re ( ρ ˉ w ) = r d r , so g = ( 2 Δ − 1 + 4 r 2 Δ − 2 ) d r 2 + 2 r 2 Δ − 1 d θ 2 g=(2\Delta^{-1}+4r^2\Delta^{-2})dr^2+2r^2\Delta^{-1}d\theta^2 g = ( 2 Δ − 1 + 4 r 2 Δ − 2 ) d r 2 + 2 r 2 Δ − 1 d θ 2 , and 2 Δ + 4 r 2 = 2 ( 1 + r 2 ) 2\Delta+4r^2=2(1+r^2) 2Δ + 4 r 2 = 2 ( 1 + r 2 ) yields Eq. (6) . □
At the origin G ( 0 ) = 2 I 2 G(0)=2I_2 G ( 0 ) = 2 I 2 and the Cartesian coefficients are analytic. The same matrix is the ordinary Cartesian Hessian of the logarithmic barrier − log ( 1 − ∣ q ∣ 2 ) -\log(1-|q|^2) − log ( 1 − ∣ q ∣ 2 ) , as direct differentiation verifies. This supplies a second algebraic characterization of the induced metric. At the origin the vanishing polar angular coefficient b ( r ) → 0 b(r)\to0 b ( r ) → 0 at r = 0 r=0 r = 0 is a coordinate artefact, not a degeneracy of the metric. In particular, extending a signed arc coordinate independently to the two Cartesian directions would produce a different metric; the correct extension is the one constructed next.
The real arc coordinate
On the real slice ρ ∈ ( − 1 , 1 ) \rho\in(-1,1) ρ ∈ ( − 1 , 1 ) define
ψ ( ρ ) = 2 ∫ 0 ρ 1 + t 2 1 − t 2 d t = 2 atanh 2 ρ 1 + ρ 2 − 2 asinh ρ . (8) \begin{aligned}
\psi(\rho)=\sqrt{2}\int_0^\rho\frac{\sqrt{1+t^2}}{1-t^2}\,dt \\
&=2\operatorname{atanh}\frac{\sqrt{2}\rho}{\sqrt{1+\rho^2}}-\sqrt{2}\operatorname{asinh}\rho. \tag*{(8)}
\end{aligned} ψ ( ρ ) = 2 ∫ 0 ρ 1 − t 2 1 + t 2 d t = 2 atanh 1 + ρ 2 2 ρ − 2 asinh ρ . (8)
Proposition 2.4 (Properties of the arc coordinate). The function ψ \psi ψ is real analytic and odd on ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) , satisfies
ψ ′ ( r ) = 2 1 + r 2 1 − r 2 , \psi'(r)=\frac{\sqrt{2}\sqrt{1+r^2}}{1-r^2}, ψ ′ ( r ) = 1 − r 2 2 1 + r 2 , is strictly increasing, and maps ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) onto R \mathbb{R} R . Its Taylor expansion at the origin is
ψ ( ρ ) = 2 ( ρ + 1 2 ρ 3 + 11 40 ρ 5 + O ( ρ 7 ) ) . (9) \psi(\rho)=\sqrt{2}\left(\rho+\frac{1}{2}\rho^3+\frac{11}{40}\rho^5+O(\rho^7)\right). \tag*{(9)} ψ ( ρ ) = 2 ( ρ + 2 1 ρ 3 + 40 11 ρ 5 + O ( ρ 7 ) ) . (9) The signed arc coordinate is ψ ( ρ ) \psi(\rho) ψ ( ρ ) , and the length from 0 0 0 to ρ \rho ρ is ∣ ψ ( ρ ) ∣ |\psi(\rho)| ∣ ψ ( ρ ) ∣ . Thus ψ \psi ψ is an isometry onto R \mathbb{R} R with the Euclidean metric.
Proof. Differentiating t = sinh u t=\sinh u t = sinh u gives d t = cosh u d u dt=\cosh u\,du d t = cosh u d u and 1 + t 2 = 1 + sinh 2 u = cosh u \sqrt{1+t^2}=\sqrt{1+\sinh^2u}=\cosh u 1 + t 2 = 1 + sinh 2 u = cosh u , so 1 + t 2 d t = cosh 2 u d u = ( 1 + sinh 2 u ) d u \sqrt{1+t^2}\,dt=\cosh^2u\,du=(1+\sinh^2u)\,du 1 + t 2 d t = cosh 2 u d u = ( 1 + sinh 2 u ) d u . With 1 − t 2 = 1 − sinh 2 u 1-t^2=1-\sinh^2u 1 − t 2 = 1 − sinh 2 u the integrand becomes ( 1 + sinh u ) / ( 1 − sinh 2 u ) (1+\sinh u)/(1-\sinh^2u) ( 1 + sinh u ) / ( 1 − sinh 2 u ) . Setting v = tanh u v=\tanh u v = tanh u in the remaining ∫ d u / ( 1 − sinh 2 u ) \int du/(1-\sinh^2u) ∫ d u / ( 1 − sinh 2 u ) integral gives ∫ d u / ( 1 − sinh 2 u ) = atanh ( 2 v ) / 2 \int du/(1-\sinh^2u)=\operatorname{atanh}(\sqrt{2}v)/\sqrt{2} ∫ d u / ( 1 − sinh 2 u ) = atanh ( 2 v ) / 2 . Evaluating both antiderivatives and using 2 v = 2 ρ / 1 + ρ 2 \sqrt{2}v=\sqrt{2}\rho/\sqrt{1+\rho^2} 2 v = 2 ρ / 1 + ρ 2 produces Eq. (8) ; both antiderivatives vanish at ρ = 0 \rho=0 ρ = 0 , and the argument of atanh \operatorname{atanh} atanh lies strictly between − 1 -1 − 1 and 1 1 1 because 2 ρ 2 < 1 + ρ 2 2\rho^2<1+\rho^2 2 ρ 2 < 1 + ρ 2 for ∣ ρ ∣ < 1 |\rho|<1 ∣ ρ ∣ < 1 . The derivative follows by differentiating the integrand, and it is positive on ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) , so ψ \psi ψ is strictly increasing. As ρ ↑ 1 \rho\uparrow1 ρ ↑ 1 the second term stays bounded while atanh → ∞ \operatorname{atanh}\to\infty atanh → ∞ , so ψ → + ∞ \psi\to+\infty ψ → + ∞ ; oddness gives the other end. Continuity and strict monotonicity then give surjectivity onto R \mathbb{R} R . Expanding each term about ρ = 0 \rho=0 ρ = 0 yields Eq. (9) . □
For the complex disk define s = ψ ( r ) > 0 s=\psi(r)>0 s = ψ ( r ) > 0 and the warped-product profile f ( s ) = 2 r / 1 − r 2 f(s)=\sqrt{2}r/\sqrt{1-r^2} f ( s ) = 2 r / 1 − r 2 with r = ψ − 1 ( s ) r=\psi^{-1}(s) r = ψ − 1 ( s ) . Then Eq. (6) reads d s 2 + f ( s ) 2 d θ 2 ds^2+f(s)^2d\theta^2 d s 2 + f ( s ) 2 d θ 2 .
Theorem 2.5 (Curvature, completeness and unique logarithms). The induced surface ( D , g ) (\mathbb{D},g) ( D , g ) is complete, simply connected and has nonpositive Gaussian curvature
K ( s ) = − f ′ ′ ( s ) f ( s ) = − r 2 ( 1 + r 2 ) 2 ∈ ( − 1 4 , 0 ] , (10) K(s)=-\frac{f''(s)}{f(s)}=-\frac{r^2}{(1+r^2)^2}\in\left(-\frac{1}{4},0\right], \tag*{(10)} K ( s ) = − f ( s ) f ′′ ( s ) = − ( 1 + r 2 ) 2 r 2 ∈ ( − 4 1 , 0 ] , (10) with K ( 0 ) = 0 K(0)=0 K ( 0 ) = 0 attained only at the origin and K → − 1 / 4 K\to-1/4 K → − 1/4 as r → 1 r\to1 r → 1 . Consequently, for every p ∈ D p\in\mathbb{D} p ∈ D the exponential map Exp p : T p D → D \operatorname{Exp}_p:T_p\mathbb{D}\to\mathbb{D} Exp p : T p D → D is a smooth global diffeomorphism and Log p \operatorname{Log}_p Log p is its inverse. For points on one radial ray the distance is ∣ ψ ( r 2 ) − ψ ( r 1 ) ∣ |\psi(r_2)-\psi(r_1)| ∣ ψ ( r 2 ) − ψ ( r 1 ) ∣ ; for opposite rays it is ψ ( r 1 ) + ψ ( r 2 ) \psi(r_1)+\psi(r_2) ψ ( r 1 ) + ψ ( r 2 ) ; for real points it is ∣ ψ ( ρ 2 ) − ψ ( ρ 1 ) ∣ |\psi(\rho_2)-\psi(\rho_1)| ∣ ψ ( ρ 2 ) − ψ ( ρ 1 ) ∣ .
Proof. Curvature. With f ( r ) = 2 r ( 1 − r 2 ) − 1 / 2 f(r)=\sqrt{2}r(1-r^2)^{-1/2} f ( r ) = 2 r ( 1 − r 2 ) − 1/2 and ψ ′ ( r ) = 2 1 + r 2 / ( 1 − r 2 ) \psi'(r)=\sqrt{2}\sqrt{1+r^2}/(1-r^2) ψ ′ ( r ) = 2 1 + r 2 / ( 1 − r 2 ) ,
f ′ ( s ) = f ′ ( r ) ψ ′ ( r ) = 1 1 − r 4 , f'(s)=\frac{f'(r)}{\psi'(r)}=\frac{1}{\sqrt{1-r^4}}, f ′ ( s ) = ψ ′ ( r ) f ′ ( r ) = 1 − r 4 1 , f ′ ′ ( s ) = 1 ψ ′ ( r ) d d r 1 1 − r 4 = 2 r 3 1 − r 2 ( 1 + r 2 ) 2 . f''(s)=\frac{1}{\psi'(r)}\frac{d}{dr}\frac{1}{\sqrt{1-r^4}}=\frac{\sqrt{2}r^3}{\sqrt{1-r^2}(1+r^2)^2}. f ′′ ( s ) = ψ ′ ( r ) 1 d r d 1 − r 4 1 = 1 − r 2 ( 1 + r 2 ) 2 2 r 3 .
For a warped product d s 2 + f ( s ) 2 d θ 2 \mathrm{d}s^2+f(s)^2\mathrm{d}\theta^2 d s 2 + f ( s ) 2 d θ 2 the only non-vanishing Christoffel symbols are Γ θ θ θ = − f f ′ \Gamma^\theta_{\theta\theta}=-ff' Γ θ θ θ = − f f ′ and Γ s θ s = f ′ / f \Gamma^s_{s\theta}=f'/f Γ s θ s = f ′ / f , and R θ s θ s = − f f ′ ′ R^s_{\theta s\theta}=-ff'' R θ s θ s = − f f ′′ , so K = − f ′ ′ / f = − r 2 / ( 1 + r 2 ) 2 K=-f''/f=-r^2/(1+r^2)^2 K = − f ′′ / f = − r 2 / ( 1 + r 2 ) 2 . The Cartesian coefficients of G G G are analytic at the origin, so this expression extends continuously with K ( 0 ) = 0 K(0)=0 K ( 0 ) = 0 . Its magnitude increases strictly on [ 0 , 1 ) [0,1) [ 0 , 1 ) and tends to 1 / 4 1/4 1/4 .
Completeness. The radial length inequality gives d ( o , x ) = ψ ( ∣ x ∣ ) d(o,x)=\psi(|x|) d ( o , x ) = ψ ( ∣ x ∣ ) : every curve has length at least the radial-coordinate change, and the radial segment attains it. Let γ : [ 0 , L ) → D \gamma:[0,L)\to\mathbb{D} γ : [ 0 , L ) → D be a curve of finite length L L L . Since g ( γ ˙ , γ ˙ ) ≥ s ˙ 2 g(\dot{\gamma},\dot{\gamma})\ge\dot{s}^2 g ( γ ˙ , γ ˙ ) ≥ s ˙ 2 for the radius coordinate s ( t ) = ψ ( ∣ γ ( t ) ∣ ) s(t)=\psi(|\gamma(t)|) s ( t ) = ψ ( ∣ γ ( t ) ∣ ) ,
L ≥ ∫ 0 L ∣ s ˙ ∣ d t ≥ ∣ lim t → L s ( t ) − s ( 0 ) ∣ . L\ge\int_0^L|\dot{s}|\,\mathrm{d}t\ge\left|\lim_{t\to L}s(t)-s(0)\right|. L ≥ ∫ 0 L ∣ s ˙ ∣ d t ≥ t → L lim s ( t ) − s ( 0 ) .
so a curve of finite length has bounded s s s . Because ψ ( r ) → + ∞ \psi(r)\to+\infty ψ ( r ) → + ∞ as r ↑ 1 r\uparrow1 r ↑ 1 , bounded s s s confines the curve to a closed Euclidean disk of radius R < 1 R<1 R < 1 . For completeness, fix a metric-Cauchy sequence. The triangle inequality gives d ( o , x n ) ≤ d ( o , x 1 ) + d ( x 1 , x n ) d(o,x_n)\le d(o,x_1)+d(x_1,x_n) d ( o , x n ) ≤ d ( o , x 1 ) + d ( x 1 , x n ) , which is bounded, and s n = ψ ( ∣ x n ∣ ) = d ( o , x n ) s_n=\psi(|x_n|)=d(o,x_n) s n = ψ ( ∣ x n ∣ ) = d ( o , x n ) , so the sequence lies in a Euclidean disk of radius R < 1 R<1 R < 1 . On that disk G ⪯ M I G\preceq MI G ⪯ M I for some finite M M M . The global bound G ⪰ 2 I G\succeq2I G ⪰ 2 I gives d ( x , y ) ≥ 2 ∣ x − y ∣ d(x,y)\ge\sqrt{2}|x-y| d ( x , y ) ≥ 2 ∣ x − y ∣ , so the sequence is Euclidean-Cauchy and converges to x x x with ∣ x ∣ ≤ R < 1 |x|\le R<1 ∣ x ∣ ≤ R < 1 . The straight segment stays in that disk; its length is at most M ∣ x n − x ∣ \sqrt{M}|x_n-x| M ∣ x n − x ∣ , proving metric convergence. Hence ( D , g ) (D,g) ( D , g ) is complete. The disk is simply connected, and K ≤ 0 K\le0 K ≤ 0 was proved above; the Cartan–Hadamard theorem then gives that Exp p \operatorname{Exp}_p Exp p is a global diffeomorphism for every p p p , and Log p \operatorname{Log}_p Log p its smooth inverse. Uniqueness of minimizing geodesics follows as well.
Distances. A radial curve has constant θ \theta θ and satisfies the geodesic equations in ( s , θ ) (s,\theta) ( s , θ ) ; its length is the variation of s s s , and no shorter curve has the same endpoints because any curve has length at least the absolute variation of s s s . A diameter extends smoothly through the origin in Cartesian coordinates, so its two halves form a single geodesic; uniqueness makes it minimizing, and its length is ψ ( r 1 ) + ψ ( r 2 ) \psi(r_1)+\psi(r_2) ψ ( r 1 ) + ψ ( r 2 ) . The real case is the restriction θ ∈ { 0 , π } \theta\in\{0,\pi\} θ ∈ { 0 , π } . □
Theorem 2.5 derives global existence and uniqueness from completeness and curvature. General geodesics carry both magnitude and phase; their computational characterization is developed in § 6 . Figure 2 displays the arc coordinate, the curvature, the numerically integrated geodesic circles and the distance ratios computed from the scalar formulas.
Figure 2. Arc coordinates and variable curvature distinguish the connection disk. (A) The signed intrinsic arc coordinate, its quotient counterpart and its linearization. (B) Gaussian curvature of the disk, compared with the constant curvature of the quotient and the disk boundary limit. (C) Geodesic circles about an off-centre point, colored by intrinsic radius. (D) Ambient and quotient distances relative to intrinsic distance for a fixed start and varying endpoint.
Reference populations and intrinsic standardization
A controlled perturbation traces a deterministic path with every complementary entry held fixed. A random law on that slice additionally specifies how the selected pair varies. The statistical results below concern this second setting. Sample covariance and cross-spectral matrices generally fluctuate in all their blocks. For those estimators Appendix § 11 propagates a specified joint perturbation through the normalization. Applying a reference law or a calibration rank across subjects would additionally require a sampling unit, a model for the complement and an exchangeable null design.
A reference law P P P is a probability measure on one declared fixed slice, transported to D \mathbb{D} D by ρ \rho ρ . The same metric on normalized slices does not imply the same probability law, and no probability statement below is derived from the geometry alone. Throughout this section o o o denotes a fixed base point and d d d the intrinsic distance of Theorem 2.5. We assume
E P d ( o , X ) 2 < ∞ , (11) \mathbb{E}_P d(o,X)^2<\infty, \tag*{(11)} E P d ( o , X ) 2 < ∞ , (11)
which holds for one base point if and only if it holds for every other, by the triangle inequality. Define F ( p ) = 1 2 E P d ( p , X ) 2 F(p)=\frac{1}{2}\mathbb{E}_P d(p,X)^2 F ( p ) = 2 1 E P d ( p , X ) 2 .
Theorem 3.1 (Centre, dispersion and mean logarithm). Under Eq. (11) the function F F F attains its minimum at exactly one point μ ∈ D \mu\in\mathbb{D} μ ∈ D . That point satisfies the first-order condition
E P Log μ X = 0 , (12) \mathbb{E}_P\operatorname{Log}_\mu X=0, \tag*{(12)} E P Log μ X = 0 , (12) its dispersion is D = E P d ( μ , X ) 2 D=\mathbb{E}_P d(\mu,X)^2 D = E P d ( μ , X ) 2 , and in any orthonormal tangent frame with coordinates v = Log μ X \mathbf{v}=\operatorname{Log}_\mu X v = Log μ X and Σ = E P v v T \Sigma=\mathbb{E}_P\mathbf{v}\mathbf{v}^{\mathsf{T}} Σ = E P v v T one has D = tr Σ D=\operatorname{tr}\Sigma D = tr Σ . The same statements hold on the real slice with its one-dimensional metric.
