Introduction
Dense retrieval ranks document embeddings x i ∈ S d − 1 x_i \in S^{d-1} x i ∈ S d − 1 by q ⊤ x i q^\top x_i q ⊤ x i . Compression replaces those vectors by a smaller index; success requires ranking quality at a stated byte budget, not merely visually separated clusters. This distinction is essential because attention-induced clustering is already well studied [1 , 2 , 3 , 4 ] , and representation collapse can harm useful information [9 ] .
We investigate three questions: whether directionality helps retrieval at a fixed budget, whether it attenuates a specified representation perturbation, and what mathematical guarantees survive nonreciprocity. The initial centroid-only hypothesis fails against competitive quantization. That failure motivates a second, explicit hypothesis: constrained skew score calibration may correct task-relevant quantization errors without the ill-conditioning allowed by a general symmetric metric.
The analysis provides three connected results: a sharp metastability exponent controlled by block mass even under exponential particle growth; a second-order attention compression estimate obtained by mass centering; and a unique skew score-correction equation with a computable optimization residual bound. Experiments evaluate both semantic retrieval and geometric nearest-neighbor search, including a ten-million-vector visual-descriptor index. The scalar probability inequalities and standard convex optimization tools used in the derivations are attributed to their established sources.
Related Work
Spherical attention and clustering. The cluster formation, cone-collapse, and metastability results of [1 , 2 , 3 ] are the closest mathematical predecessors. The cone argument in [2 Lemma 4.2 and its proof] explicitly extends beyond symmetric query–key matrices. The acute consensus estimate used here is a quantitative specialization of that established positive-interaction argument. Directional mean shift [10 , 11 ] and diffusion condensation [12 ] supply related density-based and time-inhomogeneous clustering formulations. Directed-sphere consensus has its own established literature [13 , 14 ] .
Directionality and skew geometry. Symmetry and directionality during transformer training are analyzed in [6 ] . Routing/filtering separation [7 ] and orthogonal attention [8 ] use skew structure for different objectives. The latter exponentiates a token interaction operator; our finite retrieval adapter perturbs query–document scores. These token-space operators differ from the descriptor-space correction optimized here against compressed retrieval scores. Symmetric spectral selection [5 ] has different assumptions from the directed barrier clocks studied here.
Compression and asymmetric retrieval. Two-level softmax size and dispersion corrections are studied in [36 ] . Here the aggregation analysis controls a vector-valued attention output with skew bilinear scores and explicitly accounts for spherical normalization, rather than claiming size correction as new. Product quantization [18 ] , anisotropic vector quantization [19 ] , query-aware quantization [20 ] , and reduced-rank score regression [21 ] are essential comparators. Joint query-encoder and quantizer optimization is studied in [34 ] . Our correction leaves the descriptor extractor fixed, constrains the score operator to a skew subspace, and solves a convex moment equation before rank compression. Low-rank score approximation and query-aware compression are established components. Asymmetric hashing [16 , 17 ] also shows that asymmetry requires a precise task setting; its presence alone does not establish an advantage. Learned rotations for quantization predate this work [22 ] . Token merging [15 ] supplies a practical precedent for mass-aware aggregation, but reducing tokens inside a transformer and compressing an external retrieval index are distinct tasks.
The structural Gram identities and exact duplicate aggregation are supporting lemmas. The analytic contribution concerns shielded directed event selection and a two-cluster metastability exponent under a vanishing-skew restriction. For compressed retrieval, the contribution is a constrained score-correction formulation with explicit conditioning, optimization residual, and attention aggregation error bounds. These results distinguish the scope of the present model from general spherical consensus, generic asymmetric retrieval, and unconstrained low-rank regression.
Spherical Interaction Model and Retrieval Formulation
For positive masses m j m_j m j , skew A T = − A A^{\mathsf{T}}=-A A T = − A , M = I + ε A M=I+\varepsilon A M = I + ε A , and β > 0 \beta>0 β > 0 , the continuous model is
u ˙ i = ( I − u i u i T ) ∑ j p i j u j , p i j = m j e β u i T M u j ∑ k m k e β u i T M u k . (1) \dot{u}_i=\left(I-u_i u_i^{\mathsf{T}}\right)\sum_j p_{ij}u_j,\qquad p_{ij}=\frac{m_j e^{\beta u_i^{\mathsf{T}} M u_j}}{\sum_k m_k e^{\beta u_i^{\mathsf{T}} M u_k}}. \tag*{(1)} u ˙ i = ( I − u i u i T ) j ∑ p ij u j , p ij = ∑ k m k e β u i T M u k m j e β u i T M u j . (1)
For A = 0 A=0 A = 0 and equal masses, (1) is the normalized spherical self-attention flow studied in [1 , 3 ] ; general query–key bilinear scores are formulated in [2 Section 2.2] . The exponential dot-product kernel is also related to directional mean shift [10 , 11 ] , although the projected continuous dynamics here are distinct from its discrete mode-seeking iteration. Its directed barrier is g i j = 1 − u i T M u j g_{ij}=1-u_i^{\mathsf{T}}Mu_j g ij = 1 − u i T M u j , with g i j − g j i = − 2 ε u i T A u j g_{ij}-g_{ji}=-2\varepsilon u_i^{\mathsf{T}}Au_j g ij − g j i = − 2 ε u i T A u j . Appendix § 11 gives closed ordinary/skew Gram dynamics, a shielded first-event theorem, an explicit inversion of nearest-pair order, and rotating triangles that exclude unconditional coarsening. Self-edges are explicitly excluded from the corridor’s positive external-barrier condition. Figure 1 fixes the geometric notation and distinguishes the two directed barriers between neighboring caps.
Figure 1. Our geometric notation: caps on the unit circle, with directed interactions and unequal barriers. The cap construction is visually motivated by Figure 1 of [3 ] ; this drawing is original.
The same skew form measures cycle directionality. For H i j = u i T A u j H_{ij}=u_i^{\mathsf{T}}Au_j H ij = u i T A u j , the three-cycle log affinity is 2 β ε ( H a b + H b c + H c a ) 2\beta\varepsilon(H_{ab}+H_{bc}+H_{ca}) 2 β ε ( H ab + H b c + H c a ) . For the planar rotation A = J A=J A = J , it equals − 4 β ε -4\beta\varepsilon − 4 β ε times the signed triangle area, as shown in Figure 2 . Masses and row normalizers cancel exactly; the full derivation is in Appendix § 11 .
Figure 2. The skew interaction measures the directed signed area of a triangle. The figure represents a two-dimensional linear plane; the vertices may also be constrained to a unit circle.
The first ML pilot uses a finite algorithm: u i + = normalize ( u i + h ( I − u i u i T ) ∑ j p i j u j ) u_i^{+}=\operatorname{normalize}\left(u_i+h\left(I-u_i u_i^{\mathsf{T}}\right)\sum_j p_{ij}u_j\right) u i + = normalize ( u i + h ( I − u i u i T ) ∑ j p ij u j ) . We take β = 12 \beta=12 β = 12 , h = 0.25 h=0.25 h = 0.25 , and two or four steps selected on validation. This is not an exact ODE solver and does not invoke a recursive-merger theorem. Spherical Lloyd clustering of the final vectors selects a partition with K ∈ { 128 , 256 , 512 } K\in\{128,256,512\} K ∈ { 128 , 256 , 512 } ; prototypes average the original vectors.
Document scores are prototype scores expanded through an assignment array. All documents in a cluster therefore tie, resolved by a fixed lexicographic ID order. The pilot also compares a positive-pair-moment skew learned only from fitting queries, its sign reversal, random skew, reciprocal flow, and ordinary spherical clustering. A supervised symmetric-moment control was added after the initial pilot and is labeled exploratory.
Proposition 3.1 (A centroid-only ranking obstruction). If distinct unit vectors x i , x j x_i,x_j x i , x j share a reconstructed vector, there is a unit query for which their exact score gap is ∥ x i − x j ∥ \lVert x_i-x_j\rVert ∥ x i − x j ∥ while their compressed scores tie. No centroid-only representation of that cluster preserves every query’s strict pairwise ordering.
Proof. Take q = ( x i − x j ) / ∥ x i − x j ∥ q=(x_i-x_j)/\lVert x_i-x_j\rVert q = ( x i − x j ) / ∥ x i − x j ∥ . The exact score difference is the stated norm; the reconstructed difference is zero. A fixed tie-break cannot recover both this ordering and its reversal under − q -q − q .
This is a standard distortion obstruction applied to the pilot. Improving the dynamics alone cannot restore discarded within-cluster information.
Certified Nonreciprocal Score Correction
Let X , X ^ ∈ R N × d X,\widehat{X}\in\mathbb{R}^{N\times d} X , X ∈ R N × d contain original and compressed document vectors, and let Q ∈ R s × d Q\in\mathbb{R}^{s\times d} Q ∈ R s × d contain fitting queries. Score approximation as multivariate regression is established in LoRANN [21 Sections 3.1–3.2] ; score-sensitive quantizer losses are developed in [19 Section 3, Theorem 3.2] and [20 ] . Here the codebook is fixed and the regression variable is restricted to a skew query-side residual. The resulting retrieval distortion is
J ( A ) = ∥ Q A X ^ T − Q ( X − X ^ ) T ∥ F 2 2 s N + λ 2 ∥ A ∥ F 2 , A T = − A , λ > 0. (2) \mathcal{J}(A)=\frac{\lVert Q A\widehat{X}^{\mathsf{T}}-Q(X-\widehat{X})^{\mathsf{T}}\rVert_{F}^{2}}{2sN}+\frac{\lambda}{2}\lVert A\rVert_{F}^{2},\qquad A^{\mathsf{T}}=-A,\quad\lambda>0. \tag*{(2)} J ( A ) = 2 s N ∥ Q A X T − Q ( X − X ) T ∥ F 2 + 2 λ ∥ A ∥ F 2 , A T = − A , λ > 0. (2)
Thus the target is the uncompressed score, rather than a geometric reconstruction loss that weights every direction equally. Define
C = Q T Q / s , H = X ^ T X ^ / N , D = ( X − X ^ ) T X ^ / N , E = ( X − X ^ ) T ( X − X ^ ) / N , C=Q^{\mathsf{T}}Q/s,\qquad H=\widehat{X}^{\mathsf{T}}\widehat{X}/N,\qquad D=(X-\widehat{X})^{\mathsf{T}}\widehat{X}/N,\qquad E=(X-\widehat{X})^{\mathsf{T}}(X-\widehat{X})/N, C = Q T Q / s , H = X T X / N , D = ( X − X ) T X / N , E = ( X − X ) T ( X − X ) / N ,
and write Π − ( B ) = ( B − B T ) / 2 \Pi_{-}(B)=(B-B^{\mathsf{T}})/2 Π − ( B ) = ( B − B T ) /2 . These moments are sufficient to fit the correction without storing a query-by-document score matrix.
Theorem 4.1 (Unique calibrated skew solution and residual certificate). Problem (2) has a unique minimizer A ∗ A_{\ast} A ∗ , characterized by
T ( A ∗ ) : = Π − ( C A ∗ H ) + λ A ∗ = Π − ( C D ) . (3) \mathcal{T}(A_{\ast}):=\Pi_{-}(CA_{\ast}H)+\lambda A_{\ast}=\Pi_{-}(CD). \tag*{(3)} T ( A ∗ ) := Π − ( C A ∗ H ) + λ A ∗ = Π − ( C D ) . (3) On the real skew subspace, T \mathcal{T} T is self-adjoint and positive definite, with eigenvalues in [ λ , L ] [\lambda,L] [ λ , L ] , where L = λ + ∥ C ∥ o p ∥ H ∥ o p L=\lambda+\lVert C\rVert_{\mathrm{op}}\lVert H\rVert_{\mathrm{op}} L = λ + ∥ C ∥ op ∥ H ∥ op . For any skew approximation B B B , put R B = Π − ( C B H − C D ) + λ B R_B=\Pi_{-}(CBH-CD)+\lambda B R B = Π − ( C B H − C D ) + λ B . Then
∥ B − A ∗ ∥ F ≤ ∥ R B ∥ F / λ , (4) \lVert B-A_{\ast}\rVert_{F}\leq\lVert R_B\rVert_{F}/\lambda, \tag*{(4)} ∥ B − A ∗ ∥ F ≤ ∥ R B ∥ F / λ , (4) 0 ≤ J ( B ) − J ( A ∗ ) ≤ ∥ R B ∥ F 2 / ( 2 λ ) . (5) 0\leq\mathcal{J}(B)-\mathcal{J}(A_{\ast})\leq\lVert R_B\rVert_{F}^{2}/(2\lambda). \tag*{(5)} 0 ≤ J ( B ) − J ( A ∗ ) ≤ ∥ R B ∥ F 2 / ( 2 λ ) . (5) In particular A ∗ = 0 A_{\ast}=0 A ∗ = 0 if and only if Π − ( C D ) = 0 \Pi_{-}(CD)=0 Π − ( C D ) = 0 . The bound certifies the convex problem, not global optimality of a subsequent rank restriction.
Proof. Expanding the squared norm gives
2 J ( A ) = tr ( C A H A T ) − 2 ⟨ C D , A ⟩ F + tr ( C E ) + λ ∥ A ∥ F 2 . 2\mathcal{J}(A)=\operatorname{tr}(CAHA^{\mathsf{T}})-2\langle CD,A\rangle_{F}+\operatorname{tr}(CE)+\lambda\lVert A\rVert_{F}^{2}. 2 J ( A ) = tr ( C A H A T ) − 2 ⟨ C D , A ⟩ F + tr ( C E ) + λ ∥ A ∥ F 2 .
For skew variations Z Z Z , differentiation yields d J ( A ) [ Z ] = ⟨ C A H − C D + λ A , Z ⟩ F \mathrm{d}\mathcal{J}(A)[Z]=\langle CAH-CD+\lambda A,Z\rangle_{F} d J ( A ) [ Z ] = ⟨ C A H − C D + λ A , Z ⟩ F . The orthogonal projection onto admissible variations gives (3) . For skew Z , W Z,W Z , W , symmetry of C , H C,H C , H and cyclicity of trace imply ⟨ Z , T W ⟩ F = ⟨ T Z , W ⟩ F \langle Z,TW\rangle_{F}=\langle TZ,W\rangle_{F} ⟨ Z , T W ⟩ F = ⟨ T Z , W ⟩ F . Moreover
⟨ Z , T Z ⟩ F = ∥ C 1 / 2 Z H 1 / 2 ∥ F 2 + λ ∥ Z ∥ F 2 . \langle Z,TZ\rangle_{F}=\lVert C^{1/2}ZH^{1/2}\rVert_{F}^{2}+\lambda\lVert Z\rVert_{F}^{2}. ⟨ Z , T Z ⟩ F = ∥ C 1/2 Z H 1/2 ∥ F 2 + λ ∥ Z ∥ F 2 .
The lower bound is λ ∥ Z ∥ F 2 \lambda\lVert Z\rVert_{F}^{2} λ ∥ Z ∥ F 2 and the upper bound is L ∥ Z ∥ F 2 L\lVert Z\rVert_{F}^{2} L ∥ Z ∥ F 2 . Thus J \mathcal{J} J is coercive and strictly convex on this finite-dimensional subspace and has the unique stationary minimizer stated. Set F = B − A ⋆ F=B-A_{\star} F = B − A ⋆ . Then R B = T F R_{B}=TF R B = T F , so λ ∥ F ∥ F 2 ≤ ⟨ F , R B ⟩ F ≤ ∥ F ∥ F ∥ R B ∥ F \lambda\lVert F\rVert_{F}^{2}\le\langle F,R_{B}\rangle_{F}\le\lVert F\rVert_{F}\lVert R_{B}\rVert_{F} λ ∥ F ∥ F 2 ≤ ⟨ F , R B ⟩ F ≤ ∥ F ∥ F ∥ R B ∥ F , which proves (4) , including F = 0 F=0 F = 0 . Exact quadratic expansion at the stationary point gives J ( B ) − J ( A ⋆ ) = ⟨ F , T F ⟩ F / 2 \mathcal{J}(B)-\mathcal{J}(A_{\star})=\langle F,TF\rangle_{F}/2 J ( B ) − J ( A ⋆ ) = ⟨ F , T F ⟩ F /2 . Diagonalizing the self-adjoint operator on an orthonormal skew basis gives ⟨ F , T F ⟩ F = ⟨ R B , T − 1 R B ⟩ F ≤ ∥ R B ∥ F 2 / λ \langle F,TF\rangle_{F}=\langle R_{B},T^{-1}R_{B}\rangle_{F}\le\lVert R_{B}\rVert_{F}^{2}/\lambda ⟨ F , T F ⟩ F = ⟨ R B , T − 1 R B ⟩ F ≤ ∥ R B ∥ F 2 / λ . Finally, the normal equation at zero is equivalent to Π − ( C D ) = 0 \Pi_{-}(CD)=0 Π − ( C D ) = 0 .
The classical conjugate-gradient method [37 ] solves (3) using the matrix-vector operation A ↦ Π − ( C A H ) + λ A A\mapsto\Pi_{-}(CAH)+\lambda A A ↦ Π − ( C A H ) + λ A . No inversion of a d 2 d^{2} d 2 -by-d 2 d^{2} d 2 matrix is required. General structured least squares and Sylvester solvers precede this work [29 , 30 ] ; the result above specializes that machinery to quantized score distortion and provides the residual used by the retrieval artifact. This formulation also differs from unrestricted reduced-rank score regression [21 ] by enforcing a Euclidean skew correction and its associated conditioning bound.
For a deployed rank-2 r 2r 2 r correction, a real Schur decomposition groups the nonzero skew blocks into
( 0 a j − a j 0 ) . \begin{pmatrix}0&a_j\\-a_j&0\end{pmatrix}. ( 0 − a j a j 0 ) .
We retain the r r r largest ∣ a j ∣ \lvert a_j\rvert ∣ a j ∣ and clip each to a prescribed bound b b b . The best-rank principle is the classical Eckart–Young result [38 ] , applied to paired skew singular values. This is the Frobenius-nearest skew matrix satisfying rank at most 2 r 2r 2 r and spectral norm at most b b b : the singular values occur in equal pairs, the unitarily invariant best-rank approximation retains the largest pairs, and the scalar nearest point to [ − b , b ] [-b,b] [ − b , b ] is clipping. If A ~ \widetilde{A} A is an inexact full solution with residual at most η \eta η and B B B is this projection, then
J ( B ) − J ( A ⋆ ) ≤ L 2 ( ∥ B − A ~ ∥ F + η λ ) 2 . \mathcal{J}(B)-\mathcal{J}(A_{\star})\le\frac{L}{2}\left(\lVert B-\widetilde{A}\rVert_{F}+\frac{\eta}{\lambda}\right)^{2}. J ( B ) − J ( A ⋆ ) ≤ 2 L ( ∥ B − A ∥ F + λ η ) 2 .
This follows from (4) , the triangle inequality, and the quadratic upper bound. The full solver error and the rank-compression error are therefore reported separately. Symmetric and unrestricted normal-equation controls distinguish the skew constraint from generic regression capacity.
Native-distance margin calibration
For unit original descriptors, cosine and squared-L2 exact rankings coincide. After quantization, the reconstructed norms vary, so native L2 and inner-product PQ rankings need not coincide. We therefore calibrate the score actually used by the L2 index. For a fitting query q i q_i q i , let x i + x_i^{+} x i + be its exact nearest neighbor and let x ~ i k − \widetilde{x}_{ik}^{-} x ik − be a compressed-search hard negative. Put Δ i k = x ~ i + − x ~ i k − \Delta_{ik}=\widetilde{x}_i^{+}-\widetilde{x}_{ik}^{-} Δ ik = x i + − x ik − and
t i k = q i T ( x i + − x i k − − Δ i k ) + 1 2 ( ∥ x ~ i + ∥ 2 − ∥ x ~ i k − ∥ 2 ) . t_{ik}=q_i^{\mathsf{T}}\left(x_i^{+}-x_{ik}^{-}-\Delta_{ik}\right)+\frac{1}{2}\left(\lVert\widetilde{x}_i^{+}\rVert^{2}-\lVert\widetilde{x}_{ik}^{-}\rVert^{2}\right). t ik = q i T ( x i + − x ik − − Δ ik ) + 2 1 ( ∥ x i + ∥ 2 − ∥ x ik − ∥ 2 ) .
The target includes the reconstructed-norm term; dropping it calibrates a different scoring rule. The finite margin objective is
J g a p ( A ) = 1 2 s K ∑ i , k ( q i T A Δ i k − t i k ) 2 + λ 2 ∥ A ∥ F 2 , A T = − A . \mathcal{J}_{\mathrm{gap}}(A)=\frac{1}{2sK}\sum_{i,k}\left(q_i^{\mathsf{T}}A\Delta_{ik}-t_{ik}\right)^{2}+\frac{\lambda}{2}\lVert A\rVert_{F}^{2},\qquad A^{\mathsf{T}}=-A. J gap ( A ) = 2 sK 1 i , k ∑ ( q i T A Δ ik − t ik ) 2 + 2 λ ∥ A ∥ F 2 , A T = − A .
Define L ( A ) i k = q i T A Δ i k / s K \mathcal{L}(A)_{ik}=q_i^{\mathsf{T}}A\Delta_{ik}/\sqrt{sK} L ( A ) ik = q i T A Δ ik / sK and y i k = t i k / s K y_{ik}=t_{ik}/\sqrt{sK} y ik = t ik / sK . Its unique minimizer solves
[ Π − L ∗ L + λ I ] A ∗ = Π − L ∗ y . [\Pi-\mathcal{L}^{*}\mathcal{L}+\lambda I]A_{*}=\Pi-\mathcal{L}^{*}y. [ Π − L ∗ L + λ I ] A ∗ = Π − L ∗ y .
The proof of Theorem 4.1 applies verbatim after replacing the separable moment Hessian by L ∗ L \mathcal{L}^{*}\mathcal{L} L ∗ L : its quadratic form is ∥ L Z ∥ F 2 + λ ∥ Z ∥ F 2 \|\mathcal{L}Z\|_{F}^{2}+\lambda\|Z\|_{F}^{2} ∥ L Z ∥ F 2 + λ ∥ Z ∥ F 2 , so existence, uniqueness, and the residual error bound are unchanged. The implementation evaluates this operator by grouped query–negative products without forming a design matrix.
To connect this surrogate to ranking, let γ i k = q i T ( x i + − x i k − ) \gamma_{ik}=q_i^{\mathsf{T}}(x_{i}^{+}-x_{ik}^{-}) γ ik = q i T ( x i + − x ik − ) and let γ ^ i k \hat{\gamma}_{ik} γ ^ ik be the calibrated native-L2 score gap. Expanding − ∥ q i T ( I + A ) − x ^ ∥ 2 / 2 -\|q_i^{\mathsf{T}}(I+A)-\hat{x}\|^{2}/2 − ∥ q i T ( I + A ) − x ^ ∥ 2 /2 and cancelling its query-only term gives
γ ^ i k − γ i k = q i T A Δ i k − t i k . \hat{\gamma}_{ik}-\gamma_{ik}=q_i^{\mathsf{T}}A\Delta_{ik}-t_{ik}. γ ^ ik − γ ik = q i T A Δ ik − t ik .
For any γ 0 > 0 \gamma_{0}>0 γ 0 > 0 , a deterministic finite-sample inequality is
# { ( i , k ) : γ ^ i k ≤ 0 } s K ≤ # { ( i , k ) : γ i k < γ 0 } s K + 2 J g a p ( A ) γ 0 2 . \frac{\#\{(i,k):\hat{\gamma}_{ik}\leq0\}}{sK}
\leq
\frac{\#\{(i,k):\gamma_{ik}<\gamma_{0}\}}{sK}
+\frac{2\mathcal{J}_{\mathrm{gap}}(A)}{\gamma_{0}^{2}}. sK # {( i , k ) : γ ^ ik ≤ 0 } ≤ sK # {( i , k ) : γ ik < γ 0 } + γ 0 2 2 J gap ( A ) .
Indeed, a reversed pair with γ i k ≥ γ 0 \gamma_{ik}\geq\gamma_{0} γ ik ≥ γ 0 has squared gap error at least γ 0 2 \gamma_{0}^{2} γ 0 2 ; summing and retaining the nonnegative ridge term proves the display. This bounds errors on the specified pairs and does not turn a training-pair certificate into a test-distribution guarantee.
Low-Rank Implementation and Index Budget
Let x ^ i \hat{x}_{i} x ^ i be a product-quantized reconstruction, with m m m subspaces, L L L float32 codewords per subspace, and one uint8 code per document per subspace, following the product-code representation of [18 ] . The asymmetric-distance construction in [18 ] , Section III-A keeps the query unquantized and evaluates it against compressed database vectors. For squared-L2 search its ranking score, up to a query-only constant, is
s L 2 ( q , i ) = q T x ^ i − 1 2 ∥ x ^ i ∥ 2 . s_{\mathrm{L2}}(q,i)=q^{\mathsf{T}}\hat{x}_{i}-\frac{1}{2}\|\hat{x}_{i}\|^{2}. s L2 ( q , i ) = q T x ^ i − 2 1 ∥ x ^ i ∥ 2 .
This standard identity explains why the norm term in § 4.1 cannot be dropped. Our correction changes the query before the same compressed scorer; it does not replace PQ by a new codec. The deployed scorer stores codebooks, codes, and two factors U , V ∈ R d × r U,V\in\mathbb{R}^{d\times r} U , V ∈ R d × r :
A = U V T − V U T , s A ( q , i ) = q T ( I + A ) x ^ i , ∥ A ∥ o p ≤ 1 2 . (6) A=UV^{\mathsf{T}}-VU^{\mathsf{T}},\qquad s_{A}(q,i)=q^{\mathsf{T}}(I+A)\hat{x}_{i},\qquad\|A\|_{\mathrm{op}}\leq\frac{1}{2}. \tag*{(6)} A = U V T − V U T , s A ( q , i ) = q T ( I + A ) x ^ i , ∥ A ∥ op ≤ 2 1 . (6)
Here ϵ \epsilon ϵ is absorbed into A A A , and query and document spaces are treated differently. The rank of A A A is at most 2 r 2r 2 r , not r r r . In the semantic retrieval ablation, the factors are trained using the marginal log-likelihood of judged relevant documents among all corpus documents, with temperature 20 and a factor penalty 10 − 3 ( ∥ U ∥ F 2 + ∥ V ∥ F 2 ) 10^{-3}(\|U\|_{F}^{2}+\|V\|_{F}^{2}) 1 0 − 3 ( ∥ U ∥ F 2 + ∥ V ∥ F 2 ) . Only fitting queries supply supervision. Adam uses learning rate 0.01 for 40 epochs, batch size at most 64. The operator-norm projection is computed exactly from a 2 r 2r 2 r -dimensional factor core after each update. Validation nDCG@10 selects a checkpoint, including epoch zero. Float32 norm projection is checked with 10 − 5 10^{-5} 1 0 − 5 slack; mathematical bounds apply with the actual norm, including that numerical tolerance.
