Introduction
Collision-free motion planning for multi-agent robotic systems remains an active research area because of the complexity of coordinating robot trajectories, with applications ranging from coordinated drone shows to autonomous operations on land [40 ] , in the air [2 ] , and on water [21 ] . A variety of numerical methods exist to solve this problem, from sampling and graph search [22 , 31 , 30 , 38 ] to joint assignment and planning [35 ] and hybrid search and trajectory optimization [18 ] . Trajectory optimizers use mixed-integer formulations [13 ] , spline parameterizations [36 , 39 ] , sequential convex programming (SCP) [2 , 29 , 7 , 15 ] , learning-based methods [9 , 17 ] , distributed optimization [37 , 32 ] , and separation principles [33 , 27 ] .
Jan Kamm and Antonio Terpin contributed equally to this work.
Corresponding author: Jan Kamm (jakamm@ethz.ch ).
ORCID: Jan Kamm, 0009-0005-6289-4198 ; Antonio Terpin, 0000-0002-8456-6786 ; Raffaello D’Andrea, 0000-0001-5287-7849
We consider collision-free motion planning under double-integrator dynamics, with fixed final time and endpoint positions. The squared-acceleration objective measures control effort and is equivalent to squared-thrust minimization [2 ] . This objective is commonly used for multi-agent assignment and collision avoidance [5 , 4 ] , navigation among polygonal obstacles [6 ] , and, alongside other cost terms, distributed quadrotor planning [25 ] . Various drone and floating-robot experiments have demonstrated that simplified planning models generate reference trajectories amenable to tracking by low-level controllers [2 , 25 , 21 ] . However, these approaches compute trajectories numerically because nonconvex collision-avoidance constraints make an analytical solution difficult. Even for two robots, existing theory provides no complete analytical characterization of minimum-effort transitions between prescribed endpoint states over a fixed time horizon under double-integrator dynamics.
Importantly, the two-robot case isolates the pairwise interaction that underlies fleet planning, and an analytical characterization would identify the fundamental properties of optimal collision-free motion, offering new possibilities for efficient planning. Analytical characterizations exist in other settings. Hu et al. [19 ] construct exact optimal motions for two planar agents through a geometric analysis of the separation constraint. They minimize squared velocity and leave endpoint velocities free. Kirkpatrick and Liu [23 , 24 ] give a complete geometric characterization for two robots minimizing total path length, with paths consisting of straight segments and circular arcs. For the squared-acceleration objective, Beaver and Malikopoulos [5 ] construct trajectories by successively inserting constrained arcs and solving for their joining times. Their method optimizes each agent against prescribed higher-priority trajectories and restricts sustained contact to constant relative speed. A complete characterization of the jointly optimal two-robot motion requires accounting for variable relative speed and determining which combinations of isolated contacts and sustained contact are globally optimal.
Contributions. In this paper, we provide a complete analytical characterization for the optimal trajectories of two robots with zero initial and final velocities under double-integrator dynamics and a squared-acceleration objective. We prove that the contact set consists of finitely many isolated points and boundary arcs, with cubic motion between contacts, and we derive an exact cost identity that gives a sufficient condition for global optimality (§ 2.2.1 and § 2.2.2 ). Then, we prove that every optimal relative trajectory lies in a plane, classify the minimizers into eight contact families, and show that their admissible construction images partition the endpoint space (§ 2.2.3 ). We give a certified numerical procedure that recovers the selected trajectory and its cost to any prescribed positive accuracy for algebraic endpoint data (§ 2.2.4 ). Our two-robot characterization underlies the contact-based fleet planner in [20 ] , which incorporates acceleration constraints and computes collision-free trajectories for one million robots in minutes and for thousands in a fraction of a second.
Results
We consider two robots with zero initial and final velocities, strict initial separation and terminal contact allowed. Their relative position r ( t ) r(t) r ( t ) must remain outside the interior of an exclusion ball whose radius R R R is the required separation. We call an encounter at the minimum separation a contact . The relative trajectory may touch its boundary at isolated points or follow it along an arc, representing sustained contact (Figure 1 (d–g)). Every global optimum belongs to one of eight contact families (Figure 1 (a–c)). When the endpoint segment avoids the disk’s interior, the free cubic is optimal (F). Otherwise, the relative trajectory joins distinct cubic pieces at one, two, or three isolated contacts (P, PP, PPP), or follows a single boundary arc (A), possibly with one isolated contact before (PA), after (AP), or on each side (PAP). Figure 1 (a–c) illustrates numerically how the optimal contact pattern depends on the robots’ initial and final relative positions. Every globally minimizing relative trajectory lies in a plane (Proposition 2.6). In that plane, the initial and final relative positions are ( d 0 , 0 ) (d_0,0) ( d 0 , 0 ) and ( r T , 1 , r T , 2 ) (r_{T,1},r_{T,2}) ( r T , 1 , r T , 2 ) , respectively, in units of R R R . In the following we present the key mathematical results that make this characterization precise.