Proof. Finiteness and continuity. For p p p in a compact set K K K and every x x x , d ( p , x ) 2 ≤ 2 d ( o , x ) 2 + 2 sup p ∈ K d ( o , p ) 2 d(p,x)^2\le2d(o,x)^2+2\sup_{p\in K}d(o,p)^2 d ( p , x ) 2 ≤ 2 d ( o , x ) 2 + 2 sup p ∈ K d ( o , p ) 2 by the triangle inequality, and the right side is P P P -integrable by Eq. (11) . Dominated convergence and continuity of d d d give continuity of F F F on K K K .
Coercivity. The L 2 L^2 L 2 triangle inequality gives ( E d ( p , X ) 2 ) 1 / 2 ≥ d ( p , o ) − ( E d ( o , X ) 2 ) 1 / 2 (\mathbb{E}\,d(p,X)^2)^{1/2}\ge d(p,o)-(\mathbb{E}\,d(o,X)^2)^{1/2} ( E d ( p , X ) 2 ) 1/2 ≥ d ( p , o ) − ( E d ( o , X ) 2 ) 1/2 , hence F ( p ) → ∞ F(p)\to\infty F ( p ) → ∞ as d ( p , o ) → ∞ d(p,o)\to\infty d ( p , o ) → ∞ . A complete finite-dimensional Riemannian manifold has compact closed bounded balls (Hopf–Rinow, available here because Theorem 2.5 proves completeness), so a minimum exists.
Strict convexity. Let p t p_t p t be the geodesic from p 0 p_0 p 0 to p 1 p_1 p 1 . The squared-distance inequality on a manifold with K ≤ 0 K\le0 K ≤ 0 states
F ( p t ) ≤ ( 1 − t ) F ( p 0 ) + t F ( p 1 ) − 1 2 t ( 1 − t ) d ( p 0 , p 1 ) 2 . (13) F(p_t)\le(1-t)F(p_0)+tF(p_1)-\frac{1}{2}t(1-t)d(p_0,p_1)^2. \tag*{(13)} F ( p t ) ≤ ( 1 − t ) F ( p 0 ) + tF ( p 1 ) − 2 1 t ( 1 − t ) d ( p 0 , p 1 ) 2 . (13)
It follows from the second variation of 1 2 d ( ⋅ , x ) 2 \frac{1}{2}d(\cdot,x)^2 2 1 d ( ⋅ , x ) 2 : the Hessian is at least the metric when K ≤ 0 K\le0 K ≤ 0 , including at x = p x=p x = p by smoothness of the squared distance there; integrating twice along p t p_t p t and taking expectation gives Eq. (13) . The negative quadratic term forces strict convexity, hence a unique minimizer.
First variation. For p ≠ x p\ne x p = x the gradient of 1 2 d ( p , x ) 2 \frac{1}{2}d(p,x)^2 2 1 d ( p , x ) 2 is − Log p x -\operatorname{Log}_p x − Log p x , whose norm equals d ( p , x ) ≤ d ( p , o ) + d ( o , x ) d(p,x)\le d(p,o)+d(o,x) d ( p , x ) ≤ d ( p , o ) + d ( o , x ) and is locally integrable by Eq. (11) . Dominated differentiation therefore gives ∇ F ( p ) = − E P Log p X \nabla F(p)=-\mathbb{E}_P\operatorname{Log}_p X ∇ F ( p ) = − E P Log p X , and stationarity at μ \mu μ is Eq. (12) .
Dispersion. Orthonormal coordinates preserve the norm, so D = E ∥ v ∥ 2 = E tr ( v v T ) = tr Σ D = \mathbb{E}\lVert v\rVert^{2} = \mathbb{E}\operatorname{tr}(vv^{\mathsf{T}}) = \operatorname{tr}\Sigma D = E ∥ v ∥ 2 = E tr ( v v T ) = tr Σ . The real slice is the totally geodesic submanifold θ ∈ { 0 , π } \theta\in\{0,\pi\} θ ∈ { 0 , π } of Theorem 2.5 , and the same argument applies verbatim.
A diameter is a closed totally geodesic convex subset: its complete geodesic is minimizing between every two of its points. A law supported on the real axis is fixed by conjugation of the disk, so uniqueness in Theorem 3.1 forces its centre to be fixed by that reflection and hence to lie on the diameter. For such a law the theorem reduces to
ψ ( μ ) = E P ψ ( X ) , D = Var P ψ ( X ) . (14) \psi(\mu) = \mathbb{E}_{P}\psi(X), \qquad D = \operatorname{Var}_{P}\psi(X). \tag*{(14)} ψ ( μ ) = E P ψ ( X ) , D = Var P ψ ( X ) . (14)
Definition 3.2 (Signed isolated-deviation score). For a reference law supported on the declared real diameter and an observation x x x on that diameter, assume 0 < D < ∞ 0 < D < \infty 0 < D < ∞ . The signed score is
Z S I D ( x ) = ψ ( ρ ( x ) ) − E P ψ ( X ) D , (15) Z_{\mathrm{SID}}(x) = \frac{\psi(\rho(x)) - \mathbb{E}_{P}\psi(X)}{\sqrt{D}}, \tag*{(15)} Z SID ( x ) = D ψ ( ρ ( x )) − E P ψ ( X ) , (15)
where the orientation is the declared increasing direction of the real connection.
Changing the declared orientation replaces Z Z Z by − Z -Z − Z and leaves Z 2 Z^{2} Z 2 unchanged. The sign is information: it distinguishes a connection displaced toward larger real correlation from one displaced toward smaller real correlation, and no function of Z 2 Z^{2} Z 2 can recover it.
Both ψ ( r ) \psi(r) ψ ( r ) and Fisher’s atanh ( r ) \operatorname{atanh}(r) atanh ( r ) increase strictly on the real diameter. With a fixed orientation and positive reference standard deviations, either standardized coordinate therefore gives identical one-sided ranks, including ties. Their numerical displacements differ. For two-sided deviations the ordering also depends on the fitted centre and the relative stretching on either side. A complex connection has two tangent coordinates, so this one-dimensional ordering argument does not apply to its joint score.
Proposition 3.3 (Joint score, whitening and rank). Let v = Log μ X \mathbf{v} = \operatorname{Log}_{\mu}X v = Log μ X and Σ = E v v T \Sigma= \mathbb{E}\mathbf{v}\mathbf{v}^{\mathsf{T}} Σ = E v v T . If Σ ≻ 0 \Sigma\succ0 Σ ≻ 0 and W = Σ − 1 / 2 v \mathbf{W} = \Sigma^{-1/2}\mathbf{v} W = Σ − 1/2 v , then E W = 0 \mathbb{E}\mathbf{W} = 0 E W = 0 , Cov ( W ) = I 2 \operatorname{Cov}(\mathbf{W}) = I_{2} Cov ( W ) = I 2 and, with Q = ∥ W ∥ 2 Q = \lVert\mathbf{W}\rVert^{2} Q = ∥ W ∥ 2 , E Q = 2 \mathbb{E}Q = 2 E Q = 2 . Under an orthogonal change of tangent frame
v ↦ O v v \mapsto Ov v ↦ O v and Σ ↦ O Σ O ⊤ \Sigma\mapsto O\Sigma O^\top Σ ↦ O Σ O ⊤ , so Q Q Q is unchanged although the components of v v v rotate. If Σ \Sigma Σ has rank k ∈ { 0 , 1 , 2 } k \in\{0,1,2\} k ∈ { 0 , 1 , 2 } , then v ∈ range ( Σ ) v \in\operatorname{range}(\Sigma) v ∈ range ( Σ ) almost surely and
E [ v ⊤ Σ + v ] = k . (16) \mathbb{E}\left[v^\top\Sigma^{+}v\right] = k. \tag*{(16)} E [ v ⊤ Σ + v ] = k . (16)
Equation (16) is a support score, not full-dimensional whitening.
Proof. Centring gives E W = Σ − 1 / 2 E v = 0 \mathbb{E}W=\Sigma^{-1/2}\mathbb{E}v=0 E W = Σ − 1/2 E v = 0 by Eq. (12) , and Cov ( W ) = Σ − 1 / 2 Σ Σ − 1 / 2 = I 2 \operatorname{Cov}(W)=\Sigma^{-1/2}\Sigma\Sigma^{-1/2}=I_2 Cov ( W ) = Σ − 1/2 Σ Σ − 1/2 = I 2 ; taking traces gives E Q = 2 \mathbb{E}Q=2 E Q = 2 . For the frame change, Σ − 1 / 2 \Sigma^{-1/2} Σ − 1/2 transforms to O Σ − 1 / 2 O ⊤ O\Sigma^{-1/2}O^\top O Σ − 1/2 O ⊤ , so ∥ W ∥ 2 \lVert W\rVert^2 ∥ W ∥ 2 is preserved. If Σ a = 0 \Sigma a=0 Σ a = 0 , then E ( a ⊤ v ) 2 = a ⊤ Σ a = 0 \mathbb{E}(a^\top v)^2=a^\top\Sigma a=0 E ( a ⊤ v ) 2 = a ⊤ Σ a = 0 , so a ⊤ v = 0 a^\top v=0 a ⊤ v = 0 almost surely; a finite basis of the nullspace gives the support statement. Spectral decomposition of Σ + \Sigma^+ Σ + yields E [ v ⊤ Σ + v ] = tr ( Σ + Σ ) = k . □ \mathbb{E}[v^\top\Sigma^+v]=\operatorname{tr}(\Sigma^+\Sigma)=k. \square E [ v ⊤ Σ + v ] = tr ( Σ + Σ ) = k . □
A law supported on a complete geodesic through μ \mu μ with positive dispersion has rank one; a point mass has rank zero. Conversely, rank-one covariance implies that the logarithms lie on a line, so the support lies on that geodesic. Rank zero is a point mass and admits no standardized score; infinite second moments invalidate both the centre-dispersion construction and the covariance claim. For example, place X = Exp ( T e ) X=\operatorname{Exp}(Te) X = Exp ( T e ) with a real Cauchy variable T T T ; then d ( o , X ) = ∣ T ∣ d(o,X)=\lvert T\rvert d ( o , X ) = ∣ T ∣ and Eq. (11) fails.
Remark 3.4 (Whitening is a moment identity, not a law). Let v v v be uniform on the four vectors ± 2 e 1 , ± 2 e 2 \pm\sqrt{2}e_1,\pm\sqrt{2}e_2 ± 2 e 1 , ± 2 e 2 at the origin and put X = Exp o v X=\operatorname{Exp}_o v X = Exp o v . Then the mean logarithm is zero and the covariance is I 2 I_2 I 2 , so Q = 2 Q=2 Q = 2 almost surely, yet the law of Q Q Q is a point mass, the coordinates are dependent, and Q Q Q is not chi-square. A χ 2 2 \chi^2_2 χ 2 2 law follows only from the additional hypothesis v ∼ N ( 0 , Σ ) v\sim N(0,\Sigma) v ∼ N ( 0 , Σ ) ; it is not supplied by the geometry, by finite moments, or by a central limit theorem for estimated centres.
Empirical eigensupport is not proof of population support. A rank decision made from estimated eigenvalues is a decision about the estimated covariance, and the calibration theory of § 7 is stated for exactly that frozen object. Figure 3 shows a declared reference law on the disk with its intrinsic centre, the same observations in the tangent space at that centre after population whitening, two laws that share an intrinsic covariance spectrum and differ by a rotation, and the resulting rank calibration.
Figure 3. From a reference population to calibrated intrinsic scores. (A) A Gaussian tangent law mapped to the disk, with geodesic-distance contours about its centre. (B) The same observations after population whitening; the dashed unit circle illustrates identity covariance. In A and B, color tracks each observation’s intrinsic displacement. (C) Rotated covariance ellipses with the same intrinsic spectrum. (D) Split calibration of a frozen fitted score with ties. Error bars are pointwise 95% binomial Monte Carlo intervals; the training fit is held fixed. Computational details are in the Supplementary Methods.
Ambient and quotient comparisons
Three constructions are routinely applied to matrix-valued observations, and they answer different questions. The ambient affine-invariant geometry lets every entry vary. A diagonal-scale quotient removes the scale directions. The present construction holds the complement fixed. This section computes all three on the same normalized endpoints, so that the comparison is exact rather than qualitative. Figure 4 compares them as the relative phase varies and separates the signed one-dimensional score from its rotationally invariant joint counterpart.
Figure 4. The three distances retain different information. (A) Distances between identical endpoints as their relative phase varies. (B) Ambient and quotient distances relative to intrinsic distance. (C) Ambient versus intrinsic distance for sampled pairs; color measures the relative distance deficit. (D) The joint score is even in the signed radial coordinate and therefore loses displacement direction. The distance inequalities follow from the geometric results in the text.
The ambient affine-invariant distance and logarithm are
d A ( P , Q ) = ∥ log ( P − 1 / 2 Q P − 1 / 2 ) ∥ F . (17) d_A(P,Q)=\left\lVert\log\left(P^{-1/2}QP^{-1/2}\right)\right\rVert_F. \tag*{(17)} d A ( P , Q ) = log ( P − 1/2 Q P − 1/2 ) F . (17)
Log P Q = P 1 / 2 log ( P − 1 / 2 Q P − 1 / 2 ) P 1 / 2 , (18) \operatorname{Log}_P Q=P^{1/2}\log\left(P^{-1/2}QP^{-1/2}\right)P^{1/2}, \tag*{(18)} Log P Q = P 1/2 log ( P − 1/2 Q P − 1/2 ) P 1/2 , (18)
with the real Hermitian Frobenius norm ∥ H ∥ F 2 = Re tr ( H 2 ) \lVert H\rVert_F^2=\operatorname{Re}\operatorname{tr}(H^2) ∥ H ∥ F 2 = Re tr ( H 2 ) .
Theorem 4.1 (Exact ambient comparison on one slice). For p , q ∈ D p,q\in\mathbb{D} p , q ∈ D write Δ p = 1 − ∣ p ∣ 2 \Delta_p=1-\lvert p\rvert^2 Δ p = 1 − ∣ p ∣ 2 , Δ q = 1 − ∣ q ∣ 2 \Delta_q=1-\lvert q\rvert^2 Δ q = 1 − ∣ q ∣ 2 and
ℓ = 1 2 log Δ q Δ p , h = 1 − Re ( p ‾ q ) Δ p Δ q , \ell=\frac{1}{2}\log\frac{\Delta_q}{\Delta_p},\qquad h=\frac{1-\operatorname{Re}(\overline{p}q)}{\sqrt{\Delta_p\Delta_q}}, ℓ = 2 1 log Δ p Δ q , h = Δ p Δ q 1 − Re ( p q ) , k = acosh h . (19) k=\operatorname{acosh}h. \tag*{(19)} k = acosh h . (19) Then h ≥ 1 h\ge1 h ≥ 1 , and
d A ( A ( p ) , A ( q ) ) 2 = 2 ℓ 2 + 2 k 2 . (20) d_A(A(p),A(q))^2=2\ell^2+2k^2. \tag*{(20)} d A ( A ( p ) , A ( q ) ) 2 = 2 ℓ 2 + 2 k 2 . (20) For distinct endpoints the intrinsic distance of Theorem 2.5 is strictly larger than this ambient distance.
Proof. By Theorem 2.2 only the two-by-two blocks matter. The eigenvalues λ 1 , λ 2 \lambda_1,\lambda_2 λ 1 , λ 2 of the pencil R ( p ) , R ( q ) R(p),R(q) R ( p ) , R ( q ) satisfy
λ 1 λ 2 = det R ( q ) det R ( p ) = Δ q Δ p , λ 1 + λ 2 = 2 ( 1 − Re ( p ‾ q ) ) Δ p , \lambda_1\lambda_2=\frac{\det R(q)}{\det R(p)}=\frac{\Delta_q}{\Delta_p},\qquad
\lambda_1+\lambda_2=\frac{2(1-\operatorname{Re}(\overline{p}q))}{\Delta_p}, λ 1 λ 2 = det R ( p ) det R ( q ) = Δ p Δ q , λ 1 + λ 2 = Δ p 2 ( 1 − Re ( p q )) ,
and they are positive because the pencil is congruent to an HPD matrix. Put λ 1 , 2 = Δ q / Δ p e ± μ \lambda_{1,2}=\sqrt{\Delta_q/\Delta_p}\,e^{\pm\mu} λ 1 , 2 = Δ q / Δ p e ± μ . The product condition is satisfied identically and the sum condition becomes μ + μ − 1 = 2 h \mu+\mu^{-1}=2h μ + μ − 1 = 2 h . The arithmetic-geometric mean inequality gives h ≥ 1 h\ge1 h ≥ 1 , and μ + μ − 1 = 2 h \mu+\mu^{-1}=2h μ + μ − 1 = 2 h is solved by μ = e ± k \mu=e^{\pm k} μ = e ± k . Therefore log λ 1 , 2 = ℓ ± k \log\lambda_{1,2}=\ell\pm k log λ 1 , 2 = ℓ ± k and ∑ i ( log λ i ) 2 = 2 ℓ 2 + 2 k 2 \sum_i(\log\lambda_i)^2=2\ell^2+2k^2 ∑ i ( log λ i ) 2 = 2 ℓ 2 + 2 k 2 , which is Eq. (20) .