For fixed d d d , N N N , m m m , the payload size is
B P Q = 4 d L + N m , B a d a p t e r = 4 d L + N m + 8 d r . (7) B_{\mathrm{PQ}}=4dL+Nm,\qquad B_{\mathrm{adapter}}=4dL+Nm+8dr. \tag*{(7)} B PQ = 4 d L + N m , B adapter = 4 d L + N m + 8 d r . (7)
Thus replacing L L L by L − 2 r L-2r L − 2 r pays for both float32 factors exactly:
4 d L + N m = 4 d ( L − 2 r ) + N m + 8 d r . 4dL+Nm=4d(L-2r)+Nm+8dr. 4 d L + N m = 4 d ( L − 2 r ) + N m + 8 d r .
Equation (7) counts the product-code storage of [18 ] and adds the factors in (6) ; this storage identity is derived here for the chosen representation. Our fixed-budget comparison is P Q 64 × 64 \mathrm{PQ}_{64\times64} PQ 64 × 64 versus P Q 64 × 48 \mathrm{PQ}_{64\times48} PQ 64 × 48 plus r = 8 r=8 r = 8 factors. Both use exactly 430,016 bytes on SciFact and 330,816 bytes on NFCorpus. This equality includes the adapter, not just the document codes. Common document IDs, text payloads, and encoder weights are excluded from both sides. No original document embeddings are needed for scoring. The Python reference reconstructs vectors to batch matrix products; this is not an optimized lookup-table search engine.
Symmetric U V T + V U T UV^{\mathsf{T}}+VU^{\mathsf{T}} U V T + V U T and general U V T UV^{\mathsf{T}} U V T adapters use identical parameter counts, training data, norm bounds, and checkpoints. Their maximal ranks differ, which is disclosed rather than called a rank-matched control. Unbounded training and the earlier coarse P Q 8 \mathrm{PQ}8 PQ 8 codec are retained in the results so the successful-looking variant cannot conceal failed trials. The semantic ablations are exploratory because their variants followed an inspected pilot; the SIFT experiment uses a separately fixed query partition. The finite calibration mechanism is not asserted to be the long-time limit of (1) .
What skew geometry guarantees
Proposition 5.1 (Conditioning and retrieval-error certificate). For real skew A A A with ∥ A ∥ o p ≤ b \lVert A\rVert_{\mathrm{op}}\le b ∥ A ∥ op ≤ b ,
∥ q T ( I + A ) ∥ 2 = ∥ q ∥ 2 + ∥ q T A ∥ 2 , σ min ( I + A ) ≥ 1 , cond 2 ( I + A ) ≤ 1 + b 2 . \lVert q^{\mathsf{T}}(I+A)\rVert^2=\lVert q\rVert^2+\lVert q^{\mathsf{T}}A\rVert^2,\qquad\sigma_{\min}(I+A)\ge1,\qquad\operatorname{cond}_{2}(I+A)\le\sqrt{1+b^2}. ∥ q T ( I + A ) ∥ 2 = ∥ q ∥ 2 + ∥ q T A ∥ 2 , σ m i n ( I + A ) ≥ 1 , cond 2 ( I + A ) ≤ 1 + b 2 . If ∥ x i − x ^ i ∥ ≤ r i \lVert x_i-\hat{x}_i\rVert\le r_i ∥ x i − x ^ i ∥ ≤ r i and ∥ q ∥ = 1 \lVert q\rVert=1 ∥ q ∥ = 1 , then
∣ q T ( I + A ) x i − s A ( q , i ) ∣ ≤ 1 + b 2 r i . \left|q^{\mathsf{T}}(I+A)x_i-s_A(q,i)\right|\le\sqrt{1+b^2}\,r_i. q T ( I + A ) x i − s A ( q , i ) ≤ 1 + b 2 r i . Consequently a pairwise exact calibrated margin larger than 1 + b 2 ( r i + r j ) \sqrt{1+b^2}(r_i+r_j) 1 + b 2 ( r i + r j ) preserves that pair’s order after quantization. Relative to uncalibrated cosine, the conservative error is instead b + 1 + b 2 r i b+\sqrt{1+b^2}r_i b + 1 + b 2 r i ; calibration changes the target ranking.
Proof. The cross term vanishes since q T A q = 0 q^{\mathsf{T}}Aq=0 q T A q = 0 . Apply the resulting norm identity to arbitrary vectors and use ∥ A ∥ o p ≤ b \lVert A\rVert_{\mathrm{op}}\le b ∥ A ∥ op ≤ b for both singular value bounds. Cauchy–Schwarz gives the score error. Subtract the two score errors to obtain the ranking condition. Finally ∣ q T A x i ∣ ≤ b \left|q^{\mathsf{T}}Ax_i\right|\le b q T A x i ≤ b for unit q , x i q,x_i q , x i .
For b = 1 / 2 b=1/2 b = 1/2 , the condition number is at most 5 / 2 \sqrt{5}/2 5 /2 . This is an algebraic guarantee; it does not guarantee improved nDCG or certify top-k k k relevance from unlabeled geometry. Radius certificates for the much coarser centroid pilot certify no top-10 sets in the observed data, a useful failure case rather than a reason to suppress the measurement.
Proposition 5.2 (When a skew direction can improve a smooth loss). Let L ( A ) \mathcal{L}(A) L ( A ) be differentiable at zero, with Frobenius gradient G G G , and put S = ( G − G T ) / 2 S=(G-G^{\mathsf{T}})/2 S = ( G − G T ) /2 . Over skew directions D D D with ∥ D ∥ F ≤ 1 \lVert D\rVert_F\le1 ∥ D ∥ F ≤ 1 ,
min D ⟨ G , D ⟩ F = − ∥ S ∥ F . \min_D\langle G,D\rangle_F=-\lVert S\rVert_F. D min ⟨ G , D ⟩ F = − ∥ S ∥ F . There is a strict first-order skew descent direction exactly when S ≠ 0 S\ne0 S = 0 . If the second directional derivative is bounded above by H H H along D = − S / ∥ S ∥ F D=-S/\lVert S\rVert_F D = − S / ∥ S ∥ F , then L ( t D ) ≤ L ( 0 ) − t ∥ S ∥ F + H t 2 / 2 \mathcal{L}(tD)\le\mathcal{L}(0)-t\lVert S\rVert_F+Ht^2/2 L ( t D ) ≤ L ( 0 ) − t ∥ S ∥ F + H t 2 /2 .
Proof. Symmetric matrices are Frobenius-orthogonal to skew matrices, so ⟨ G , D ⟩ F = ⟨ S , D ⟩ F \langle G,D\rangle_F=\langle S,D\rangle_F ⟨ G , D ⟩ F = ⟨ S , D ⟩ F . Cauchy–Schwarz gives the minimum and its attaining direction. Taylor’s integral remainder gives the finite-step bound. The S = 0 S=0 S = 0 case excludes strict first-order descent only; it says nothing about possible second-order improvements.
This is a standard projected-gradient fact, used as a falsifiable mechanism. It explains why directionality cannot be assumed beneficial for arbitrary data. It does not establish that the factorized nonconvex optimizer reaches the best skew matrix, or that skew is better than an unconstrained adapter.
Particle Growth and the Metastable Time Scale
For two narrow blocks, normalized interaction weights depend on total block mass rather than the number of identical or nearby particles. This distinction leads to Theorem L.1 : if the block-mass ratio is e o ( β ) e^{o(\beta)} e o ( β ) , the limiting center angle is acute, and the skew pairing of centers vanishes, the first cross-block clock has exponent 1 − cos ϕ 1-\cos\phi 1 − cos ϕ , without a restriction on particle-count growth. Its proof uses a common weighted projective kernel to obtain subexponential intrablock relaxation, then a uniform nonlinear contraction and a cross-block Gram comparison. The block-mass fractions add to one in the final comparison, eliminating the minimum-individual-mass factor. For exactly collapsed blocks with mass ratio e β a e^{\beta a} e β a , the exponent changes to [ 1 − cos ϕ − ∣ a ∣ ] + [1-\cos\phi-|a|]_{+} [ 1 − cos ϕ − ∣ a ∣ ] + . This identifies an explicit boundary of the count-invariant regime.
Figure 3 shows the explicit three-point geometry underlying the two-clock example: u 1 = ( 1 , 0 , 0 ) u_1=(1,0,0) u 1 = ( 1 , 0 , 0 ) , u 2 = ( cos 0.8 , 0 , sin 0.8 ) u_2=(\cos0.8,0,\sin0.8) u 2 = ( cos 0.8 , 0 , sin 0.8 ) , and u 3 = ( cos 1.2 , sin 1.2 , 0 ) u_3=(\cos1.2,\sin1.2,0) u 3 = ( cos 1.2 , sin 1.2 , 0 ) . For the planar skew rotation and ε = 0.6 \varepsilon=0.6 ε = 0.6 , the favorable directed pair is ( 3 , 1 ) (3,1) ( 3 , 1 ) , although ( 1 , 2 ) (1,2) ( 1 , 2 ) is initially closest. The corridor theorem and Corollary K.5 establish the first-event selection and second-event exponent, respectively.
Figure 3. Exact three-dimensional point coordinates from (31) , shown in an orthographic projection. The blue arc is the shielded great-circle corridor; the dashed line projects u 2 u_2 u 2 onto its plane.
Experimental Evaluation
Data, separation, and metrics
We use BEIR [23 ] SciFact (5,183 documents) and NFCorpus (3,633 documents), one frozen all-MiniLM-L6-v2 encoder [25 ] , mask-aware mean pooling, maximum 256 wordpieces, and normalized 384-dimensional embeddings. Corpus and model revisions and downloaded-file SHA256 values are recorded. SciFact has 648 fitting, 161 validation, and 300 test queries; NFCorpus has 2,072 fitting, 518 validation, and 323 test queries. A fixed 20% query-ID holdout from the official training queries supplies validation; official test IDs are disjoint. Official development sets are not used. The corpus is transductively available for unsupervised codebook fitting, but test qrels are never used for fitting or checkpoint selection.
We report nDCG@10 with exponential relevance gains and Recall@100, averaging three seeds. Confidence intervals bootstrap whole test queries after averaging initialization seeds, with 5,000 resamples; these intervals do not estimate uncertainty across encoders or data sets. No multiple-comparison adjustment is applied, so the exploratory intervals are descriptive. Encoder pretraining overlap with BEIR cannot be excluded, and the semantic adapter uses labels while plain PQ does not. A separate fresh benchmark is needed for a confirmatory generalization claim. MTEB [24 ] motivates evaluation across tasks rather than treating clustering quality as retrieval quality.
Semantic Retrieval and Coarsening Ablations
Table 1. Fixed-byte retrieval comparison. Values are means over three seeds. PQ48 alone uses fewer bytes; all three adapters and PQ64 have identical payload sizes within each dataset.
On SciFact, the skew-minus-byte-matched-PQ nDCG difference is +0.0086, with a query-bootstrap 95% interval [-0.0038, +0.0207]. On NFCorpus, the skew-minus-byte-matched-PQ nDCG difference is +0.0003, with a query-bootstrap 95% interval [-0.0057, +0.0061]. The SciFact point estimate is positive, but neither fixed-byte confidence interval excludes zero. The general adapter is competitive, and the matched adapter controls do not establish a unique benefit of skew structure in this fine-codec regime. The ancillary records include the coarse PQ8 codec and unbounded-adapter ablations. The float32 document-vector payload is reduced by a factor of 18.51 on SciFact and 16.87 on NFCorpus. These ratios exclude common encoder, IDs, and text storage and do not describe end-to-end system memory.
Table 2. Coarsening ablation on SciFact: nDCG@10 by prototype budget and local perturbation stability.
Joint Particle-Count and Temperature Scaling
The growing-system theorem predicts a count-independent exponent for balanced near-collision blocks, even if N β N_{\beta} N β is exponential in β \beta β . We integrate the full smooth particle ODE with distinct, noncollapsed initial points and N β = ⌈ e γ β ⌉ N_{\beta} = \lceil e^{\gamma\beta} \rceil N β = ⌈ e γ β ⌉ for γ ∈ { 0.15 , 0.25 } \gamma\in\{0.15, 0.25\} γ ∈ { 0.15 , 0.25 } . The center angle is 0.8, the threshold is 0.3, and the skew operator is nonzero with zero restriction to the center plane. Initial off-plane perturbations have radius 0.025 / β 0.025/\sqrt{\beta} 0.025/ β . Each block has total mass one; no forced collisions or quotient replacement are used in this experiment. A second regime gives individual log masses a range 0.2 β 2 0.2\beta^{2} 0.2 β 2 within each block before normalizing each block to total mass one. It violates the older minimum-individual-mass scaling condition while preserving the block-mass hypothesis of Theorem L.1 .
Table 3. Full-flow joint scaling, G = 1 − cos ( 0.8 ) = 0.30329329 G = 1 - \cos(0.8) = 0.30329329 G = 1 − cos ( 0.8 ) = 0.30329329 . Total block masses are balanced. Extreme intrablock weights have logarithmic range 0.2 β 2 0.2\beta^{2} 0.2 β 2 .
The diagnostic log ( β T β ) / β \log(\beta T_{\beta})/\beta log ( β T β ) / β removes the leading logarithmic endpoint factor to expose finite-temperature convergence; the theorem itself concerns log T β / β \log T_{\beta}/\beta log T β / β . These finite computations support the stated joint-limit regime and do not prove it. The exact-block mass-imbalance check separately distinguishes the limiting rates 1 − cos ( 0.8 ) 1 - \cos(0.8) 1 − cos ( 0.8 ) and 1 − cos ( 0.8 ) − 0.1 1 - \cos(0.8) - 0.1 1 − cos ( 0.8 ) − 0.1 when the block-mass ratio grows as e 0.1 β e^{0.1\beta} e 0.1 β .
The Gaussian perturbation test adds noise of expected norm approximately 0.01 to corpus embeddings, renormalizes, and measures output displacement amplification and top-10 neighbor overlap on a deterministic 256-point sample. It is a specified local stress test, not evidence for stability under domain shift, different encoders, or adversarial perturbations. The reciprocal and directed flows show similar attenuation; a small measured attenuation difference does not establish a directionality advantage.
Million-Vector Visual Retrieval
We use all 1,000,000 SIFT-128 database descriptors from the public ANN-Benchmarks dataset [33 ] . Each descriptor is normalized to unit length. We recompute exact cosine top-100 ground truth using IndexFlatIP; the dataset’s supplied Euclidean neighbors are not reused. The 10,000 queries are split by fixed order into 2,000 fitting, 2,000 validation, and 6,000 test queries. No language model or descriptor fine-tuning is used in this experiment.
Each FAISS [32 ] index uses 16 eight-bit PQ subspaces, trained on a fixed 65,536-vector corpus sample. OPQ [31 ] uses 15 outer iterations before a separately trained PQ codebook. Three PQ initialization seeds are evaluated. Calibration uses only fitting queries and the same corpus sample; validation nearest-neighbor hit@10 selects λ ∈ { 10 − 6 , 10 − 5 , 10 − 4 } \lambda\in\{10^{-6}, 10^{-5}, 10^{-4}\} λ ∈ { 1 0 − 6 , 1 0 − 5 , 1 0 − 4 } . Corrections have spectral norm at most 0.5 and rank at most 16. The general rank-16 control stores twice as many factor entries as the canonical skew representation; the table includes that actual storage difference.
Table 4. Normalized SIFT1M: exact-nearest-neighbor hit@10/100, top-10 set recall, persistent serialized index plus correction size, and measured query time. Means over three codebook seeds.
Hit@k k k is the fraction of test queries whose exact nearest neighbor appears among the first k k k returned documents. Set recall@10 is the fraction of the exact top-10 set recovered. Search uses exhaustive compressed lookup-table scans, not reconstructed full vectors. Timings include the OPQ query rotation and factored correction, exclude metric evaluation, use four FAISS threads, and take the median of three full-test searches after warm-up. They are amortized batch-query times on the recorded host, not single-request tail latency.
For PQ, the skew-minus-base hit@10 difference is +0.0031, with query-bootstrap 95% interval [+0.0008, +0.0053].
For OPQ, the skew-minus-base hit@10 difference is +0.0081, with query-bootstrap 95% interval [+0.0059, +0.0104].
These intervals average seeds before resampling test queries and do not quantify between-dataset uncertainty. Neither a smaller calibration loss nor a positive point estimate is treated as a universal retrieval advantage. The score-calibration theorem certifies its convex distortion objective; measured neighbor recovery remains a separate empirical quantity.
Native-distance controls. Because the exact descriptors are unit length, squared-L2 and cosine exact rankings agree. Quantized reconstructions have varying norms, so their compressed L2 and inner-product rankings can differ. The following native-distance controls are therefore required before interpreting a calibration gain.
Table 5. Strong native-L2 controls and calibrated hard-negative margins on SIFT1M. Query correction is included in timing; all values average three codebook seeds.
The margin experiment uses 16 compressed-search hard negatives per fitting query and the norm-aware target in § 4.1 . Validation chooses among ridge penalties 10 − 5 10^{-5} 1 0 − 5 , 10 − 4 10^{-4} 1 0 − 4 , 10 − 3 10^{-3} 1 0 − 3 ; the smaller initial penalty grid failed the solver stopping criterion and was replaced by a diagonally preconditioned solve before this final evaluation. All final residuals and validation grids are retained. Native L2 changes recovery much more than the skew score correction. Margin calibration improves its target alignment, while symmetric and general margin controls remain stronger in this experiment; no unique empirical advantage of skew is inferred.
The margin-skew minus native PQ-L2 hit@10 difference is +0.0022, with query-bootstrap interval [+0.0002, +0.0043].
Table 6. Rank sensitivity of the score-distortion skew correction for native PQ-L2, seed 0. The ridge penalty is fixed to its validation-selected value.
Ten-Million-Vector Search and Memory Compression
We ingest the first 10,000,000 real BIGANN SIFT descriptors from the public billion-scale benchmark archive, without duplicating the one-million-vector dataset. The source prefix is obtained by an exact HTTP byte range, its length and SHA256 are recorded, and the descriptors are normalized in streaming blocks. Exact cosine top-100 neighbors are recomputed for all 10,000 public queries. The same fixed 2,000/2,000/6,000 fitting/validation/test partition is used. The public BIGANN query file has no exact descriptor overlap with the SIFT1M query file in this run.
The engineered index is native FAISS [32 ] IVF4096,PQ16x8 with squared-L2 scoring. It is trained on 262,144 sampled corpus descriptors and ingested in blocks of 200,000. Search uses an Apple M4 Pro host with 48 GiB RAM, macOS 26.3, and four CPU threads. We report a single built index and a probe-budget sweep, rather than treating probe settings as independent seeds. Serialized size includes vector IDs, inverted lists, codebooks, and coarse centers; it is a persistent payload measurement, not process peak RAM.
Table 7. BIGANN10M IVF-PQ search: 6,000 held-out queries. Median of three timed batch searches, four CPU threads.
The serialized index occupies 231.04 MiB, compared with 4.77 GiB for raw float32 vectors. Training takes 17.21 seconds and streaming ingestion 66.30 seconds on the recorded host. Ground-truth construction is excluded from both build timings.
The end-to-end skew margin control is also run through IVF routing and candidate scoring. Its validation protocol includes the identity update; validation selects that identity here. Accordingly its test rankings are unchanged and no ten-million-vector calibration gain is claimed. This guards against attributing index-parameter effects to a learned correction.
Mass-centered attention memory. To test the finite-compression theorem directly, we aggregate each of the 4,096 IVF cells by its exact count and the arithmetic mean of its original normalized keys. A streamed pass computes exact attention over all 10 million keys for 1,024 held-out queries at three inverse temperatures, with a fixed rank-two skew interaction of strength 0.3. The comparison retains identical cell masses and contrasts the exact barycenters, normalized barycenters, and the original sample-trained coarse centroids. It therefore isolates centering/normalization effects rather than changing the partition.
Table 8. Ten-million-key attention memory: mean Euclidean output error against exact streamed attention. All three methods retain 4,096 centers and the same cell masses.
The center-and-count payload is 2.031 MiB, versus 5.12 GB of original key vectors. This compresses an attention-memory readout, not document identities or exact nearest-neighbor rankings. The measured error increases at high inverse temperature, consistently with the explicit temperature factors in Theorem 8.1 ; the small-radius theorem is not interpreted as a uniformly small error bound for arbitrary IVF cells.
Computational Complexity and Evaluation Scope
The dense coarsening reference requires O ( T N 2 d ) O(TN^{2}d) O ( T N 2 d ) arithmetic and O ( N 2 ) O(N^{2}) O ( N 2 ) attention storage. The deployed correction is evaluated through compressed FAISS indices, using O ( d r ) O(dr) O ( d r ) query-side work for a rank-2 r 2r 2 r factorization. § 7.4 measures complete lookup-table search separately from metric evaluation, with native PQ and optimized product quantization (OPQ) controls. Quality, persistent index size, and measured query time are reported jointly; reconstruction error alone is not used as a retrieval claim.
Second-Order Mass-Centered Attention Compression
Exact mass aggregation becomes quantitatively stronger when the compression center cancels the first moment. In a block I a I_a I a , put m a = ∑ j ∈ I a m j m_a=\sum_{j\in I_a}m_j m a = ∑ j ∈ I a m j , w a j = m j / m a w_{aj}=m_j/m_a w aj = m j / m a and c a = ∑ j w a j u j c_a=\sum_j w_{aj}u_j c a = ∑ j w aj u j , without normalizing c a c_a c a . Assume ∥ u j − c a ∥ ≤ r a ≤ r \lVert u_j-c_a\rVert\le r_a\le r ∥ u j − c a ∥ ≤ r a ≤ r and ∥ u j ∥ = 1 \lVert u_j\rVert=1 ∥ u j ∥ = 1 . Let p ˉ a ( x ) \bar p_a(x) p ˉ a ( x ) be the softmax using masses m a m_a m a and scores β x ⊤ M c a \beta x^\top M c_a β x ⊤ M c a for a unit query x x x . The factor m a m_a m a is the mass correction already used for proportional attention in token merging [15 Section 3, Eq. (1)] ; for unit item masses it adds log ∣ I a ∣ \log|I_a| log ∣ I a ∣ to the cluster logit. Size and dispersion corrections in two-level softmax are analyzed in [36 Sections 3–4] . The theorem below controls the vector output and the additional error caused by normalizing the barycenter.
Theorem 8.1 (Centered log-partition and attention error). Write K = ∥ M ∥ o p K=\lVert M\rVert_{\mathrm{op}} K = ∥ M ∥ op and h = β K r h=\beta Kr h = β K r . The exact block probabilities are obtained from p ˉ a \bar p_a p ˉ a by adding logit corrections δ a \delta_a δ a with
0 ≤ δ a = log ∑ j ∈ I a w a j e β x ⊤ M ( u j − c a ) ≤ h 2 / 2. 0\leq\delta_a=\log\sum_{j\in I_a}w_{aj}e^{\beta x^\top M(u_j-c_a)}\leq h^2/2. 0 ≤ δ a = log j ∈ I a ∑ w aj e β x ⊤ M ( u j − c a ) ≤ h 2 /2. If μ ( x ) \mu(x) μ ( x ) is the full attention output and μ ˉ ( x ) = ∑ a p ˉ a ( x ) c a \bar\mu(x)=\sum_a\bar p_a(x)c_a μ ˉ ( x ) = ∑ a p ˉ a ( x ) c a , then
∥ μ ( x ) − μ ˉ ( x ) ∥ ≤ β K r 2 + min { 2 , e h 2 / 2 − 1 } . (8) \lVert\mu(x)-\bar\mu(x)\rVert\leq\beta Kr^2+\min\{2,e^{h^2/2}-1\}. \tag*{(8)} ∥ μ ( x ) − μ ˉ ( x )∥ ≤ β K r 2 + min { 2 , e h 2 /2 − 1 } . (8)
For fixed β \beta β , K K K , this is O ( r 2 ) O(r^{2}) O ( r 2 ) rather than the uncentered O ( r ) O(r) O ( r ) bound. The order r 2 r^{2} r 2 cannot be uniformly improved.
Proof. In each block the residuals z j = u j − c a z_j = u_j - c_a z j = u j − c a have weighted mean zero. The scalar Y j = β x T M z j Y_j = \beta x^{\mathsf{T}}Mz_j Y j = β x T M z j lies in [ − h , h ] [-h,h] [ − h , h ] and has mean zero. Jensen’s inequality gives the lower bound. Hoeffding’s bounded-variable log-moment inequality [28 ] , with range length 2 h 2h 2 h , gives log E e Y ≤ ( 2 h ) 2 / 8 = h 2 / 2 \log\mathbb{E}e^{Y} \le(2h)^{2}/8 = h^{2}/2 log E e Y ≤ ( 2 h ) 2 /8 = h 2 /2 . Factoring the exact block numerator as m a e β x T M c a e δ a m_a e^{\beta x^{\mathsf{T}}Mc_a}e^{\delta_a} m a e β x T M c a e δ a proves the probability formula. If D = h 2 / 2 D = h^{2}/2 D = h 2 /2 , normalization multiplies each p ˉ a \bar{p}_a p ˉ a by a ratio in [ e − D , e D ] [e^{-D},e^{D}] [ e − D , e D ] , so the probability ℓ 1 \ell^{1} ℓ 1 error is at most min ( 2 , e D − 1 ) \min(2,e^{D}-1) min ( 2 , e D − 1 ) .
It remains to compare the within-block, tilted mean to c a c_a c a . Define
ν a ( t ) = ∑ j w a j e t x T M u j u j ∑ j w a j e t x T M u j . \nu_a(t) = \frac{\sum_j w_{aj}e^{t x^{\mathsf{T}}Mu_j}u_j}{\sum_j w_{aj}e^{t x^{\mathsf{T}}Mu_j}}. ν a ( t ) = ∑ j w aj e t x T M u j ∑ j w aj e t x T M u j u j .