Figure 1. Two-robot contact structure in relative coordinates. (a–c) Numerical contact-family map for R = T = 1 R = T = 1 R = T = 1 : transition surfaces and slices at d 0 = 1.15 d_0 = 1.15 d 0 = 1.15 and d 0 = 1.343 d_0 = 1.343 d 0 = 1.343 ; the inset magnifies the narrow PPP region. The cylinder excludes ∥ r T ∥ < 1 \lVert r_T \rVert< 1 ∥ r T ∥ < 1 . (d–g) Representative trajectories with R = 1 R = 1 R = 1 . Blue denotes free motion; red dots and segments mark isolated contacts and boundary arcs. Filled and open dots mark initial and final positions.
Problem formulation
Fix a time horizon T > 0 T>0 T > 0 . For integers d ≥ 1 d\geq1 d ≥ 1 and k ≥ 0 k\geq0 k ≥ 0 , L 2 ( ( 0 , T ) ; R d ) L^2((0,T);\mathbb{R}^d) L 2 (( 0 , T ) ; R d ) is the space of square-integrable functions, and H k ( ( 0 , T ) ; R d ) H^k((0,T);\mathbb{R}^d) H k (( 0 , T ) ; R d ) consists of functions whose weak derivatives through order k k k are square-integrable [1 Chapters 2–3] . We write C k ( [ 0 , T ] ; R d ) C^k([0,T];\mathbb{R}^d) C k ([ 0 , T ] ; R d ) for functions with k k k continuous derivatives. For z , w ∈ R d z,w\in\mathbb{R}^d z , w ∈ R d and f ∈ L 2 ( ( 0 , T ) ; R d ) f\in L^2((0,T);\mathbb{R}^d) f ∈ L 2 (( 0 , T ) ; R d ) , set ⟨ z , w ⟩ : = ∑ ℓ = 1 d z ℓ w ℓ \langle z,w\rangle:=\sum_{\ell=1}^{d}z_\ell w_\ell ⟨ z , w ⟩ := ∑ ℓ = 1 d z ℓ w ℓ , ∥ z ∥ : = ⟨ z , z ⟩ 1 / 2 \lVert z\rVert:=\langle z,z\rangle^{1/2} ∥ z ∥ := ⟨ z , z ⟩ 1/2 , and ∥ f ∥ L 2 ( ( 0 , T ) ; R d ) : = ( ∫ 0 T ∥ f ( t ) ∥ 2 d t ) 1 / 2 \lVert f\rVert_{L^2((0,T);\mathbb{R}^d)}:=\left(\int_0^T\lVert f(t)\rVert^2\,\mathrm{d}t\right)^{1/2} ∥ f ∥ L 2 (( 0 , T ) ; R d ) := ( ∫ 0 T ∥ f ( t ) ∥ 2 d t ) 1/2 . We omit the norm subscript when the argument makes its meaning unambiguous. We write ∣ A ∣ \lvert A\rvert ∣ A ∣ for the cardinality of a set A A A and ∣ a ∣ \lvert a\rvert ∣ a ∣ for the absolute value of a scalar a a a . The vectors e 1 , … , e d e_1,\ldots,e_d e 1 , … , e d denote the orthonormal coordinate vectors of the chosen spatial frame. Square brackets enclose function or trajectory arguments; parentheses indicate evaluation at a time or a finite-dimensional argument. Write D t k D_t^k D t k for the k k k th distributional time derivative, including D t 1 D_t^1 D t 1 for the first. Ordinary time derivatives are denoted by r ˙ \dot r r ˙ , r ¨ \ddot r r ¨ , r ( 3 ) r^{(3)} r ( 3 ) , and r ( 4 ) r^{(4)} r ( 4 ) , with the same convention for other trajectories. For Sobolev trajectories, these symbols denote the corresponding weak derivatives. Dots on normalized trajectories refer to their displayed time variable. Time differentiation acts componentwise on vector-valued functions. We write M + ( ( 0 , T ] ) \mathcal{M}_+((0,T]) M + (( 0 , T ]) for finite nonnegative Radon measures and δ s \delta_s δ s for the Dirac measure with unit mass at s s s . The support supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) consists of points whose every neighborhood in [ 0 , T ] [0,T] [ 0 , T ] has positive μ \mu μ -measure. For a function with finite one-sided limits, write f ( t − ) : = lim s ↑ t f ( s ) f(t^-):=\lim_{s\uparrow t}f(s) f ( t − ) := lim s ↑ t f ( s ) , f ( t + ) : = lim s ↓ t f ( s ) f(t^+):=\lim_{s\downarrow t}f(s) f ( t + ) := lim s ↓ t f ( s ) , and [ [ f ] ] t : = f ( t + ) − f ( t − ) [\![f]\!]_t:=f(t^+)-f(t^-) [ [ f ] ] t := f ( t + ) − f ( t − ) . The minus and plus signs indicate limits from the left and right, respectively; f ( 0 + ) f(0^+) f ( 0 + ) and f ( T − ) f(T^-) f ( T − ) are the endpoint limits from inside the time interval. A right-continuous representative takes the value f ( t + ) f(t^+) f ( t + ) at each interior time. In measure integrals, unmarked limits include endpoints, while a lower limit a + a^+ a + excludes a a a and an upper limit b − b^- b − excludes b b b . For example, ∫ 0 + t f ( s ) d μ ( s ) \int_{0^+}^{t}f(s)\,\mathrm{d}\mu(s) ∫ 0 + t f ( s ) d μ ( s ) integrates over 0 < s ≤ t 0<s\leq t 0 < s ≤ t , whereas ∫ 0 + T − f ( s ) d μ ( s ) \int_{0^+}^{T^-}f(s)\,\mathrm{d}\mu(s) ∫ 0 + T − f ( s ) d μ ( s ) integrates over 0 < s < T 0<s<T 0 < s < T .