Every slice path is an ambient path, so d A ≤ d d_A\le d d A ≤ d . Suppose equality held for distinct endpoints. Then the unique minimizing slice geodesic would be an ambient minimizing geodesic. The ambient geodesic equation is A ′ ′ = A ′ A − 1 A ′ A''=A'A^{-1}A' A ′′ = A ′ A − 1 A ′ , obtained by differentiating the exponential in Eq. (18) . At any point where z ′ ≠ 0 z'\ne0 z ′ = 0 , the first diagonal entry satisfies
e 1 ∗ A ′ ′ e 1 = e 1 ∗ A ′ A − 1 A ′ e 1 > 0 , e_1^*A''e_1=e_1^*A'A^{-1}A'e_1>0, e 1 ∗ A ′′ e 1 = e 1 ∗ A ′ A − 1 A ′ e 1 > 0 ,
because A ′ e 1 ≠ 0 A'e_1\ne0 A ′ e 1 = 0 and A − 1 ≻ 0 A^{-1}\succ0 A − 1 ≻ 0 . A slice curve has constant first diagonal entry, hence e 1 ∗ A ′ ′ e 1 = 0 e_1^*A''e_1=0 e 1 ∗ A ′′ e 1 = 0 identically. A nonconstant slice geodesic has constant positive speed. Since z z z is its only varying coordinate, z ′ ≠ 0 z'\ne0 z ′ = 0 at every parameter value, yielding the contradiction. □ \square □
For example, the equal-weight ambient mean of R ( r ) R(r) R ( r ) and R ( − r ) R(-r) R ( − r ) , 0 < r < 1 0<r<1 0 < r < 1 , is 1 − r 2 I 2 \sqrt{1-r^2}\,I_2 1 − r 2 I 2 because the matrices commute, whereas the intrinsic slice mean is R ( 0 ) = I 2 R(0)=I_2 R ( 0 ) = I 2 . The ambient mean has changed both diagonal entries; it has left the slice. At the slice reference point
log R ( ρ ) = 1 2 log ( 1 − r 2 ) I 2 + atanh r r ( 0 ρ ρ ‾ 0 ) . (21) \log R(\rho)=\frac{1}{2}\log(1-r^2)I_2+\frac{\operatorname{atanh}r}{r}
\begin{pmatrix}
0 & \rho\\
\overline{\rho} & 0
\end{pmatrix}. \tag*{(21)} log R ( ρ ) = 2 1 log ( 1 − r 2 ) I 2 + r atanh r ( 0 ρ ρ 0 ) . (21)
with the ratio equal to one at the origin. The edge component is the ambient radial vector of length 2 atanh r \sqrt{2}\operatorname{atanh} r 2 atanh r , while the intrinsic slice logarithm has length ψ ( r ) \psi(r) ψ ( r ) . Their first-order agreement follows from ψ ( r ) = 2 r + O ( r 3 ) \psi(r)=\sqrt{2}r+O(r^3) ψ ( r ) = 2 r + O ( r 3 ) and does not identify their finite-displacement scores or their covariances.
The two-variable diagonal-scale quotient
Complex correlation quotient geometry was formulated by David and Gu [12 ] ; the real inverse-hyperbolic-tangent distance and the associated anomaly scores are established in [13 , 39 ] , and curvature restrictions on the full quotient-affine manifold in [40 ] . Here we calculate the two-variable comparator in the same normalization as Eq. (2) , so that the three geometries can be compared directly.
Proposition 4.2 (Exact complex two-variable quotient). Quotienting H ( 2 ) \mathbb{H}(2) H ( 2 ) by positive real diagonal congruence gives, on ρ = r e i θ \rho=re^{i\theta} ρ = r e i θ ,
d s Q 2 = 2 d r 2 ( 1 − r 2 ) 2 + 2 r 2 d θ 2 1 − r 2 . (22) \mathrm{d}s_Q^2=\frac{2\,\mathrm{d}r^2}{(1-r^2)^2}+\frac{2r^2\,\mathrm{d}\theta^2}{1-r^2}. \tag*{(22)} d s Q 2 = ( 1 − r 2 ) 2 2 d r 2 + 1 − r 2 2 r 2 d θ 2 . (22) Its Gaussian curvature is the constant − 1 / 2 -1/2 − 1/2 , and its distance is d Q ( p , q ) = 2 k d_Q(p,q)=\sqrt{2}\,k d Q ( p , q ) = 2 k with k k k from Eq. (19) . Consequently
d A 2 = d Q 2 + 2 ℓ 2 (23) d_A^2=d_Q^2+2\ell^2 \tag*{(23)} d A 2 = d Q 2 + 2 ℓ 2 (23) on the normalized representatives, and the real restriction is 2 ∣ atanh p − atanh q ∣ \sqrt{2}\lvert\operatorname{atanh}p-\operatorname{atanh}q\rvert 2 ∣ atanh p − atanh q ∣ .
Proof. At ρ = r \rho=r ρ = r the diagonal-scale directions are spanned by R R R and J = diag ( 1 , − 1 ) J=\operatorname{diag}(1,-1) J = diag ( 1 , − 1 ) . A radial edge derivative U U U satisfies g ( U , R ) = − 2 r / ( 1 − r 2 ) g(U,R)=-2r/(1-r^2) g ( U , R ) = − 2 r / ( 1 − r 2 ) , g ( R , R ) = 2 g(R,R)=2 g ( R , R ) = 2 and g ( U , J ) = 0 g(U,J)=0 g ( U , J ) = 0 ; the phase derivative is orthogonal to both. The span is orthogonal in this basis because g ( R , J ) = tr ( R − 1 J ) = 0 g(R,J)=\operatorname{tr}(R^{-1}J)=0 g ( R , J ) = tr ( R − 1 J ) = 0 . Removing the radial projection of the vertical directions subtracts 2 r 2 / ( 1 − r 2 ) 2 2r^2/(1-r^2)^2 2 r 2 / ( 1 − r 2 ) 2 from the radial coefficient of Eq. (6) and leaves the angular coefficient unchanged. Rephasing transports the calculation to every ρ \rho ρ , which gives Eq. (22) . Put u = 2 atanh r u=\sqrt{2}\operatorname{atanh}r u = 2 atanh r ; the warped-product profile is 2 sinh ( u / 2 ) \sqrt{2}\sinh(u/\sqrt{2}) 2 sinh ( u / 2 ) , whose second derivative divided by itself is 1 / 2 1/2 1/2 , so the curvature is − 1 / 2 -1/2 − 1/2 . Equivalently, the map
X ( ρ ) = ( 1 , x , y ) 1 − r 2 (24) X(\rho)=\frac{(1,x,y)}{\sqrt{1-r^2}} \tag*{(24)} X ( ρ ) = 1 − r 2 ( 1 , x , y ) (24)
pulls the metric 2 ( − d X 0 2 + d X 1 2 + d X 2 2 ) 2(-\mathrm{d}X_0^2+\mathrm{d}X_1^2+\mathrm{d}X_2^2) 2 ( − d X 0 2 + d X 1 2 + d X 2 2 ) on the future unit hyperboloid back to Eq. (22) . The Lorentz inner product of X ( p ) X(p) X ( p ) and X ( q ) X(q) X ( q ) is − h -h − h . The intersection of the hyperboloid with their two-plane is a geodesic of length 2 k \sqrt{2}k 2 k , parameterized by
sinh ( ( 1 − t ) k ) sinh k X ( p ) + sinh ( t k ) sinh k X ( q ) , 0 ≤ t ≤ 1 , \frac{\sinh((1-t)k)}{\sinh k}X(p)+\frac{\sinh(tk)}{\sinh k}X(q),\qquad0\le t\le1, sinh k sinh (( 1 − t ) k ) X ( p ) + sinh k sinh ( t k ) X ( q ) , 0 ≤ t ≤ 1 ,
and completeness with negative curvature makes it the unique shortest path. Equation (23) is then Eq. (20) with the factor 2 ℓ 2 2\ell^2 2 ℓ 2 split off. Restricting to a diameter gives the stated real formula.
This calculation concerns the two-variable Schur block, not the quotient of a full n n n -variable correlation matrix, and confers neither constant curvature nor uniqueness on the latter; the general limitations are established in [38 , 40 ] . For implementation the hyperboloid logarithm is V = k [ X ( q ) − h X ( p ) ] / sinh k V=k[X(q)-hX(p)]/\sinh k V = k [ X ( q ) − h X ( p )] / sinh k with the continuous zero-distance limit, pulled back by d ρ = ( d X 1 , d X 2 ) / X 0 − ( X 1 , X 2 ) d X 0 / X 0 2 \mathrm{d}\rho=(\mathrm{d}X_1,\mathrm{d}X_2)/X_0-(X_1,X_2)\mathrm{d}X_0/X_0^2 d ρ = ( d X 1 , d X 2 ) / X 0 − ( X 1 , X 2 ) d X 0 / X 0 2 .
Flat coordinate comparators
The log-Euclidean comparator uses the isometric Hermitian vectorization of log A \log A log A with Euclidean centring and covariance [2 , 3 ] . It extends to complex HPD because the principal Hermitian logarithm is a smooth bijection onto the space of Hermitian matrices, with inverse the matrix exponential. The log-Cholesky construction in the real case is due to Lin 30; the complex extension used here is the following pullback, which we derive rather than attribute to [30 ] .
Let A = L L ∗ A=LL^* A = L L ∗ be the Cholesky factorization with positive real diagonal and define c ( A ) c(A) c ( A ) to contain log L j j \log L_{jj} log L j j together with the unweighted real and imaginary parts of L i j L_{ij} L ij for i > j i>j i > j . These coordinates are a diffeomorphism onto R n 2 \mathbb{R}^{n^2} R n 2 : exponentiating the diagonal and filling the strict lower triangle reconstructs L L L and then A A A , and the positive pivots in the recursive Cholesky formula make both the map and its inverse smooth. Euclidean distance, straight-line coordinate geodesics, coordinate averaging and covariance follow by this explicit pullback. There is no 2 \sqrt{2} 2 weight here, because the coordinate metric is defined by this choice rather than by Hermitian Frobenius vectorization. On a fixed slice A 11 = a A_{11}=a A 11 = a is constant, so log L 11 = 1 2 log a \log L_{11}=\frac{1}{2}\log a log L 11 = 2 1 log a is removed before covariance fitting.
Log-Cholesky depends on the ordering of the variables, and arbitrary rescalings change its strict-triangle coordinates differently from its log diagonal. For dimensional data a fixed declared unit scale is required before taking logarithms or choosing this coordinate metric, and no universal invariance of its whitened score is presumed.
Invariances of the declared slice
The slice carries a group of transformations that preserve its normalized geometry exactly. Knowing that group, and knowing what it does not contain, is what separates a statement about a geometric object from a statement about a particular numerical matrix.
Proposition 5.1 (Slice-preserving congruences). Let T = diag ( d 1 , d 2 , W ) T=\operatorname{diag}(d_1,d_2,W) T = diag ( d 1 , d 2 , W ) with d 1 , d 2 ∈ C ∖ { 0 } d_1,d_2\in\mathbb{C}\setminus\{0\} d 1 , d 2 ∈ C ∖ { 0 } and W W W invertible on the complement. Then T T T maps one fixed-complement slice onto another, and
α ′ = ∣ d 1 ∣ 2 α , β ′ = ∣ d 2 ∣ 2 β , γ ′ = d 1 d 2 ‾ γ , \alpha'=|d_1|^2\alpha,\qquad\beta'=|d_2|^2\beta,\qquad\gamma'=d_1\overline{d_2}\gamma, α ′ = ∣ d 1 ∣ 2 α , β ′ = ∣ d 2 ∣ 2 β , γ ′ = d 1 d 2 γ , ρ ′ = e i φ ρ , e i φ = d 1 d 2 ‾ ∣ d 1 d 2 ∣ . \rho'=e^{i\varphi}\rho,\qquad e^{i\varphi}=\frac{d_1\overline{d_2}}{|d_1d_2|}. ρ ′ = e i φ ρ , e i φ = ∣ d 1 d 2 ∣ d 1 d 2 .
The induced action on the disk is a rotation. It is an isometry of Eq. (5) . Swapping the two selected labels conjugates ρ \rho ρ ; a permutation that transports the selected pair and the complement only relabels the construction. In the real case, positive rescaling leaves ρ \rho ρ unchanged while real rescaling with d 1 d 2 < 0 d_1d_2<0 d 1 d 2 < 0 reverses the signed coordinate.
Proof. Under T T T the blocks become C ′ = W C W ∗ C'=WCW^* C ′ = W C W ∗ , E ′ = diag ( d 1 , d 2 ) E W ∗ E'=\operatorname{diag}(d_1,d_2)EW^* E ′ = diag ( d 1 , d 2 ) E W ∗ and α ′ = ∣ d 1 ∣ 2 α \alpha'=|d_1|^2\alpha α ′ = ∣ d 1 ∣ 2 α , b ′ = ∣ d 2 ∣ 2 b b'=|d_2|^2b b ′ = ∣ d 2 ∣ 2 b . Then H ′ = E ′ C ′ − 1 E ′ ∗ = diag ( d 1 , d 2 ) E ∗ C − 1 E diag ( d 1 ‾ , d 2 ‾ ) H'=E'C'^{-1}E'^*= \operatorname{diag}(d_1,d_2)E^*C^{-1}E\operatorname{diag}(\overline{d_1},\overline{d_2}) H ′ = E ′ C ′ − 1 E ′ ∗ = diag ( d 1 , d 2 ) E ∗ C − 1 E diag ( d 1 , d 2 ) , because W W W and W − 1 W^{-1} W − 1 cancel in E ′ C ′ − 1 E ′ ∗ E'C'^{-1}E'^* E ′ C ′ − 1 E ′ ∗ . Reading off the entries gives the first three identities of Eq. (25), and
ρ ′ = d 1 d 2 ‾ ( z − γ ) ∣ d 1 ∣ 2 α ∣ d 2 ∣ 2 β = d 1 d 2 ‾ ∣ d 1 d 2 ∣ ρ . \rho'=\frac{d_1\overline{d_2}(z-\gamma)}{\sqrt{|d_1|^2\alpha\,|d_2|^2\beta}}=\frac{d_1\overline{d_2}}{|d_1d_2|}\rho. ρ ′ = ∣ d 1 ∣ 2 α ∣ d 2 ∣ 2 β d 1 d 2 ( z − γ ) = ∣ d 1 d 2 ∣ d 1 d 2 ρ .
A rotation of the disk preserves r = ∣ ρ ∣ r=|\rho| r = ∣ ρ ∣ , and by Eq. (6) it preserves G G G . Swapping the two selected labels exchanges a ↔ b a\leftrightarrow b a ↔ b and sends z ↦ z ‾ z\mapsto\overline{z} z ↦ z , hence ρ ↦ ρ ‾ \rho\mapsto\overline{\rho} ρ ↦ ρ . For real factors, e i φ = sign ( d 1 d 2 ) e^{i\varphi}=\operatorname{sign}(d_1d_2) e i φ = sign ( d 1 d 2 ) . The signed coordinate reverses exactly when d 1 d 2 < 0 d_1d_2<0 d 1 d 2 < 0 .
If the reference law is pushed forward by the same action, its centre, logarithms and covariance push forward under the differential, so the dispersion and the scalar Q Q Q of Proposition 3.3 are invariant while signed or frame components obey their rotation or reflection rules. Using a different reference law without pushing the original one forward is a different statistical problem and carries no such invariance.
Proposition 5.2 (Fixed complementary regressors). Fix R 0 R_0 R 0 of compatible size and invertible W W W , and let T = ( D 0 0 W ) T=\begin{pmatrix}D&0\\0&W\end{pmatrix} T = ( D 0 0 W ) with D = diag ( d 1 , d 2 ) D=\operatorname{diag}(d_1,d_2) D = diag ( d 1 , d 2 ) (or a selected-label permutation followed by that diagonal). Then the transformed cross block is ( D E + R 0 C ) W ∗ (DE+R_0C)W^* ( D E + R 0 C ) W ∗ and remains fixed across the slice. For diagonal D D D the normalized connection transforms as in Eq. (25); an additional swap of the selected labels composes that rotation with complex conjugation.
Proof. The cross block of T A T ∗ TAT^* T A T ∗ in the selected rows is D E + R 0 C DE+R_0C D E + R 0 C times W ∗ W^* W ∗ , which does not depend on z z z . Expanding the Schur complement of the transformed matrix and cancelling the three terms introduced by $R_0
leaves S ′ = D S D ∗ S^{\prime}=DSD^{*} S ′ = D S D ∗ . For diagonal D D D , normalizing by diag ( ∣ d 1 ∣ α , ∣ d 2 ∣ β ) \operatorname{diag}(|d_1|^{\sqrt{\alpha}},|d_2|^{\sqrt{\beta}}) diag ( ∣ d 1 ∣ α , ∣ d 2 ∣ β ) gives Eq. (25). A swap exchanges the two diagonal entries and conjugates the normalized off-diagonal entry, as in Proposition 5.1 .
Proposition 5.2 describes a sufficient family, not an exhaustive classification of slice-preserving transformations.
Remark 5.3 (What ambient invariance does not transfer). Generic mixing and inversion preserve the ambient affine-invariant distance but do not preserve the isolated-entry constraint. The real Hadamard rotation T = 2 − 1 / 2 ( 1 1 1 − 1 ) T=2^{-1/2}\begin{pmatrix}1&1\\1&-1\end{pmatrix} T = 2 − 1/2 ( 1 1 1 − 1 ) sends R ( t ) R(t) R ( t ) to diag ( 1 + t , 1 − t ) \operatorname{diag}(1+t,1-t) diag ( 1 + t , 1 − t ) , so a varying edge has become a varying diagonal. Likewise
R ( t ) − 1 = ( 1 − t 2 ) − 1 ( 1 − t − t 1 ) R(t)^{-1}=(1-t^2)^{-1}\begin{pmatrix}1&-t\\-t&1\end{pmatrix} R ( t ) − 1 = ( 1 − t 2 ) − 1 ( 1 − t − t 1 )
has varying diagonal entries: its first diagonal derivative vanishes at t = 0 t=0 t = 0 but its second derivative does not. A pointwise inverse normalization that happens to reproduce − t -t − t is therefore not a constant slice-preserving matrix action.
These examples prohibit unrestricted transfer of ambient invariance to a labelled connection score. An isolated tensor connection retains its physical frame, and covariance or coherence conditioning by itself establishes no directed causal interpretation. The conditional spectral quantities built on the same Schur block carry a directional reading only when a directional model is supplied on top of them, as in partial-coherence tests for Granger causality [20 , 36 ] .
The induced geometry supplies a fixed-complement setting for established tools from correlation geometry, Fréchet averaging, tangent whitening and exchangeable rank calibration [12 , 27 , 28 , 31 , 39 , 40 ] .
Geodesics, logarithms and parallel transport
Theorem 2.5 guarantees that every pair of points is joined by one minimizing geodesic. This section turns that existence statement into a computable formula: the geodesic problem reduces to one scalar root on two exactly integrable legs, and the logarithm is recovered from the same data. Figure 5 shows three endpoint configurations from one base point, the turning branch at a wider separation, the three distances on identical endpoints and the scalar reduction against independent shooting.
Figure 5. Geodesics admit a scalar reduction. (A) Diameter, monotone-radius and turning paths from a common base point. (B) The minimum radius of a turning path, compared with a monotone path. Heat contours in A and B indicate intrinsic distance from the origin. (C) Intrinsic, ambient and quotient distances on identical endpoints. (D) Distance from the scalar reduction agrees with independent shooting of the Cartesian geodesic equation; color indicates distance.
Christoffel symbols and the geodesic equation
Write x = ( x 1 , x 2 ) x=(x_1,x_2) x = ( x 1 , x 2 ) , q = ∣ x ∣ 2 q=|x|^2 q = ∣ x ∣ 2 , and let δ \delta δ denote the Kronecker symbol.