Direct differentiation gives ν a ′ ( t ) = Cov a , t ( u ) M T x \nu_a'(t) = \operatorname{Cov}_{a,t}(u)M^{\mathsf{T}}x ν a ′ ( t ) = Cov a , t ( u ) M T x . The covariance is positive semidefinite and
∥ Cov a , t ( u ) ∥ o p ≤ tr Cov a , t ( u ) ≤ E a , t ∥ u − c a ∥ 2 ≤ r 2 . \left\|\operatorname{Cov}_{a,t}(u)\right\|_{\mathrm{op}} \le\operatorname{tr}\operatorname{Cov}_{a,t}(u) \le\mathbb{E}_{a,t}\|u-c_a\|^{2} \le r^{2}. ∥ Cov a , t ( u ) ∥ op ≤ tr Cov a , t ( u ) ≤ E a , t ∥ u − c a ∥ 2 ≤ r 2 .
Hence ∥ ν a ( β ) − c a ∥ ≤ β K r 2 \|\nu_a(\beta)-c_a\| \le\beta Kr^{2} ∥ ν a ( β ) − c a ∥ ≤ β K r 2 . Sum this estimate with the exact block probabilities and then bound the probability replacement using ∥ c a ∥ ≤ 1 \|c_a\| \le1 ∥ c a ∥ ≤ 1 . This proves (8) .
For sharpness take one equal-mass block with u ± = ( cos θ , ± sin θ ) u_{\pm}=(\cos\theta,\pm\sin\theta) u ± = ( cos θ , ± sin θ ) , M = I M=I M = I , and x = ( 0 , 1 ) x=(0,1) x = ( 0 , 1 ) . The center is ( cos θ , 0 ) (\cos\theta,0) ( cos θ , 0 ) , r = sin θ r=\sin\theta r = sin θ , and the exact second output coordinate is sin θ tanh ( β sin θ ) \sin\theta\tanh(\beta\sin\theta) sin θ tanh ( β sin θ ) . For fixed β > 0 \beta>0 β > 0 its ratio to r 2 r^{2} r 2 tends to β \beta β , so a uniform o ( r 2 ) o(r^{2}) o ( r 2 ) output bound is impossible. Also δ = log cosh ( β r ) = β 2 r 2 / 2 + O ( r 4 ) \delta=\log\cosh(\beta r)=\beta^{2}r^{2}/2+O(r^{4}) δ = log cosh ( β r ) = β 2 r 2 /2 + O ( r 4 ) , attaining the second-order log-partition coefficient asymptotically.
If a spherical representative is required, use v a = c a / ∥ c a ∥ v_a=c_a/\|c_a\| v a = c a /∥ c a ∥ whenever r < 1 r<1 r < 1 . The unit-key identity 1 − ∥ c a ∥ 2 = ∑ j w a j ∥ u j − c a ∥ 2 ≤ r 2 1-\|c_a\|^{2}=\sum_j w_{aj}\|u_j-c_a\|^{2}\le r^{2} 1 − ∥ c a ∥ 2 = ∑ j w aj ∥ u j − c a ∥ 2 ≤ r 2 implies ∥ v a − c a ∥ = 1 − ∥ c a ∥ ≤ r 2 \|v_a-c_a\|=1-\|c_a\|\le r^{2} ∥ v a − c a ∥ = 1 − ∥ c a ∥ ≤ r 2 . Applying the logit perturbation estimate once more adds at most r 2 + min ( 2 , e 2 β K r 2 − 1 ) r^{2}+\min(2,e^{2\beta Kr^{2}}-1) r 2 + min ( 2 , e 2 β K r 2 − 1 ) to (8) . This explicitly accounts for normalization rather than treating a Euclidean barycenter as a unit vector. The aggregation estimate controls attention-output error, whereas the calibrated quadratic controls retrieval-score distortion. Their joint use is explicit: a partition first fixes masses and representatives; its centered residuals then enter D D D and E E E in the calibration objective. The residual certificate quantifies the additional optimization error. These quantities can therefore be measured separately instead of treating small cluster radius as an unqualified guarantee of ranking preservation.
Verification and Reproducibility
The additional equation audit checks 60 independently assembled skew-basis systems (maximum solution difference 9.76e-13), automatic derivatives and centered finite differences (maximum directional error 5.47e-13), and 240 centered-attention configurations. A directed-rounding interval residual enclosure is included for a small exact-float input system. It bounds the optimization gap for that instance, rather than claiming interval certification of the entire large benchmark. A further 300 heterogeneous-mass configurations test the projective common-kernel diameter inequality used by the growing-block theorem; the maximum projective differential-identity error is 6.66e-16, with no contraction-bound violations. These finite tests complement the analytic argument. The analytic ODE theorems have handwritten proofs and are not fully formalized. The checked file formal/Algebra.v proves nine kernels: finite-sum Gram product algebra; duplicate-mass summation; planar cycle-area algebra; skew-ratio cancellation under s 2 = 1 − c 2 s^2 = 1 - c^2 s 2 = 1 − c 2 ; the transverse-contraction constant inequality; a pairwise ranking-gap inequality; the algebraic relative-clock comparison; cap-minimum positivity; and the projective coordinate cancellation. Rocq 9.1.1 compilation and the independent coqchk library checker both succeed. No project-defined axioms or admitted proofs are used. The assumption log explicitly lists standard classical-real and functional-extensionality dependencies. We do not call these proofs axiom-free.
Lean/mathlib is a viable alternative for a larger analytic development, with axiom auditing via #print axioms [26 ] ; Rocq’s small kernel serves the corresponding check here [27 ] . Neither system automatically certifies a prose theorem by recognizing its displayed equations. The formal statement must faithfully encode the hypotheses and analysis. SymPy cancellation, randomized differential identities, finite differences for the rotating spectrum, a second ODE integration method, and counterexample tests supply complementary evidence but not substitutes for that formalization. The audit runner fails on missing evidence, admitted proofs, broken byte budgets, test-label overlap, or numerical identity failures; unresolved conjectures remain explicitly unresolved.
Discussion
The spherical flow and compressed scorer use the same skew bilinear geometry, but their guarantees concern different objects: metastable event times, attention output approximation, and query–document ranking distortion. The second-order aggregation bound and the certified calibration objective provide explicit links between these objects without identifying a finite query update with an infinite-time coarsening limit. The empirical comparisons are interpreted within the tested descriptors, index budgets, and query splits. The full-flow prefactor requires stronger synchronization than its logarithmic exponent and is not inferred from a numerical fit.
Closed Gram Geometry and Supporting Analysis
Fix integers n , d ≥ 2 n,d \ge2 n , d ≥ 2 , masses m i > 0 m_i>0 m i > 0 , β > 0 \beta>0 β > 0 , a real skew-symmetric A ∈ R d × d A \in\mathbb{R}^{d\times d} A ∈ R d × d , and ε ∈ R \varepsilon\in\mathbb{R} ε ∈ R . Set M = I d + ε A M=I_d+\varepsilon A M = I d + ε A . For U = ( u 1 , … , u n ) U=(u_1,\ldots,u_n) U = ( u 1 , … , u n ) with u i ∈ S d − 1 u_i \in\mathbb{S}^{d-1} u i ∈ S d − 1 , define
u ˙ i = ( I d − u i u i T ) ∑ j = 1 n p i j ( U ) u j , p i j ( U ) = m i exp ( β u i T M u j ) ∑ ℓ = 1 n m ℓ exp ( β u i T M u ℓ ) . (9) \dot{u}_i=(I_d-u_i u_i^{\mathsf T})\sum_{j=1}^{n}p_{ij}(U)u_j,\qquad p_{ij}(U)=\frac{m_i\exp(\beta u_i^{\mathsf T}Mu_j)}{\sum_{\ell=1}^{n}m_\ell\exp(\beta u_i^{\mathsf T}Mu_\ell)}. \tag*{(9)} u ˙ i = ( I d − u i u i T ) j = 1 ∑ n p ij ( U ) u j , p ij ( U ) = ∑ ℓ = 1 n m ℓ exp ( β u i T M u ℓ ) m i exp ( β u i T M u j ) . (9)
All solutions exist uniquely for all t ≥ 0 t\ge0 t ≥ 0 on the compact manifold ( S d − 1 ) n (\mathbb{S}^{d-1})^n ( S d − 1 ) n .
The directed barrier is
g i j ( U ) = 1 − u i T M u j = 1 − C i j − ε H i j , p i j p i i = m j m i e − β g i j . (10) g_{ij}(U)=1-u_i^{\mathsf T}Mu_j=1-C_{ij}-\varepsilon H_{ij},\qquad\frac{p_{ij}}{p_{ii}}=\frac{m_j}{m_i}e^{-\beta g_{ij}}. \tag*{(10)} g ij ( U ) = 1 − u i T M u j = 1 − C ij − ε H ij , p ii p ij = m i m j e − β g ij . (10)
where C i j = u i T u j C_{ij}=u_i^{\mathsf T}u_j C ij = u i T u j and H i j = u i T A u j H_{ij}=u_i^{\mathsf T}Au_j H ij = u i T A u j . In particular C = C T C=C^{\mathsf T} C = C T , H = − H T H=-H^{\mathsf T} H = − H T , and
g i j − g j i = − 2 ε H i j . (11) g_{ij}-g_{ji}=-2\varepsilon H_{ij}. \tag*{(11)} g ij − g j i = − 2 ε H ij . (11)
Although both transition probabilities are positive, their scales need not be comparable.
Simultaneous congruence of two Gram matrices
Write U ∈ R n × d U\in\mathbb{R}^{n\times d} U ∈ R n × d for the matrix with rows u i T u_i^{\mathsf T} u i T . Set
P = ( p i j ) , b i = ∑ j p i j C i j , B = P − diag ( b 1 , … , b n ) . (12) P=(p_{ij}),\qquad b_i=\sum_j p_{ij}C_{ij},\qquad B=P-\operatorname{diag}(b_1,\ldots,b_n). \tag*{(12)} P = ( p ij ) , b i = j ∑ p ij C ij , B = P − diag ( b 1 , … , b n ) . (12)
Then U ˙ = B U \dot{U}=BU U ˙ = B U exactly. This simple matrix identity leads to a useful invariant geometry.
Theorem A.1 (Closed Gram dynamics and preserved pencil). The pair ( C , H ) = ( U U T , U A U T ) (C,H)=(UU^{\mathsf T},UAU^{\mathsf T}) ( C , H ) = ( U U T , U A U T ) obeys the closed system
C ˙ = B C + C B T , H ˙ = B H + H B T . (13) \dot{C}=BC+CB^{\mathsf T},\qquad\dot{H}=BH+HB^{\mathsf T}. \tag*{(13)} C ˙ = B C + C B T , H ˙ = B H + H B T . (13) The coefficient matrix B B B is determined by C , H C,H C , H alone through
P i j = m j e β ( C i j + ε H i j ) ∑ k m k e β ( C i k + ε H i k ) . (14) P_{ij}=\frac{m_j e^{\beta(C_{ij}+\varepsilon H_{ij})}}{\sum_k m_k e^{\beta(C_{ik}+\varepsilon H_{ik})}}. \tag*{(14)} P ij = ∑ k m k e β ( C ik + ε H ik ) m j e β ( C ij + ε H ij ) . (14) For every finite t t t there is an invertible Φ ( t ) \Phi(t) Φ ( t ) with
C ( t ) = Φ ( t ) C ( 0 ) Φ ( t ) T , H ( t ) = Φ ( t ) H ( 0 ) Φ ( t ) T . (15) C(t)=\Phi(t)C(0)\Phi(t)^{\mathsf T},\qquad H(t)=\Phi(t)H(0)\Phi(t)^{\mathsf T}. \tag*{(15)} C ( t ) = Φ ( t ) C ( 0 ) Φ ( t ) T , H ( t ) = Φ ( t ) H ( 0 ) Φ ( t ) T . (15) Consequently rank U ( t ) \operatorname{rank} U(t) rank U ( t ) , rank H ( t ) \operatorname{rank} H(t) rank H ( t ) , and the common row span of U ( t ) U(t) U ( t ) are preserved at finite times. If C ( 0 ) C(0) C ( 0 ) is nonsingular, the eigenvalues of C ( t ) − 1 H ( t ) C(t)^{-1}H(t) C ( t ) − 1 H ( t ) are constant in time.
Proof. Differentiate U U T UU^{\mathsf T} U U T and U A U T UAU^{\mathsf T} U A U T , using U ˙ = B U \dot{U}=BU U ˙ = B U , to obtain (13) . The fundamental solution Φ ˙ = B ( t ) Φ \dot{\Phi}=B(t)\Phi Φ ˙ = B ( t ) Φ , Φ ( 0 ) = I n \Phi(0)=I_n Φ ( 0 ) = I n , is invertible at finite times and satisfies U ( t ) = Φ ( t ) U ( 0 ) U(t)=\Phi(t)U(0) U ( t ) = Φ ( t ) U ( 0 ) . Hence (15) . For invertible C C C , C ( t ) − 1 H ( t ) = Φ ( t ) − T [ C ( 0 ) − 1 H ( 0 ) ] Φ ( t ) T C(t)^{-1}H(t)=\Phi(t)^{-\mathsf T}[C(0)^{-1}H(0)]\Phi(t)^{\mathsf T} C ( t ) − 1 H ( t ) = Φ ( t ) − T [ C ( 0 ) − 1 H ( 0 )] Φ ( t ) T , which is a similarity transform. The remaining assertions follow from invertibility of Φ \Phi Φ .
Remark A.2. For n > d n>d n > d , C C C is necessarily singular and the statement concerning C − 1 H C^{-1}H C − 1 H is unavailable. The simultaneous congruence (15) , and hence the algebraic matrix pencil H − λ C H-\lambda C H − λ C , nevertheless remain meaningful. This invariant is structural, not a general formula for merger order.
Cyclic affinity equals a directed geometric area
Define the logarithmic affinity of a three-cycle by
A a b c = log p a b p b c p c a p b a p c b p a c . (16) \mathcal{A}_{abc}=\log\frac{p_{ab}p_{bc}p_{ca}}{p_{ba}p_{cb}p_{ac}}. \tag*{(16)} A ab c = log p ba p c b p a c p ab p b c p c a . (16)
All mass and normalization factors cancel:
A a b c = 2 β ε ( H a b + H b c + H c a ) = 2 β ε ( u a − u c ) T A ( u b − u c ) . (17) \mathcal{A}_{abc}=2\beta\varepsilon\left(H_{ab}+H_{bc}+H_{ca}\right)=2\beta\varepsilon\left(u_a-u_c\right)^{\mathsf T}A\left(u_b-u_c\right). \tag*{(17)} A ab c = 2 β ε ( H ab + H b c + H c a ) = 2 β ε ( u a − u c ) T A ( u b − u c ) . (17)
If d = 2 d=2 d = 2 and A = J = ( 0 − 1 1 0 ) A=J=\begin{pmatrix}0&-1\\1&0\end{pmatrix} A = J = ( 0 1 − 1 0 ) , the last expression is − 4 β ε -4\beta\varepsilon − 4 β ε times the signed Euclidean area of the triangle ( u a , u b , u c ) (u_a,u_b,u_c) ( u a , u b , u c ) . Thus the instantaneous Kolmogorov cycle criterion becomes an oriented area condition . It concerns reversibility of the instantaneous positive Markov matrix P ( U ) P(U) P ( U ) , not time reversibility of the nonlinear ODE.
A global consensus region without a gradient structure
Despite the lack of reciprocal weights, an acute configuration is globally attracted to consensus. This is a quantitative specialization of prior cone-collapse results, not a new consensus mechanism: the proof of [2 Lemma 4.2] explicitly extends to general query–key matrices. Lemma K.1 gives the standard projective-coordinate extension to a whole open hemisphere.
Theorem B.1 (Acute-cone consensus). If u i ( 0 ) T u j ( 0 ) > 0 u_i(0)^{\mathsf T}u_j(0)>0 u i ( 0 ) T u j ( 0 ) > 0 for all i , j i,j i , j , then every pair satisfies u i ( t ) T u j ( t ) → 1 u_i(t)^{\mathsf T}u_j(t)\to1 u i ( t ) T u j ( t ) → 1 as t → ∞ t\to\infty t → ∞ . More precisely, letting
ρ ( t ) = min i , j u i ( t ) T u j ( t ) , κ β = m min ∑ j m j exp ( − 2 β ∥ M ∥ o p ) , (18) \rho(t)=\min_{i,j}u_i(t)^{\mathsf T}u_j(t),\qquad\kappa_\beta=\frac{m_{\min}}{\sum_jm_j}\exp\left(-2\beta\lVert M\rVert_{\mathrm{op}}\right), \tag*{(18)} ρ ( t ) = i , j min u i ( t ) T u j ( t ) , κ β = ∑ j m j m m i n exp ( − 2 β ∥ M ∥ op ) , (18) one has ρ ( t ) ≥ ρ ( 0 ) > 0 \rho(t)\ge\rho(0)>0 ρ ( t ) ≥ ρ ( 0 ) > 0 and, at almost every t t t ,
ρ ˙ ( t ) ≥ 2 κ β ( 1 − ρ ( t ) 2 ) . (19) \dot{\rho}(t)\ge2\kappa_\beta\left(1-\rho(t)^2\right). \tag*{(19)} ρ ˙ ( t ) ≥ 2 κ β ( 1 − ρ ( t ) 2 ) . (19) In particular, 1 − ρ ( t ) 1-\rho(t) 1 − ρ ( t ) decays exponentially in t t t for each fixed β \beta β .
Proof. For c i j = u i T u j c_{ij}=u_i^{\mathsf T}u_j c ij = u i T u j direct differentiation gives
c ˙ i j = ∑ ℓ p i ℓ ( c ℓ j − c i ℓ c i j ) + ∑ ℓ p j ℓ ( c i ℓ − c j ℓ c i j ) . (20) \dot{c}_{ij}=\sum_{\ell}p_{i\ell}\left(c_{\ell j}-c_{i\ell}c_{ij}\right)+\sum_{\ell}p_{j\ell}\left(c_{i\ell}-c_{j\ell}c_{ij}\right). \tag*{(20)} c ˙ ij = ℓ ∑ p i ℓ ( c ℓ j − c i ℓ c ij ) + ℓ ∑ p j ℓ ( c i ℓ − c j ℓ c ij ) . (20)
Choose an active minimizing off-diagonal pair ( i , j ) (i,j) ( i , j ) . If ρ ≥ 0 \rho\ge0 ρ ≥ 0 , then c ℓ j ≥ ρ c_{\ell j}\ge\rho c ℓ j ≥ ρ and c i ℓ ≤ 1 c_{i\ell}\le1 c i ℓ ≤ 1 , so each term in the first sum is nonnegative; the second sum is handled identically. Retaining only ℓ = j \ell=j ℓ = j in the first sum and ℓ = i \ell=i ℓ = i in the second yields c ˙ i j ≥ ( p i j + p j i ) ( 1 − ρ 2 ) \dot{c}_{ij}\ge(p_{ij}+p_{ji})(1-\rho^2) c ˙ ij ≥ ( p ij + p j i ) ( 1 − ρ 2 ) . For every configuration, p i j ≥ κ β p_{ij}\ge\kappa_\beta p ij ≥ κ β . The finite minimum of smooth functions is absolutely continuous and satisfies the lower differential estimate almost everywhere. Scalar comparison gives (19) and the conclusion.
The exactly integrable two-body subsystem
For n = 2 n=2 n = 2 , set c = u 1 T u 2 c=u_1^{\mathsf{T}}u_2 c = u 1 T u 2 and h = u 1 T A u 2 h=u_1^{\mathsf{T}}Au_2 h = u 1 T A u 2 . Writing q 12 = p 12 q_{12}=p_{12} q 12 = p 12 and q 21 = p 21 q_{21}=p_{21} q 21 = p 21 , the equations reduce to
c ˙ = ( q 12 + q 21 ) ( 1 − c 2 ) , h ˙ = − ( q 12 + q 21 ) c h . (21) \dot{c}=(q_{12}+q_{21})(1-c^2),\qquad\dot{h}=-(q_{12}+q_{21})ch. \tag*{(21)} c ˙ = ( q 12 + q 21 ) ( 1 − c 2 ) , h ˙ = − ( q 12 + q 21 ) c h . (21)
Thus the signed ratio
α = h ( t ) 1 − c ( t ) 2 = h ( 0 ) 1 − c ( 0 ) 2 (22) \alpha=\frac{h(t)}{\sqrt{1-c(t)^2}}=\frac{h(0)}{\sqrt{1-c(0)^2}} \tag*{(22)} α = 1 − c ( t ) 2 h ( t ) = 1 − c ( 0 ) 2 h ( 0 ) (22)
is conserved whenever − 1 < c < 1 -1<c<1 − 1 < c < 1 . The stronger statement with the sign of α \alpha α follows from (21) ; merely conserving its square would lose the orientation.
For c 0 ∈ ( − 1 , 1 ) c_0\in(-1,1) c 0 ∈ ( − 1 , 1 ) , define q 12 ( c ) q_{12}(c) q 12 ( c ) and q 21 ( c ) q_{21}(c) q 21 ( c ) by substituting h = α 1 − c 2 h=\alpha\sqrt{1-c^2} h = α 1 − c 2 into
q 12 = 1 1 + ( m 1 / m 2 ) e β ( 1 − c − ε h ) , q 21 = 1 1 + ( m 2 / m 1 ) e β ( 1 − c + ε h ) . (23) q_{12}=\frac{1}{1+(m_1/m_2)e^{\beta(1-c-\varepsilon h)}},\qquad q_{21}=\frac{1}{1+(m_2/m_1)e^{\beta(1-c+\varepsilon h)}}. \tag*{(23)} q 12 = 1 + ( m 1 / m 2 ) e β ( 1 − c − ε h ) 1 , q 21 = 1 + ( m 2 / m 1 ) e β ( 1 − c + ε h ) 1 . (23)
The exact hitting time of c ∗ ∈ ( c 0 , 1 ) c_\ast\in(c_0,1) c ∗ ∈ ( c 0 , 1 ) is
T β ( c ∗ ) = ∫ c 0 c ∗ d c ( 1 − c 2 ) ( q 12 ( c ) + q 21 ( c ) ) . (24) T_\beta(c_\ast)=\int_{c_0}^{c_\ast}\frac{\mathrm{d}c}{(1-c^2)(q_{12}(c)+q_{21}(c))}. \tag*{(24)} T β ( c ∗ ) = ∫ c 0 c ∗ ( 1 − c 2 ) ( q 12 ( c ) + q 21 ( c )) d c . (24)
Proposition C.1 (Two-body exponential clock). For fixed masses, c 0 c_0 c 0 , c ∗ c_\ast c ∗ , A A A , ε \varepsilon ε ,
lim β → ∞ log T β ( c ∗ ) β = [ 1 − c 0 − ∣ ε h 0 ∣ ] + . (25) \lim_{\beta\to\infty}\frac{\log T_\beta(c_\ast)}{\beta}=[1-c_0-|\varepsilon h_0|]_+. \tag*{(25)} β → ∞ lim β log T β ( c ∗ ) = [ 1 − c 0 − ∣ ε h 0 ∣ ] + . (25) If B 0 = 1 − c 0 − ∣ ε h 0 ∣ > 0 B_0=1-c_0-|\varepsilon h_0|>0 B 0 = 1 − c 0 − ∣ ε h 0 ∣ > 0 , the hitting time is exponentially long with rate B 0 B_0 B 0 .
Proof. Uniformly for c c c in the compact interval [ c 0 , c ∗ ] [c_0,c_\ast] [ c 0 , c ∗ ] , the inverse two-sided transition rate has logarithmic exponential rate equal to [ f ( c ) ] + [f(c)]_+ [ f ( c ) ] + where f ( c ) = 1 − c − ∣ ε α ∣ 1 − c 2 f(c)=1-c-|\varepsilon\alpha|\sqrt{1-c^2} f ( c ) = 1 − c − ∣ ε α ∣ 1 − c 2 . The function f f f is convex and f ( 1 ) = 0 f(1)=0 f ( 1 ) = 0 . Hence its maximum on [ c 0 , c ∗ ] [c_0,c_\ast] [ c 0 , c ∗ ] cannot exceed max ( f ( c 0 ) , 0 ) \max(f(c_0),0) max ( f ( c 0 ) , 0 ) ; if f ( c 0 ) > 0 f(c_0)>0 f ( c 0 ) > 0 the maximum is attained at c 0 c_0 c 0 , and otherwise the positive part vanishes throughout. Applying the Laplace principle to (24) , whose nonexponential factor is bounded above and below, proves (25) . □
Directed corridor selection of a first threshold event
A first merger means the first crossing of a prescribed positive angular threshold. Exact collision at finite time cannot occur by uniqueness of the smooth flow.
Definition D.1 (Geometric isolation corridor). Select an ordered pair ( a , b ) (a,b) ( a , b ) with initial angle θ 0 ∈ ( 0 , π ) \theta_0\in(0,\pi) θ 0 ∈ ( 0 , π ) , a threshold θ ∗ ∈ ( 0 , θ 0 ) \theta_\ast\in(0,\theta_0) θ ∗ ∈ ( 0 , θ 0 ) , and initial points v i = u i ( 0 ) v_i=u_i(0) v i = u i ( 0 ) . Let γ \gamma γ be the shorter oriented great-circle arc from v a v_a v a to v b v_b v b . Define C a b \mathcal{C}_{ab} C ab as the compact family of configurations in which u a , u b u_a,u_b u a , u b lie on γ \gamma γ in their original order, satisfy θ ∗ ≤ dist S ( u a , u b ) ≤ θ 0 \theta_\ast\leq\operatorname{dist}_{S}(u_a,u_b)\leq\theta_0 θ ∗ ≤ dist S ( u a , u b ) ≤ θ 0 , and every u j u_j u j with j ∉ { a , b } j\notin\{a,b\} j ∈ / { a , b } is fixed at v j v_j v j . The pair has a strictly shielded corridor if, with
B 0 = min { g a b ( U 0 ) , g b a ( U 0 ) } > 0 , (26) B_0=\min\{g_{ab}(U_0),g_{ba}(U_0)\}>0, \tag*{(26)} B 0 = min { g ab ( U 0 ) , g ba ( U 0 )} > 0 , (26) there are constants Δ , σ > 0 \Delta,\sigma>0 Δ , σ > 0 such that throughout C a b \mathcal{C}_{ab} C ab :
(i) Every off-diagonal directed edge ( i , j ) (i,j) ( i , j ) , with i ≠ j i\ne j i = j , other than ( a , b ) , ( b , a ) (a,b),(b,a) ( a , b ) , ( b , a ) has g i j ≥ B 0 + Δ g_{ij}\geq B_0+\Delta g ij ≥ B 0 + Δ .
(ii) Every unordered pair other than { a , b } \{a,b\} { a , b } has angular separation at least θ ∗ + σ \theta_\ast+\sigma θ ∗ + σ .
The full off-diagonal-edge condition in (i) is deliberately stronger than needed; it makes the proof transparent and includes all possible competitors.
Theorem D.2 (First-merger selection and sharp exponential rate). Under theorem D.1, for all sufficiently large β \beta β , the pair { a , b } \{a,b\} { a , b } is the unique first pair to reach angle θ ∗ \theta_* θ ∗ . If τ β \tau_\beta τ β denotes this first hitting time, then
lim β → ∞ log τ β β = B 0 . (27) \lim_{\beta\to\infty}\frac{\log\tau_\beta}{\beta}=B_0. \tag*{(27)} β → ∞ lim β log τ β = B 0 . (27) The theorem holds for arbitrary fixed positive masses and any fixed skew-symmetric A A A .
Proof. We provide the bootstrap details, since a bare comparison of initial barriers is insufficient. Write K = ∥ M ∥ o p K=\lVert M\rVert_{\mathrm{op}} K = ∥ M ∥ op . Each barrier is 2 K 2K 2 K -Lipschitz in the max norm of configurations. In a fixed sufficiently small tubular neighborhood of C a b C_{ab} C ab , all external barriers exceed B 0 + 3 Δ / 4 B_0+3\Delta/4 B 0 + 3Δ/4 and all external angular gaps exceed θ ∗ + σ / 2 \theta_*+\sigma/2 θ ∗ + σ /2 .
The two moving points on the same oriented great circle have u a T A u b = α sin θ u_a^{\mathsf T}Au_b=\alpha\sin\theta u a T A u b = α sin θ with a fixed signed α \alpha α determined by the plane and its orientation. Thus the favorable internal barrier is
f ( cos θ ) = 1 − cos θ − ∣ ε α ∣ sin θ . (28) f(\cos\theta)=1-\cos\theta-\lvert\varepsilon\alpha\rvert\sin\theta. \tag*{(28)} f ( cos θ ) = 1 − cos θ − ∣ ε α ∣ sin θ . (28)
As in theorem C.1, this is a convex function of cos θ \cos\theta cos θ , equals B 0 B_0 B 0 at the original separation, and equals zero at coincident points. It is therefore bounded above by B 0 B_0 B 0 along the corridor. In a sufficiently small tube, min ( g a b , g b a ) ≤ B 0 + Δ / 8 \min(g_{ab},g_{ba})\le B_0+\Delta/8 min ( g ab , g ba ) ≤ B 0 + Δ/8 .
Because p i j / p i i = ( m j / m i ) e − β g i j p_{ij}/p_{ii}=(m_j/m_i)e^{-\beta g_{ij}} p ij / p ii = ( m j / m i ) e − β g ij , uniform positivity of masses yields the estimates
q ( t ) : = p a b ( t ) + p b a ( t ) ≥ c e − β ( B 0 + Δ / 8 ) , η β : = max i ∑ j ≠ i ( i , j ) ∉ { ( a , b ) , ( b , a ) } p i j ≤ C e − β ( B 0 + 3 Δ / 4 ) . (29) q(t):=p_{ab}(t)+p_{ba}(t)\ge c e^{-\beta(B_0+\Delta/8)},\qquad
\eta_\beta:=\max_i\sum_{\substack{j\ne i\\(i,j)\notin\{(a,b),(b,a)\}}}p_{ij}\le C e^{-\beta(B_0+3\Delta/4)}. \tag*{(29)} q ( t ) := p ab ( t ) + p ba ( t ) ≥ c e − β ( B 0 + Δ/8 ) , η β := i max j = i ( i , j ) ∈ / {( a , b ) , ( b , a )} ∑ p ij ≤ C e − β ( B 0 + 3Δ/4 ) . (29)
More precisely, in the second bound we sum over edges outside the chosen two directed edges, including edges emanating from external points. This notational convention will be used below. The internal two-body vector field moves u a , u b u_a,u_b u a , u b toward one another along their common great circle with nonnegative scalar speeds p a b sin θ p_{ab}\sin\theta p ab sin θ and p b a sin θ p_{ba}\sin\theta p ba sin θ . The external field has max norm at most C η β C\eta_\beta C η β .
As long as the solution remains in the tube and θ ∈ [ θ ∗ , θ 0 ] \theta\in[\theta_*,\theta_0] θ ∈ [ θ ∗ , θ 0 ] , the angular equation obeys
θ ˙ ≤ − q sin θ + C η β ≤ − 1 2 q sin θ < 0 (30) \dot{\theta}\le-q\sin\theta+C\eta_\beta\le-\frac{1}{2}q\sin\theta<0 \tag*{(30)} θ ˙ ≤ − q sin θ + C η β ≤ − 2 1 q sin θ < 0 (30)
for large β \beta β . The crossing time is consequently at most C e β ( B 0 + Δ / 8 ) Ce^{\beta(B_0+\Delta/8)} C e β ( B 0 + Δ/8 ) . If E E E is the maximal normal deviation, then D + E ≤ q ( 1 − c ) E + C η β D+E\le q(1-c)E+C\eta_\beta D + E ≤ q ( 1 − c ) E + C η β . On the fixed interval [ θ ∗ , θ 0 ] ⊂ ( 0 , π ) [\theta_*,\theta_0]\subset(0,\pi) [ θ ∗ , θ 0 ] ⊂ ( 0 , π ) , one has ( 1 − c ) / sin θ = tan ( θ / 2 ) ≤ tan ( θ 0 / 2 ) (1-c)/\sin\theta=\tan(\theta/2)\le\tan(\theta_0/2) ( 1 − c ) / sin θ = tan ( θ /2 ) ≤ tan ( θ 0 /2 ) . Hence ∫ q ( 1 − c ) d t ≤ C ∫ q sin θ d t = O ( 1 ) \int q(1-c)\,dt\le C\int q\sin\theta\,dt=O(1) ∫ q ( 1 − c ) d t ≤ C ∫ q sin θ d t = O ( 1 ) before threshold crossing, including when c c c is initially negative. Gronwall therefore controls normal amplification uniformly. During this period external points, and the normal deviations of u a , u b u_a,u_b u a , u b away from the oriented great circle, move by at most C e − 5 β Δ / 8 Ce^{-5\beta\Delta/8} C e − 5 β Δ/8 . Tangential components of the internal field point into the arc, while the integrated external error is of the same exponentially small order. A first-exit bootstrap closes the tube assumption. All competing angular gaps stay above θ ∗ \theta_* θ ∗ , proving selection.
To sharpen the clock, the distance of the true solution from C a b C_{ab} C ab is O ( e − 5 β Δ / 8 ) O(e^{-5\beta\Delta/8}) O ( e − 5 β Δ/8 ) . Hence the favorable internal barrier is at most B 0 + O ( e − 5 β Δ / 8 ) B_0+O(e^{-5\beta\Delta/8}) B 0 + O ( e − 5 β Δ/8 ) and q ≥ c e − β B 0 q\ge ce^{-\beta B_0} q ≥ c e − β B 0 for large β \beta β . The angular comparison improves the bound to τ β ≤ C e β B 0 \tau_\beta\le Ce^{\beta B_0} τ β ≤ C e β B 0 .
Conversely, for arbitrary ν ∈ ( 0 , B 0 ) \nu\in(0,B_0) ν ∈ ( 0 , B 0 ) a sufficiently small neighborhood of U 0 U_0 U 0 has every off-diagonal directed barrier at least B 0 − ν B_0-\nu B 0 − ν . Within this neighborhood all particle speeds are at most C ν e − β ( B 0 − ν ) C_\nu e^{-\beta(B_0-\nu)} C ν e − β ( B 0 − ν ) . A fixed positive displacement is necessary before the angular threshold can be reached, so τ β ≥ c ν e β ( B 0 − ν ) \tau_\beta\ge c_\nu e^{\beta(B_0-\nu)} τ β ≥ c ν e β ( B 0 − ν ) . Taking logarithms, then β → ∞ \beta\to\infty β → ∞ and ν ↓ 0 \nu\downarrow0 ν ↓ 0 , gives (27) . □ \square □
An analytic three-point inversion of Euclidean proximity
Consider n = d = 3 n=d=3 n = d = 3 , m i = 1 m_i=1 m i = 1 , and
ε = 0.6 , A = ( 0 − 1 0 1 0 0 0 0 0 ) , (31) \varepsilon=0.6,\qquad A=\begin{pmatrix}0&-1&0\\1&0&0\\0&0&0\end{pmatrix}, \tag*{(31)} ε = 0.6 , A = 0 1 0 − 1 0 0 0 0 0 , (31)
u 1 = ( 1 , 0 , 0 ) u_1=(1,0,0) u 1 = ( 1 , 0 , 0 ) , u 2 = ( cos 0.8 , 0 , sin 0.8 ) u_2=(\cos0.8,0,\sin0.8) u 2 = ( cos 0.8 , 0 , sin 0.8 ) , u 3 = ( cos 1.2 , sin 1.2 , 0 ) u_3=(\cos1.2,\sin1.2,0) u 3 = ( cos 1.2 , sin 1.2 , 0 ) .
Angles here and in the figures are measured in radians. The three initial angles are
θ 12 = 0.8 , θ 13 = 1.2 , θ 23 = arccos ( cos 0.8 cos 1.2 ) ≈ 1.3156. (32) \theta_{12}=0.8,\qquad\theta_{13}=1.2,\qquad\theta_{23}=\arccos(\cos0.8\cos1.2)\approx1.3156. \tag*{(32)} θ 12 = 0.8 , θ 13 = 1.2 , θ 23 = arccos ( cos 0.8 cos 1.2 ) ≈ 1.3156. (32)
Thus the closest pair is ( 1 , 2 ) (1,2) ( 1 , 2 ) , not ( 1 , 3 ) (1,3) ( 1 , 3 ) . But the favorable edge of ( 1 , 3 ) (1,3) ( 1 , 3 ) has barrier
B 0 = g 31 ( U 0 ) = 1 − cos 1.2 − 0.6 sin 1.2 ≈ 0.07841879. (33) B_0=g_{31}(U_0)=1-\cos1.2-0.6\sin1.2\approx0.07841879. \tag*{(33)} B 0 = g 31 ( U 0 ) = 1 − cos 1.2 − 0.6 sin 1.2 ≈ 0.07841879. (33)
By contrast g 12 = g 21 = 1 − cos 0.8 ≈ 0.30329329 g_{12}=g_{21}=1-\cos0.8\approx0.30329329 g 12 = g 21 = 1 − cos 0.8 ≈ 0.30329329 .
For arbitrary moving points v v v on the x y xy x y -arc 0 ≤ ϕ ≤ 1.2 0\le\phi\le1.2 0 ≤ ϕ ≤ 1.2 , the external scores against the fixed u 2 u_2 u 2 satisfy
v T M u 2 = cos 0.8 ( cos ϕ + 0.6 sin ϕ ) ≤ cos 0.8 1 + 0.6 2 , (34) v^{\mathsf T}M u_2=\cos0.8(\cos\phi+0.6\sin\phi)\le\cos0.8\sqrt{1+0.6^2}, \tag*{(34)} v T M u 2 = cos 0.8 ( cos ϕ + 0.6 sin ϕ ) ≤ cos 0.8 1 + 0. 6 2 , (34)
while u 2 T M v = cos 0.8 ( cos ϕ − 0.6 sin ϕ ) ≤ cos 0.8 u_2^{\mathsf T}Mv=\cos0.8(\cos\phi-0.6\sin\phi)\le\cos0.8 u 2 T M v = cos 0.8 ( cos ϕ − 0.6 sin ϕ ) ≤ cos 0.8 . Consequently
g e x t e r n a l ≥ 1 − cos 0.8 1 + 0.6 2 ≈ 0.187507 > B 0 . (35) g_{\mathrm{external}}\ge1-\cos0.8\sqrt{1+0.6^2}\approx0.187507>B_0. \tag*{(35)} g external ≥ 1 − cos 0.8 1 + 0. 6 2 ≈ 0.187507 > B 0 . (35)
All external angular gaps are at least 0.8 0.8 0.8 . With θ ∗ = 0.3 \theta_\ast=0.3 θ ∗ = 0.3 , theorem D.2 applies and guarantees
τ 13 ( 0.3 ) < min { τ 12 ( 0.3 ) , τ 23 ( 0.3 ) } for all sufficiently large β , (36) \tau_{13}(0.3)<\min\{\tau_{12}(0.3),\tau_{23}(0.3)\}\quad\text{for all sufficiently large }\beta, \tag*{(36)} τ 13 ( 0.3 ) < min { τ 12 ( 0.3 ) , τ 23 ( 0.3 )} for all sufficiently large β , (36)
with lim β → ∞ β − 1 log τ 13 ( 0.3 ) = B 0 \lim_{\beta\to\infty}\beta^{-1}\log\tau_{13}(0.3)=B_0 lim β → ∞ β − 1 log τ 13 ( 0.3 ) = B 0 .
Along the isolated directed corridor with angular separation θ \theta θ , the favorable internal barrier is g 31 ( θ ) = 1 − cos θ − 0.6 sin θ g_{31}(\theta)=1-\cos\theta-0.6\sin\theta g 31 ( θ ) = 1 − cos θ − 0.6 sin θ , and its maximum on [ 0.3 , 1.2 ] [0.3,1.2] [ 0.3 , 1.2 ] is its value at θ = 1.2 \theta=1.2 θ = 1.2 . Its opposite-directed counterpart is g 13 ( θ ) = 1 − cos θ + 0.6 sin θ g_{13}(\theta)=1-\cos\theta+0.6\sin\theta g 13 ( θ ) = 1 − cos θ + 0.6 sin θ . Figure 4 compares these two functions with the uniform external-barrier lower bound in (35) .
Figure 4. The entire corridor, rather than the initial ordering alone, certifies the first event. The external barrier stays strictly above the favorable internal barrier’s initial maximum.
Theorem E.1 (Two separate exponential clocks in the explicit configuration). For the initial data (31) let t 1 ( β ) t_1(\beta) t 1 ( β ) be the first crossing of angle 0.3 0.3 0.3 and
t 2 ( β ) = inf { t : min ( θ 12 ( t ) , θ 23 ( t ) ) ≤ 0.3 } . (37) t_2(\beta)=\inf\{t:\min(\theta_{12}(t),\theta_{23}(t))\le0.3\}. \tag*{(37)} t 2 ( β ) = inf { t : min ( θ 12 ( t ) , θ 23 ( t )) ≤ 0.3 } . (37) Then t 2 ( β ) < ∞ t_2(\beta)<\infty t 2 ( β ) < ∞ for every β > 0 \beta>0 β > 0 , and with G = 1 − cos ( 0.8 ) 1 + 0.6 2 G=1-\cos(0.8)\sqrt{1+0.6^2} G = 1 − cos ( 0.8 ) 1 + 0. 6 2 ,
lim inf β → ∞ log t 2 ( β ) β ≥ G , lim inf β → ∞ log ( t 2 / t 1 ) β ≥ G − B 0 > 0. (38) \liminf_{\beta\to\infty}\frac{\log t_2(\beta)}{\beta}\ge G,\qquad\liminf_{\beta\to\infty}\frac{\log(t_2/t_1)}{\beta}\ge G-B_0>0. \tag*{(38)} β → ∞ lim inf β log t 2 ( β ) ≥ G , β → ∞ lim inf β log ( t 2 / t 1 ) ≥ G − B 0 > 0. (38) Thus, the first event and any event involving the external third point are separated on the exponential scale. This earlier lower bound is sharpened to the exact exponent by Corollary K.5 below.
Proof. The three initial scalar products are positive, so t 2 < ∞ t_2<\infty t 2 < ∞ by theorem B.1. It remains to prove the lower bound. Let A \mathcal{A} A denote the entire planar arc of angles [ 0 , 1.2 ] [0,1.2] [ 0 , 1.2 ] in the x y xy x y -plane, and fix any a < G a<G a < G . In a small fixed neighborhood of configurations in which u 1 , u 3 ∈ A u_1,u_3\in\mathcal{A} u 1 , u 3 ∈ A and u 2 u_2 u 2 retains its initial position, every external directed barrier is at least G − δ G-\delta G − δ for arbitrarily small fixed δ > 0 \delta>0 δ > 0 . Their total transition force is bounded by η = C e − β ( G − δ ) \eta=Ce^{-\beta(G-\delta)} η = C e − β ( G − δ ) .
Here it is essential to control the two internal points for a time much longer than their first event. Write z i z_i z i for their coordinates normal to the x y xy x y -plane, and let E = max ( ∣ z 1 ∣ , ∣ z 3 ∣ ) E=\max(|z_1|,|z_3|) E = max ( ∣ z 1 ∣ , ∣ z 3 ∣ ) . The internal portion of their equations is u ˙ 1 = p 13 ( u 3 − c u 1 ) \dot{u}_1=p_{13}(u_3-cu_1) u ˙ 1 = p 13 ( u 3 − c u 1 ) and u ˙ 3 = p 31 ( u 1 − c u 3 ) \dot{u}_3=p_{31}(u_1-cu_3) u ˙ 3 = p 31 ( u 1 − c u 3 ) , where c = u 1 u 3 c=u_1u_3 c = u 1 u 3 . At any active maximum of ∣ z i ∣ |z_i| ∣ z i ∣ , positivity of p i j p_{ij} p ij gives the upper Dini-derivative bound
D + E ≤ q ( 1 − c ) E + C η , q = p 13 + p 31 . (39) D^+E\le q(1-c)E+C\eta,\qquad q=p_{13}+p_{31}. \tag*{(39)} D + E ≤ q ( 1 − c ) E + C η , q = p 13 + p 31 . (39)
Meanwhile f = 1 − c f=1-c f = 1 − c satisfies f ˙ = − q ( 1 − c 2 ) + R \dot{f}=-q(1-c^2)+R f ˙ = − q ( 1 − c 2 ) + R , with ∣ R ∣ ≤ C η |R|\le C\eta ∣ R ∣ ≤ C η . As long as c ≥ 0 c\ge0 c ≥ 0 ,
∫ 0 t q ( s ) ( 1 − c ( s ) ) d s ≤ ∫ 0 t q ( s ) ( 1 − c ( s ) 2 ) d s ≤ f ( 0 ) + C η t . (40) \int_0^t q(s)(1-c(s))\,\mathrm{d}s\le\int_0^t q(s)(1-c(s)^2)\,\mathrm{d}s\le f(0)+C\eta t. \tag*{(40)} ∫ 0 t q ( s ) ( 1 − c ( s )) d s ≤ ∫ 0 t q ( s ) ( 1 − c ( s ) 2 ) d s ≤ f ( 0 ) + C η t . (40)
Gronwall therefore yields E ( t ) ≤ C η t exp ( f ( 0 ) + C η t ) E(t)\le C\eta t\exp(f(0)+C\eta t) E ( t ) ≤ C η t exp ( f ( 0 ) + C η t ) .
The planar angular coordinates of u 1 , u 3 u_1,u_3 u 1 , u 3 satisfy cooperative equations: at the smaller angle the internal contribution is nonnegative, and at the larger it is nonpositive. The external contributions are O ( η ) O(\eta) O ( η ) while the planar radii stay bounded away from zero. Thus the smaller angle remains above − C η t -C\eta t − C η t and the larger below 1.2 + C η t 1.2+C\eta t 1.2 + C η t . Also ∥ u 2 ( t ) − u 2 ( 0 ) ∥ ≤ C η t \lVert u_2(t)-u_2(0)\rVert\le C\eta t ∥ u 2 ( t ) − u 2 ( 0 )∥ ≤ C η t . These estimates close a tubular bootstrap for 0 ≤ t ≤ e β a 0 \le t \le e^{\beta a} 0 ≤ t ≤ e β a whenever δ < G − a \delta< G-a δ < G − a , since η e β a → 0 \eta e^{\beta a} \to0 η e β a → 0 . Throughout this interval, u 2 u_2 u 2 stays at angular distance at least 0.8 − o ( 1 ) > 0.3 0.8-o(1)>0.3 0.8 − o ( 1 ) > 0.3 from both other points. Consequently t 2 ≥ e β a t_2 \ge e^{\beta a} t 2 ≥ e β a for all large β \beta β . Letting a ↑ G a \uparrow G a ↑ G proves the first inequality in (38) ; the second follows from theorem D.2.
For the isolated internal pair, the boundary Laplace method yields the leading prefactor
τ 13 ( 0.3 ) ∼ e β B 0 β sin ( 1.2 ) [ sin ( 1.2 ) − 0.6 cos ( 1.2 ) ] ≈ 1.50137 β e 0.07841879 β . (41) \tau_{13}(0.3) \sim\frac{e^{\beta B_0}}{\beta\sin(1.2)\left[\sin(1.2)-0.6\cos(1.2)\right]} \approx\frac{1.50137}{\beta}e^{0.07841879\beta}. \tag*{(41)} τ 13 ( 0.3 ) ∼ β sin ( 1.2 ) [ sin ( 1.2 ) − 0.6 cos ( 1.2 ) ] e β B 0 ≈ β 1.50137 e 0.07841879 β . (41)
The reason is that the maximal favorable barrier occurs uniquely at the initial endpoint and its derivative there is strictly positive. All external effects in the corridor are exponentially smaller, even after multiplication by β \beta β . A complete relative phase-speed comparison is needed to transfer this isolated endpoint calculation to the full ODE; the logarithmic first-event theorem alone does not imply the prefactor. Equation (41) should therefore be read as an isolated-pair asymptotic and a full-flow prediction pending that comparison.
Rotating relative equilibria and the absence of universal coarsening
Set d = 2 d=2 d = 2 , n = 3 n=3 n = 3 , equal masses, and A = J A=J A = J . At equally spaced angles
θ j ( t ) = ω t + 2 π j 3 , j = 0 , 1 , 2 , (42) \theta_j(t)=\omega t+\frac{2\pi j}{3},\qquad j=0,1,2, \tag*{(42)} θ j ( t ) = ω t + 3 2 π j , j = 0 , 1 , 2 , (42)
the angular ODE induced by (9) gives
ω = − 3 e − β / 2 sinh ( 3 β ε / 2 ) e β + 2 e − β / 2 cosh ( 3 β ε / 2 ) . (43) \omega=-\frac{\sqrt{3}e^{-\beta/2}\sinh\left(\sqrt{3}\beta\varepsilon/2\right)}{e^\beta+2e^{-\beta/2}\cosh\left(\sqrt{3}\beta\varepsilon/2\right)}. \tag*{(43)} ω = − e β + 2 e − β /2 cosh ( 3 β ε /2 ) 3 e − β /2 sinh ( 3 β ε /2 ) . (43)
For ε ≠ 0 \varepsilon\ne0 ε = 0 this is a nonconstant periodic orbit, with u i T u j = − 1 / 2 u_i^{\mathsf T}u_j=-1/2 u i T u j = − 1/2 for all i ≠ j i\ne j i = j . It is a rotating relative equilibrium , not a merger. In particular no unconditional theorem claiming that all separated configurations coarsen can hold. Nor can the full vector field be the gradient of a time-independent scalar potential for a positive-definite Riemannian metric, because a nontrivial periodic orbit would contradict strict energy monotonicity.
For ∣ β ε ∣ ≪ 1 \lvert\beta\varepsilon\rvert\ll1 ∣ β ε ∣ ≪ 1 , the angular velocity obeys
ω = − 3 β ε 2 e − 3 β / 2 1 + 2 e − 3 β / 2 + O ( ( β ε ) 3 e − 3 β / 2 ) , (44) \omega=-\frac{3\beta\varepsilon}{2}\frac{e^{-3\beta/2}}{1+2e^{-3\beta/2}}+O\left((\beta\varepsilon)^3e^{-3\beta/2}\right), \tag*{(44)} ω = − 2 3 β ε 1 + 2 e − 3 β /2 e − 3 β /2 + O ( ( β ε ) 3 e − 3 β /2 ) , (44)
exhibiting a potentially exponentially long rotation time scale, distinct from an exponentially long merger time. Figure 5 depicts the resulting relative equilibrium.
Figure 5. A nontrivial periodic orbit in the nonreciprocal three-body flow. Arrows indicate the direction for one sign of ε \varepsilon ε ; reversing ε \varepsilon ε reverses rotation.
Exact rotating-triangle spectrum
The angular equation is θ ˙ i = ∑ j p i j sin ( θ j − θ i ) \dot{\theta}_i=\sum_j p_{ij}\sin(\theta_j-\theta_i) θ ˙ i = ∑ j p ij sin ( θ j − θ i ) . The rotating triangular relative equilibrium has an exact spectrum; its neutral mode must not be confused with asymptotic stability.
Theorem F.1 (Strict relative instability of the rotating triangle). For every β > 0 \beta>0 β > 0 and ε ∈ R \varepsilon\in\mathbb{R} ε ∈ R , the equilateral relative equilibrium on $\mathbb{S}^1 has one zero angular eigenvalue (common rotation) and two eigenvalues with strictly positive real part. Let
r = 3 2 , a = e − β ( 3 / 2 + ε r ) 1 + e − β ( 3 / 2 + ε r ) + e − β ( 3 / 2 − ε r ) , b = e − β ( 3 / 2 − ε r ) 1 + e − β ( 3 / 2 + ε r ) + e − β ( 3 / 2 − ε r ) . (45) r=\frac{\sqrt{3}}{2},\qquad a=\frac{e^{-\beta(3/2+\varepsilon r)}}{1+e^{-\beta(3/2+\varepsilon r)}+e^{-\beta(3/2-\varepsilon r)}},\qquad b=\frac{e^{-\beta(3/2-\varepsilon r)}}{1+e^{-\beta(3/2+\varepsilon r)}+e^{-\beta(3/2-\varepsilon r)}}. \tag*{(45)} r = 2 3 , a = 1 + e − β ( 3/2 + ε r ) + e − β ( 3/2 − ε r ) e − β ( 3/2 + ε r ) , b = 1 + e − β ( 3/2 + ε r ) + e − β ( 3/2 − ε r ) e − β ( 3/2 − ε r ) . (45) Then ω = r ( a − b ) \omega=r(a-b) ω = r ( a − b ) and, with
ℓ + = a [ − 1 2 + β ( − r + ε 2 ) ( r − ω ) ] , (46) \ell_+=a\left[-\frac{1}{2}+\beta\left(-r+\frac{\varepsilon}{2}\right)(r-\omega)\right], \tag*{(46)} ℓ + = a [ − 2 1 + β ( − r + 2 ε ) ( r − ω ) ] , (46) ℓ − = b [ − 1 2 + β ( r + ε 2 ) ( − r − ω ) ] , (47) \ell_-=b\left[-\frac{1}{2}+\beta\left(r+\frac{\varepsilon}{2}\right)(-r-\omega)\right], \tag*{(47)} ℓ − = b [ − 2 1 + β ( r + 2 ε ) ( − r − ω ) ] , (47) the eigenvalues are
λ 0 = 0 , λ ± = − 3 2 ( ℓ + + ℓ − ) ± i 3 2 ( ℓ + − ℓ − ) . (48) \lambda_0=0,\qquad\lambda_\pm=-\frac{3}{2}(\ell_++\ell_-)\pm i\frac{\sqrt{3}}{2}(\ell_+-\ell_-). \tag*{(48)} λ 0 = 0 , λ ± = − 2 3 ( ℓ + + ℓ − ) ± i 2 3 ( ℓ + − ℓ − ) . (48) In fact Re λ ± ≥ 3 4 ( a + b ) > 0 \operatorname{Re}\lambda_\pm\ge\frac{3}{4}(a+b)>0 Re λ ± ≥ 4 3 ( a + b ) > 0 .