Lemma 6.1 (Levi-Civita symbols). For the metric (5) the Christoffel symbols of the Levi-Civita connection are
Γ j k i = x j δ i k + x k δ i j 1 − q + x i δ j k 1 + q . (25) \Gamma^{i}_{jk}=\frac{x_j\delta_{ik}+x_k\delta_{ij}}{1-q}+\frac{x_i\delta_{jk}}{1+q}. \tag*{(25)} Γ j k i = 1 − q x j δ ik + x k δ ij + 1 + q x i δ j k . (25) The geodesic equation is therefore
x ¨ = − 2 ( x ⋅ x ˙ ) x ˙ 1 − q − ∣ x ˙ ∣ 2 x 1 + q . (26) \ddot{x}=-\frac{2(x\cdot\dot{x})\dot{x}}{1-q}-\frac{|\dot{x}|^2x}{1+q}. \tag*{(26)} x ¨ = − 1 − q 2 ( x ⋅ x ˙ ) x ˙ − 1 + q ∣ x ˙ ∣ 2 x . (26)
Proof. The Levi-Civita connection is the unique torsion-free connection compatible with G G G , so it suffices to verify those two properties for Eq. (25) and to lower an index. Symmetry in j , k j,k j , k is immediate from Eq. (25) , which is torsion-free.
For compatibility, put λ = 2 / ( 1 − q ) \lambda=2/(1-q) λ = 2/ ( 1 − q ) and η = 4 / ( 1 − q ) 2 \eta=4/(1-q)^2 η = 4/ ( 1 − q ) 2 , so that G i j = λ δ i j + η x i x j G_{ij}=\lambda\delta_{ij}+\eta x_ix_j G ij = λ δ ij + η x i x j . Then
∂ m λ = 4 x m ( 1 − q ) 2 , ∂ m η = 16 x m ( 1 − q ) 3 . \partial_m\lambda=\frac{4x_m}{(1-q)^2},\qquad\partial_m\eta=\frac{16x_m}{(1-q)^3}. ∂ m λ = ( 1 − q ) 2 4 x m , ∂ m η = ( 1 − q ) 3 16 x m .
and
∂ k G i j = 4 x k δ i j ( 1 − q ) 2 + η ( δ i k x j + x i δ j k ) + 16 x i x j x k ( 1 − q ) 3 . \partial_kG_{ij}=\frac{4x_k\delta_{ij}}{(1-q)^2}+\eta(\delta_{ik}x_j+x_i\delta_{jk})+\frac{16x_ix_jx_k}{(1-q)^3}. ∂ k G ij = ( 1 − q ) 2 4 x k δ ij + η ( δ ik x j + x i δ j k ) + ( 1 − q ) 3 16 x i x j x k .
On the other hand G i l Γ j k l = λ Γ i j k + η x i x l Γ j k l G_{il}\Gamma^l_{jk}=\lambda\Gamma_{ijk}+\eta x_ix_l\Gamma^l_{jk} G i l Γ j k l = λ Γ ij k + η x i x l Γ j k l , and
x l Γ j k l = 2 x j x k 1 − q + q δ j k 1 + q . x_l\Gamma^l_{jk}=\frac{2x_jx_k}{1-q}+\frac{q\delta_{jk}}{1+q}. x l Γ j k l = 1 − q 2 x j x k + 1 + q q δ j k .
Adding the corresponding expression with i i i and j j j interchanged, and using η = λ 2 \eta=\lambda^2 η = λ 2 together with ∂ k λ = 4 x k / ( 1 − q ) 2 \partial_k\lambda=4x_k/(1-q)^2 ∂ k λ = 4 x k / ( 1 − q ) 2 and ∂ k η = 16 x k / ( 1 − q ) 3 \partial_k\eta=16x_k/(1-q)^3 ∂ k η = 16 x k / ( 1 − q ) 3 , reproduces exactly ∂ k G i j \partial_kG_{ij} ∂ k G ij : the δ i k \delta_{ik} δ ik coefficient becomes 2 x j / ( 1 − q ) 2 + 2 x j / [ ( 1 − q ) ( 1 + q ) ] = 4 x j / ( 1 − q 2 ) 2x_j/(1-q)^2+2x_j/[(1-q)(1+q)]=4x_j/(1-q^2) 2 x j / ( 1 − q ) 2 + 2 x j / [( 1 − q ) ( 1 + q )] = 4 x j / ( 1 − q 2 ) ; the δ i j \delta_{ij} δ ij coefficient becomes 4 x k / ( 1 − q ) 2 4x_k/(1-q)^2 4 x k / ( 1 − q ) 2 , and the cubic term becomes 4 η x i x j x k / ( 1 − q ) = 16 x i x j x k / ( 1 − q ) 3 4\eta x_ix_jx_k/(1-q)=16x_ix_jx_k/(1-q)^3 4 η x i x j x k / ( 1 − q ) = 16 x i x j x k / ( 1 − q ) 3 . Hence ∂ k G i j = G i l Γ j k l + G j l Γ i k l \partial_kG_{ij}=G_{il}\Gamma^l_{jk}+G_{jl}\Gamma^l_{ik} ∂ k G ij = G i l Γ j k l + G j l Γ ik l , which is metric compatibility. By the fundamental theorem of Riemannian geometry this is the Levi-Civita connection of G G G .
Contracting Eq. (25) with x ˙ j x ˙ k \dot{x}^j\dot{x}^k x ˙ j x ˙ k gives
Γ j k i x ˙ j x ˙ k = 2 ( x ⋅ x ˙ ) x ˙ i 1 − q + ∣ x ˙ ∣ 2 x i 1 + q , \Gamma^i_{jk}\dot{x}^j\dot{x}^k=\frac{2(x\cdot\dot{x})\dot{x}^i}{1-q}+\frac{|\dot{x}|^2x^i}{1+q}, Γ j k i x ˙ j x ˙ k = 1 − q 2 ( x ⋅ x ˙ ) x ˙ i + 1 + q ∣ x ˙ ∣ 2 x i ,
and x ¨ = − Γ j k i x ˙ j x ˙ k \ddot{x}=-\Gamma^i_{jk}\dot{x}^j\dot{x}^k x ¨ = − Γ j k i x ˙ j x ˙ k is Eq. (26) .
The Cartesian form is regular everywhere in the open disk; the apparent singularity in the polar symbols at r = 0 r=0 r = 0 is the coordinate artefact already noted after Theorem 2.3 . For reference,
Γ r r r = r 1 + r 2 + 2 r 1 − r 2 , Γ θ θ r = − r 1 + r 2 , \Gamma^r_{rr}=\frac{r}{1+r^2}+\frac{2r}{1-r^2},\qquad\Gamma^r_{\theta\theta}=-\frac{r}{1+r^2}, Γ r r r = 1 + r 2 r + 1 − r 2 2 r , Γ θ θ r = − 1 + r 2 r ,
Γ r θ θ = 1 r ( 1 − r 2 ) . (27) \Gamma^\theta_{r\theta}=\frac{1}{r(1-r^2)}. \tag*{(27)} Γ r θ θ = r ( 1 − r 2 ) 1 . (27)
First integrals and the scalar reduction
Because G G G is invariant under rotations, the geodesic flow has two first integrals. Parametrize a geodesic by arc length, so g ( x ˙ , x ˙ ) = 1 g(\dot{x},\dot{x})=1 g ( x ˙ , x ˙ ) = 1 , and write s = ψ ( r ) s=\psi(r) s = ψ ( r ) and w = f ( s ) = 2 r / 1 − r 2 w=f(s)=\sqrt{2}\,r/\sqrt{1-r^2} w = f ( s ) = 2 r / 1 − r 2 .
Lemma 6.2 (Clairaut constant and energy). Along a unit-speed geodesic the quantity
c : = f ( s ) 2 θ ˙ = w 2 θ ˙ (28) c:=f(s)^2\dot{\theta}=w^2\dot{\theta} \tag*{(28)} c := f ( s ) 2 θ ˙ = w 2 θ ˙ (28) is constant, and
s ˙ 2 + c 2 w 2 = 1. (29) \dot{s}^{\,2}+\frac{c^2}{w^2}=1. \tag*{(29)} s ˙ 2 + w 2 c 2 = 1. (29)
In particular w ≥ ∣ c ∣ w \ge|c| w ≥ ∣ c ∣ wherever the geodesic is defined, with equality exactly at a radial turning point, and a non-radial geodesic has at most one turning point. The constant c c c is signed, carrying the direction of rotation.
Proof. In polar coordinates the Lagrangian of the metric (6) is L = 1 2 ( a ( r ) r ˙ 2 + b ( r ) θ ˙ 2 ) \mathcal{L} = \frac{1}{2}\left(a(r)\dot{r}^{2} + b(r)\dot{\theta}^{2}\right) L = 2 1 ( a ( r ) r ˙ 2 + b ( r ) θ ˙ 2 ) and does not depend on θ \theta θ , so p θ = b ( r ) θ ˙ p_\theta= b(r)\dot{\theta} p θ = b ( r ) θ ˙ is conserved. Since b ( r ) = 2 r 2 1 − r 2 = f ( s ) 2 b(r) = \frac{2r^{2}}{1-r^{2}} = f(s)^{2} b ( r ) = 1 − r 2 2 r 2 = f ( s ) 2 , this is Eq. (28) . Unit speed gives a r ˙ 2 + b θ ˙ 2 = 1 a\dot{r}^{2} + b\dot{\theta}^{2} = 1 a r ˙ 2 + b θ ˙ 2 = 1 , and a = ψ ′ 2 a = \psi'^{2} a = ψ ′ 2 together with s ˙ = ψ ′ r ˙ \dot{s} = \psi'\dot{r} s ˙ = ψ ′ r ˙ turns this into Eq. (29) . Equation (30) forces w 2 ≥ c 2 w^{2} \ge c^{2} w 2 ≥ c 2 , and equality forces s ˙ = 0 \dot{s} = 0 s ˙ = 0 . The second derivative of s s s at a turning point is s ¨ = f f s θ ˙ 2 > 0 \ddot{s} = f f_s \dot{\theta}^{2} > 0 s ¨ = f f s θ ˙ 2 > 0 whenever c ≠ 0 c \ne0 c = 0 , so the radius has a strict minimum there and cannot turn twice.
Set h ( w ) = d s / d w h(w) = ds/dw h ( w ) = d s / d w , the inverse of d w / d s dw/ds d w / d s . From w = 2 r / 1 − r 2 w = \sqrt{2}r/\sqrt{1-r^{2}} w = 2 r / 1 − r 2 one computes r ( w ) = w / w 2 + 2 r(w) = w/\sqrt{w^{2}+2} r ( w ) = w / w 2 + 2 and
h ( w ) = 1 − r ( w ) 4 = 2 1 + w 2 w 2 + 2 . (30) h(w) = \sqrt{1-r(w)^{4}} = \frac{2\sqrt{1+w^{2}}}{w^{2}+2}. \tag*{(30)} h ( w ) = 1 − r ( w ) 4 = w 2 + 2 2 1 + w 2 . (30)
The function h h h is positive, satisfies h ( 0 ) = 1 h(0) = 1 h ( 0 ) = 1 , and is strictly decreasing for w > 0 w > 0 w > 0 .
Lemma 6.3 (Closed-form geodesic legs). For 0 < c ≤ W 0 < c \le W 0 < c ≤ W define
A ( c , W ) = ∫ c W c h ( w ) w w 2 − c 2 d w (31) A(c,W) = \int_{c}^{W} \frac{c h(w)}{w\sqrt{w^{2}-c^{2}}}\,dw \tag*{(31)} A ( c , W ) = ∫ c W w w 2 − c 2 c h ( w ) d w (31) = ∫ 0 arccos ( c / W ) h ( c sec t ) d t , (32) = \int_{0}^{\arccos(c/W)} h(c\sec t)\,dt, \tag*{(32)} = ∫ 0 a r c c o s ( c / W ) h ( c sec t ) d t , (32) L ( c , W ) = ∫ c W w h ( w ) w 2 − c 2 d w (33) L(c,W) = \int_{c}^{W} \frac{w h(w)}{\sqrt{w^{2}-c^{2}}}\,dw \tag*{(33)} L ( c , W ) = ∫ c W w 2 − c 2 w h ( w ) d w (33) = ∫ 0 acosh ( W / c ) c cosh u h ( c cosh u ) d u . (34) = \int_{0}^{\operatorname{acosh}(W/c)} c\cosh u\,h(c\cosh u)\,du. \tag*{(34)} = ∫ 0 acosh ( W / c ) c cosh u h ( c cosh u ) d u . (34) With
D = W 2 − c 2 , a = 1 + c 2 , D = \sqrt{W^{2}-c^{2}}, \qquad a = \sqrt{1+c^{2}}, D = W 2 − c 2 , a = 1 + c 2 , b = 2 + c 2 , v = D 1 + W 2 . (35) b = \sqrt{2+c^{2}}, \qquad v = \frac{D}{\sqrt{1+W^{2}}}. \tag*{(35)} b = 2 + c 2 , v = 1 + W 2 D . (35)
both legs are elementary:
A ( c , W ) = arctan v c + c b atanh v b , (36) A(c,W)=\arctan\frac{v}{c}+\frac{c}{b}\operatorname{atanh}\frac{v}{b}, \tag*{(36)} A ( c , W ) = arctan c v + b c atanh b v , (36)
L ( c , W ) = 2 ( asinh D a − 1 b atanh v b ) . (37) L(c,W)=2\left(\operatorname{asinh}\frac{D}{a}-\frac{1}{b}\operatorname{atanh}\frac{v}{b}\right). \tag*{(37)} L ( c , W ) = 2 ( asinh a D − b 1 atanh b v ) . (37)
Moreover A ( 0 , W ) = π / 2 A(0,W)=\pi/2 A ( 0 , W ) = π /2 , L ( 0 , W ) = ψ ( r ( W ) ) L(0,W)=\psi(r(W)) L ( 0 , W ) = ψ ( r ( W )) , A ( c , c ) = L ( c , c ) = 0 A(c,c)=L(c,c)=0 A ( c , c ) = L ( c , c ) = 0 , and both legs are continuous on { 0 ≤ c ≤ W , W > 0 } \{0\le c\le W,\ W>0\} { 0 ≤ c ≤ W , W > 0 } together with the limits c ↓ 0 c\downarrow0 c ↓ 0 at fixed W > 0 W>0 W > 0 . The angular leg has no continuous extension to ( 0 , 0 ) (0,0) ( 0 , 0 ) : along c = 0 c=0 c = 0 its value is π / 2 \pi/2 π /2 , whereas along c = W > 0 c=W>0 c = W > 0 it is zero. The length leg does extend continuously there with value zero.
Proof. The substitution w = c sec t w=c\sec t w = c sec t in Eq. (32) gives d w = w tan t d t dw=w\tan t\,dt d w = w tan t d t and w 2 − c 2 = c tan t \sqrt{w^2-c^2}=c\tan t w 2 − c 2 = c tan t , hence d w / ( w w 2 − c 2 ) = d t / c dw/(w\sqrt{w^2-c^2})=dt/c d w / ( w w 2 − c 2 ) = d t / c and the second form. The substitution w = c cosh u w=c\cosh u w = c cosh u in Eq. (34) gives d w / w 2 − c 2 = d u dw/\sqrt{w^2-c^2}=du d w / w 2 − c 2 = d u and the second form there.
For Eq. (36) , differentiate F ( W ) = arctan ( v / c ) + ( c / b ) atanh ( v / b ) F(W)=\arctan(v/c)+(c/b)\operatorname{atanh}(v/b) F ( W ) = arctan ( v / c ) + ( c / b ) atanh ( v / b ) with respect to W W W . Since D ′ = W / D D'=W/D D ′ = W / D and v ′ = ( W / D − W v / 1 + W 2 ) / 1 + W 2 v'=(W/D-Wv/\sqrt{1+W^2})/\sqrt{1+W^2} v ′ = ( W / D − W v / 1 + W 2 ) / 1 + W 2 , an algebraic simplification gives
F ′ ( W ) = c h ( W ) W W 2 − c 2 , (38) F'(W)=\frac{c h(W)}{W\sqrt{W^2-c^2}}, \tag*{(38)} F ′ ( W ) = W W 2 − c 2 c h ( W ) , (38)
which is the integrand of Eq. (32) evaluated at the upper limit, and F ( c ) = 0 F(c)=0 F ( c ) = 0 . Hence F = A ( c , ⋅ ) F=A(c,\cdot) F = A ( c , ⋅ ) . For Eq. (37) , the same computation for G ( W ) = 2 ( asinh ( D / a ) − ( 1 / b ) atanh ( v / b ) ) G(W)=2(\operatorname{asinh}(D/a)-(1/b)\operatorname{atanh}(v/b)) G ( W ) = 2 ( asinh ( D / a ) − ( 1/ b ) atanh ( v / b )) gives
G ′ ( W ) = W h ( W ) W 2 − c 2 , (39) G'(W)=\frac{W h(W)}{\sqrt{W^2-c^2}}, \tag*{(39)} G ′ ( W ) = W 2 − c 2 W h ( W ) , (39)
which is the integrand of Eq. (34) , and G ( c ) = 0 G(c)=0 G ( c ) = 0 . For fixed W > 0 W>0 W > 0 , extend the angular integrand by zero to [ 0 , π / 2 ] [0,\pi/2] [ 0 , π /2 ] . It is bounded by one and converges pointwise almost everywhere to one as c ↓ 0 c\downarrow0 c ↓ 0 , so dominated convergence gives A → π / 2 A\to\pi/2 A → π /2 . For the length leg, set w = c 2 + ( W 2 − c 2 ) v 2 w=\sqrt{c^2+(W^2-c^2)v^2} w = c 2 + ( W 2 − c 2 ) v 2 to obtain L = W 2 − c 2 ∫ 0 1 h ( w ( v ) ) d v L=\sqrt{W^2-c^2}\int_0^1h(w(v))\,dv L = W 2 − c 2 ∫ 0 1 h ( w ( v )) d v . Dominated convergence gives L ( 0 , W ) = ∫ 0 W h ( w ) d w = ψ ( r ( W ) ) L(0,W)=\int_0^W h(w)\,dw=\psi(r(W)) L ( 0 , W ) = ∫ 0 W h ( w ) d w = ψ ( r ( W )) . Also 0 ≤ L ( c , W ) ≤ W 0\le L(c,W)\le W 0 ≤ L ( c , W ) ≤ W , which proves continuity of L L L at ( 0 , 0 ) (0,0) ( 0 , 0 ) . At c = W c=W c = W both integrals have zero length, and both closed forms reduce to arctan 0 + 0 = 0 \arctan0+0=0 arctan 0 + 0 = 0 and 2 ( asinh 0 − 0 ) = 0 2(\operatorname{asinh}0-0)=0 2 ( asinh 0 − 0 ) = 0 .