Proof. At each equilateral configuration, the two off-diagonal derivatives in the angular Jacobian are ℓ + , ℓ − \ell_+,\ell_- ℓ + , ℓ − . Rotational equivariance makes its diagonal entry − ( ℓ + + ℓ − ) -(\ell_++\ell_-) − ( ℓ + + ℓ − ) and its circulant eigenvalues give (48) . A rearrangement yields
ℓ + + ℓ − = − ( 1 2 + 3 β 4 ) ( a + b ) + β ω 2 + β ε ω 2 ( 1 − a − b ) . (49) \ell_++\ell_-=-\left(\frac{1}{2}+\frac{3\beta}{4}\right)(a+b)+\beta\omega^2+\frac{\beta\varepsilon\omega}{2}(1-a-b). \tag*{(49)} ℓ + + ℓ − = − ( 2 1 + 4 3 β ) ( a + b ) + β ω 2 + 2 β ε ω ( 1 − a − b ) . (49)
Since ε ω ≤ 0 \varepsilon\omega\le0 ε ω ≤ 0 , a + b ≤ 1 a+b\le1 a + b ≤ 1 , and ω 2 = 3 4 ( a − b ) 2 ≤ 3 4 ( a + b ) \omega^2=\frac{3}{4}(a-b)^2\le\frac{3}{4}(a+b) ω 2 = 4 3 ( a − b ) 2 ≤ 4 3 ( a + b ) , we conclude ℓ + + ℓ − ≤ − ( a + b ) / 2 \ell_++\ell_-\le-(a+b)/2 ℓ + + ℓ − ≤ − ( a + b ) /2 . The asserted strict real-part bound follows.
Remark F.2. On S 1 \mathbb{S}^1 S 1 with three points there are precisely three angular eigenvalues: one zero and the conjugate pair. There is no additional independent transverse mode. In the symmetric case ε = 0 \varepsilon=0 ε = 0 , the two nonzero eigenvalues coincide and are positive. Their smallness for large β \beta β describes a weakly unstable , rather than attracting, triangular orbit.
What happens after the first event?
The next question is whether the dynamics admits a rigorous recursive reduction. We can prove the exact reduction on a collision manifold, but not yet the requisite shadowing of this manifold by near-collisions.
Proposition G.1 (Exact weighted quotient on collision strata). Let P = { I 1 , … , I k } P=\{I_1,\ldots,I_k\} P = { I 1 , … , I k } be a partition of { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } . The stratum
M P = { U : u i = v a for each i ∈ I a } (50) \mathcal{M}_P=\{U:u_i=v_a\text{ for each }i\in I_a\} \tag*{(50)} M P = { U : u i = v a for each i ∈ I a } (50) is flow-invariant. On this stratum the exact reduced equations for v a v_a v a have the form (9) , with the combined masses
m ~ a = ∑ i ∈ I a m i . (51) \widetilde{m}_a=\sum_{i\in I_a}m_i. \tag*{(51)} m a = i ∈ I a ∑ m i . (51)
Proof. If u i = u j u_i=u_j u i = u j , the i i i th and j j j th rows of the attention kernel have the same denominator and induce the same velocity; uniqueness maintains equality. Terms in the same block of a numerator add as ∑ j ∈ I b m j e β v a T M v b = m ~ b e β v a T M v b \sum_{j\in I_b}m_j e^{\beta v_a^{\mathsf T}M v_b}=\widetilde{m}_b e^{\beta v_a^{\mathsf T}M v_b} ∑ j ∈ I b m j e β v a T M v b = m b e β v a T M v b , giving the reduced system.
Proposition G.2 (Uniform transverse linearization at complete consensus). At U ∗ = ( v , … , v ) U_*=(v,\ldots,v) U ∗ = ( v , … , v ) , with ξ i ∈ T v S d − 1 \xi_i\in T_vS^{d-1} ξ i ∈ T v S d − 1 , the linearized flow is
ξ ˙ i = ξ ˉ − ξ i , ξ ˉ = ∑ j m j ξ ˉ j ∑ j m j . (52) \dot{\xi}_i=\bar{\xi}-\xi_i,\qquad\bar{\xi}=\frac{\sum_j m_j\bar{\xi}_j}{\sum_j m_j}. \tag*{(52)} ξ ˙ i = ξ ˉ − ξ i , ξ ˉ = ∑ j m j ∑ j m j ξ ˉ j . (52) The mean mode has eigenvalue zero and all n − 1 n-1 n − 1 transverse modes have eigenvalue − 1 -1 − 1 , independently of β \beta β , A A A , ε \varepsilon ε .
Proof. At consensus the weighted probabilities are p i j = m j / ∑ k m k p_{ij}=m_j/\sum_k m_k p ij = m j / ∑ k m k . The derivative of p i j p_{ij} p ij multiplies v v v , which is removed by I − v v T I-vv^{\mathsf T} I − v v T . The derivative of the projection contributes − ξ i -\xi_i − ξ i , while differentiation of the value vectors contributes ξ ˉ \bar{\xi} ξ ˉ . This gives (52) .
Partial-collision stability and a boundary-layer prefactor
What the normal exponent really is
The eigenvalue − 1 -1 − 1 from theorem G.2 is not the exact normal eigenvalue of a partially collided block. Here is the correct general formula.
Proposition H.1 (Exact normal linearization of a two-point collision stratum). At a configuration with u a = u b = v u_a=u_b=v u a = u b = v (but other points arbitrary), let
w j = m j e β v T M u j ∑ k m k e β v T M u k , μ = ∑ j w j u j , P v = I − v v T . (53) w_j=\frac{m_j e^{\beta v^{\mathsf T}M u_j}}{\sum_k m_k e^{\beta v^{\mathsf T}M u_k}},\qquad\mu=\sum_j w_j u_j,\qquad P_v=I-vv^{\mathsf T}. \tag*{(53)} w j = ∑ k m k e β v T M u k m j e β v T M u j , μ = j ∑ w j u j , P v = I − v v T . (53) The covariant linearization of the relative perturbation ξ ∈ T v S d − 1 \xi\in T_vS^{d-1} ξ ∈ T v S d − 1 is
L v ξ = − ( v T μ ) ξ + β P v [ ∑ j w j ( u j − μ ) ( M ( u j − μ ) ) T ] ξ . (54) L_v\xi=-(v^{\mathsf T}\mu)\xi+\beta P_v\left[\sum_j w_j(u_j-\mu)(M(u_j-\mu))^{\mathsf T}\right]\xi. \tag*{(54)} L v ξ = − ( v T μ ) ξ + β P v [ j ∑ w j ( u j − μ ) ( M ( u j − μ ) ) T ] ξ . (54) If η = ∑ j ∉ { a , b } w j \eta=\sum_{j\notin\{a,b\}}w_j η = ∑ j ∈ / { a , b } w j , then
∥ L v + I ∥ ≤ 2 η + 4 β ∥ M ∥ op η . (55) \lVert L_v+I\rVert\le2\eta+4\beta\lVert M\rVert_{\operatorname{op}}\eta. \tag*{(55)} ∥ L v + I ∥ ≤ 2 η + 4 β ∥ M ∥ op η . (55) Thus a shield with ( 1 + β ) η = o ( 1 ) (1+\beta)\eta=o(1) ( 1 + β ) η = o ( 1 ) makes the normal eigenvalues − 1 + o ( 1 ) -1+o(1) − 1 + o ( 1 ) , not identically − 1 -1 − 1 .
Proof. With the key configuration U U U frozen, write f ( x ; U ) = ( I − x x T ) μ ( x ; U ) f(x;U)=(I-xx^{\mathsf T})\mu(x;U) f ( x ; U ) = ( I − x x T ) μ ( x ; U ) , μ ( x ; U ) = ∑ j w j ( x ; U ) u j \mu(x;U)=\sum_j w_j(x;U)u_j μ ( x ; U ) = ∑ j w j ( x ; U ) u j . The difference of the two vector fields is exactly f ( u a ; U ) − f ( u b ; U ) f(u_a;U)-f(u_b;U) f ( u a ; U ) − f ( u b ; U ) , so key variations cancel to first order in the relative mode. Differentiating the row-wise softmax gives
D x μ ( v ; U ) = β ∑ j w j ( u j − μ ) ( M ( u j − μ ) ) T . (56) D_x\mu(v;U)=\beta\sum_j w_j(u_j-\mu)(M(u_j-\mu))^{\mathsf T}. \tag*{(56)} D x μ ( v ; U ) = β j ∑ w j ( u j − μ ) ( M ( u j − μ ) ) T . (56) The tangent projection of D x f ( v ; U ) D_xf(v;U) D x f ( v ; U ) is (54) . Since the internal key vectors are v v v , we have 1 − v T μ ≤ 2 η 1-v^{\mathsf T}\mu\le2\eta 1 − v T μ ≤ 2 η and ∑ j w j ∥ u j − μ ∥ 2 = 1 − ∥ μ ∥ 2 ≤ 4 η \sum_j w_j\lVert u_j-\mu\rVert^2=1-\lVert\mu\rVert^2\le4\eta ∑ j w j ∥ u j − μ ∥ 2 = 1 − ∥ μ ∥ 2 ≤ 4 η . The operator bound follows.
Remark H.2 (Correct Lyapunov scaling). For V = 1 2 ∥ u a − u b ∥ 2 V=\frac{1}{2}\lVert u_a-u_b\rVert^2 V = 2 1 ∥ u a − u b ∥ 2 and an exact normal law ξ ˙ = − ξ \dot{\xi}=-\xi ξ ˙ = − ξ , one has V ˙ = − 2 V \dot{V}=-2V V ˙ = − 2 V , not − V -V − V . More importantly, the exact collision stratum is invariant: there is no arbitrary additive external forcing in the relative equation, because identical queries have identical velocities. A robust nonlinear tube bound still requires a uniform estimate of D x f D_xf D x f in a neighborhood, as well as control of the quotient dynamics; theorem H.1 alone is not a global shadowing theorem.
The rotating inner pair produces a nonvanishing exponential bias
Return to the three-point configuration (31) and first suppress its external point . Write u 1 = ( cos φ , sin φ , 0 ) u_1=(\cos\varphi,\sin\varphi,0) u 1 = ( cos φ , sin φ , 0 ) and u 3 = ( cos ( φ + θ ) , sin ( φ + θ ) , 0 ) u_3=(\cos(\varphi+\theta),\sin(\varphi+\theta),0) u 3 = ( cos ( φ + θ ) , sin ( φ + θ ) , 0 ) , starting at ( φ , θ ) = ( 0 , 1.2 ) (\varphi,\theta)=(0,1.2) ( φ , θ ) = ( 0 , 1.2 ) . The two internal transition probabilities are
p + ( θ ) = 1 1 + e β ( 1 − cos θ + ε sin θ ) , p − ( θ ) = 1 1 + e β ( 1 − cos θ − ε sin θ ) . (57) p_{+}(\theta)=\frac{1}{1+e^{\beta(1-\cos\theta+\varepsilon\sin\theta)}},\qquad p_{-}(\theta)=\frac{1}{1+e^{\beta(1-\cos\theta-\varepsilon\sin\theta)}}. \tag*{(57)} p + ( θ ) = 1 + e β ( 1 − c o s θ + ε s i n θ ) 1 , p − ( θ ) = 1 + e β ( 1 − c o s θ − ε s i n θ ) 1 . (57)
Consequently φ ˙ = p + sin θ \dot{\varphi}=p_{+}\sin\theta φ ˙ = p + sin θ , θ ˙ = − ( p + + p − ) sin θ \dot{\theta}=-(p_{+}+p_{-})\sin\theta θ ˙ = − ( p + + p − ) sin θ . The final center angle of the isolated pair is exactly
Φ β = ∫ 0 1.2 p + ( θ ) p + ( θ ) + p − ( θ ) d θ . (58) \Phi_{\beta}=\int_{0}^{1.2}\frac{p_{+}(\theta)}{p_{+}(\theta)+p_{-}(\theta)}\,\mathrm{d}\theta. \tag*{(58)} Φ β = ∫ 0 1.2 p + ( θ ) + p − ( θ ) p + ( θ ) d θ . (58)
Proposition H.3 (A finite bias left by the inner boundary layer). For fixed ε > 0 \varepsilon>0 ε > 0 and initial θ 0 ∈ ( 0 , π ) \theta_0\in(0,\pi) θ 0 ∈ ( 0 , π ) ,
lim β → ∞ β Φ β = log 2 ε . (59) \lim_{\beta\to\infty}\beta\Phi_{\beta}=\frac{\log2}{\varepsilon}. \tag*{(59)} β → ∞ lim β Φ β = ε log 2 . (59)
Proof. Set x = β θ x=\beta\theta x = β θ in (58) . For each fixed x ≥ 0 x\geq0 x ≥ 0 , the numerator-to-total ratio converges to ( 1 + e ε x ) − 1 (1+e^{\varepsilon x})^{-1} ( 1 + e ε x ) − 1 . The ratio is bounded by 2 e − β ε sin θ 2e^{-\beta\varepsilon\sin\theta} 2 e − β ε s i n θ ; since sin θ ≥ c θ 0 θ \sin\theta\geq c_{\theta_0}\theta sin θ ≥ c θ 0 θ on [ 0 , θ 0 ] [0,\theta_0] [ 0 , θ 0 ] for θ 0 < π \theta_0<\pi θ 0 < π , dominated convergence applies. Thus β Φ β → ∫ 0 ∞ ( 1 + e ε x ) − 1 d x = ( log 2 ) / ε \beta\Phi_{\beta}\to\int_{0}^{\infty}(1+e^{\varepsilon x})^{-1}\mathrm{d}x=(\log2)/\varepsilon β Φ β → ∫ 0 ∞ ( 1 + e ε x ) − 1 d x = ( log 2 ) / ε .
The displacement Φ β = O ( β − 1 ) \Phi_{\beta}=O(\beta^{-1}) Φ β = O ( β − 1 ) vanishes geometrically, but β ε Φ β → log 2 \beta\varepsilon\Phi_{\beta}\to\log2 β ε Φ β → log 2 and hence its influence on a directed exponential weight does not vanish .
A rigorous prefactor for the exactly collapsed quotient
Let v β = ( cos Φ β , sin Φ β , 0 ) v_{\beta}=(\cos\Phi_{\beta},\sin\Phi_{\beta},0) v β = ( cos Φ β , sin Φ β , 0 ) and w = ( cos 0.8 , 0 , sin 0.8 ) w=(\cos0.8,0,\sin0.8) w = ( cos 0.8 , 0 , sin 0.8 ) . Consider an exact two-body quotient initialized at ( v β , w ) (v_{\beta},w) ( v β , w ) , with masses ( 2 , 1 ) (2,1) ( 2 , 1 ) and the same M M M . Let T β r e d T_{\beta}^{\mathrm{red}} T β red denote its hitting time of angle 0.3 0.3 0.3 .
Theorem H.4 (Matched prefactor for the exact reduced model). For ε = 0.6 \varepsilon=0.6 ε = 0.6 and the configuration above, set c = cos 0.8 c=\cos0.8 c = cos 0.8 , s = sin 0.8 s=\sin0.8 s = sin 0.8 , G = 1 − c G=1-c G = 1 − c . Then
lim β → ∞ T β r e d e − β G = 1 s 2 ( 1 2 2 c + 2 2 − c ) ≈ 0.95054656. (60) \lim_{\beta\to\infty}T_{\beta}^{\mathrm{red}}e^{-\beta G}=\frac{1}{s^2\left(\frac{1}{2}2^c+2\,2^{-c}\right)}\approx0.95054656. \tag*{(60)} β → ∞ lim T β red e − β G = s 2 ( 2 1 2 c + 2 2 − c ) 1 ≈ 0.95054656. (60)
Proof. The initial ordinary and skew scalar products satisfy
c β = c cos Φ β = c + O ( β − 2 ) , h β = c sin Φ β = c log 2 ε β + o ( β − 1 ) . (61) c_{\beta}=c\cos\Phi_{\beta}=c+O(\beta^{-2}),\qquad h_{\beta}=c\sin\Phi_{\beta}=\frac{c\log2}{\varepsilon\beta}+o(\beta^{-1}). \tag*{(61)} c β = c cos Φ β = c + O ( β − 2 ) , h β = c sin Φ β = εβ c log 2 + o ( β − 1 ) . (61)
Thus the two directed barriers at the initial endpoint are G ∓ c ( log 2 ) / β + o ( β − 1 ) G\mp c(\log2)/\beta+o(\beta^{-1}) G ∓ c ( log 2 ) / β + o ( β − 1 ) . With masses ( 2 , 1 ) (2,1) ( 2 , 1 ) , their two probabilities obey
e β G ( q v w + q w v ) ⟶ K = 1 2 2 c + 2 2 − c . (62) e^{\beta G}(q_{vw}+q_{wv})\longrightarrow K=\frac{1}{2}2^c+2\,2^{-c}. \tag*{(62)} e β G ( q v w + q w v ) ⟶ K = 2 1 2 c + 2 2 − c . (62)
The two-body first integral (22) preserves the skew product to relative order O ( β − 1 ) O(\beta^{-1}) O ( β − 1 ) in the O ( β − 1 ) O(\beta^{-1}) O ( β − 1 ) endpoint layer. In (24) , the maximal exponential barrier occurs at its initial endpoint with derivative − 1 + O ( β − 1 ) -1+O(\beta^{-1}) − 1 + O ( β − 1 ) in the c c c coordinate. The endpoint Laplace method gives T β r e d ∼ e β G / [ β ( 1 − c 2 ) K ] T_{\beta}^{\mathrm{red}}\sim e^{\beta G}/[\beta(1-c^2)K] T β red ∼ e β G / [ β ( 1 − c 2 ) K ] , proving the claim.
Remark H.5 (Crucial scope distinction). This theorem is exact for the constructed reduced two-body dynamics . To identify T β r e d T_{\beta}^{\mathrm{red}} T β red with the second-event clock t 2 t_2 t 2 of the original three-point flow up to 1 + o ( 1 ) 1+o(1) 1 + o ( 1 ) , one must still prove quantitative post-event shadowing and a controlled event crossing. The observed diagnostic Q β = β t 2 e − β G Q_{\beta}=\beta t_2e^{-\beta G} Q β = β t 2 e − β G approaching 1 from above at the available β \beta β does not imply a limiting value of 1: the reduced model predicts the different limit 0.95054656.
Table 9. The full three-body numerical diagnostic compared to the exact two-body quotient quadrature; neither finite table implies equality of the asymptotic limits.
Figure 6 extends the quotient quadrature to larger β \beta β and displays the distinction between the reduced-model limit and finite full-flow measurements.
Figure 6. The boundary layer produces an order-one skew-weight correction. The numerically calculated exact quotient tends toward its proved limit, while finite-β \beta β full dynamics closely follow the reduced clock. A proof of uniform full-flow matching remains open.
The remaining nonlinear shadowing bridge
The first threshold is defined at angle 0.3, so at the event time itself V = 1 − cos ( 0.3 ) V=1-\cos(0.3) V = 1 − cos ( 0.3 ) : the trajectory is not yet within o ( β − 1 ) o(\beta^{-1}) o ( β − 1 ) of the collision stratum. A separate post-threshold relaxation interval is necessary. Once the directed inner rate is bounded below, reaching a polynomially small transverse scale takes O ( log β ) O(\log\beta) O ( log β ) time; reaching an exponentially small one may take O ( β ) O(\beta) O ( β ) time. Thus the fast time variable τ = β t \tau=\beta t τ = β t cannot be asserted universally from the collision eigenvalue, which is of order one under a shield.
A promising route for proving equality between full and quotient prefactors is to use the integrated interaction rate rather than a global Lipschitz estimate. For the exact two-body quotient, ψ ˙ = − q ( ψ ) sin ψ \dot{\psi}=-q(\psi)\sin\psi ψ ˙ = − q ( ψ ) sin ψ gives
∫ t 0 T q ( ψ ( t ) ) d t = ∫ ψ ( T ) ψ ( t 0 ) d ψ sin ψ = O ( 1 ) (63) \int_{t_0}^{T}q(\psi(t))\,\mathrm{d}t=\int_{\psi(T)}^{\psi(t_0)}\frac{\mathrm{d}\psi}{\sin\psi}=O(1) \tag*{(63)} ∫ t 0 T q ( ψ ( t )) d t = ∫ ψ ( T ) ψ ( t 0 ) sin ψ d ψ = O ( 1 ) (63)
when ψ \psi ψ stays away from 0 , π 0,\pi 0 , π . Hence the integral of a Jacobian bound of order β q \beta q β q is only O ( β ) O(\beta) O ( β ) despite the exponential physical time interval. If a transverse discrepancy has already decreased to e − K β e^{-K\beta} e − K β with sufficiently large K K K , the corresponding variation-of-constants estimate can potentially be closed. This is an approach , not yet a verified full-flow shadowing theorem; the required coupling and event-transversality inequalities must be supplied.
What the Gram pencil does not prove
When n = d n=d n = d and U U U is invertible, C − 1 H = U − T A U T C^{-1}H=U^{-\mathsf{T}}AU^{\mathsf{T}} C − 1 H = U − T A U T is similar to the fixed matrix A A A . Thus its spectrum, for instance { i , − i , 0 } \{i,-i,0\} { i , − i , 0 } when A A A is the coordinate-plane rotation in three dimensions, is determined by A A A itself and is not a new angular-momentum law or a merger-order selector. For n < d n<d n < d the conserved pencil describes the restriction of A A A to the time-independent row span; for n > d n>d n > d it is singular and requires a reduced pencil. Finite-time congruence preserves rank, but a rank-deficient limit is possible because the congruence transformations can become ill-conditioned as t → ∞ t\to\infty t → ∞ .
Open problem H.6 (Rigorous second-merger shadowing). Suppose a first pair has entered a neighborhood of its collision stratum, while the remaining directed barriers stay separated. Find verifiable geometric and spectral hypotheses that imply, on a time interval of length exp ( β H ) \exp(\beta H) exp ( β H ) ,
sup 0 ≤ t ≤ e β H dist ( U ( t ) , M P ) = o ( β − 1 ) , (64) \sup_{0\leq t\leq e^{\beta H}} \operatorname{dist}(U(t),\mathcal{M}_{P})=o(\beta^{-1}), \tag*{(64)} 0 ≤ t ≤ e β H sup dist ( U ( t ) , M P ) = o ( β − 1 ) , (64) and that the induced quotient trajectory is sufficiently accurate to preserve the identity and exponent of the second threshold crossing. Can the hypotheses be propagated after a second crossing?
The scale o ( β − 1 ) o(\beta^{-1}) o ( β − 1 ) is motivated by the exponential sensitivity exp ( − β g ) \exp(-\beta g) exp ( − β g ) to barrier errors: an O ( β − 1 ) O(\beta^{-1}) O ( β − 1 ) displacement may induce an order-one multiplicative change in transition rates. Even if (64) holds, a separate transversality condition is needed to ensure that a small perturbation of a trajectory preserves the identity of a near-simultaneous merger.
Numerical two-event separation
We integrated (9) in the concrete example (31) using a high-accuracy adaptive DOP853 solver with stable row-wise softmax and event detection at angle 0.3. The first event is ( 1 , 3 ) (1,3) ( 1 , 3 ) . The second event is the first time another pair crosses the same threshold, with no forced collision or merging applied in the ODE. Table 10 and Figure 7 report the measured separation of these two events.
Table 10. Numerical threshold events for (31); these data are not a theorem about the second-event asymptotic.
Figure 7. Two distinct numerical clocks. The first clock is covered by theorem D.2; the rapidly growing second clock still lacks a rigorous general geometric selection theorem.
The sharp second-clock exponent
The exact weighted quotient in theorem G.1 suggests a specific exponent for the second event. If the first pair becomes a stable two-mass block at a point converging to u 1 ( 0 ) u_1(0) u 1 ( 0 ) as β → ∞ \beta\to\infty β → ∞ , then its angle from the remaining point is asymptotically 0.8 0.8 0.8 , and its limiting directed skew product is zero. This motivated the formerly open identity, now proved in Corollary K.5 ,
lim β → ∞ log t 2 ( β ) β = 1 − cos 0.8 ≈ 0.30329329. (65) \lim_{\beta\to\infty} \frac{\log t_2(\beta)}{\beta} = 1-\cos0.8 \approx0.30329329. \tag*{(65)} β → ∞ lim β log t 2 ( β ) = 1 − cos 0.8 ≈ 0.30329329. (65)
The diagnostic Q β = β t 2 exp [ − β ( 1 − cos 0.8 ) ] Q_\beta= \beta t_2 \exp[-\beta(1-\cos0.8)] Q β = β t 2 exp [ − β ( 1 − cos 0.8 )] takes the values 1.174, 1.067, 1.029 and 1.011 at β = 20 , 30 , 40 , 50 \beta= 20,30,40,50 β = 20 , 30 , 40 , 50 . The rigorously calculated exact-quotient prefactor in theorem H.4 is 0.95054656, not 1 and not the naive zero-skew quotient value 1 / [ 5 2 sin 2 ( 0.8 ) ] ≈ 0.77730 1/[\frac{5}{2}\sin^2(0.8)] \approx0.77730 1/ [ 2 5 sin 2 ( 0.8 )] ≈ 0.77730 . That numerical match, even at very high relative accuracy, does not yet prove that the full three-body t 2 t_2 t 2 has the same prefactor. The full-flow second exponent is now given by Corollary K.5 ; the prefactor remains unresolved.
A proof audit of second-clock preservation
The claim that a small normal distance from a collision stratum automatically preserves the next hitting-time prefactor is false. We separate a proved transverse estimate , a proved conditional scalar comparison , and the still unresolved tangential coupling of the full and quotient dynamics. This distinction is essential when the desired relative error is 1 + o ( 1 ) 1+o(1) 1 + o ( 1 ) rather than merely the correct exponential rate.
A nonlinear contraction bound in a shielded tube
Fix an ordered pair of indices a ≠ b a \ne b a = b , put u = u a u=u_a u = u a , v = u b v=u_b v = u b , and let
V ( t ) = 1 − u ( t ) T v ( t ) = 1 2 ∥ u ( t ) − v ( t ) ∥ 2 , K = ∥ M ∥ o p ≥ 1. (66) V(t)=1-u(t)^{\mathsf T}v(t)=\frac{1}{2}\lVert u(t)-v(t)\rVert^2,\qquad K=\lVert M\rVert_{\mathrm{op}}\ge1. \tag*{(66)} V ( t ) = 1 − u ( t ) T v ( t ) = 2 1 ∥ u ( t ) − v ( t ) ∥ 2 , K = ∥ M ∥ op ≥ 1. (66)
For a frozen configuration U U U , extend the attention law to query points x ∈ R d x \in\mathbb{R}^d x ∈ R d by
w j ( x ; U ) = m j e β x T M u j ∑ ℓ m ℓ e β x T M u ℓ , μ ( x ; U ) = ∑ j w j ( x ; U ) u j . (67) w_j(x;U) = \frac{m_j e^{\beta x^\mathsf{T} M u_j}}{\sum_\ell m_\ell e^{\beta x^\mathsf{T} M u_\ell}}, \qquad\mu(x;U) = \sum_j w_j(x;U)u_j. \tag*{(67)} w j ( x ; U ) = ∑ ℓ m ℓ e β x T M u ℓ m j e β x T M u j , μ ( x ; U ) = j ∑ w j ( x ; U ) u j . (67)
Define the segment shield
η ( t ) = sup 0 ≤ s ≤ 1 ∑ j ∉ { a , b } w j ( s u ( t ) + ( 1 − s ) v ( t ) ; U ( t ) ) . (68) \eta(t) = \sup_{0 \le s \le1} \sum_{j \notin\{a,b\}} w_j(su(t) + (1-s)v(t);U(t)). \tag*{(68)} η ( t ) = 0 ≤ s ≤ 1 sup j ∈ / { a , b } ∑ w j ( s u ( t ) + ( 1 − s ) v ( t ) ; U ( t )) . (68)
Note that this is an assumption on all intermediate queries, stronger than checking the external weights only at u u u and v v v .
Theorem I.1 (Explicit nonlinear transverse contraction under a segment shield). Suppose β ≥ 1 \beta\ge1 β ≥ 1 and, for an interval [ t 0 , T ] [t_0,T] [ t 0 , T ] , the exact solution of (9) obeys
V ( t 0 ) ≤ 1 32 β K , η ( t ) ≤ 1 64 ( 1 + β K ) ( t ∈ [ t 0 , T ] ) . (69) V(t_0) \le\frac{1}{32\beta K}, \qquad\eta(t) \le\frac{1}{64(1+\beta K)} \quad(t \in[t_0,T]). \tag*{(69)} V ( t 0 ) ≤ 32 β K 1 , η ( t ) ≤ 64 ( 1 + β K ) 1 ( t ∈ [ t 0 , T ]) . (69) Then
V ( t ) ≤ V ( t 0 ) e − ( t − t 0 ) , ∥ u ( t ) − v ( t ) ∥ ≤ 2 V ( t 0 ) e − ( t − t 0 ) / 2 (70) V(t) \le V(t_0)e^{-(t-t_0)}, \qquad\lVert u(t)-v(t)\rVert\le\sqrt{2V(t_0)}e^{-(t-t_0)/2} \tag*{(70)} V ( t ) ≤ V ( t 0 ) e − ( t − t 0 ) , ∥ u ( t ) − v ( t )∥ ≤ 2 V ( t 0 ) e − ( t − t 0 ) /2 (70) for every t ∈ [ t 0 , T ] t \in[t_0,T] t ∈ [ t 0 , T ] . The time T T T may depend on β \beta β and may be exponentially large; the only long-time assumption is that the segment shield continues to hold.
Proof. Let D = u − v D=u-v D = u − v , and suppress the fixed configuration U ( t ) U(t) U ( t ) in the notation. Direct differentiation of (66) using (9) gives the exact identity
V ˙ = D T [ μ ( u ) − μ ( v ) ] − V [ u T μ ( u ) + v T μ ( v ) ] . (71) \dot{V}=D^\mathsf{T}[\mu(u)-\mu(v)]-V[u^\mathsf{T}\mu(u)+v^\mathsf{T}\mu(v)]. \tag*{(71)} V ˙ = D T [ μ ( u ) − μ ( v )] − V [ u T μ ( u ) + v T μ ( v )] . (71)
For a query x x x , the softmax Jacobian is the covariance operator
D x μ ( x ; U ) = β ( ∑ j w j ( x ) u j u j T − μ ( x ) μ ( x ) T ) M T . (72) D_x\mu(x;U)=\beta\left(\sum_j w_j(x)u_j u_j^\mathsf{T}-\mu(x)\mu(x)^\mathsf{T}\right)M^\mathsf{T}. \tag*{(72)} D x μ ( x ; U ) = β ( j ∑ w j ( x ) u j u j T − μ ( x ) μ ( x ) T ) M T . (72)
On the segment between u u u and v v v , use v v v as a comparison center. Since the covariance is positive semidefinite and its operator norm does not exceed its trace,
∥ ∑ j w j ( x ) u j u j T − μ ( x ) μ ( x ) T ∥ o p ≤ ∑ j w j ( x ) ∥ u j − v ∥ 2 ≤ 2 V + 4 η ( t ) . (73) \left\lVert\sum_j w_j(x)u_j u_j^\mathsf{T}-\mu(x)\mu(x)^\mathsf{T}\right\rVert_{\mathrm{op}}
\le\sum_j w_j(x)\lVert u_j-v\rVert^2
\le2V+4\eta(t). \tag*{(73)} j ∑ w j ( x ) u j u j T − μ ( x ) μ ( x ) T op ≤ j ∑ w j ( x ) ∥ u j − v ∥ 2 ≤ 2 V + 4 η ( t ) . (73)
By the mean-value formula,
∥ μ ( u ) − μ ( v ) ∥ ≤ β K ( 2 V + 4 η ) ∥ D ∥ . (74) \lVert\mu(u)-\mu(v)\rVert\le\beta K(2V+4\eta)\lVert D\rVert. \tag*{(74)} ∥ μ ( u ) − μ ( v )∥ ≤ β K ( 2 V + 4 η ) ∥ D ∥ . (74)
The total weight on the two internal keys is at least 1 − η 1-\eta 1 − η , each internal key has inner product at least 1 − V 1-V 1 − V with its own endpoint query, and external keys have inner product at least − 1 -1 − 1 . Hence
u T μ ( u ) ≥ 1 − V − 2 η , v T μ ( v ) ≥ 1 − V − 2 η . (75) u^\mathsf{T}\mu(u)\ge1-V-2\eta,\qquad v^\mathsf{T}\mu(v)\ge1-V-2\eta. \tag*{(75)} u T μ ( u ) ≥ 1 − V − 2 η , v T μ ( v ) ≥ 1 − V − 2 η . (75)
Combining (71) –(75) and ∥ D ∥ 2 = 2 V \lVert D\rVert^2=2V ∥ D ∥ 2 = 2 V yields
V ˙ ≤ [ − 2 + 2 V + 4 η + 4 β K V + 8 β K η ] V . (76) \dot{V}\le[-2+2V+4\eta+4\beta KV+8\beta K\eta]V. \tag*{(76)} V ˙ ≤ [ − 2 + 2 V + 4 η + 4 β K V + 8 β K η ] V . (76)
As long as V ≤ 1 / ( 32 β K ) V\le1/(32\beta K) V ≤ 1/ ( 32 β K ) , the hypotheses (69) make the bracket strictly smaller than − 1 -1 − 1 . The resulting decay preserves the assumed tube bound, closing the first-exit bootstrap. Integrating V ˙ ≤ − V \dot{V}\le-V V ˙ ≤ − V proves (70) . ∎
Remark I.2 (What this theorem does and does not establish). The exact collision stratum is invariant, so the relative equation has no compulsory additive O ( η ) O(\eta) O ( η ) forcing term. Theorem I.1 is a genuine uniform-in-time normal estimate. It does not establish the segment shield for the target three-body trajectory, nor does it compare the motion along the collision stratum with the quotient trajectory. The initial threshold 0.3 in our example has V = 1 − cos ( 0.3 ) ≈ 0.04466 V=1-\cos(0.3)\approx0.04466 V = 1 − cos ( 0.3 ) ≈ 0.04466 , which for arbitrarily large β \beta β does not satisfy (69) ; a separate entrance/relaxation estimate is required. Once inside the tube, an additional 3 log β 3\log\beta 3 log β of contraction gives ∥ u − v ∥ = O ( β − 2 ) \lVert u-v\rVert=O(\beta^{-2}) ∥ u − v ∥ = O ( β − 2 ) ; obtaining e − c β e^{-c\beta} e − c β from a polynomially small initial distance generally requires an O ( β ) O(\beta) O ( β ) interval, not merely O ( log β ) O(\log\beta) O ( log β ) .
Normal proximity alone cannot determine a prefactor
The next proposition gives a counterexample to inferring second-clock preservation from dist ( U , M P ) = o ( β − 1 ) \operatorname{dist}(U,\mathcal{M}_{P})=o(\beta^{-1}) dist ( U , M P ) = o ( β − 1 ) alone, even when that distance is exactly zero.
Proposition I.3 (Tangential β − 1 \beta^{-1} β − 1 shifts change a metastable clock). Consider the symmetric two-center quotient (ε = 0 \varepsilon=0 ε = 0 ) with fixed positive masses m 1 , m 2 m_1,m_2 m 1 , m 2 , initial angular separation ψ 0 ∈ ( 0 , π ) \psi_0\in(0,\pi) ψ 0 ∈ ( 0 , π ) , and fixed threshold ψ ∗ ∈ ( 0 , ψ 0 ) \psi_\ast\in(0,\psi_0) ψ ∗ ∈ ( 0 , ψ 0 ) . Let T β ( ψ 0 ) T_\beta(\psi_0) T β ( ψ 0 ) be its hitting time, and let a ∈ R a\in\mathbb{R} a ∈ R be fixed. Then
lim β → ∞ T β ( ψ 0 + a / β ) T β ( ψ 0 ) = e a sin ψ 0 (77) \lim_{\beta\to\infty}\frac{T_\beta(\psi_0+a/\beta)}{T_\beta(\psi_0)}=e^{a\sin\psi_0} \tag*{(77)} β → ∞ lim T β ( ψ 0 ) T β ( ψ 0 + a / β ) = e a s i n ψ 0 (77)
Both initial states may be realized exactly on the same two-block collision stratum. Their normal distances to that stratum vanish identically.
Proof. The angular equation of the symmetric quotient is ψ ˙ = − q β ( ψ ) sin ψ \dot{\psi}=-q_\beta(\psi)\sin\psi ψ ˙ = − q β ( ψ ) sin ψ , with
q β ( ψ ) = ( m 2 m 1 + m 1 m 2 ) e − β ( 1 − cos ψ ) ( 1 + o ( 1 ) ) (78) q_\beta(\psi)=\left(\frac{m_2}{m_1}+\frac{m_1}{m_2}\right)e^{-\beta(1-\cos\psi)}(1+o(1)) \tag*{(78)} q β ( ψ ) = ( m 1 m 2 + m 2 m 1 ) e − β ( 1 − c o s ψ ) ( 1 + o ( 1 )) (78)
uniformly in a fixed neighborhood of ψ 0 \psi_0 ψ 0 . Endpoint Laplace asymptotics therefore give, with k = m 2 / m 1 + m 1 / m 2 k=m_2/m_1+m_1/m_2 k = m 2 / m 1 + m 1 / m 2 ,
T β ( ψ 0 ) ∼ e β ( 1 − cos ψ 0 ) β k sin 2 ψ 0 . (79) T_\beta(\psi_0)\sim\frac{e^{\beta(1-\cos\psi_0)}}{\beta k\sin^2\psi_0}. \tag*{(79)} T β ( ψ 0 ) ∼ β k sin 2 ψ 0 e β ( 1 − c o s ψ 0 ) . (79)
Since β [ cos ψ 0 − cos ( ψ 0 + a / β ) ] → a sin ψ 0 \beta[\cos\psi_0-\cos(\psi_0+a/\beta)]\to a\sin\psi_0 β [ cos ψ 0 − cos ( ψ 0 + a / β )] → a sin ψ 0 , taking the ratio proves (77) . Exact block collisions supply quotient initial conditions lying on the stratum with zero transverse distance. □
Figure 8 illustrates the two initial phases in Proposition I.3, both on the exact collision stratum.
Figure 8. Normal control and tangential clock control are different. Two configurations can lie exactly on the collision stratum yet have different leading hitting-time constants because their quotient phases differ by O ( β − 1 ) O(\beta^{-1}) O ( β − 1 ) . Schematic, not a numerical trajectory.
A correct conditional criterion for relative clock preservation
It is useful to formulate the scalar step that a future full-flow shadowing theorem would actually need to verify. Unlike the desired full-to-quotient reduction, the following is elementary and rigorous.
Lemma I.4 (Phase-velocity comparison and endpoint stability). Fix 0 < ψ ∗ < ψ 0 < π 0 < \psi_{*} < \psi_{0} < \pi 0 < ψ ∗ < ψ 0 < π . Suppose two monotonically decreasing angular phases traverse [ ψ ∗ , ψ 0 ] [\psi_{*},\psi_{0}] [ ψ ∗ , ψ 0 ] with autonomous positive speed functions F β ( ψ ) F_{\beta}(\psi) F β ( ψ ) and F ˉ β ( ψ ) \bar{F}_{\beta}(\psi) F ˉ β ( ψ ) , and
ρ β : = sup ψ ∈ [ ψ ∗ , ψ 0 ] ∣ F β ( ψ ) F ˉ β ( ψ ) − 1 ∣ ⟶ 0. (80) \rho_{\beta} := \sup_{\psi\in[\psi_{*},\psi_{0}]} \left|\frac{F_{\beta}(\psi)}{\bar{F}_{\beta}(\psi)} - 1\right| \longrightarrow0. \tag*{(80)} ρ β := ψ ∈ [ ψ ∗ , ψ 0 ] sup F ˉ β ( ψ ) F β ( ψ ) − 1 ⟶ 0. (80) If both start at ψ 0 \psi_{0} ψ 0 , their hitting times satisfy T β / T ˉ β = 1 + O ( ρ β ) T_{\beta}/\bar{T}_{\beta} = 1 + O(\rho_{\beta}) T β / T ˉ β = 1 + O ( ρ β ) . The same conclusion holds for an initial phase perturbation o ( β − 1 ) o(\beta^{-1}) o ( β − 1 ) , provided, on a fixed endpoint neighborhood, F ˉ β ( ψ ) = a β ( ψ ) e − β f ( ψ ) \bar{F}_{\beta}(\psi) = a_{\beta}(\psi)e^{-\beta f(\psi)} F ˉ β ( ψ ) = a β ( ψ ) e − β f ( ψ ) , where a β a_{\beta} a β is uniformly bounded above and below by positive constants on the entire phase interval and locally regular at the upper endpoint, f ∈ C 2 f \in C^{2} f ∈ C 2 has its unique global maximum on the interval at the upper endpoint, and f ′ ( ψ 0 ) > 0 f'(\psi_{0}) > 0 f ′ ( ψ 0 ) > 0 .
Proof. With common endpoints, T β = ∫ ψ ∗ ψ 0 F β ( ψ ) − 1 d ψ T_{\beta} = \int_{\psi_{*}}^{\psi_{0}} F_{\beta}(\psi)^{-1}\,d\psi T β = ∫ ψ ∗ ψ 0 F β ( ψ ) − 1 d ψ and T ˉ β = ∫ ψ ∗ ψ 0 F ˉ β ( ψ ) − 1 d ψ \bar{T}_{\beta} = \int_{\psi_{*}}^{\psi_{0}} \bar{F}_{\beta}(\psi)^{-1}\,d\psi T ˉ β = ∫ ψ ∗ ψ 0 F ˉ β ( ψ ) − 1 d ψ . By (80) , T ˉ β / ( 1 + ρ β ) ≤ T β ≤ T ˉ β / ( 1 − ρ β ) \bar{T}_{\beta}/(1+\rho_{\beta}) \leq T_{\beta} \leq\bar{T}_{\beta}/(1-\rho_{\beta}) T ˉ β / ( 1 + ρ β ) ≤ T β ≤ T ˉ β / ( 1 − ρ β ) for large β \beta β . With an upper endpoint shift δ β = o ( β − 1 ) \delta_{\beta} = o(\beta^{-1}) δ β = o ( β − 1 ) , the additional endpoint interval contributes at most O ( ∣ δ β ∣ e β f ( ψ 0 ) + O ( β ∣ δ β ∣ ) ) O(|\delta_{\beta}|e^{\beta f(\psi_{0})+O(\beta|\delta_{\beta}|)}) O ( ∣ δ β ∣ e β f ( ψ 0 ) + O ( β ∣ δ β ∣ ) ) , whereas endpoint Laplace gives T ˉ β ≍ β − 1 e β f ( ψ 0 ) \bar{T}_{\beta} \asymp\beta^{-1}e^{\beta f(\psi_{0})} T ˉ β ≍ β − 1 e β f ( ψ 0 ) . The relative difference is O ( β ∣ δ β ∣ ) = o ( 1 ) O(\beta|\delta_{\beta}|) = o(1) O ( β ∣ δ β ∣ ) = o ( 1 ) . □
The remaining quantitative tangential estimate
Let r ( t ) r(t) r ( t ) denote the normal error and z ( t ) z(t) z ( t ) a properly synchronized tangential separation of the projected full trajectory and the quotient. A natural but not yet verified differential inequality would be
D + r ≤ − κ r , D + z ≤ L β ( t ) z + K β ( t ) r . (81) D^{+}r \leq-\kappa r,\qquad D^{+}z \leq L_{\beta}(t)z + K_{\beta}(t)r. \tag*{(81)} D + r ≤ − κ r , D + z ≤ L β ( t ) z + K β ( t ) r . (81)
If it were established along the entire phase interval up to the second event, the standard variation-of-constants estimate would give
z ( t ) ≤ exp ( ∫ t 0 t L β + ( s ) d s ) [ z ( t 0 ) + ∫ t 0 t K β ( s ) r ( s ) d s ] . (82) z(t) \leq\exp\left(\int_{t_{0}}^{t} L_{\beta}^{+}(s)\,ds\right)\left[z(t_{0})+\int_{t_{0}}^{t}K_{\beta}(s)r(s)\,ds\right]. \tag*{(82)} z ( t ) ≤ exp ( ∫ t 0 t L β + ( s ) d s ) [ z ( t 0 ) + ∫ t 0 t K β ( s ) r ( s ) d s ] . (82)
A sufficient additional requirement for o ( β − 1 ) o(\beta^{-1}) o ( β − 1 ) quotient closeness is that the right-hand side be o ( β − 1 ) o(\beta^{-1}) o ( β − 1 ) uniformly up to the second event. In an exactly monotone two-body quotient, ∫ q d t = ∫ d ψ / sin ψ = O ( 1 ) \int q\,dt = \int d\psi/\sin\psi= O(1) ∫ q d t = ∫ d ψ / sin ψ = O ( 1 ) on an interval bounded away from 0 0 0 and π \pi π , but ∫ β q d t = O ( β ) \int\beta q\,dt = O(\beta) ∫ β q d t = O ( β ) , not O ( 1 ) O(1) O ( 1 ) . This observation alone does not verify (81) for the full system or guarantee a sufficiently small prefactor in (82) .
One must also distinguish two different barriers in the concrete example. The initial second-clock exponent candidate is G ∗ = 1 − cos ( 0.8 ) ≈ 0.30329329 G_{*} = 1 - \cos(0.8) \approx0.30329329 G ∗ = 1 − cos ( 0.8 ) ≈ 0.30329329 , whereas the earlier *uniform external shield lower bound in the first corridor is G s h ≈ 0.187507 G_{\mathrm{sh}} \approx0.187507 G sh ≈ 0.187507 . Neither can be used unchanged as a uniform bound on the entire subsequent trajectory: as the second angle approaches 0.3 0.3 0.3 , its symmetric base barrier becomes 1 − cos ( 0.3 ) ≈ 0.044664 1-\cos(0.3) \approx0.044664 1 − cos ( 0.3 ) ≈ 0.044664 . In addition, a tube estimate only through T = e β H T=e^{\beta H} T = e β H for a fixed H < G ∗ H<G_* H < G ∗ ends exponentially earlier than a clock of size β − 1 e β G ∗ \beta^{-1}e^{\beta G_*} β − 1 e β G ∗ . The endpoint derivative f ′ ( ψ 0 ) > 0 f'(\psi_0)>0 f ′ ( ψ 0 ) > 0 controls Laplace asymptotics but does not supply a β \beta β -uniform lower bound on the physical crossing speed, which itself can be exponentially small.
Open problem I.6 (Full-flow tangential matching and second-event prefactor). For the three-point configuration (31) , prove the coupled bounds needed to apply Lemma I.4 up to the second threshold and determine whether
lim β → ∞ β t 2 ( β ) e − β ( 1 − cos 0.8 ) = 1 sin 2 ( 0.8 ) ( 1 2 2 cos 0.8 + 2 2 − cos 0.8 ) (83) \lim_{\beta\to\infty}\beta t_2(\beta)e^{-\beta(1-\cos0.8)}
=
\frac{1}{\sin^2(0.8)\left(\frac{1}{2}2^{\cos0.8}+2\,2^{-\cos0.8}\right)} \tag*{(83)} β → ∞ lim β t 2 ( β ) e − β ( 1 − c o s 0.8 ) = sin 2 ( 0.8 ) ( 2 1 2 c o s 0.8 + 2 2 − c o s 0.8 ) 1 (83)
The right-hand side is proved for the exactly collapsed two-body quotient only by Theorem H.4 . The full-flow identity remains unproved. A proposed recursive merger theorem must additionally propagate its geometric gap and synchronization assumptions across each stage.
Collapse-layer estimates and a conditional phase-matching mechanism
We refine the proof checkpoints from section I. Two different claims must be kept apart: estimates for a perturbed two-center normal form , which can be proved under explicit error bounds, and verification that the original three-dimensional flow satisfies those error bounds with a sufficiently accurate moving coordinate chart. Only the former is completed here.
An almost-conserved skew invariant with an explicit stopping scale
Let u , v ∈ S d − 1 u,v\in S^{d-1} u , v ∈ S d − 1 and write c = u T v c=u^{\mathsf T}v c = u T v , h = u T A v h=u^{\mathsf T}Av h = u T A v , s = 1 − c 2 s=\sqrt{1-c^2} s = 1 − c 2 and α = h / s \alpha=h/s α = h / s whenever s > 0 s>0 s > 0 . Consider the perturbed two-center equations
c ˙ = q ( 1 − c 2 ) + r c , h ˙ = − q c h + r h , ∣ r c ∣ + ∣ r h ∣ ≤ C η β ( t ) . (84) \dot{c}=q(1-c^2)+r_c,\qquad\dot{h}=-qch+r_h,\qquad|r_c|+|r_h|\le C\eta_\beta(t). \tag*{(84)} c ˙ = q ( 1 − c 2 ) + r c , h ˙ = − q c h + r h , ∣ r c ∣ + ∣ r h ∣ ≤ C η β ( t ) . (84)
The coefficient q q q is nonnegative and may depend on the full configuration. The exact two-center cancellation survives in the following sense.
Lemma J.1 (Skew invariant under external forcing). Assume (84) with ∣ α ∣ ≤ ∥ A ∥ o p |\alpha|\le\lVert A\rVert_{\mathrm{op}} ∣ α ∣ ≤ ∥ A ∥ op , as is automatically true for points on the sphere. Then
α ˙ = r h s + c h r c s 3 , ∣ α ˙ ∣ ≤ C η β ( t ) ( 1 s + ∥ A ∥ o p s 2 ) . (85) \dot{\alpha}=\frac{r_h}{s}+\frac{ch\,r_c}{s^3},\qquad|\dot{\alpha}|\le C\eta_\beta(t)\left(\frac{1}{s}+\frac{\lVert A\rVert_{\mathrm{op}}}{s^2}\right). \tag*{(85)} α ˙ = s r h + s 3 c h r c , ∣ α ˙ ∣ ≤ C η β ( t ) ( s 1 + s 2 ∥ A ∥ op ) . (85) In particular, if s ( t ) ≥ β − N s(t)\ge\beta^{-N} s ( t ) ≥ β − N for t ∈ [ t a , t b ] t\in[t_a,t_b] t ∈ [ t a , t b ] and η β ( t ) ≤ C 0 e − β G s h \eta_\beta(t)\le C_0e^{-\beta G_{\mathrm{sh}}} η β ( t ) ≤ C 0 e − β G sh , then
∣ α ( t b ) − α ( t a ) ∣ ≤ C 1 ( t b − t a ) ( β N + β 2 N ) e − β G s h . (86) |\alpha(t_b)-\alpha(t_a)|\le C_1(t_b-t_a)(\beta^N+\beta^{2N})e^{-\beta G_{\mathrm{sh}}}. \tag*{(86)} ∣ α ( t b ) − α ( t a ) ∣ ≤ C 1 ( t b − t a ) ( β N + β 2 N ) e − β G sh . (86) Thus an O ( log β ) O(\log\beta) O ( log β ) collapse window stopped at s = β − 3 s=\beta^{-3} s = β − 3 has exponentially small skew-invariant drift, conditional on the stated external shield.
Proof. Differentiate h / s h/s h / s and use s ˙ = − c c ˙ / s \dot{s}=-c\dot{c}/s s ˙ = − c c ˙ / s . The terms involving q q q cancel identically. Since the restriction of a skew form to the plane spanned by u u u and v v v satisfies ∣ u T A v ∣ ≤ ∥ A ∥ o p 1 − c 2 |u^{\mathsf T}Av|\le\lVert A\rVert_{\mathrm{op}}\sqrt{1-c^2} ∣ u T A v ∣ ≤ ∥ A ∥ op 1 − c 2 , one has ∣ α ∣ ≤ ∥ A ∥ o p |\alpha|\le\lVert A\rVert_{\mathrm{op}} ∣ α ∣ ≤ ∥ A ∥ op . Substituting ∣ h ∣ ≤ ∥ A ∥ o p s |h|\le\lVert A\rVert_{\mathrm{op}}s ∣ h ∣ ≤ ∥ A ∥ op s into the remaining terms yields (85); integration gives (86).