Proposition 6.4 (Existence and uniqueness of the reduction). Fix non-degenerate endpoints: radii r 1 , r 2 > 0 r_1,r_2>0 r 1 , r 2 > 0 and signed angular separation δ \delta δ with 0 < ∣ δ ∣ < π 0<|\delta|<\pi 0 < ∣ δ ∣ < π . Put W i = w ( r i ) W_i=w(r_i) W i = w ( r i ) and order W 1 ≤ W 2 W_1\le W_2 W 1 ≤ W 2 , and let δ ∗ = A ( W 1 , W 2 ) \delta_* = A(W_1,W_2) δ ∗ = A ( W 1 , W 2 ) . Then a unique geodesic joins the endpoints, the magnitude of its Clairaut constant is the unique c ∈ ( 0 , W 1 ] c\in(0,W_1] c ∈ ( 0 , W 1 ] solving
A ( c , W 2 ) − A ( c , W 1 ) = ∣ δ ∣ , ∣ δ ∣ ≤ δ ∗ , (40) A(c,W_2)-A(c,W_1)=|\delta|,\qquad|\delta|\le\delta_*, \tag*{(40)} A ( c , W 2 ) − A ( c , W 1 ) = ∣ δ ∣ , ∣ δ ∣ ≤ δ ∗ , (40) A ( c , W 2 ) + A ( c , W 1 ) = ∣ δ ∣ , ∣ δ ∣ ≥ δ ∗ . (41) A(c,W_2)+A(c,W_1)=|\delta|,\qquad|\delta|\ge\delta_*. \tag*{(41)} A ( c , W 2 ) + A ( c , W 1 ) = ∣ δ ∣ , ∣ δ ∣ ≥ δ ∗ . (41) On the first branch the radius varies monotonically and the distance is L ( c , W 2 ) − L ( c , W 1 ) L(c,W_2)-L(c,W_1) L ( c , W 2 ) − L ( c , W 1 ) ; on the strict second branch ∣ δ ∣ > δ ∗ |\delta|>\delta_* ∣ δ ∣ > δ ∗ the geodesic has one interior radial minimum, and its distance is L ( c , W 2 ) + L ( c , W 1 ) L(c,W_2)+L(c,W_1) L ( c , W 2 ) + L ( c , W 1 ) . At ∣ δ ∣ = δ ∗ |\delta|=\delta_* ∣ δ ∣ = δ ∗ the two branch equations coincide and both formulas give the same distance.
Proof. For W 1 < W 2 W_1<W_2 W 1 < W 2 the monotone angle is
M ( c ) = ∫ W 1 W 2 c h ( w ) w w 2 − c 2 d w . M(c)=\int_{W_1}^{W_2}\frac{c h(w)}{w\sqrt{w^2-c^2}}\,dw. M ( c ) = ∫ W 1 W 2 w w 2 − c 2 c h ( w ) d w .
It is strictly increasing for 0 < c < W 1 0<c<W_1 0 < c < W 1 , since the derivative of its integrand with respect to c c c is h ( w ) ( w 2 − c 2 ) − 3 / 2 > 0 h(w)(w^2-c^2)^{-3/2}>0 h ( w ) ( w 2 − c 2 ) − 3/2 > 0 . Continuity at the endpoints gives M ( 0 ) = 0 M(0)=0 M ( 0 ) = 0 and M ( W 1 ) = δ ∗ M(W_1)=\delta_* M ( W 1 ) = δ ∗ . If W 1 = W 2 W_1=W_2 W 1 = W 2 , then M ≡ 0 M\equiv0 M ≡ 0 and δ ∗ = 0 \delta_*=0 δ ∗ = 0 , so no nonzero angle uses the monotone branch.
For fixed W W W , the transformed angular integral in Eq. (32) is strictly decreasing in c c c : its upper limit decreases, and h ( c sec t ) h(c\sec t) h ( c sec t ) decreases pointwise on the common interval. Thus T ( c ) = A ( c , W 2 ) + A ( c , W 1 ) T(c)=A(c,W_2)+A(c,W_1) T ( c ) = A ( c , W 2 ) + A ( c , W 1 ) decreases strictly from π \pi π at c = 0 c=0 c = 0 to δ ∗ \delta_* δ ∗ at c = W 1 c=W_1 c = W 1 . These ranges prove existence and uniqueness of the stated scalar root and identify the branch without a separate monotonicity assumption.
Conversely, prescribe the unit tangent at the smaller-radius endpoint with angular component c / W 1 c/W_1 c / W 1 and radial component ± 1 − c 2 / W 1 2 \pm\sqrt{1-c^2/W_1^2} ± 1 − c 2 / W 1 2 , taking the outward sign for M M M and inward sign for T T T . The smooth Cartesian geodesic equation has a unique solution. Its conserved quantities give the stated angular integrals. In the turning case the inward part reaches w = c w=c w = c in finite length; the positive radial acceleration there gives its smooth outward continuation. The root equation places the other endpoint at precisely its prescribed angle and radius. Theorem 2.5 makes this geodesic minimizing.
The energy identity gives the length element d t = h ( w ) ∣ d w ∣ / w 2 − c 2 dt=h(w)|dw|/\sqrt{w^2-c^2} d t = h ( w ) ∣ d w ∣/ w 2 − c 2 , hence the difference or sum of the length legs. At ∣ δ ∣ = δ ∗ |\delta|=\delta_* ∣ δ ∣ = δ ∗ , c = W 1 c=W_1 c = W 1 : the radial minimum is the smaller-radius endpoint, with no interior turning point, and both length expressions agree. The sign of δ \delta δ fixes the angular direction. □ \square □
Degenerate configurations are handled separately: coincident points have d = 0 d=0 d = 0 ; equal phases give d = ∣ ψ ( r 2 ) − ψ ( r 1 ) ∣ d=|\psi(r_2)-\psi(r_1)| d = ∣ ψ ( r 2 ) − ψ ( r 1 ) ∣ ; opposite phases give d = ψ ( r 1 ) + ψ ( r 2 ) d=\psi(r_1)+\psi(r_2) d = ψ ( r 1 ) + ψ ( r 2 ) ; an endpoint at the origin gives d = ψ ( r o t h e r ) d=\psi(r_{\mathrm{other}}) d = ψ ( r other ) by Theorem 2.5 . Neither wrapping at ± π \pm\pi ± π nor exchanging the endpoints produces a second segment.
For stable evaluation, the monotone branch should be integrated over [ W 1 , W 2 ] [W_1,W_2] [ W 1 , W 2 ] rather than formed as a difference of nearly equal legs, and Eq. (34) can be evaluated on the fixed unit interval by the change of variable w ( v ) = c 2 + ( W 2 − c 2 ) v 2 w(v)=\sqrt{c^2+(W^2-c^2)v^2} w ( v ) = c 2 + ( W 2 − c 2 ) v 2 :
L ( c , W ) = W 2 − c 2 ∫ 0 1 h ( w ( v ) ) d v , (42) L(c,W)=\sqrt{W^2-c^2}\int_0^1h(w(v))\,dv, \tag*{(42)} L ( c , W ) = W 2 − c 2 ∫ 0 1 h ( w ( v )) d v , (42)
which also has the correct limit at c = 0 c=0 c = 0 .
Logarithms and parallel transport
At a non-zero base point p p p use the orthonormal frame e s = ∂ s e_s=\partial_s e s = ∂ s , e θ = f − 1 ∂ θ e_\theta=f^{-1}\partial_\theta e θ = f − 1 ∂ θ , and let W p = w ( r p ) W_p=w(r_p) W p = w ( r p ) and d d d be the distance of Proposition 6.4. Writing c c c for the solved Clairaut constant and ε = sign δ \varepsilon=\operatorname{sign}\delta ε = sign δ , the logarithm is
Log p ( q ) = d ( σ 1 − c 2 W p 2 e s + ε c W p e θ ) (43) \operatorname{Log}_p(q)=d\left(\sigma\sqrt{1-\frac{c^2}{W_p^2}}\,e_s+\varepsilon\frac{c}{W_p}\,e_\theta\right) \tag*{(43)} Log p ( q ) = d ( σ 1 − W p 2 c 2 e s + ε W p c e θ ) (43)
where σ = + 1 \sigma=+1 σ = + 1 if the radius initially increases and σ = − 1 \sigma=-1 σ = − 1 if it decreases; a turning segment starts inward, and a monotone segment uses sign ( r q − r p ) \operatorname{sign}(r_q-r_p) sign ( r q − r p ) . At c = W p c=W_p c = W p the radial component vanishes. At p = 0 p=0 p = 0 the Cartesian frame 2 − 1 / 2 ∂ x 1 , 2 − 1 / 2 ∂ x 2 2^{-1/2}\partial_{x_1},2^{-1/2}\partial_{x_2} 2 − 1/2 ∂ x 1 , 2 − 1/2 ∂ x 2 is used instead, with components ψ ( r q ) ( cos θ q , sin θ q ) \psi(r_q)(\cos\theta_q,\sin\theta_q) ψ ( r q ) ( cos θ q , sin θ q ) , and coincident points have zero logarithm.
The Cartesian initial velocity is recovered from Eq. (43) by writing u s , u θ u_s,u_\theta u s , u θ for its orthonormal components and using
v = u s a ( r p ) ( cos θ p , sin θ p ) + u θ 1 − r p 2 2 ( − sin θ p , cos θ p ) . v=\frac{u_s}{\sqrt{a(r_p)}}(\cos\theta_p,\sin\theta_p)+\frac{u_\theta\sqrt{1-r_p^2}}{\sqrt{2}}(-\sin\theta_p,\cos\theta_p). v = a ( r p ) u s ( cos θ p , sin θ p ) + 2 u θ 1 − r p 2 ( − sin θ p , cos θ p ) .
Integrating Eq. (26) with this initial velocity to time one computes Exp p \operatorname{Exp}_p Exp p . Algorithm 1 summarizes the distance and logarithm evaluation.
For parallel transport, the connection identities away from the origin are ∇ e s = f s d θ e θ \nabla e_s=f_s\,d\theta\,e_\theta ∇ e s = f s d θ e θ and ∇ e θ = − f s d θ e s \nabla e_\theta=-f_s\,d\theta\,e_s ∇ e θ = − f s d θ e s , so a transported field V = u e s + v e θ V=ue_s+ve_\theta V = u e s + v e θ solves
u ′ = f s θ ˙ v , v ′ = − f s θ ˙ u (44) u'=f_s\dot{\theta}v,\qquad v'=-f_s\dot{\theta}u \tag*{(44)} u ′ = f s θ ˙ v , v ′ = − f s θ ˙ u (44)
a rotation through − ∫ f s d θ -\int f_s\,d\theta − ∫ f s d θ . At the origin the nonsingular Cartesian equation V ˙ i + Γ j k i x ˙ j V k = 0 \dot{V}^i+\Gamma^i_{jk}\dot{x}^jV^k=0 V ˙ i + Γ j k i x ˙ j V k = 0 is used. Metric compatibility proves that transported tangent inner products are preserved along the curve. The columns of G − 1 / 2 G^{-1/2} G − 1/2 supply an orthonormal frame at a single point; they do not by themselves implement transport between two points separated along a curve, and no such identification is used here.
Numerical error bounds
The following conditional bounds describe a route to validated numerical enclosures. The accompanying implementation returns floating-point estimates with diagnostics; its precision checks do not implement interval certification.
For a continuous non-negative integrand on a finite interval, subdividing the interval and enclosing the integrand on each cell by outward-rounded interval arithmetic bounds the integral by the sum of cell widths times the lower and upper bounds. The transformed legs of Lemma 6.3 have smooth integrands on their closed intervals for c > 0 c>0 c > 0 , and Eq. (42) remains continuous at c = 0 c=0 c = 0 , so refinement gives mathematical convergence with directed arithmetic. Branch decisions require disjoint certified angle intervals; overlapping comparisons must be refined or reported as unresolved.
Algorithm 1: Intrinsic distance and logarithm on the connection disk
Input: normalized endpoints p , q ∈ D p,q\in D p , q ∈ D . Output: intrinsic distance d d d and Cartesian logarithm v = Log p ( q ) v=\operatorname{Log}_p(q) v = Log p ( q ) . For z ≠ 0 z\ne0 z = 0 , let e ( z ) = ( Re z , Im z ) / ∣ z ∣ e(z)=(\operatorname{Re}z,\operatorname{Im}z)/|z| e ( z ) = ( Re z , Im z ) /∣ z ∣ .
Reject non-finite endpoints or ∣ p ∣ ≥ 1 |p|\ge1 ∣ p ∣ ≥ 1 or ∣ q ∣ ≥ 1 |q|\ge1 ∣ q ∣ ≥ 1 . If p = q p=q p = q , return ( d , v ) = ( 0 , 0 ) (d,v)=(0,0) ( d , v ) = ( 0 , 0 ) . Set r p = ∣ p ∣ r_p=|p| r p = ∣ p ∣ and r q = ∣ q ∣ r_q=|q| r q = ∣ q ∣ .
If p = 0 p=0 p = 0 , return d = ψ ( r q ) d=\psi(r_q) d = ψ ( r q ) and v = e ( q ) ψ ( r q ) / 2 v=e(q)\psi(r_q)/\sqrt{2} v = e ( q ) ψ ( r q ) / 2 . If q = 0 q=0 q = 0 , return d = ψ ( r p ) d=\psi(r_p) d = ψ ( r p ) and v = − d e ( p ) / a ( r p ) v=-d\,e(p)/\sqrt{a(r_p)} v = − d e ( p ) / a ( r p ) .
Set δ = Arg ( q / p ) ∈ ( − π , π ] \delta=\operatorname{Arg}(q/p)\in(-\pi,\pi] δ = Arg ( q / p ) ∈ ( − π , π ] , α = ∣ δ ∣ \alpha=|\delta| α = ∣ δ ∣ , e r = e ( p ) e_r=e(p) e r = e ( p ) and e θ = ( − e r , 2 , e r , 1 ) e_\theta=(-e_{r,2},e_{r,1}) e θ = ( − e r , 2 , e r , 1 ) .
If α = 0 \alpha=0 α = 0 , set s = ψ ( r q ) − ψ ( r p ) s=\psi(r_q)-\psi(r_p) s = ψ ( r q ) − ψ ( r p ) and return d = ∣ s ∣ d=|s| d = ∣ s ∣ , v = s e r / a ( r p ) v=s\,e_r/\sqrt{a(r_p)} v = s e r / a ( r p ) . If α = π \alpha=\pi α = π , return d = ψ ( r p ) + ψ ( r q ) d=\psi(r_p)+\psi(r_q) d = ψ ( r p ) + ψ ( r q ) and v = − d e r / a ( r p ) v=-d\,e_r/\sqrt{a(r_p)} v = − d e r / a ( r p ) .
Set W p = w ( r p ) W_p=w(r_p) W p = w ( r p ) , W q = w ( r q ) W_q=w(r_q) W q = w ( r q ) , W s = min ( W p , W q ) W_s=\min(W_p,W_q) W s = min ( W p , W q ) , W l = max ( W p , W q ) W_l=\max(W_p,W_q) W l = max ( W p , W q ) and α ∗ = A ( W s , W l ) \alpha_* = A(W_s,W_l) α ∗ = A ( W s , W l ) .
Select the monotone branch if α < α ∗ \alpha<\alpha_* α < α ∗ and the turning branch otherwise. At equality set c = W s c=W_s c = W s . Otherwise solve on the full bracket [ 0 , W s ] [0,W_s] [ 0 , W s ] : A ( c , W l ) − A ( c , W s ) = α A(c,W_l)-A(c,W_s)=\alpha A ( c , W l ) − A ( c , W s ) = α for the monotone branch, or A ( c , W l ) + A ( c , W s ) = α A(c,W_l)+A(c,W_s)=\alpha A ( c , W l ) + A ( c , W s ) = α for the turning branch.
On the monotone branch set d = L ( c , W l ) − L ( c , W s ) d=L(c,W_l)-L(c,W_s) d = L ( c , W l ) − L ( c , W s ) and σ = sign ( r q − r p ) \sigma=\operatorname{sign}(r_q-r_p) σ = sign ( r q − r p ) . On the turning branch set d = L ( c , W l ) + L ( c , W s ) d=L(c,W_l)+L(c,W_s) d = L ( c , W l ) + L ( c , W s ) and σ = − 1 \sigma=-1 σ = − 1 .
Form u s = σ d 1 − c 2 / W p 2 u_s=\sigma d\sqrt{1-c^2/W_p^2} u s = σ d 1 − c 2 / W p 2 and u θ = sign ( δ ) c / W p u_\theta=\operatorname{sign}(\delta)c/W_p u θ = sign ( δ ) c / W p . Return v = u s e r / a ( r p ) + u θ 1 − r p 2 e θ / 2 v=u_se_r/\sqrt{a(r_p)}+u_\theta\sqrt{1-r_p^2}\,e_\theta/\sqrt{2} v = u s e r / a ( r p ) + u θ 1 − r p 2 e θ / 2 together with d d d .
Numerical evaluation. The branch inequalities above specify exact arithmetic. Evaluate the legs and radial limits stably, and increase precision for close endpoints, small radii or unresolved branch comparisons. Resolve c c c and the logarithm components before rounding to machine precision. Return an available result when a bracket, finite value or requested tolerance cannot be resolved. Precision refinement is distinct from the interval enclosures discussed in § 6 .
A complementary a-posteriori route uses a validated enclosure of the Cartesian flow. If q ^ = Exp p ( v ^ ) \hat{q}=\operatorname{Exp}_p(\hat{v}) q ^ = Exp p ( v ^ ) is enclosed and d ( q ^ , q ) ≤ Δ d(\hat{q},q)\le\Delta d ( q ^ , q ) ≤ Δ is certified, then
∥ v ^ − Log p ( q ) ∥ p ≤ Δ , (45) \lVert\hat{v}-\operatorname{Log}_p(q)\rVert_p\le\Delta, \tag*{(45)} ∥ v ^ − Log p ( q ) ∥ p ≤ Δ , (45)
because on a Hadamard manifold the exponential expands distances and its inverse is one-Lipschitz. If both the enclosure and q q q lie in a Euclidean disk of radius R < 1 R<1 R < 1 with Euclidean separation at most ϵ \epsilon ϵ , the straight segment gives Δ ≤ a ( R ) ϵ \Delta\le\sqrt{a(R)}\,\epsilon Δ ≤ a ( R ) ϵ .