Remark J.2 (Scope of the O ( log β ) O(\log\beta) O ( log β ) collapse estimate). For the isolated pair of (31) , the favorable barrier f ( θ ) = 1 − cos θ − 0.6 sin θ f(\theta)=1-\cos\theta-0.6\sin\theta f ( θ ) = 1 − cos θ − 0.6 sin θ is strictly negative on ( 0 , 0.3 ] (0,0.3] ( 0 , 0.3 ] . Hence the favorable transition probability is at least 1 / 2 1/2 1/2 , and the isolated q q q is at least 1 / 2 1/2 1/2 on that interval. A perturbed solution inherits a positive rate only if its external interactions and its angular drift are controlled. Under such a bootstrap one obtains a logarithmic relaxation from the 0.3 0.3 0.3 threshold to a polynomially small angular scale. This does not itself prove that the requisite shield holds for the full system throughout that window.
Angular reparameterization removes the physical metastable clock
The next lemma is an elementary but useful mechanism for relative phase matching. It deliberately specifies all the assumptions needed for its scalar application.
Lemma J.3 (Conditional phase-defect estimate). Let 0 < θ min < θ 0 < π 0<\theta_{\min}<\theta_0<\pi 0 < θ m i n < θ 0 < π and p , q p,q p , q be nonnegative functions on [ θ min , θ 0 ] [\theta_{\min},\theta_0] [ θ m i n , θ 0 ] with p ≤ q p\le q p ≤ q and q > 0 q>0 q > 0 . Suppose a perturbed planar pair obeys
θ ˙ = − q ( θ ) sin θ + r θ ( t ) , φ ˙ = p ( θ ) sin θ + r φ ( t ) (87) \dot{\theta}=-q(\theta)\sin\theta+r_\theta(t),\qquad\dot{\varphi}=p(\theta)\sin\theta+r_\varphi(t) \tag*{(87)} θ ˙ = − q ( θ ) sin θ + r θ ( t ) , φ ˙ = p ( θ ) sin θ + r φ ( t ) (87) while θ \theta θ decreases strictly from θ 0 \theta_0 θ 0 to θ min \theta_{\min} θ m i n , and ∣ r θ ∣ + ∣ r φ ∣ ≤ η ( θ ) |r_\theta|+|r_\varphi|\le\eta(\theta) ∣ r θ ∣ + ∣ r φ ∣ ≤ η ( θ ) with η ( θ ) ≤ 1 2 q ( θ ) sin θ \eta(\theta)\le\frac{1}{2}q(\theta)\sin\theta η ( θ ) ≤ 2 1 q ( θ ) sin θ . Let φ i s o \varphi_{\mathrm{iso}} φ iso solve the unperturbed angular system with the same initial φ \varphi φ and the same initial θ \theta θ . Then at any common angular coordinate,
∣ d φ d θ + p ( θ ) q ( θ ) ∣ ≤ 4 η ( θ ) q ( θ ) sin θ , ∣ φ − φ i s o ∣ ≤ 4 ∫ θ θ 0 η ( a ) q ( a ) sin a d a . (88) \left|\frac{\mathrm{d}\varphi}{\mathrm{d}\theta}+\frac{p(\theta)}{q(\theta)}\right|\le\frac{4\eta(\theta)}{q(\theta)\sin\theta},\qquad|\varphi-\varphi_{\mathrm{iso}}|\le4\int_{\theta}^{\theta_0}\frac{\eta(a)}{q(a)\sin a}\,\mathrm{d}a. \tag*{(88)} d θ d φ + q ( θ ) p ( θ ) ≤ q ( θ ) sin θ 4 η ( θ ) , ∣ φ − φ iso ∣ ≤ 4 ∫ θ θ 0 q ( a ) sin a η ( a ) d a . (88)
Proof. By (87) , d φ / d θ = − ( p sin θ + r φ ) / ( q sin θ − r θ ) \mathrm{d}\varphi/\mathrm{d}\theta=-(p\sin\theta+r_\varphi)/(q\sin\theta-r_\theta) d φ / d θ = − ( p sin θ + r φ ) / ( q sin θ − r θ ) . Subtract − p / q -p/q − p / q and use p ≤ q p\le q p ≤ q and ∣ r θ ∣ ≤ 1 2 q sin θ |r_\theta|\le\frac{1}{2}q\sin\theta ∣ r θ ∣ ≤ 2 1 q sin θ . The numerator is bounded by q ( ∣ r θ ∣ + ∣ r φ ∣ ) q(|r_\theta|+|r_\varphi|) q ( ∣ r θ ∣ + ∣ r φ ∣ ) and the denominator below by 1 2 q 2 sin θ \frac{1}{2}q^2\sin\theta 2 1 q 2 sin θ , yielding the displayed factor 4 4 4 as a valid conservative constant. Integrate in the monotone angular coordinate.
Corollary J.4 (Phase bias under two separated error regions). Under the hypotheses of theorem J.3 , let θ min = β − 3 \theta_{\min}=\beta^{-3} θ m i n = β − 3 and θ 0 = 1.2 \theta_0=1.2 θ 0 = 1.2 . Suppose, with constants independent of β \beta β , that
η ( θ ) q ( θ ) ≤ { C e − β Δ , 0.3 ≤ θ ≤ 1.2 , C e − β G s h , β − 3 ≤ θ ≤ 0.3 , Δ , G s h > 0. (89) \frac{\eta(\theta)}{q(\theta)}\le
\begin{cases}
C e^{-\beta\Delta}, & 0.3\le\theta\le1.2,\\
C e^{-\beta G_{\mathrm{sh}}}, & \beta^{-3}\le\theta\le0.3,
\end{cases}
\qquad\Delta,G_{\mathrm{sh}}>0. \tag*{(89)} q ( θ ) η ( θ ) ≤ { C e − β Δ , C e − β G sh , 0.3 ≤ θ ≤ 1.2 , β − 3 ≤ θ ≤ 0.3 , Δ , G sh > 0. (89) Then the discrepancy from the complete isolated terminal phase Φ β \Phi_\beta Φ β obeys
∣ φ ( β − 3 ) − Φ β ∣ ≤ C ′ e − β Δ + C ′ e − β G s h log β + β − 3 = o ( β − 1 ) . (90) |\varphi(\beta^{-3})-\Phi_\beta|\le C'e^{-\beta\Delta}+C'e^{-\beta G_{\mathrm{sh}}}\log\beta+\beta^{-3}=o(\beta^{-1}). \tag*{(90)} ∣ φ ( β − 3 ) − Φ β ∣ ≤ C ′ e − β Δ + C ′ e − β G sh log β + β − 3 = o ( β − 1 ) . (90) Here the common initial phase is set to zero; the sign convention for φ \varphi φ is that of (58) .
Proof. On the fixed outer interval, ∫ 0.3 1.2 ( sin θ ) − 1 d θ \int_{0.3}^{1.2}(\sin\theta)^{-1}\mathrm{d}\theta ∫ 0.3 1.2 ( sin θ ) − 1 d θ is finite. On [ β − 3 , 0.3 ] [\beta^{-3},0.3] [ β − 3 , 0.3 ] , ∫ ( sin θ ) − 1 d θ = O ( log β ) \int(\sin\theta)^{-1}\mathrm{d}\theta=O(\log\beta) ∫ ( sin θ ) − 1 d θ = O ( log β ) . Apply (88) and (89) . The residual isolated phase from 0 0 0 to β − 3 \beta^{-3} β − 3 is at most β − 3 \beta^{-3} β − 3 because 0 ≤ p / q ≤ 1 0\le p/q\le1 0 ≤ p / q ≤ 1 .
Figure 9 displays the sign change of the isolated inner barrier and the limited scope of the initial corridor shield.
Figure 9. The favorable isolated inner barrier becomes negative in the post-threshold interval (shaded). Its rate is no longer exponentially slow there. The horizontal shield is a geometric bound from the initial corridor , not a proved uniform shield until the second event.
Why this still does not establish the full second prefactor
The two lemmas make the remaining analytical task narrower, not automatic. For the original three-dimensional flow one must construct a moving angular chart and prove that its actual coefficients differ from the isolated p , q p,q p , q in a manner absorbed by r θ , r φ r_\theta,r_\varphi r θ , r φ with (89) . A bound on α \alpha α alone does not control the absolute common phase: a simultaneous rotation leaves α \alpha α unchanged. After collapse, one must also control the tangential drift between the full block center and the exact weighted quotient through the second event. The first-corridor shield G s h G_{\mathrm{sh}} G sh cannot be used unchanged over that entire trajectory.
In particular, a tentative center-equation remainder should retain potential contributions O ( β ∥ ζ ∥ 2 ) O(\beta\|\zeta\|^2) O ( β ∥ ζ ∥ 2 ) from internal nonreciprocal weights, rather than assuming an unverified coefficient-free O ( ∥ ζ ∥ 2 ) O(\|\zeta\|^2) O ( ∥ ζ ∥ 2 ) bound for the relative displacement ζ \zeta ζ . A time-integrated, phase-dependent bound on that remainder and the competing event speeds is the next proof checkpoint. Consequently, the full-flow limit in (83) remains a conjectural identification , despite the precise isolated-pair and quotient calculations.
A sharp second exponent without prefactor shadowing
We establish matching logarithmic upper and lower bounds for two-cluster interaction times using positive interactions and invariant cones. The argument does not infer a prefactor from normal proximity. The cone and projective-coordinate steps below use the positive-interaction mechanism in [2 Lemma 4.2 and its proof] and [14 ] . The sharp clock conclusion requires the separate barrier estimates proved here; it is not a consequence of consensus alone.
Lemma K.1 (Positive projective coordinates). Consider u ˙ i = ∑ j p i j ( u j − ( u i ⊤ u j ) u i ) \dot{u}_i=\sum_j p_{ij}(u_j-(u_i^\top u_j)u_i) u ˙ i = ∑ j p ij ( u j − ( u i ⊤ u j ) u i ) with continuous p i j ≥ 0 p_{ij}\ge0 p ij ≥ 0 and ∑ j p i j = 1 \sum_j p_{ij}=1 ∑ j p ij = 1 . If a unit vector e e e satisfies a i ( 0 ) : = e ⊤ u i ( 0 ) ≥ a ∗ > 0 a_i(0):=e^\top u_i(0)\ge a_\ast>0 a i ( 0 ) := e ⊤ u i ( 0 ) ≥ a ∗ > 0 , then a i ( t ) ≥ a ∗ a_i(t)\ge a_\ast a i ( t ) ≥ a ∗ . For any fixed linear map L L L , the coordinates z i = L u i / a i z_i=Lu_i/a_i z i = L u i / a i obey
z ˙ i = ∑ j p i j a j a i ( z j − z i ) . \dot{z}_i=\sum_j p_{ij}\frac{a_j}{a_i}(z_j-z_i). z ˙ i = j ∑ p ij a i a j ( z j − z i ) . Consequently their convex hull is nonincreasing in time, and max i ∥ z i ( t ) ∥ ≤ max i ∥ z i ( 0 ) ∥ \max_i\|z_i(t)\|\le\max_i\|z_i(0)\| max i ∥ z i ( t ) ∥ ≤ max i ∥ z i ( 0 ) ∥ .
Proof. At an active minimum of the a i a_i a i , put b i = ∑ j p i j u i ⊤ u j ≤ 1 b_i=\sum_j p_{ij}u_i^\top u_j\le1 b i = ∑ j p ij u i ⊤ u j ≤ 1 . Then a ˙ i = ∑ j p i j a j − b i a i ≥ ( 1 − b i ) a i ≥ 0 \dot{a}_i=\sum_j p_{ij}a_j-b_i a_i\ge(1-b_i)a_i\ge0 a ˙ i = ∑ j p ij a j − b i a i ≥ ( 1 − b i ) a i ≥ 0 . The finite-minimum derivative rule gives invariance of a i ≥ a ∗ a_i\ge a_\ast a i ≥ a ∗ . Differentiating L u i / a i Lu_i/a_i L u i / a i cancels the b i b_i b i terms and yields the displayed equation. Every supporting linear functional has nonincreasing maximum and nondecreasing minimum for this cooperative equation, proving convex-hull invariance. This is the standard projective consensus mechanism, related to the cone-collapse argument of [2 ] and the directed-sphere analysis of [14 ] ; no priority is claimed for the coordinate lemma.
Theorem K.2 (Zero-skew two-cluster clock). Fix a finite number of particles, positive masses, a skew matrix A A A , and ε ∈ R \varepsilon\in\mathbb{R} ε ∈ R . Let v , w ∈ S d − 1 v,w \in S^{d-1} v , w ∈ S d − 1 satisfy
v T w = cos ϕ , 0 < θ ∗ < ϕ < π / 2 , v T A w = 0. v^{\mathsf{T}}w=\cos\phi,\qquad0<\theta_*<\phi<\pi/2,\qquad v^{\mathsf{T}}Aw=0. v T w = cos ϕ , 0 < θ ∗ < ϕ < π /2 , v T A w = 0. For each β \beta β , initialize the smooth normalized interaction flow in two nonempty blocks I , J I,J I , J such that
r β : = max { max i ∈ I ∥ u i ( 0 ) − v ∥ , max j ∈ J ∥ u j ( 0 ) − w ∥ } ⟶ 0. r_\beta:=\max\left\{\max_{i\in I}\lVert u_i(0)-v\rVert,\max_{j\in J}\lVert u_j(0)-w\rVert\right\}\longrightarrow0. r β := max { i ∈ I max ∥ u i ( 0 ) − v ∥ , j ∈ J max ∥ u j ( 0 ) − w ∥ } ⟶ 0. Let T β T_\beta T β be the first crossing of angle θ ∗ \theta_* θ ∗ by a pair with one index in each block. Then
lim β → ∞ β − 1 log T β = G : = 1 − cos ϕ . \lim_{\beta\to\infty}\beta^{-1}\log T_\beta=G:=1-\cos\phi. β → ∞ lim β − 1 log T β = G := 1 − cos ϕ . No o ( β − 1 ) o(\beta^{-1}) o ( β − 1 ) initial error, transverse eigenvalue assumption, or full-to-quotient prefactor identification is required.
Proof. Write K = ∥ I + ε A ∥ o p K=\lVert I+\varepsilon A\rVert_{\mathrm{op}} K = ∥ I + ε A ∥ op . For the lower bound, fix 0 < a < G 0<a<G 0 < a < G and a sufficiently small δ > 0 \delta>0 δ > 0 . Initially each block lies in the angular δ \delta δ -cap about its own center for all sufficiently large β \beta β . Bootstrap both blocks in their respective 2 δ 2\delta 2 δ -caps. Since v T A w = 0 v^{\mathsf{T}}Aw=0 v T A w = 0 , every cross-block barrier in this tube is at least G − 4 K δ G-4K\delta G − 4 K δ . The self term in each softmax denominator implies that the total external weight in any row is at most η β = C exp [ − β ( G − 4 K δ ) ] \eta_\beta=C\exp[-\beta(G-4K\delta)] η β = C exp [ − β ( G − 4 K δ )] . At a minimum of v T u i v^{\mathsf{T}}u_i v T u i within I I I , every internal term is nonnegative, as long as that minimum is positive; external terms have absolute value at most 2 η β 2\eta_\beta 2 η β . Hence
min i ∈ I v T u i ( t ) ≥ cos δ − 2 η β t . \min_{i\in I}v^{\mathsf{T}}u_i(t)\ge\cos\delta-2\eta_\beta t. i ∈ I min v T u i ( t ) ≥ cos δ − 2 η β t .
The same estimate holds for the w w w -block. Choose δ \delta δ with a < G − 4 K δ a<G-4K\delta a < G − 4 K δ and ϕ − 4 δ > θ ∗ \phi-4\delta>\theta_* ϕ − 4 δ > θ ∗ . On 0 ≤ t ≤ e β a 0\le t\le e^{\beta a} 0 ≤ t ≤ e β a , the right-hand error tends to zero, so a first-exit argument preserves both 2 δ 2\delta 2 δ -caps. Every cross angle exceeds θ ∗ \theta_* θ ∗ ; thus T β ≥ e β a T_\beta\ge e^{\beta a} T β ≥ e β a . Let a ↑ G a\uparrow G a ↑ G .
For the upper bound, let E = span { v , w } E=\operatorname{span}\{v,w\} E = span { v , w } and e = ( v + w ) / ∥ v + w ∥ e=(v+w)/\lVert v+w\rVert e = ( v + w ) / ∥ v + w ∥ . At the initial time e T u i ≥ a ∗ > 0 e^{\mathsf{T}}u_i\ge a_*>0 e T u i ≥ a ∗ > 0 uniformly for large β \beta β . Apply Lemma K.1 with L = P E ⊥ L=P_{E^\perp} L = P E ⊥ . The initial normal coordinates are O ( r β ) O(r_\beta) O ( r β ) , hence ∥ P E ⊥ u i ( t ) ∥ ≤ C r β \lVert P_{E^\perp}u_i(t)\rVert\le Cr_\beta ∥ P E ⊥ u i ( t )∥ ≤ C r β at all times. The restriction of A A A to E E E vanishes as a bilinear form: its values on the basis ( v , w ) (v,w) ( v , w ) are zero by skew symmetry and v T A w = 0 v^{\mathsf{T}}Aw=0 v T A w = 0 . Decomposition into E E E and E ⊥ E^\perp E ⊥ gives
∣ u i ( t ) T A u j ( t ) ∣ ≤ C r β for all i , j , t . \left|u_i(t)^{\mathsf{T}}Au_j(t)\right|\le Cr_\beta\qquad\text{for all }i,j,t. u i ( t ) T A u j ( t ) ≤ C r β for all i , j , t .
Initially ρ β ( 0 ) = min i , j u i ( 0 ) T u j ( 0 ) = cos ϕ + O ( r β ) > 0 \rho_\beta(0)=\min_{i,j}u_i(0)^{\mathsf{T}}u_j(0)=\cos\phi+O(r_\beta)>0 ρ β ( 0 ) = min i , j u i ( 0 ) T u j ( 0 ) = cos ϕ + O ( r β ) > 0 . The active-pair calculation for acute consensus yields ρ β ( t ) ≥ ρ β ( 0 ) \rho_\beta(t)\ge\rho_\beta(0) ρ β ( t ) ≥ ρ β ( 0 ) . All logits before multiplication by β \beta β therefore lie between ρ β ( 0 ) − C r β \rho_\beta(0)-Cr_\beta ρ β ( 0 ) − C r β and 1 + C r β 1+Cr_\beta 1 + C r β . Consequently every transition probability is at least
k β : = m min ∑ j m j exp { − β [ 1 − ρ β ( 0 ) + C ′ r β ] } . k_\beta:=\frac{m_{\min}}{\sum_jm_j}\exp\{-\beta[1-\rho_\beta(0)+C'r_\beta]\}. k β := ∑ j m j m m i n exp { − β [ 1 − ρ β ( 0 ) + C ′ r β ]} .
For an active minimizing pair, ρ ˙ β ≥ 2 k β ( 1 − ρ β 2 ) \dot{\rho}_\beta\ge2k_\beta(1-\rho_\beta^2) ρ ˙ β ≥ 2 k β ( 1 − ρ β 2 ) almost everywhere. Integration to cos θ ∗ \cos\theta_* cos θ ∗ bounds the time by C / k β C/k_\beta C / k β , where C C C is uniform for large β \beta β . In particular all cross pairs have reached the threshold by that time. Thus lim sup β − 1 log T β ≤ 1 − cos ϕ \limsup\beta^{-1}\log T_\beta\le1-\cos\phi lim sup β − 1 log T β ≤ 1 − cos ϕ . Together with the lower bound this proves the claim.
Lemma K.4 (One-sided entrance in the explicit three-point example). Use A = ( 0 − 1 0 1 0 0 0 0 0 ) A = \begin{pmatrix} 0 & -1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 0 \end{pmatrix} A = 0 1 0 − 1 0 0 0 0 0 , ε = 0.6 \varepsilon= 0.6 ε = 0.6 , equal masses, and the initial points
v = ( 1 , 0 , 0 ) , w = ( cos 0.8 , 0 , sin 0.8 ) , u 3 ( 0 ) = ( cos 1.2 , sin 1.2 , 0 ) . v=(1,0,0),\qquad w=(\cos0.8,0,\sin0.8),\qquad u_{3}(0)=(\cos1.2,\sin1.2,0). v = ( 1 , 0 , 0 ) , w = ( cos 0.8 , 0 , sin 0.8 ) , u 3 ( 0 ) = ( cos 1.2 , sin 1.2 , 0 ) .
For each fixed δ ∈ ( 0 , 0.3 ) \delta\in(0,0.3) δ ∈ ( 0 , 0.3 ) , let s β , δ s_{\beta,\delta} s β , δ be the first time ∠ ( u 1 , u 3 ) = δ \angle(u_{1},u_{3})=\delta ∠ ( u 1 , u 3 ) = δ . Then
β − 1 log s β , δ ⟶ B 0 , u 1 ( s β , δ ) → v , u 2 ( s β , δ ) → w , dist S ( u 3 ( s β , δ ) , v ) ≤ δ + o ( 1 ) , \beta^{-1}\log s_{\beta,\delta}\longrightarrow B_{0},\qquad u_{1}(s_{\beta,\delta})\to v,\qquad u_{2}(s_{\beta,\delta})\to w,\qquad\operatorname{dist}_{S}(u_{3}(s_{\beta,\delta}),v)\leq\delta+o(1), β − 1 log s β , δ ⟶ B 0 , u 1 ( s β , δ ) → v , u 2 ( s β , δ ) → w , dist S ( u 3 ( s β , δ ) , v ) ≤ δ + o ( 1 ) ,
where B 0 = 1 − cos 1.2 − 0.6 sin 1.2 B_{0}=1-\cos1.2-0.6\sin1.2 B 0 = 1 − cos 1.2 − 0.6 sin 1.2 .
Proof. The same off-diagonal corridor shield applies with threshold δ \delta δ :
G s h = 1 − cos 0.8 1 + 0.6 2 > B 0 . G_{\mathrm{sh}}=1-\cos0.8\sqrt{1+0.6^{2}}>B_{0}. G sh = 1 − cos 0.8 1 + 0. 6 2 > B 0 .
Every competing angular gap is at least 0.8 0.8 0.8 . The corridor bootstrap gives the first clock, a tube error o ( β − 1 ) o(\beta^{-1}) o ( β − 1 ) that is in fact exponentially small for fixed δ \delta δ , and an exponentially small time integral of the external row weights.
To establish the additional anchoring assertion, write b = 1 − u 1 T u 3 ≥ 0 b=1-u_{1}^{\mathsf T}u_{3}\geq0 b = 1 − u 1 T u 3 ≥ 0 and a = − ε u 1 T A u 3 a=-\varepsilon u_{1}^{\mathsf T}Au_{3} a = − ε u 1 T A u 3 . Inside that tube, while θ = ∠ ( u 1 , u 3 ) ∈ [ δ , 1.2 ] \theta=\angle(u_{1},u_{3})\in[\delta,1.2] θ = ∠ ( u 1 , u 3 ) ∈ [ δ , 1.2 ] , a = 0.6 sin θ + o ( β − 1 ) ≥ 0.3 sin δ a=0.6\sin\theta+o(\beta^{-1})\geq0.3\sin\delta a = 0.6 sin θ + o ( β − 1 ) ≥ 0.3 sin δ for large β \beta β . After division by the self terms put x = e − β ( b + a ) x=e^{-\beta(b+a)} x = e − β ( b + a ) , y = e − β ( b − a ) y=e^{-\beta(b-a)} y = e − β ( b − a ) and let s 1 , s 3 s_{1},s_{3} s 1 , s 3 denote the external contributions to the two row denominators. The shield gives s 1 , s 3 = o ( 1 ) s_{1},s_{3}=o(1) s 1 , s 3 = o ( 1 ) , and
p 13 p 31 = x ( 1 + y + s 3 ) y ( 1 + x + s 1 ) ≤ e − 2 β a + e − β ( b + a ) + s 3 e − 2 β a ≤ C e − β a . \frac{p_{13}}{p_{31}}=\frac{x(1+y+s_{3})}{y(1+x+s_{1})}\leq e^{-2\beta a}+e^{-\beta(b+a)}+s_{3}e^{-2\beta a}\leq Ce^{-\beta a}. p 31 p 13 = y ( 1 + x + s 1 ) x ( 1 + y + s 3 ) ≤ e − 2 β a + e − β ( b + a ) + s 3 e − 2 β a ≤ C e − β a .
If q = p 13 + p 31 q=p_{13}+p_{31} q = p 13 + p 31 , then θ ˙ = − q sin θ + O ( η β ) \dot{\theta}=-q\sin\theta+O(\eta_{\beta}) θ ˙ = − q sin θ + O ( η β ) and
∫ 0 s β , δ q sin θ d t ≤ 1.2 + C ∫ η β d t = O ( 1 ) . \int_{0}^{s_{\beta,\delta}}q\sin\theta\,dt\leq1.2+C\int\eta_{\beta}\,dt=O(1). ∫ 0 s β , δ q sin θ d t ≤ 1.2 + C ∫ η β d t = O ( 1 ) .
Therefore the internal contribution to the displacement of u 1 u_{1} u 1 is at most C e − 0.3 β sin δ Ce^{-0.3\beta\sin\delta} C e − 0.3 β s i n δ ; its external contribution tends to zero as well. The displacement of u 2 u_{2} u 2 is entirely external and tends to zero. The triangle inequality gives the last assertion.