For a finite training sample set F m ( p ) = ( 2 m ) − 1 ∑ i d ( p , q i ) 2 F_m(p)=(2m)^{-1}\sum_i d(p,q_i)^2 F m ( p ) = ( 2 m ) − 1 ∑ i d ( p , q i ) 2 . On this surface ∇ F m ≃ I \nabla F_m\simeq I ∇ F m ≃ I and ∇ F m = − m − 1 ∑ i Log p ( q i ) \nabla F_m=-m^{-1}\sum_i\operatorname{Log}_p(q_i) ∇ F m = − m − 1 ∑ i Log p ( q i ) by the proof of Theorem 3.1 . If a certified gradient enclosure proves ∥ ∇ F m ( p ) ∥ ≤ ϵ \lVert\nabla F_m(p)\rVert\le\epsilon ∥ ∇ F m ( p )∥ ≤ ϵ , strong monotonicity along the segment from p p p to the empirical minimizer gives d ( p , μ ^ m ) ≤ ϵ d(p,\hat{\mu}_m)\le\epsilon d ( p , μ ^ m ) ≤ ϵ . A centre certificate alone is not a score certificate: errors in transport, log components, empirical covariance, eigenvalue classification and rank decisions must also be propagated.
Estimation, calibration and independent-replicate inference
Every statement in this section refers to the fixed-complement reference law P P P of § 3 , and uses the geometry proved in Theorem 2.5 . Throughout, X 1 , … , X m X_1,\ldots,X_m X 1 , … , X m are independent with law P P P , and μ ^ m \hat{\mu}_m μ ^ m denotes their exact empirical intrinsic minimizer.
Consistency
Theorem 7.1 (Consistency of the empirical centre). Under Eq. (11) , the empirical minimizer exists and is unique almost surely, and μ ^ m → μ \hat{\mu}_m \to\mu μ ^ m → μ almost surely.
Proof. Existence and uniqueness are Theorem 3.1 applied to the empirical measure. Write F m , 2 ( y ) = m − 1 ∑ i d ( y , X i ) 2 F_{m,2}(y)=m^{-1}\sum_i d(y,X_i)^2 F m , 2 ( y ) = m − 1 ∑ i d ( y , X i ) 2 , so that the population half-loss is F F F and F m , 2 → 2 F F_{m,2}\to2F F m , 2 → 2 F pointwise by the strong law.
Localization. Minimality of μ ^ m \hat{\mu}_m μ ^ m together with the L 2 L^2 L 2 triangle inequality gives d ( o , μ ^ m ) ≤ 2 ( F m , 2 ( o ) ) 1 / 2 d(o,\hat{\mu}_m)\le2(F_{m,2}(o))^{1/2} d ( o , μ ^ m ) ≤ 2 ( F m , 2 ( o ) ) 1/2 , and the right side is bounded eventually almost surely by the strong law. Choose the deterministic radius R 0 > max { d ( o , μ ) , 2 ( E d ( o , X ) 2 ) 1 / 2 } R_0>\max\{d(o,\mu),2(\mathbb{E}d(o,X)^2)^{1/2}\} R 0 > max { d ( o , μ ) , 2 ( E d ( o , X ) 2 ) 1/2 } . The empirical minimizers eventually lie in this ball almost surely.
Uniform law on that ball. For a , b a,b a , b in the ball,
∣ d ( a , X ) 2 − d ( b , X ) 2 ∣ ≤ d ( a , b ) ( 2 d ( o , X ) + 2 R 0 ) \lvert d(a,X)^2-d(b,X)^2\rvert\le d(a,b)\left(2d(o,X)+2R_0\right) ∣ d ( a , X ) 2 − d ( b , X ) 2 ∣ ≤ d ( a , b ) ( 2 d ( o , X ) + 2 R 0 ) by the triangle inequality, and the right side is P P P -integrable. A finite net, the strong law at the net points, and then a decreasing net mesh give sup d ( o , y ) ≤ R 0 ∣ F m , 2 ( y ) − 2 F ( y ) ∣ → 0 \sup_{d(o,y)\le R_0}\lvert F_{m,2}(y)-2F(y)\rvert\to0 sup d ( o , y ) ≤ R 0 ∣ F m , 2 ( y ) − 2 F ( y )∣ → 0 almost surely.
Identification of the limit. Any subsequential limit of μ ^ m \hat{\mu}_m μ ^ m is a minimizer of F F F by the uniform law, and Theorem 3.1 makes that minimizer unique, so every convergent subsequence has limit μ \mu μ . Compactness of the localization ball then gives convergence of the full sequence.
The exact empirical centre satisfies the sample mean-logarithm equation by the finite-sum first variation. A stopped numerical centre does not: it requires its own error certificate, and a small reported gradient is not equality.
Covariance consistency follows by an explicit plug-in argument, and the estimated centre is handled by continuity rather than left implicit. Fix a smooth orthonormal frame near μ \mu μ and let v ( y , x ) v(y,x) v ( y , x ) denote log coordinates in that frame. Assume continuity of y ↦ v ( y , x ) y\mapsto v(y,x) y ↦ v ( y , x ) for almost every x x x and an envelope B B B with sup y ∈ K ∥ v ( y , x ) ∥ 2 ≤ B ( x ) \sup_{y\in K}\lVert v(y,x)\rVert^2\le B(x) sup y ∈ K ∥ v ( y , x ) ∥ 2 ≤ B ( x ) and E B ( X ) < ∞ \mathbb{E}B(X)<\infty E B ( X ) < ∞ on a compact neighbourhood K K K of μ \mu μ . These conditions hold here: smoothness of the global logarithm gives continuity, and for K ⊂ { y : d ( y , μ ) ≤ R } K \subset\{y : d(y,\mu) \le R\} K ⊂ { y : d ( y , μ ) ≤ R } the envelope B ( x ) = 2 d ( μ , x ) 2 + 2 R 2 B(x) = 2d(\mu,x)^2 + 2R^2 B ( x ) = 2 d ( μ , x ) 2 + 2 R 2 is integrable by Eq. (11) .
Define the population second moment in the same frame,
Σ ( y ) = E P v ( y , X ) v ( y , X ) T , (46) \Sigma(y) = \mathbb{E}_{P}v(y,X)v(y,X)^{\mathsf{T}}, \tag*{(46)} Σ ( y ) = E P v ( y , X ) v ( y , X ) T , (46)
with Σ = Σ ( μ ) \Sigma= \Sigma(\mu) Σ = Σ ( μ ) . For y k → y y_k \to y y k → y the integrands converge pointwise by continuity and are dominated by B B B , so dominated convergence gives Σ ( y k ) → Σ ( y ) \Sigma(y_k) \to\Sigma(y) Σ ( y k ) → Σ ( y ) : the map Σ ( ⋅ ) \Sigma(\cdot) Σ ( ⋅ ) is continuous on K K K , uniformly in no further sense being needed. For m ≥ 2 m \ge2 m ≥ 2 , define the empirical covariance
Σ ^ m = 1 m − 1 ∑ i = 1 m ( v ^ i − v ˉ m ) ( v ^ i − v ˉ m ) T , v ^ i = v ( μ ^ m , X i ) , (47) \widehat{\Sigma}_m = \frac{1}{m-1}\sum_{i=1}^{m}(\hat{v}_i-\bar{v}_m)(\hat{v}_i-\bar{v}_m)^{\mathsf{T}}, \qquad\hat{v}_i = v(\hat{\mu}_m,X_i), \tag*{(47)} Σ m = m − 1 1 i = 1 ∑ m ( v ^ i − v ˉ m ) ( v ^ i − v ˉ m ) T , v ^ i = v ( μ ^ m , X i ) , (47)
put f y ( x ) = v ( y , x ) v ( y , x ) T f_y(x) = v(y,x)v(y,x)^{\mathsf{T}} f y ( x ) = v ( y , x ) v ( y , x ) T and M ϵ ( x ) = sup d ( y , μ ) ≤ ϵ ∥ f y ( x ) − f μ ( x ) ∥ M_{\epsilon}(x) = \sup_{d(y,\mu)\le\epsilon}\lVert f_y(x)-f_\mu(x)\rVert M ϵ ( x ) = sup d ( y , μ ) ≤ ϵ ∥ f y ( x ) − f μ ( x )∥ , restricting the supremum to K K K . Continuity gives M ϵ ( x ) → 0 M_{\epsilon}(x) \to0 M ϵ ( x ) → 0 pointwise, while M ϵ ≤ 2 B M_{\epsilon} \le2B M ϵ ≤ 2 B gives E M ϵ ( X ) → 0 \mathbb{E}M_{\epsilon}(X) \to0 E M ϵ ( X ) → 0 by dominated convergence. For each fixed ϵ \epsilon ϵ , once μ ^ m \hat{\mu}_m μ ^ m enters this ball,
∥ 1 m ∑ i f μ ^ m ( X i ) − 1 m ∑ i f μ ( X i ) ∥ ≤ 1 m ∑ i M ϵ ( X i ) . \left\lVert\frac{1}{m}\sum_i f_{\hat{\mu}_m}(X_i)-\frac{1}{m}\sum_i f_\mu(X_i)\right\rVert\le\frac{1}{m}\sum_i M_{\epsilon}(X_i). m 1 i ∑ f μ ^ m ( X i ) − m 1 i ∑ f μ ( X i ) ≤ m 1 i ∑ M ϵ ( X i ) .
The strong law on a countable sequence ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 makes the left side converge almost surely to zero. The supremum is measurable because it equals the supremum on a countable dense subset of the compact parameter ball. The fixed-centre second moment converges to Σ \Sigma Σ by the strong law. At the exact empirical centre, v ˉ m = 0 \bar{v}_m = 0 v ˉ m = 0 by the first-order equation, so the factor m / ( m − 1 ) → 1 m/(m-1) \to1 m / ( m − 1 ) → 1 yields Σ ^ m → Σ \widehat{\Sigma}_m \to\Sigma Σ m → Σ almost surely. Fitting the centre on the same data does not imply that this estimator is unbiased.
A central limit theorem on the induced surface
The following is the chart estimating-equation argument of [9 , 10 ] , specialized to this surface with its derivative conditions verified rather than assumed.
Lemma 7.2 (Hessian comparison on the induced disk). Let t = d ( p , x ) t = d(p,x) t = d ( p , x ) and f x ( p ) = 1 2 d ( p , x ) 2 f_x(p) = \frac{1}{2}d(p,x)^2 f x ( p ) = 2 1 d ( p , x ) 2 . In an orthonormal tangent frame with the first axis along the geodesic from p p p to x x x ,
I ⪯ Hess f x ( p ) ⪯ ( 1 + t 2 ) I . (48) I \preceq\operatorname{Hess} f_x(p) \preceq\left(1+\frac{t}{2}\right)I. \tag*{(48)} I ⪯ Hess f x ( p ) ⪯ ( 1 + 2 t ) I . (48) Proof. The radial eigenvalue is exactly one, because along the radial geodesic f x = r 2 / 2 f_x = r^2/2 f x = r 2 /2 . For the transverse eigenvalue, let J J J be the Jacobi field along the unit-speed segment from x x x to p p p with J ( 0 ) = 0 J(0) = 0 J ( 0 ) = 0 and J ( t ) = v J(t) = v J ( t ) = v , ∣ v ∣ = 1 \lvert v\rvert= 1 ∣ v ∣ = 1 . On a surface,
Hess f x ( v , v ) = t ∫ 0 t ( ∣ J ′ ∣ 2 − K J 2 ) d s , \operatorname{Hess} f_x(v,v) = t\int_0^t(\lvert J'\rvert^2-KJ^2)\,\mathrm{d}s, Hess f x ( v , v ) = t ∫ 0 t (∣ J ′ ∣ 2 − K J 2 ) d s , the index form of J J J , and J J J minimizes it because there are no conjugate points. Since − 1 / 4 ≤ K ≤ 0 -1/4 \le K \le0 − 1/4 ≤ K ≤ 0 , comparison with the spaces of constant curvature − 1 / 4 -1/4 − 1/4 and 0 0 0 gives
1 ≤ Hess f x ( v , v ) ≤ t 2 coth t 2 . 1 \le\operatorname{Hess} f_x(v,v) \le\frac{t}{2}\coth\frac{t}{2}. 1 ≤ Hess f x ( v , v ) ≤ 2 t coth 2 t . The inequality u coth u ≤ 1 + u u\coth u \le1+u u coth u ≤ 1 + u for u > 0 u>0 u > 0 (from coth u = 1 + 2 / ( e 2 u − 1 ) ≤ 1 + 1 / u \coth u = 1 + 2/(e^{2u}-1) \le1+1/u coth u = 1 + 2/ ( e 2 u − 1 ) ≤ 1 + 1/ u ) gives ( t / 2 ) coth ( t / 2 ) ≤ 1 + t / 2 (t/2)\coth(t/2) \le1+t/2 ( t /2 ) coth ( t /2 ) ≤ 1 + t /2 . Radial and transverse cross terms vanish by the first variation, so the bound holds in every direction. At t = 0 t=0 t = 0 the Hessian is the identity and the bound is attained.
Theorem 7.3 (Sandwich central limit theorem). Take normal coordinates θ \theta θ at μ \mu μ , so θ 0 = 0 \theta_0=0 θ 0 = 0 , and let ℓ ( θ , x ) = d ( exp μ θ , x ) 2 / 2 \ell(\theta,x)=d(\exp_\mu\theta,x)^2/2 ℓ ( θ , x ) = d ( exp μ θ , x ) 2 /2 with G = ∂ θ ℓ G=\partial_\theta\ell G = ∂ θ ℓ and H = ∂ θ G H=\partial_\theta G H = ∂ θ G . Assume that exact consistent empirical minimizers lie eventually in the chart interior, that ℓ \ell ℓ is twice continuously differentiable there for almost every x x x , that E ∥ G ( 0 , X ) ∥ 2 < ∞ \mathbb{E}\lVert G(0,X)\rVert^2<\infty E ∥ G ( 0 , X ) ∥ 2 < ∞ and E ∥ H ( 0 , X ) ∥ < ∞ \mathbb{E}\lVert H(0,X)\rVert<\infty E ∥ H ( 0 , X )∥ < ∞ , that
E sup ∥ θ ∥ ≤ ϵ ∥ H ( θ , X ) − H ( 0 , X ) ∥ ⟶ 0 ( ϵ ↓ 0 ) , (49) \mathbb{E}\sup_{\lVert\theta\rVert\le\epsilon}\lVert H(\theta,X)-H(0,X)\rVert\longrightarrow0 \qquad(\epsilon\downarrow0), \tag*{(49)} E ∥ θ ∥ ≤ ϵ sup ∥ H ( θ , X ) − H ( 0 , X )∥ ⟶ 0 ( ϵ ↓ 0 ) , (49) and that A = E H ( 0 , X ) A=\mathbb{E}H(0,X) A = E H ( 0 , X ) is nonsingular. Then
m θ ^ m ⟹ N ( 0 , A − 1 Σ A − T ) , (50) \sqrt{m}\hat{\theta}_m \Longrightarrow N(0,A^{-1}\Sigma A^{-\mathsf{T}}), \tag*{(50)} m θ ^ m ⟹ N ( 0 , A − 1 Σ A − T ) , (50) Σ = E G ( 0 , X ) G ( 0 , X ) T . \Sigma= \mathbb{E}G(0,X)G(0,X)^{\mathsf{T}}. Σ = E G ( 0 , X ) G ( 0 , X ) T . Proof. Taylor expansion in integral form about θ = 0 \theta=0 θ = 0 gives, at the empirical minimizer,
0 = 1 m ∑ i = 1 m G ( 0 , X i ) + A m m θ ^ m , 0 = \frac{1}{\sqrt{m}}\sum_{i=1}^{m}G(0,X_i)+A_m\sqrt{m}\hat{\theta}_m, 0 = m 1 i = 1 ∑ m G ( 0 , X i ) + A m m θ ^ m , A m = 1 m ∑ i = 1 m ∫ 0 1 H ( t θ ^ m , X i ) d t . A_m = \frac{1}{m}\sum_{i=1}^{m}\int_0^1 H(t\hat{\theta}_m,X_i)\,\mathrm{d}t. A m = m 1 i = 1 ∑ m ∫ 0 1 H ( t θ ^ m , X i ) d t . To handle the dependence of θ ^ m \hat{\theta}_m θ ^ m on the same sample, put B ϵ ( x ) = sup ∥ θ ∥ ≤ ϵ ∥ H ( θ , x ) − H ( 0 , x ) ∥ B_{\epsilon}(x)=\sup_{\lVert\theta\rVert\le\epsilon}\lVert H(\theta,x)-H(0,x)\rVert B ϵ ( x ) = sup ∥ θ ∥ ≤ ϵ ∥ H ( θ , x ) − H ( 0 , x )∥ . On { ∥ θ ^ m ∥ ≤ ϵ } \{\lVert\hat{\theta}_m\rVert\le\epsilon\} {∥ θ ^ m ∥ ≤ ϵ } , ∥ A m − m − 1 ∑ i H ( 0 , X i ) ∥ ≤ m − 1 ∑ i B ϵ ( X i ) \lVert A_m-m^{-1}\sum_iH(0,X_i)\rVert\le m^{-1}\sum_iB_{\epsilon}(X_i) ∥ A m − m − 1 ∑ i H ( 0 , X i )∥ ≤ m − 1 ∑ i B ϵ ( X i ) . For any η > 0 \eta>0 η > 0 , Markov’s inequality bounds the limsup probability that this norm exceeds η \eta η by E B ϵ ( X ) / η \mathbb{E}B_{\epsilon}(X)/\eta E B ϵ ( X ) / η , since consistency makes the complementary event vanish. Letting ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 and using Eq. (49) , followed by the strong law for H ( 0 , X ) H(0,X) H ( 0 , X ) , proves A m → A A_m\to A A m → A in probability. At θ = 0 \theta=0 θ = 0 the gradient is G ( 0 , X ) = − Log μ X G(0,X)=-\operatorname{Log}_{\mu}X G ( 0 , X ) = − Log μ X by Theorem 3.1 , so E G ( 0 , X ) = 0 \mathbb{E}G(0,X)=0 E G ( 0 , X ) = 0 and Cov G ( 0 , X ) = Σ \operatorname{Cov}G(0,X)=\Sigma Cov G ( 0 , X ) = Σ . The multivariate central limit theorem for independent identically distributed summands gives m − 1 / 2 ∑ i G ( 0 , X i ) ⟹ N ( 0 , Σ ) m^{-1/2}\sum_iG(0,X_i)\Longrightarrow N(0,\Sigma) m − 1/2 ∑ i G ( 0 , X i ) ⟹ N ( 0 , Σ ) , and Slutsky’s theorem with A m → A A_m\to A A m → A gives Eq. (50) .
Using d 2 d^2 d 2 instead of d 2 / 2 d^2/2 d 2 /2 multiplies G G G and Σ \Sigma Σ by four, leaving the sandwich unchanged. The limit is not Σ / m \Sigma/m Σ/ m in general: curvature enters through A A A . Lemma 7.2 supplies Eq. (49) for this surface, because on a compact normal-coordinate ball the chart Hessian is bounded by C ( 1 + d ( μ , x ) ) C(1+d(\mu,x)) C ( 1 + d ( μ , x )) with C C C depending only on the ball and E ( d 2 / 2 ) < ∞ \mathbb{E}(d^2/2)<\infty E ( d 2 /2 ) < ∞ ; see Appendix § 13 . At the centre A ⪰ I A\succeq I A ⪰ I , so A A A is finite and nonsingular. On the real arc, A = I A=I A = I exactly.
No conclusion about individual residuals follows. Equation (50) describes the sampling distribution of an estimated centre; it does not make an individual observation Gaussian, and it does not make the joint score of Remark 3.4 chi-square.
Corollary 7.4 (Accuracy required of a numerical centre). Let F m ( y ) = ( 2 m ) − 1 ∑ i d ( y , X i ) 2 F_m(y)=(2m)^{-1}\sum_i d(y,X_i)^2 F m ( y ) = ( 2 m ) − 1 ∑ i d ( y , X i ) 2 . If a computed centre
μ ~ m \tilde{\mu}_m μ ~ m satisfies an exact residual bound ∥ ∇ F m ( μ ~ m ) ∥ ≤ ϵ m \lVert\nabla F_m(\tilde{\mu}_m)\rVert\le\epsilon_m ∥ ∇ F m ( μ ~ m )∥ ≤ ϵ m , then d ( μ ~ m , μ ^ m ) ≤ ϵ m d(\tilde{\mu}_m,\hat{\mu}_m) \le\epsilon_m d ( μ ~ m , μ ^ m ) ≤ ϵ m . Thus ϵ m → 0 \epsilon_m \to0 ϵ m → 0 almost surely preserves strong consistency, and m ϵ m → 0 \sqrt{m}\epsilon_m \to0 m ϵ m → 0 in probability preserves the limit in Eq. (50) .
Proof. Strong convexity gives ∥ ∇ F m ( y ) ∥ ≥ d ( y , μ ^ m ) \lVert\nabla F_m(y)\rVert\ge d(y,\hat{\mu}_m) ∥ ∇ F m ( y )∥ ≥ d ( y , μ ^ m ) by integrating its Hessian along the geodesic from the minimizer. The logarithm at μ \mu μ is one-Lipschitz on a Hadamard manifold, so the normal-coordinate difference between the computed and exact centres is at most ϵ m \epsilon_m ϵ m . The conclusions follow from consistency and Slutsky’s theorem. □
A fixed floating-point stopping threshold is a finite-sample diagnostic. Applying this corollary requires a bound on the exact residual, including errors in the computed logarithms.
Frozen scores and the finite-sample rank
In practice a score is a frozen numerical procedure fitted on training data, not an ideal geometric quantity. The following theorem is stated for such a procedure and needs only measurability and exchangeability.
Let T \mathcal{T} T denote the training data together with every deterministic choice made from them: the centre, the frame, the covariance, the spectral thresholds, the retained projector P ^ \hat{P} P ^ and the pseudoinverse Σ ^ + \hat{\Sigma}^{+} Σ ^ + . These define one frozen measurable function s T s_{\mathcal{T}} s T on the slice.
For example, fix an eigenvalue threshold from training and retain the corresponding eigenspace. For a new tangent vector v v v , report v ⊤ Σ ^ + v v^{\top}\hat{\Sigma}^{+}v v ⊤ Σ ^ + v on that space and the residual ∥ ( I − P ^ ) v ∥ \lVert(I-\hat{P})v\rVert ∥( I − P ^ ) v ∥ separately. A scalar calibration rule may assign + ∞ +\infty + ∞ when this residual exceeds a tolerance fixed at training time. Use the identical rule for each calibration and test point. At rank zero only the support-departure rule has a meaning. A failed logarithm or centre fit supplies no score.
Theorem 7.5 (Finite calibration rank). Conditional on ( T , c ) (\mathcal{T},c) ( T , c ) , where c c c denotes the declared complement, assume that n c a l n_{\mathrm{cal}} n cal calibration observations and one null test observation are exchangeable with the same law, and that the frozen measurable score is defined for all of them. Write s i s_i s i for the calibration scores and s ∗ s_* s ∗ for the score of the null test observation. Then, with N c = n c a l + 1 N_c=n_{\mathrm{cal}}+1 N c = n cal + 1 ,
p = 1 + ∑ i = 1 n c a l 1 { s i ≥ s ∗ } N c (51) p=\frac{1+\sum_{i=1}^{n_{\mathrm{cal}}}\mathbf{1}\{s_i\ge s_*\}}{N_c} \tag*{(51)} p = N c 1 + ∑ i = 1 n cal 1 { s i ≥ s ∗ } (51) is super-uniform:
Pr ( p ≤ a ∣ T , c ) ≤ ⌊ a N c ⌋ N c ≤ a , 0 ≤ a ≤ 1. (52) \Pr(p\le a\mid\mathcal{T},c)\le\frac{\lfloor aN_c\rfloor}{N_c}\le a,\qquad0\le a\le1. \tag*{(52)} Pr ( p ≤ a ∣ T , c ) ≤ N c ⌊ a N c ⌋ ≤ a , 0 ≤ a ≤ 1. (52)
Proof. For the multiset of N c N_c N c scores put R j = # { i : s i ≥ s j } R_j=\#\{i:s_i\ge s_j\} R j = # { i : s i ≥ s j } and arrange the tie blocks in descending order, including a top block of + ∞ +\infty + ∞ values when present. In a block of size b b b with h h h strictly higher observations, every member has R j = h + b R_j=h+b R j = h + b . Hence among all indices at most k k k satisfy R j ≤ k R_j\le k R j ≤ k : such blocks can only end within the first k k k positions of the ordering. Exchangeability makes the test index uniform over the N c N_c N c positions conditional on the unordered score multiset, so Pr ( R ∗ ≤ k ∣ T , c ) ≤ k / N c \Pr(R_*\le k\mid\mathcal{T},c)\le k/N_c Pr ( R ∗ ≤ k ∣ T , c ) ≤ k / N c . Since p ≤ a p\le a p ≤ a implies R ∗ ≤ ⌊ a N c ⌋ R_*\le\lfloor aN_c\rfloor R ∗ ≤ ⌊ a N c ⌋ , Eq. (52) follows. Ties are handled by the relation ≥ \ge ≥ , and if all scores equal + ∞ +\infty + ∞ then p = 1 p=1 p = 1 . □
This is the split-rank argument underlying the split-conformal principle[28 ] . The guarantee averages over the calibration sample and the new null, conditional on the fit and the complement; it is not conditional on realized calibration scores. Shared calibration makes many test ranks dependent without spoiling each marginal guarantee. Bonferroni over valid marginal ranks controls family-wise error by the union bound; no automatic Benjamini–Hochberg claim follows.
The theorem applies to any defined common computational function, even one that differs from the ideal geometry. It does not certify numerical accuracy of the centre, logarithm, covariance or quadratic score. Solver tolerances and refinement agreement are separate diagnostics. A computational support infinity is an output definition, whereas an arithmetic failure is unavailable and must not be relabelled as infinity; if the procedure is partial on the reference law, the total-function hypothesis has not been established, and selecting only successful computations cannot restore the unconditional guarantee.
Independent replicates and multiplicity
For independent planned clusters, bounded-variable concentration gives conservative uncertainty intervals even when observations within a cluster are dependent. Holm’s step-down rule controls familywise error for a family of valid marginal equality tests without requiring independence between their ranks[22 , 23 ] . Supplementary Methods S8 gives the bounds and their derivations, including paired comparisons and unavailable outcomes. These guarantees use complete planned clusters as the independent units.
Varying complements
The preceding theory conditions on a fixed complement. When the complement is itself estimated, the normalized connection inherits derivatives from every surrounding entry; Appendix B gives the exact differential. Two consequences matter here. First, a fixed population matrix used for oracle normalization does not make the sample complement fixed, and oracle coordinates can leave the admissible disk, so no conditional slice calibration guarantee is available in that regime. Second, if a matrix estimator satisfies m ( a m − a 0 ) ⇒ N ( 0 , Ω ) \sqrt{m}(a_m-a_0)\Rightarrow N(0,\Omega) m ( a m − a 0 ) ⇒ N ( 0 , Ω ) in a fixed real vectorization, then for any smooth real functional φ \varphi φ of the normalized connection the delta method gives m [ φ ( a m ) − φ ( a 0 ) ] ⇒ N ( 0 , ∇ φ ⊤ Ω ∇ φ ) \sqrt{m}[\varphi(a_m)-\varphi(a_0)]\Rightarrow N(0,\nabla\varphi^{\top}\Omega\nabla\varphi) m [ φ ( a m ) − φ ( a 0 )] ⇒ N ( 0 , ∇ φ ⊤ Ω∇ φ ) , provided the limiting variance is positive and consistently estimated. A standardized normal limit requires those two additional conditions and is not a consequence of smoothness alone.
Numerical illustrations
The numerical examples examine the geodesic formulas and illustrate intrinsic statistics under declared reference laws. The Supplementary Methods give the sampling designs, solver tolerances, replication counts and numerical tables.
Geodesics and curvature
Independent shooting of the Cartesian geodesic equation agrees with the scalar distance reduction over the tested interior range (Figure 5 D and Figure 6 A). The largest relative discrepancy is 3.05 × 10 − 13 3.05 \times10^{-13} 3.05 × 1 0 − 13 . Shooting starts from a Euclidean initial guess and evaluates length through the metric norm of its fitted initial velocity, independently of the scalar formula.
Figure 6. Numerical accuracy and intrinsic centre estimation. (A) Cumulative relative distance error against independent shooting; values below 10 − 17 10^{-17} 1 0 − 17 are placed at that plotting floor. (B) Curvature recovered from geodesic circles approaches the exact values, marked at zero circle radius. (C) RMS centre error with pointwise 95% bootstrap intervals and a fitted logarithmic slope. (D) Empirical distributions of m d ( μ ^ m , 0 ) \sqrt{m}d(\hat{\mu}_m,0) m d ( μ ^ m , 0 ) across sample sizes. Sample counts, solver settings and numerical tables are provided in the Supplementary Methods.
Curvature estimates from geodesic-circle circumferences approach the values of Theorem 2.5 as the circle radius decreases (Figure 6 B). The finite-radius correction is most visible at the origin, where the limiting curvature is zero. Together, these comparisons assess the implementation on an interior domain; the behaviour arbitrarily close to the singular boundary requires the numerical representations discussed in Appendix § 12 .
Intrinsic statistics
For the centre experiment, a rotationally symmetric Gaussian law is mapped to the disk by Y ↦ tanh ( ∥ Y ∥ ) Y / ∥ Y ∥ Y \mapsto\tanh(\lVert Y\rVert)Y/\lVert Y\rVert Y ↦ tanh (∥ Y ∥) Y / ∥ Y ∥ . Its population centre is the origin, and Appendix § 13 verifies the finite second intrinsic moment. The empirical centre error is compatible with the m − 1 / 2 m^{-1/2} m − 1/2 scaling of Theorem 7.3 : the fitted logarithmic slope is − 0.490 -0.490 − 0.490 , with a 95% bootstrap interval [ − 0.511 , − 0.471 ] [-0.511,-0.471] [ − 0.511 , − 0.471 ] (Figure 6 C). Distributions of the scaled errors show the associated finite-sample variation (Figure 6 D).
Figure 3 separates the roles of population geometry and score calibration. A Gaussian tangent law illustrates covariance whitening and rotations that preserve the covariance spectrum. For calibration, a training sample from a uniform disk supplies a fitted centre and covariance, which are then held fixed. Independent calibration and test samples evaluate the resulting quadratic score. Rounding retains ties and illustrates the conservative rank rule under the exchangeable design of Theorem 7.5 . The plotted uncertainty intervals quantify Monte Carlo variation conditional on the training fit.
For a covariance-level comparison, four-point tangent laws give known centres and allow direct evaluation of the population sandwich matrix. Across the three laws and two sample sizes in Supplementary Methods S9, relative covariance discrepancies ranged from 0.065 to 0.164. Coverage of nominal 95% ellipses constructed with the population sandwich ranged from 0.938 to 0.963. These experiments assess the covariance prediction with known population parameters. Practical ellipses estimated from a single training sample introduce additional uncertainty.
Discussion
Fixing the surrounding entries of a Hermitian positive-definite matrix gives an isolated connection a complete, two-dimensional geometry. The fixed Schur congruence identifies its admissible disk and removes the ambient dimension from the induced tangent metric. The variable nonpositive curvature guarantees unique geodesics, global logarithms and a unique reference centre under a finite second intrinsic moment. Conservation of angular momentum then reduces geodesic computation to one scalar equation with elementary angular and length integrals.
The constraint determines the meaning of the distance. Ambient affine-invariant paths can change diagonal entries, whereas quotient paths identify matrices under diagonal rescaling. On the fixed-complement slice, distinct endpoints have a strictly larger intrinsic distance than their ambient distance. Correlation quotients and transformed anomaly scores provide the established background [12 , 13 , 39 , 40 ] . The present construction specifies the path constraint explicitly and supplies its exact metric and geodesic reduction. The real restriction also retains an orientation through its signed arc coordinate. Log-Euclidean and log-Cholesky constructions provide alternative coordinate geometries [2 , 3 , 30 ] whose paths answer different questions.
The statistical interpretation requires a reference distribution on the declared slice. A common normalized geometry across slices does not identify a common probability law. Fréchet means, tangent covariance and the sandwich limit give population summaries and uncertainty for an estimated centre. Covariance whitening fixes the first two tangent moments but leaves the individual score distribution dependent on the reference law. Exchangeable rank calibration provides a finite-sample guarantee for a fitted scoring rule when the training, calibration and test design meets the stated assumptions. The numerical examples illustrate these distinctions with a declared Gaussian tangent law and an independent calibration experiment.
The framework applies directly to controlled connection perturbations with fixed surrounding entries. Covariance estimates from different subjects or samples generally have varying complements. Their normalized coordinates inherit variation from all matrix entries, and their sampling laws require a separate argument. Appendix B gives the relevant differential and a delta-method limit under a matrix-estimator central limit theorem. The slice-preserving invariances likewise depend on the declared frame and on transporting the reference law consistently. Generic variable mixing changes the isolated-entry problem.
Numerical evaluation becomes sensitive near the excluded boundary, where intrinsic distance diverges. Stable arc evaluation and explicit solver status checks are therefore part of the implementation. The reported shooting and curvature checks compare numerical evaluations with the derived formulas over a specified interior range. Extending the representation through log-gap coordinates and validating error enclosures would support computations closer to singular matrices. Applied comparisons of fixed-complement scores with other matrix summaries require a sampling design matched to the intended population.
Acknowledgements
This work was supported in part by the STI 2030–Major Projects (2022ZD0208500); the National Natural Science Foundation of China (W2411084); the Key Research and Development Projects of the Science and Technology Department of Chengdu (2024-YF08-00072-GX); and the Chengdu Science and Technology Bureau Program (2022-GH02-00042-HZ). Additional support was provided by the University of Electronic Science and Technology of China through the CNS Program (Y0301902610100201).
Competing interests
The authors declare no competing interests.
Code availability
The implementation and scripts for reproducing the numerical examples are maintained at https://github.com/ronald1129/sid-hpd-intrinsic . Repository access is available to reviewers on request.
Exact ambient support reduction
The fixed Schur congruence of Theorem 2.2 also gives a computational reduction for the unrestricted ambient comparator.
Proposition A.1 (The ambient comparator lives in a two-by-two block). Let P P P be a fixed-complement law with finite ambient second moment . In the coordinates supplied by the congruence, the unrestricted ambient Fréchet mean of P P P is diag ( M 2 , I n − 2 ) \operatorname{diag}(M_2,I_{n-2}) diag ( M 2 , I n − 2 ) with M 2 M_2 M 2 the HPD(2) ambient mean of the two-by-two blocks. Every whitened ambient logarithm has only a Hermitian two-by-two block, so full ambient tangent covariance standardization is isometrically representable in four block coordinates (three in the real SPD case), independently of n n n . The fixed slice has an additional symmetry: the whitened logarithms actually lie in a linear subspace of dimension at most three in the complex case and at most two in the real case. Both reductions retain the complete ambient tangent vector and its covariance quadratic score.
Proof. Let J = diag ( I 2 , − I n − 2 ) J=\operatorname{diag}(I_2,-I_{n-2}) J = diag ( I 2 , − I n − 2 ) . The map A ↦ J A J A\mapsto JAJ A ↦ J A J is an isometry of the ambient manifold with Eq. (2) and fixes every observation of the slice, since those observations equal diag ( R ( ρ ) , I n − 2 ) \operatorname{diag}(R(\rho),I_{n-2}) diag ( R ( ρ ) , I n − 2 ) in these coordinates. A unique ambient Fréchet mean commutes with any isometry fixing the data, so the mean is block diagonal.