:::
Corollary K.5 (The full-flow second exponent). Let t 2 ( β ) t_{2}(\beta) t 2 ( β ) be the first time either ( 1 , 2 ) (1,2) ( 1 , 2 ) or ( 2 , 3 ) (2,3) ( 2 , 3 ) reaches angle 0.3 0.3 0.3 in this example, without forced collisions. Then
lim β → ∞ log t 2 ( β ) β = 1 − cos 0.8 ≈ 0.30329329. \boxed{\lim_{\beta\to\infty}\frac{\log t_{2}(\beta)}{\beta}=1-\cos0.8\approx0.30329329.} β → ∞ lim β log t 2 ( β ) = 1 − cos 0.8 ≈ 0.30329329. Proof. The entrance lemma supplies, for each fixed δ \delta δ , two blocks in O ( δ ) O(\delta) O ( δ ) -neighborhoods of v , w v,w v , w at a time of logarithmic exponent B 0 B_{0} B 0 . Moreover v T A w = 0 v^{\mathsf T}Aw=0 v T A w = 0 and ∠ ( v , w ) = 0.8 < π / 2 \angle(v,w)=0.8<\pi/2 ∠ ( v , w ) = 0.8 < π /2 . For any 0 < a < G = 1 − cos 0.8 0<a<G=1-\cos0.8 0 < a < G = 1 − cos 0.8 , choose δ \delta δ sufficiently small and apply the cap argument of Theorem K.2 from the entrance time: the cross barriers are at least G − C δ G-C\delta G − C δ , and the two blocks remain separated for a further e β a e^{\beta a} e β a time. This gives lim inf β → ∞ β − 1 log t 2 ≥ G \liminf_{\beta\to\infty}\beta^{-1}\log t_{2}\geq G lim inf β → ∞ β − 1 log t 2 ≥ G . For the upper bound, the projective argument of that theorem, now with initial normal error O ( δ ) + o ( 1 ) O(\delta)+o(1) O ( δ ) + o ( 1 ) , bounds the remaining time by C δ e β ( G + C ′ δ + o ( 1 ) ) C_{\delta}e^{\beta(G+C'\delta+o(1))} C δ e β ( G + C ′ δ + o ( 1 )) . Adding the entrance time contributes exponent at most B 0 < G B_{0}<G B 0 < G . Let δ ↓ 0 \delta\downarrow0 δ ↓ 0 to obtain the matching upper bound. There is no earlier cross event before entrance because the corridor keeps the external angles above $0.8-o(1)>0.3.
The constant 0.95054656 0.95054656 0.95054656 for β t 2 e − β G \beta t_{2}e^{-\beta G} β t 2 e − β G remains an unproved full-flow prefactor prediction. The corollary resolves the logarithmic exponent, not that stronger limit. In particular a vanishing angular error can suffice for this exponent while an O ( β − 1 ) O(\beta^{-1}) O ( β − 1 ) tangential error can still change the prefactor by an order-one factor.
Growing Systems Governed by Block Mass
Particle count can be separated from the metastable clock more sharply than a minimum-individual-mass estimate permits. Let I β , J β I_{\beta}, J_{\beta} I β , J β be two nonempty blocks, with total masses M I , M J M_I, M_J M I , M J , and define R β = max ( M I / M J , M J / M I ) R_{\beta}=\max(M_I/M_J,M_J/M_I) R β = max ( M I / M J , M J / M I ) . The number of particles in each block is arbitrary and finite for each β \beta β .
Theorem L.1 (Block-mass clock). Fix 0 < θ ∗ < π / 2 0<\theta_*<\pi/2 0 < θ ∗ < π /2 and positive individual masses. Let ε \varepsilon ε be fixed, A β ⊤ = − A β A_{\beta}^{\top}=-A_{\beta} A β ⊤ = − A β , sup β ∥ A β ∥ o p < ∞ \sup_{\beta}\lVert A_{\beta}\rVert_{\mathrm{op}}<\infty sup β ∥ A β ∥ op < ∞ , and let unit centers v β , w β v_{\beta},w_{\beta} v β , w β satisfy v β ⊤ A β w β = 0 v_{\beta}^{\top}A_{\beta}w_{\beta}=0 v β ⊤ A β w β = 0 and ∠ ( v β , w β ) → ϕ ∈ ( θ ∗ , π / 2 ) \angle(v_{\beta},w_{\beta})\to\phi\in(\theta_*,\pi/2) ∠ ( v β , w β ) → ϕ ∈ ( θ ∗ , π /2 ) . Suppose the maximal initial distance from each particle to its block center is r β → 0 r_{\beta}\to0 r β → 0 , and
β − 1 log R β ⟶ 0. \beta^{-1}\log R_{\beta}\longrightarrow0. β − 1 log R β ⟶ 0. Then the first cross-block angle-θ ∗ \theta_* θ ∗ event satisfies
lim β → ∞ β − 1 log T β = 1 − cos ϕ . \lim_{\beta\to\infty}\beta^{-1}\log T_{\beta}=1-\cos\phi. β → ∞ lim β − 1 log T β = 1 − cos ϕ . No bound on β − 1 log ∣ I β ∪ J β ∣ \beta^{-1}\log|I_{\beta}\cup J_{\beta}| β − 1 log ∣ I β ∪ J β ∣ or on individual mass ratios is required. In particular, balanced blocks may contain ⌈ e γ β ⌉ \lceil e^{\gamma\beta}\rceil ⌈ e γ β ⌉ particles, for any fixed γ > 0 \gamma>0 γ > 0 .
Proof. All constants below depend only on a compact neighborhood of the limiting angle, the fixed threshold, and the uniform operator-norm bound; they do not depend on particle count or dimension. Put G = 1 − cos ϕ G=1-\cos\phi G = 1 − cos ϕ .
Lower bound using total block mass. Fix a < G a<G a < G and a small fixed δ > 0 \delta>0 δ > 0 . While each block stays in the angular 2 δ 2\delta 2 δ -cap about its center, a row’s internal scores differ from 1 1 1 by at most C δ C\delta C δ , and cross scores differ from cos ϕ \cos\phi cos ϕ by at most C δ + o ( 1 ) C\delta+o(1) C δ + o ( 1 ) . Summing the entire internal block in the denominator gives
∑ j ∉ I β p i j ≤ R β e − β ( G − C δ − o ( 1 ) ) , i ∈ I β . \sum_{j\notin I_{\beta}}p_{ij}\le R_{\beta}e^{-\beta(G-C\delta-o(1))},\qquad i\in I_{\beta}. j ∈ / I β ∑ p ij ≤ R β e − β ( G − C δ − o ( 1 )) , i ∈ I β .
and the analogous estimate holds for J β J_{\beta} J β . At the minimum of the center scalar product within a block, internal terms are nonnegative and the external contribution is bounded below by minus twice this row weight. Choose δ \delta δ with a < G − C δ a<G-C\delta a < G − C δ and ϕ − 4 δ > θ ∗ \phi-4\delta>\theta_* ϕ − 4 δ > θ ∗ . The integrated external force to time e β a e^{\beta a} e β a tends to zero, since log R β = o ( β ) \log R_{\beta}=o(\beta) log R β = o ( β ) . The cap first-exit argument then closes, giving T β ≥ e β a T_{\beta}\ge e^{\beta a} T β ≥ e β a .
Subexponential entrance to small block diameter. Set α β = r β + β − 1 / 4 \alpha_{\beta}=\sqrt{r_{\beta}}+\beta^{-1/4} α β = r β + β − 1/4 . Bootstrap the two blocks inside 2 α β 2\alpha_{\beta} 2 α β -caps. Their external row weights satisfy
η β ≤ R β e − β ( G − O ( α β ) − o ( 1 ) ) = e − β G + o ( β ) . \eta_{\beta}\le R_{\beta}e^{-\beta(G-O(\alpha_{\beta})-o(1))}=e^{-\beta G+o(\beta)}. η β ≤ R β e − β ( G − O ( α β ) − o ( 1 )) = e − β G + o ( β ) .
For the v v v -block use a i = v ⊤ u i a_i=v^{\top}u_i a i = v ⊤ u i and z i = u i / a i z_i=u_i/a_i z i = u i / a i . Inside these caps, a i ≥ 1 / 2 a_i\ge1/2 a i ≥ 1/2 . Its internal projective coefficients obey
p i j a j a i ≥ κ β m j M I , κ β = c e − C β α β , i , j ∈ I β . p_{ij}\frac{a_j}{a_i}\ge\kappa_{\beta}\frac{m_j}{M_I},\qquad\kappa_{\beta}=ce^{-C\beta\alpha_{\beta}},\qquad i,j\in I_{\beta}. p ij a i a j ≥ κ β M I m j , κ β = c e − C β α β , i , j ∈ I β .
Indeed, all internal score differences are O ( α β ) O(\alpha_{\beta}) O ( α β ) and the external numerator is negligible relative to the full internal denominator. The external projective forcing has norm at most C η β C\eta_{\beta} C η β ; the other block’s projective coordinates are bounded because ϕ < π / 2 \phi<\pi/2 ϕ < π /2 .
Let D I = max i , j ∈ I ∥ z i − z j ∥ D_I=\max_{i,j\in I}\lVert z_i-z_j\rVert D I = max i , j ∈ I ∥ z i − z j ∥ . At an active nonzero diameter, its direction is a supporting direction at both endpoints. Subtract the common positive kernel κ β m j / M I \kappa_\beta m_j/M_I κ β m j / M I from each row. Remaining internal coefficients are nonnegative and cannot enlarge that supporting range. This is a common-kernel consensus contraction, consistent with the positive-system approach of [35 ] . The common kernel contracts the range at rate κ β \kappa_\beta κ β . Thus, in upper Dini derivatives,
D I + ≤ − κ β D I + C η β , D I ( t ) ≤ D I ( 0 ) e − κ β t + C η β / κ β . D_I^+ \le-\kappa_\beta D_I+C\eta_\beta,\qquad D_I(t)\le D_I(0)e^{-\kappa_\beta t}+C\eta_\beta/\kappa_\beta. D I + ≤ − κ β D I + C η β , D I ( t ) ≤ D I ( 0 ) e − κ β t + C η β / κ β .
The same estimate holds for J J J . The inverse map z ↦ z / ∥ z ∥ z\mapsto z/\lVert z\rVert z ↦ z / ∥ z ∥ is uniformly Lipschitz on v ⊤ z = 1 v^\top z=1 v ⊤ z = 1 , so these projective diameter bounds also control spherical-particle diameters. Take s β = β / κ β = e o ( β ) s_\beta=\beta/\kappa_\beta=e^{o(\beta)} s β = β / κ β = e o ( β ) . Then D I ( s β ) + D J ( s β ) ≤ e − c 0 β D_I(s_\beta)+D_J(s_\beta)\le e^{-c_0\beta} D I ( s β ) + D J ( s β ) ≤ e − c 0 β for a fixed c 0 > 0 c_0>0 c 0 > 0 . The cap bootstrap is valid to that time: the integrated external force is e − β G + o ( β ) e^{-\beta G+o(\beta)} e − β G + o ( β ) , much smaller than the cap boundary margin cos ( r β + o ( r β ) ) − cos ( 2 α β ) \cos(r_\beta+o(r_\beta))-\cos(2\alpha_\beta) cos ( r β + o ( r β )) − cos ( 2 α β ) , which is bounded below by a fixed multiple of α β 2 ≥ β − 1 / 2 \alpha_\beta^2\ge\beta^{-1/2} α β 2 ≥ β − 1/2 for large β \beta β .
Persistence until the cross event. The global bisector projective argument bounds all components normal to span ( v , w ) \operatorname{span}(v,w) span ( v , w ) by C r β Cr_\beta C r β for all times. The restriction of A β A_\beta A β to this plane vanishes, so all skew pairings are O ( r β ) O(r_\beta) O ( r β ) . Before the first cross event, cross scalar products are at most cos θ ∗ \cos\theta_* cos θ ∗ . With the blocks already exponentially narrow, every query on a segment joining two same-block particles consequently has external weight at most R β e − β ( 1 − cos θ ∗ − o ( 1 ) ) R_\beta e^{-\beta(1-\cos\theta_*-o(1))} R β e − β ( 1 − c o s θ ∗ − o ( 1 )) .
For V i j = 1 − u i ⊤ u j V_{ij}=1-u_i^\top u_j V ij = 1 − u i ⊤ u j within a block let V = max i , j V i j V=\max_{i,j}V_{ij} V = max i , j V ij . Freeze the configuration and write μ ( x ) = ∑ ℓ p ℓ ( x ) u ℓ \mu(x)=\sum_\ell p_\ell(x)u_\ell μ ( x ) = ∑ ℓ p ℓ ( x ) u ℓ and Δ = u i − u j \Delta=u_i-u_j Δ = u i − u j . Direct differentiation gives the two identities
V ˙ i j = Δ ⊤ [ μ ( u i ) − μ ( u j ) ] − V i j [ u i ⊤ μ ( u i ) + u j ⊤ μ ( u j ) ] , D x μ = β Cov p ( x ) ( u ) M ⊤ . \dot V_{ij}=\Delta^\top[\mu(u_i)-\mu(u_j)]-V_{ij}[u_i^\top\mu(u_i)+u_j^\top\mu(u_j)],\qquad D_x\mu=\beta\operatorname{Cov}_{p(x)}(u)M^\top. V ˙ ij = Δ ⊤ [ μ ( u i ) − μ ( u j )] − V ij [ u i ⊤ μ ( u i ) + u j ⊤ μ ( u j )] , D x μ = β Cov p ( x ) ( u ) M ⊤ .
On each segment the covariance trace is bounded by the second moment about any same-block key, giving ∥ Cov p ( x ) ( u ) ∥ o p ≤ 2 V + 4 η \lVert\operatorname{Cov}_{p(x)}(u)\rVert_{\mathrm{op}}\le2V+4\eta ∥ Cov p ( x ) ( u ) ∥ op ≤ 2 V + 4 η . At either endpoint, internal keys have scalar product at least 1 − V 1-V 1 − V , so u i ⊤ μ ( u i ) ≥ 1 − V − 2 η u_i^\top\mu(u_i)\ge1-V-2\eta u i ⊤ μ ( u i ) ≥ 1 − V − 2 η . The mean-value formula and ∥ Δ ∥ 2 = 2 V i j \lVert\Delta\rVert^2=2V_{ij} ∥ Δ ∥ 2 = 2 V ij therefore give, at an active maximum,
D + V ≤ [ − 2 + 2 V + 4 η + 4 β K V + 8 β K η ] V , K = ∥ I + ε A β ∥ o p . D^+V\le[-2+2V+4\eta+4\beta KV+8\beta K\eta]V,\qquad K=\lVert I+\varepsilon A_\beta\rVert_{\mathrm{op}}. D + V ≤ [ − 2 + 2 V + 4 η + 4 β K V + 8 β K η ] V , K = ∥ I + ε A β ∥ op .
At s β s_\beta s β both β V \beta V β V and β η \beta\eta β η tend to zero. A first-exit argument makes the bracket at most − 1 -1 − 1 up to the cross event, so the block diameters remain exponentially small throughout this interval.
Matching upper clock. The initial global Gram minimum is ρ ( 0 ) = cos ϕ + o ( 1 ) > 0 \rho(0)=\cos\phi+o(1)>0 ρ ( 0 ) = cos ϕ + o ( 1 ) > 0 and remains nondecreasing. Before the first cross event an active global minimum is a cross pair, since internal diameters are exponentially small. For an active pair i ∈ I , j ∈ J i\in I,j\in J i ∈ I , j ∈ J , retain every ℓ ∈ J \ell\in J ℓ ∈ J in its first Gram sum and every ℓ ∈ I \ell\in I ℓ ∈ I in its second. Internal narrowness yields
ρ ˙ ≥ ( P i J + P j I ) ( 1 − ρ 2 − o ( 1 ) ) . \dot\rho\ge(P_{iJ}+P_{jI})(1-\rho^2-o(1)). ρ ˙ ≥ ( P i J + P j I ) ( 1 − ρ 2 − o ( 1 )) .
The remaining terms are nonnegative by the active-minimum argument. All scores lie between cos ϕ − o ( 1 ) \cos\phi-o(1) cos ϕ − o ( 1 ) and 1 + o ( 1 ) 1+o(1) 1 + o ( 1 ) ; hence
P i J ≥ M J M I + M J e − β ( G + o ( 1 ) ) , P j I ≥ M I M I + M J e − β ( G + o ( 1 ) ) . P_{iJ}\ge\frac{M_J}{M_I+M_J}e^{-\beta(G+o(1))},\qquad P_{jI}\ge\frac{M_I}{M_I+M_J}e^{-\beta(G+o(1))}. P i J ≥ M I + M J M J e − β ( G + o ( 1 )) , P j I ≥ M I + M J M I e − β ( G + o ( 1 )) .
The mass fractions add to one. While ρ ≤ cos θ ∗ \rho\le\cos\theta_* ρ ≤ cos θ ∗ , the last factor in the derivative estimate is at least sin 2 θ ∗ / 2 \sin^2\theta_*/2 sin 2 θ ∗ /2 for large β \beta β . Integration bounds the remaining time by C e β ( G + o ( 1 ) ) Ce^{\beta(G+o(1))} C e β ( G + o ( 1 )) . Adding the subexponential entrance time s β s_\beta s β proves the upper bound and completes the proof. □ \square □
Corollary L.2 (Count invariance and an exponential-imbalance boundary). Exact replication of any particle, with its mass split among the copies, leaves every distinct trajectory unchanged. More generally, arbitrarily growing balanced near-collision blocks satisfy Theorem L.1 . For exactly collapsed zero-skew blocks with M J / M I = e β a M_J/M_I=e^{\beta a} M J / M I = e β a , a ∈ R a\in\mathbb{R} a ∈ R , the exponent instead equals [ 1 − cos ϕ − ∣ a ∣ ] + [1-\cos\phi-|a|]_+ [ 1 − cos ϕ − ∣ a ∣ ] + .
Proof. Equal queries have equal velocities, and their numerator weights add to the original mass, proving exact replication invariance by uniqueness. For two exact blocks, the quotient probabilities are
q I J = 1 1 + e β ( 1 − cos ψ − a ) , q J I = 1 1 + e β ( 1 − cos ψ + a ) . q_{IJ} = \frac{1}{1+e^{\beta(1-\cos\psi-a)}}, \qquad q_{JI} = \frac{1}{1+e^{\beta(1-\cos\psi+a)}}. q I J = 1 + e β ( 1 − c o s ψ − a ) 1 , q J I = 1 + e β ( 1 − c o s ψ + a ) 1 .
The angular hitting integral has inverse-rate exponent [ 1 − cos ψ − ∣ a ∣ ] + [1-\cos\psi-|a|]_+ [ 1 − cos ψ − ∣ a ∣ ] + , whose maximum on [ θ ∗ , ϕ ] [\theta_*,\phi] [ θ ∗ , ϕ ] is its value at ϕ \phi ϕ . The endpoint Laplace principle gives the displayed rate. This also shows why a bound based on the smallest individual mass confounds particle replication with physical block imbalance.
Uniformity, Growing Systems, and Verification Details
Theorem M.1 (Dimension-free triangular-array clock). Fix 0 < θ ∗ < π / 2 0<\theta_*<\pi/2 0 < θ ∗ < π /2 . The zero-skew two-cluster clock theorem remains valid for a family with N = N β N=N_\beta N = N β particles, dimension d = d β d=d_\beta d = d β , and positive masses m i , β m_{i,\beta} m i , β and skew matrices A β A_\beta A β , provided
1 β log ∑ i m i , β min i m i , β → 0 , sup β ∥ A β ∥ o p < ∞ , v β ⊤ A β w β = 0 , ∠ ( v β , w β ) → ϕ ∈ ( θ ∗ , π / 2 ) , \frac{1}{\beta}\log\frac{\sum_i m_{i,\beta}}{\min_i m_{i,\beta}}\to0,\qquad\sup_\beta\lVert A_\beta\rVert_{\mathrm{op}}<\infty,\qquad v_\beta^\top A_\beta w_\beta=0,\qquad\angle(v_\beta,w_\beta)\to\phi\in(\theta_*,\pi/2), β 1 log min i m i , β ∑ i m i , β → 0 , β sup ∥ A β ∥ op < ∞ , v β ⊤ A β w β = 0 , ∠ ( v β , w β ) → ϕ ∈ ( θ ∗ , π /2 ) , and the maximal initial block radius tends to zero. Here ε \varepsilon ε is fixed. Its conclusion is β − 1 log T β → 1 − cos ϕ \beta^{-1}\log T_\beta\to1-\cos\phi β − 1 log T β → 1 − cos ϕ .
Proof. Fix a < 1 − cos ϕ a<1-\cos\phi a < 1 − cos ϕ and choose a fixed sufficiently small cap radius δ \delta δ . The cross-barrier estimate is uniform because the operator norms are bounded. The total external row weight is at most
∑ i m i , β min i m i , β e − β ( 1 − cos ϕ − O ( δ ) − o ( 1 ) ) . \frac{\sum_i m_{i,\beta}}{\min_i m_{i,\beta}}e^{-\beta(1-\cos\phi-O(\delta)-o(1))}. min i m i , β ∑ i m i , β e − β ( 1 − c o s ϕ − O ( δ ) − o ( 1 )) .
Its integral to e β a e^{\beta a} e β a tends to zero by the subexponential mass-ratio assumption. The cap first-exit argument is unchanged and gives the lower bound.
For the upper bound, choose the bisector of the two centers. Its scalar products with all initial particles exceed a common constant a ∗ > 0 a_*>0 a ∗ > 0 , since ϕ < π / 2 \phi<\pi/2 ϕ < π /2 and the initial block radii tend to zero. The projective normal-coordinate argument bounds every normal component by r β / a ∗ r_\beta/a_* r β / a ∗ ; convex-hull invariance and operator-norm estimates have no factors of N N N or d d d . Thus all skew pairings are O ( r β ) O(r_\beta) O ( r β ) uniformly. The initial minimum Gram entry is cos ϕ + o ( 1 ) \cos\phi+o(1) cos ϕ + o ( 1 ) and stays nondecreasing. The row-wise transition lower bound is
k β = min i m i , β ∑ i m i , β e − β ( 1 − cos ϕ + o ( 1 ) ) . k_\beta=\frac{\min_i m_{i,\beta}}{\sum_i m_{i,\beta}}e^{-\beta(1-\cos\phi+o(1))}. k β = ∑ i m i , β min i m i , β e − β ( 1 − c o s ϕ + o ( 1 )) .
The acute comparison integrates explicitly as
T β ≤ arctanh ( cos θ ∗ ) − arctanh ( ρ β ( 0 ) ) 2 k β . T_\beta\le\frac{\operatorname{arctanh}(\cos\theta_*)-\operatorname{arctanh}(\rho_\beta(0))}{2k_\beta}. T β ≤ 2 k β arctanh ( cos θ ∗ ) − arctanh ( ρ β ( 0 )) .
when ρ β ( 0 ) < cos θ ∗ \rho_\beta(0)<\cos\theta_* ρ β ( 0 ) < cos θ ∗ ; the numerator is bounded uniformly for large β \beta β . Taking logarithms and using the mass-ratio assumption gives the matching upper rate. For each fixed β \beta β , only a finite minimum is taken, so the standard absolutely continuous minimum rule applies even though the number of coordinates varies with β \beta β .
The mass condition cannot simply be discarded. For two zero-skew centers with initial angle ϕ \phi ϕ , let m 2 / m 1 = e β a m_2/m_1=e^{\beta a} m 2 / m 1 = e β a , where 0 < a < 1 − cos ϕ 0<a<1-\cos\phi 0 < a < 1 − cos ϕ . The two-body angular integral has rate [ 1 − cos ϕ − a ] + [1-\cos\phi-a]_+ [ 1 − cos ϕ − a ] + , because the favorable transition contains the exponent 1 − cos ψ − a 1-\cos\psi-a 1 − cos ψ − a . Its maximal positive value on the crossing interval occurs at ϕ \phi ϕ . Thus exponential mass imbalance changes the clock even though the centers and skew pairing are unchanged.
For completeness, the acute Gram minimum used above satisfies the following direct inequality. At a minimizing off-diagonal pair ( i , j ) (i,j) ( i , j ) with c i j = ρ ≥ 0 c_{ij}=\rho\ge0 c ij = ρ ≥ 0 ,
c ˙ i j = ∑ ℓ p i ℓ ( c ℓ j − c i ℓ ρ ) + ∑ ℓ p j ℓ ( c i ℓ − c j ℓ ρ ) . \dot{c}_{ij}=\sum_{\ell}p_{i\ell}(c_{\ell j}-c_{i\ell}\rho)+\sum_{\ell}p_{j\ell}(c_{i\ell}-c_{j\ell}\rho). c ˙ ij = ℓ ∑ p i ℓ ( c ℓ j − c i ℓ ρ ) + ℓ ∑ p j ℓ ( c i ℓ − c j ℓ ρ ) .
Each summand is nonnegative since c ℓ j ≥ ρ c_{\ell j}\ge\rho c ℓ j ≥ ρ and c i ℓ ≤ 1 c_{i\ell}\le1 c i ℓ ≤ 1 . Keeping ℓ = j \ell=j ℓ = j and ℓ = i \ell=i ℓ = i gives ρ ˙ ≥ ( p i j + p j i ) ( 1 − ρ 2 ) ≥ 2 k β ( 1 − ρ 2 ) \dot{\rho}\ge(p_{ij}+p_{ji})(1-\rho^2)\ge2k\beta(1-\rho^2) ρ ˙ ≥ ( p ij + p j i ) ( 1 − ρ 2 ) ≥ 2 k β ( 1 − ρ 2 ) almost everywhere. This proves both invariance and the integrated bound, with no gradient-flow assumption.
The numerical verification compares the projected normal equation against an explicitly assembled skew-basis linear system, compares its derivative with automatic differentiation and centered finite differences, and tests the second-order attention bound over multiple dimensions, temperatures, skew strengths, masses, and radii. These tests check the implemented equations and constants. The analytic theorems are not claimed to be fully formalized by the finite-dimensional algebra lemmas.
A reproducible integration specification
For validation, set Y ( t ) = ( u 1 ( t ) , u 2 ( t ) , u 3 ( t ) ) ∈ R 9 Y(t)=(u_1(t),u_2(t),u_3(t))\in\mathbb{R}^9 Y ( t ) = ( u 1 ( t ) , u 2 ( t ) , u 3 ( t )) ∈ R 9 and evaluate each row softmax as
s i j = β u i T M u j , a i = max j s i j , p i j = e s i j − a i ∑ k e s i k − a i . (91) s_{ij}=\beta u_i^{\mathsf T}Mu_j,\qquad a_i=\max_j s_{ij},\qquad p_{ij}=\frac{e^{s_{ij}-a_i}}{\sum_k e^{s_{ik}-a_i}}. \tag*{(91)} s ij = β u i T M u j , a i = j max s ij , p ij = ∑ k e s ik − a i e s ij − a i . (91)
Integrate u ˙ i = ∑ j p i j [ u j − ( u i T u j ) u i ] \dot{u}_i=\sum_jp_{ij}[u_j-(u_i^{\mathsf T}u_j)u_i] u ˙ i = ∑ j p ij [ u j − ( u i T u j ) u i ] with DOP853 using relative tolerance 10 − 9 10^{-9} 1 0 − 9 and absolute tolerance 10 − 11 10^{-11} 1 0 − 11 . Trigger a crossing when u i T u j − cos ( 0.3 ) u_i^{\mathsf T}u_j-\cos(0.3) u i T u j − cos ( 0.3 ) changes from negative to positive, adjusting for floating-point drift by normalizing the dot product in the event function. Continue without modifying the state after the first event; the next previously untriggered pair defines the second event. The numerical observations concern only the specified finite parameter values.