For a block-diagonal candidate diag ( B , D ) \operatorname{diag}(B,D) diag ( B , D ) its squared distance to an observation splits as d A ( B , R ) 2 + d A ( D , I ) 2 d_A(B,R)^2+d_A(D,I)^2 d A ( B , R ) 2 + d A ( D , I ) 2 , because the matrix logarithm of a block diagonal matrix is block diagonal. Minimizing the second term forces D = I D=I D = I , and the first term is minimized by B = M 2 B=M_2 B = M 2 , the HPD(2) ambient mean. The same block formula gives whitened logarithms diag ( log ( M 2 − 1 / 2 R M 2 − 1 / 2 ) , 0 ) \operatorname{diag}(\log(M_2^{-1/2}RM_2^{-1/2}),0) diag ( log ( M 2 − 1/2 R M 2 − 1/2 ) , 0 ) .
For the sharper support bound, write S = ( 0 1 1 0 ) S=\begin{pmatrix}0&1\\1&0\end{pmatrix} S = ( 0 1 1 0 ) and consider the isometry ι ( B ) = S B ˉ S \iota(B)=S\bar{B}S ι ( B ) = S B ˉ S of HPD(2). Every R ( ρ ) R(\rho) R ( ρ ) is fixed by ι \iota ι . Uniqueness of the ambient mean gives ι ( M 2 ) = M 2 \iota(M_2)=M_2 ι ( M 2 ) = M 2 . This involution commutes with products and real spectral matrix functions; hence each H ρ = log ( M 2 − 1 / 2 R ( ρ ) M 2 − 1 / 2 ) H_\rho=\log(M_2^{-1/2}R(\rho)M_2^{-1/2}) H ρ = log ( M 2 − 1/2 R ( ρ ) M 2 − 1/2 ) is fixed by ι \iota ι as well. A Hermitian fixed point has the form H ρ = ( t u + i v u − i v t ) H_\rho=\begin{pmatrix}t&u+iv\\u-iv&t\end{pmatrix} H ρ = ( t u − i v u + i v t ) , with t , u , v ∈ R t,u,v\in\mathbb{R} t , u , v ∈ R . Its Frobenius norm is the Euclidean norm of 2 ( t , u , v ) \sqrt{2}(t,u,v) 2 ( t , u , v ) , proving the three-coordinate representation. In the real case v = 0 v=0 v = 0 , giving two coordinates. The actual covariance rank may be smaller, depending on the reference law.
Hermitian vectorization preserves Frobenius inner products and therefore covariance eigenvalues, pseudo-inverse quadratic scores and the support-departure norm. Under the original congruence, if M ′ = T M T ∗ M'=TMT^* M ′ = T M T ∗ then O = M ′ − 1 / 2 T M 1 / 2 O=M'^{-1/2}TM^{1/2} O = M ′ − 1/2 T M 1/2 is unitary and whitened logarithms transform by H ↦ O H O ∗ H\mapsto OHO^* H ↦ O H O ∗ , which is an orthogonal change of whitened coordinates. The reduction therefore preserves the original full ambient comparator. □
This reduction does not replace whole-matrix scores by an edge projection: it is valid only because every observation in a replication shares one fixed complement. It does not apply to the varying-complement regime of Appendix § 11 , and it cannot manufacture full-rank population covariance from a singular empirical covariance. It also does not imply that the HPD(2) ambient mean has unit diagonals, as the two-point counterexample of § 4 shows. Figure A1 records the retained-coordinate counts before and after the congruence.
Figure A1. Block storage and ambient logarithm support are distinct. (A) Real coordinate counts for a Hermitian matrix of order three and its selected block. (B) Conjugate-swap symmetry confines the whitened ambient logarithms to at most three coordinates in the complex case and two in the real case, independently of matrix dimension. These support bounds refine the four- and three-coordinate block storage representations of Proposition A.1.
Differentiating the normalization when the complement varies
Geometric conditioning and covariance-estimator sampling are different regimes, and § 7 uses the transition between them. For an interior HPD matrix let H = E C − 1 E ∗ H=EC^{-1}E^* H = E C − 1 E ∗ and let u , v u,v u , v denote the first and second rows of E E E . A Hermitian variation gives, by d C − 1 = − C − 1 ( d C ) C − 1 dC^{-1}=-C^{-1}(dC)C^{-1} d C − 1 = − C − 1 ( d C ) C − 1 and the product rule,
d H = ( d E ) C − 1 E ∗ + E C − 1 d E ∗ − E C − 1 ( d C ) C − 1 E ∗ , (53) dH=(dE)C^{-1}E^{*}+EC^{-1}dE^{*}-EC^{-1}(dC)C^{-1}E^{*}, \tag*{(53)} d H = ( d E ) C − 1 E ∗ + E C − 1 d E ∗ − E C − 1 ( d C ) C − 1 E ∗ , (53)
d α = d a − 2 Re [ ( d u ) C − 1 u ∗ ] + u C − 1 ( d C ) C − 1 u ∗ , (54) d\alpha=da-2\operatorname{Re}[(du)C^{-1}u^{*}]+uC^{-1}(dC)C^{-1}u^{*}, \tag*{(54)} d α = d a − 2 Re [( d u ) C − 1 u ∗ ] + u C − 1 ( d C ) C − 1 u ∗ , (54)
d β = d b − 2 Re [ ( d v ) C − 1 v ∗ ] + v C − 1 ( d C ) C − 1 v ∗ , (55) d\beta=db-2\operatorname{Re}[(dv)C^{-1}v^{*}]+vC^{-1}(dC)C^{-1}v^{*}, \tag*{(55)} d β = d b − 2 Re [( d v ) C − 1 v ∗ ] + v C − 1 ( d C ) C − 1 v ∗ , (55)
d γ = ( d u ) C − 1 v ∗ + u C − 1 d v ∗ − u C − 1 ( d C ) C − 1 v ∗ , (56) d\gamma=(du)C^{-1}v^{*}+uC^{-1}dv^{*}-uC^{-1}(dC)C^{-1}v^{*}, \tag*{(56)} d γ = ( d u ) C − 1 v ∗ + u C − 1 d v ∗ − u C − 1 ( d C ) C − 1 v ∗ , (56)
d ρ = d z − d γ α β − ρ 2 ( d α α + d β β ) . (57) d\rho=\frac{dz-d\gamma}{\sqrt{\alpha\beta}}-\frac{\rho}{2}\left(\frac{d\alpha}{\alpha}+\frac{d\beta}{\beta}\right). \tag*{(57)} d ρ = α β d z − d γ − 2 ρ ( α d α + β d β ) . (57)
For n = 2 n=2 n = 2 the terms involving H H H are deleted. These expressions are real derivatives of the complex normalized coordinate, obtained by differentiating Eq. (4) and Definition 2.1 entrywise.
For real SPD sampling, φ ( A ) = ψ ( ρ ( A ) ) \varphi(A)=\psi(\rho(A)) φ ( A ) = ψ ( ρ ( A )) is smooth on the interior. If a matrix estimator satisfies m ( a m − a 0 ) ⇒ N ( 0 , Ω ) \sqrt{m}(\mathbf{a}_m-\mathbf{a}_0)\Rightarrow N(0,\Omega) m ( a m − a 0 ) ⇒ N ( 0 , Ω ) in a fixed real vectorization, then the delta method gives
m [ φ ( a m ) − φ ( a 0 ) ] ⇒ N ( 0 , ∇ φ T Ω ∇ φ ) , (58) \sqrt{m}\,[\varphi(\mathbf{a}_m)-\varphi(\mathbf{a}_0)]\Rightarrow N\left(0,\nabla\varphi^{\mathsf{T}}\Omega\nabla\varphi\right), \tag*{(58)} m [ φ ( a m ) − φ ( a 0 )] ⇒ N ( 0 , ∇ φ T Ω∇ φ ) , (58)
provided the limiting variance is positive and consistently estimated. For complex sampling one uses a declared smooth two-component function, for example an intrinsic logarithm at a fixed interior reference followed by its orthonormal coordinates; its Jacobian supplies a multivariate delta-method limit. No canonical signed complex scalar is implied. If the reference centre is itself estimated, its derivative and dependence must be included in the joint limit rather than held fixed silently.
For centred real Gaussian data the sample covariance fluctuation has entries Cov ( m S a b , m S c d ) → M a c M b d + M a d M b c \operatorname{Cov}(\sqrt{m}S_{ab},\sqrt{m}S_{cd})\to M_{ac}M_{bd}+M_{ad}M_{bc} Cov ( m S ab , m S c d ) → M a c M b d + M a d M b c . For proper complex Gaussian data, writing S = m − 1 ∑ x x ∗ S=m^{-1}\sum xx^{*} S = m − 1 ∑ x x ∗ before the asymptotically negligible centring adjustment,
E [ δ S a b δ S c d ‾ ] = m − 1 M a c M d b , (59) \mathbb{E}[\delta S_{ab}\overline{\delta S_{cd}}]=m^{-1}M_{ac}M_{db}, \tag*{(59)} E [ δ S ab δ S c d ] = m − 1 M a c M d b , (59)
E [ δ S a b δ S c d ] = m − 1 M a d M c b , (60) \mathbb{E}[\delta S_{ab}\delta S_{cd}]=m^{-1}M_{ad}M_{cb}, \tag*{(60)} E [ δ S ab δ S c d ] = m − 1 M a d M c b , (60)
where δ S = S − M \delta S=S-M δ S = S − M . These follow by pairing Gaussian fourth moments and subtracting the product of means; Hermitian redundancy must be respected when assembling a real covariance matrix. They are sampling covariances, not intrinsic individual reference laws. A fixed population matrix used for oracle normalization does not make the sample complement fixed, and oracle coordinates can leave the admissible disk; such failures are retained and receive no conditional slice guarantee. Figure A2 shows how the normalized coordinate inherits the slower of the two dependencies along an interpolation of the complement.
Figure A2. The complement affects both sensitivity and normalization. (A) Normalized sensitivity to the selected entry and elasticity with respect to the complement scale. (B) The coefficients α \alpha α , β \beta β and α β \sqrt{\alpha\beta} α β along the same declared interpolation. Both panels evaluate the differential formulas in Appendix B; only the complement scale varies.
Stable evaluation of the arc coordinate
The closed form (8) should not be evaluated by taking the inverse hyperbolic tangent of a number that has already been rounded to one. For 0 ≤ r < 1 0\le r<1 0 ≤ r < 1 put
u = 2 r 1 + r 2 , Δ = ( 1 − r ) ( 1 + r ) , (61) u=\frac{\sqrt{2}\,r}{\sqrt{1+r^{2}}},\qquad\Delta=(1-r)(1+r), \tag*{(61)} u = 1 + r 2 2 r , Δ = ( 1 − r ) ( 1 + r ) , (61)
so that 1 − u 2 = Δ / ( 1 + r 2 ) 1-u^{2}=\Delta/(1+r^{2}) 1 − u 2 = Δ/ ( 1 + r 2 ) by direct expansion. Since 2 atanh u = log ( 1 + u ) − log ( 1 − u ) 2\operatorname{atanh}u=\log(1+u)-\log(1-u) 2 atanh u = log ( 1 + u ) − log ( 1 − u ) and log ( 1 − u 2 ) = log ( 1 + u ) + log ( 1 − u ) \log(1-u^{2})=\log(1+u)+\log(1-u) log ( 1 − u 2 ) = log ( 1 + u ) + log ( 1 − u ) , one has 2 atanh u = 2 log ( 1 + u ) − log ( 1 − u 2 ) 2\operatorname{atanh}u=2\log(1+u)-\log(1-u^{2}) 2 atanh u = 2 log ( 1 + u ) − log ( 1 − u 2 ) , and therefore
ψ ( r ) = 2 log ( 1 + u ) + log ( 1 + r 2 ) − log Δ − 2 asinh r . (62) \psi(r)=2\log(1+u)+\log(1+r^{2})-\log\Delta-\sqrt{2}\operatorname{asinh}r. \tag*{(62)} ψ ( r ) = 2 log ( 1 + u ) + log ( 1 + r 2 ) − log Δ − 2 asinh r . (62)
Every term in Eq. (62) is finite for 0 ≤ r < 1 0\le r<1 0 ≤ r < 1 ; only − log Δ -\log\Delta − log Δ grows, and it grows exactly as the boundary is approached, which is the divergence identified in Theorem 2.5. Evaluating with log-one-plus for small arguments and using the odd symmetry of ψ \psi ψ for negative real coordinates avoids the cancellation that affects the original form.
Near the origin a series or an enclosed quadrature is preferable when subtraction loss matters. The boundary gap must be represented rather than silently clipped: if an input has already been rounded to r = 1 r=1 r = 1 , the admissible finite point is no longer identified by that floating-point value. Retaining a log gap or higher precision preserves the declared latent input and requires an explicit conversion policy between numerical representations. The implementation rejects an inverse arc value when its residual cannot be resolved in the chosen floating-point representation. Figure A3 displays the rearranged form against the boundary gap together with its leading logarithm.
Figure A3. Stable arc evaluation near the excluded boundary. (A) The rearranged arc coordinate and its leading logarithm as the boundary gap vanishes. (B) The gaps 1 − u 1-u 1 − u and 1 − r 1-r 1 − r , where u = 2 r / 1 + r 2 u=\sqrt{2}r/\sqrt{1+r^2} u = 2 r / 1 + r 2 is the direct inverse-hyperbolic-tangent argument. Keeping the boundary gap explicit avoids evaluating that function at a rounded argument of one.
Moment conditions and the verification of the central limit theorem
Bounded chart derivatives
Theorem 7.3 assumes that the chart Hessian is locally uniformly continuous with an integrable envelope. Lemma 7.2 supplies the required bound.
On a compact normal-coordinate ball B B B about μ \mu μ , the metric coefficients, their first derivatives and the Christoffel symbols are bounded by constants that depend only on B B B . The ordinary coordinate Hessian of ℓ \ell ℓ is the covariant Hessian plus a connection-coefficient term times the gradient, and the gradient has norm d ( p , x ) d(p,x) d ( p , x ) . Lemma 7.2 therefore gives
∥ Hess θ ℓ ( θ , x ) ∥ ≤ C B ( 1 + d ( μ , x ) ) (63) \lVert\operatorname{Hess}_{\theta}\ell(\theta,x)\rVert\le C_{B}\left(1+d(\mu,x)\right) \tag*{(63)} ∥ Hess θ ℓ ( θ , x )∥ ≤ C B ( 1 + d ( μ , x ) ) (63)
uniformly for θ \theta θ in the ball, with C B C_{B} C B depending only on $B. The loss is smooth in p p p for every x x x because the global logarithm of Theorem 2.5 has no cut locus. Pointwise continuity of the coordinate Hessian, together with the envelope (63) and E d ( μ , X ) ≤ ( E d 2 ) 1 / 2 < ∞ \mathbb{E}d(\mu,X)\le(\mathbb{E}d^{2})^{1/2}<\infty E d ( μ , X ) ≤ ( E d 2 ) 1/2 < ∞ , verifies the local uniform continuity condition (49) by dominated convergence.
At the centre the expected Hessian satisfies A ⪰ I A\succeq I A ⪰ I by Lemma 7.2, so A A A is finite and nonsingular. Consistency and the sandwich limit of Theorem 7.3 therefore hold on this disk for independent identically distributed observations with finite second intrinsic moment. On the real arc A = 1 A=1 A = 1 exactly.
Declared generating laws
Lemma D.1 (Intrinsic radius versus Euclidean radius). For the latent sampler ρ = tanh ( ∥ Y ∥ ) Y / ∥ Y ∥ \rho=\tanh(\lVert Y\rVert)Y/\lVert Y\rVert ρ = tanh (∥ Y ∥) Y / ∥ Y ∥ , with value zero at Y = 0 Y=0 Y = 0 ,
2 ∥ Y ∥ ≤ d ( 0 , ρ ) = ψ ( tanh ∥ Y ∥ ) ≤ 2 ∥ Y ∥ . (64) \sqrt{2}\,\lVert Y\rVert\le d(0,\rho)=\psi(\tanh\lVert Y\rVert)\le2\lVert Y\rVert. \tag*{(64)} 2 ∥ Y ∥ ≤ d ( 0 , ρ ) = ψ ( tanh ∥ Y ∥) ≤ 2 ∥ Y ∥ . (64) Proof. Write v = ∥ Y ∥ v=\lVert Y\rVert v = ∥ Y ∥ and differentiate:
d d v ψ ( tanh v ) = ψ ′ ( tanh v ) sech 2 v \frac{\mathrm{d}}{\mathrm{d}v}\psi(\tanh v)=\psi'(\tanh v)\operatorname{sech}^{2}v d v d ψ ( tanh v ) = ψ ′ ( tanh v ) sech 2 v = 2 1 + tanh 2 v 1 − tanh 2 v ( 1 − tanh 2 v ) =\frac{\sqrt{2}\sqrt{1+\tanh^{2}v}}{1-\tanh^{2}v}\left(1-\tanh^{2}v\right) = 1 − tanh 2 v 2 1 + tanh 2 v ( 1 − tanh 2 v ) = 2 1 + tanh 2 v , =\sqrt{2}\sqrt{1+\tanh^{2}v}, = 2 1 + tanh 2 v , using ψ ′ ( r ) = 2 1 + r 2 / ( 1 − r 2 ) \psi'(r)=\sqrt{2}\sqrt{1+r^{2}}/(1-r^{2}) ψ ′ ( r ) = 2 1 + r 2 / ( 1 − r 2 ) from Proposition 2.4. Since 0 ≤ tanh 2 v ≤ 1 0\le\tanh^{2}v\le1 0 ≤ tanh 2 v ≤ 1 , the derivative lies in [ 2 , 2 ] [\sqrt{2},2] [ 2 , 2 ] ; integrating from 0 0 0 and using ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 gives Eq. (64) .
Consequently Eq. (11) holds for every latent law with finite second Euclidean moment. Gaussian laws, finite Gaussian mixtures and centred laws with exponential components all qualify. A multivariate t ν t_{\nu} t ν law has radial tail density proportional to v − ν − 1 v^{-\nu-1} v − ν − 1 , so its second moment is finite if and only if ν > 2 \nu>2 ν > 2 . A t 5 t_{5} t 5 law satisfies Eq. (11) ; a t 1 t_{1} t 1 law does not. For the latter, the finite second-moment centre, covariance and central-limit construction used here does not apply; this does not exclude other notions of location. A near-boundary family with r ≤ 1 − 10 − 5 r\le1-10^{-5} r ≤ 1 − 1 0 − 5 has bounded intrinsic distance and therefore also satisfies Eq. (11) .
These are analytic checks of the declared generators, not measured outcomes. No corresponding moment assertion is made for the varying-complement oracle mapping of Appendix § 11 , which can be undefined outside the admissible disk, and none is made for a law whose second intrinsic moment has not been verified.