Two-robot model
Let r , b ∈ H 2 ( ( 0 , T ) ; R d ) r,b \in H^2((0,T); \mathbb{R}^d) r , b ∈ H 2 (( 0 , T ) ; R d ) , d ≥ 2 d \ge2 d ≥ 2 , denote the position of the first robot relative to the second and their barycenter, respectively. Both trajectories have prescribed endpoint positions r 0 , r T r_0,r_T r 0 , r T and b 0 , b T b_0,b_T b 0 , b T , and zero initial and final velocities. The endpoint data satisfy ∥ r 0 ∥ > R \lVert r_0\rVert> R ∥ r 0 ∥ > R and ∥ r T ∥ ≥ R \lVert r_T\rVert\ge R ∥ r T ∥ ≥ R , where R > 0 R > 0 R > 0 is the minimum separation. The cost is 2 ∥ b ¨ ∥ 2 + 1 2 ∥ r ¨ ∥ 2 2\lVert\ddot{b}\rVert^2 + \frac{1}{2}\lVert\ddot{r}\rVert^2 2 ∥ b ¨ ∥ 2 + 2 1 ∥ r ¨ ∥ 2 . Since separation constrains only r r r , the barycentric and relative trajectories minimize their costs independently. The unique optimal barycenter is the cubic Hermite interpolant [14 Section 2.5] b H ( t ) = b 0 + ( b T − b 0 ) t ^ 2 ( 3 − 2 t ^ ) b^H(t)=b_0+(b_T-b_0)\hat{t}^2(3-2\hat{t}) b H ( t ) = b 0 + ( b T − b 0 ) t ^ 2 ( 3 − 2 t ^ ) , where t ^ = t / T \hat{t}=t/T t ^ = t / T . The relative trajectory solves
min r 1 2 ∥ r ¨ ∥ 2 s.t. r ( 0 ) = r 0 , r ( T ) = r T , r ˙ ( 0 ) = r ˙ ( T ) = 0 , ∥ r ( t ) ∥ ≥ R , t ∈ [ 0 , T ] . (1) \min_r \frac{1}{2}\lVert\ddot{r}\rVert^2 \quad\text{s.t. } r(0)=r_0,\ r(T)=r_T,\ \dot{r}(0)=\dot{r}(T)=0,\ \lVert r(t)\rVert\ge R,\ t\in[0,T]. \tag*{(1)} r min 2 1 ∥ r ¨ ∥ 2 s.t. r ( 0 ) = r 0 , r ( T ) = r T , r ˙ ( 0 ) = r ˙ ( T ) = 0 , ∥ r ( t )∥ ≥ R , t ∈ [ 0 , T ] . (1)
Let M ∗ \mathcal{M}^* M ∗ denote its set of global minimizers.
Remark 2.1 (Nondimensionalization). For t ^ = t / T ∈ [ 0 , 1 ] \hat{t}=t/T\in[0,1] t ^ = t / T ∈ [ 0 , 1 ] , define r ^ ( t ^ ) : = r ( T t ^ ) / R \hat{r}(\hat{t}):=r(T\hat{t})/R r ^ ( t ^ ) := r ( T t ^ ) / R . The inverse r ( t ) = R r ^ ( t / T ) r(t)=R\hat{r}(t/T) r ( t ) = R r ^ ( t / T ) identifies feasible trajectories with those of (1) at T = R = 1 T=R=1 T = R = 1 , with endpoint positions r 0 / R r_0/R r 0 / R , r T / R r_T/R r T / R and zero initial and final velocities. Moreover, ∥ r ¨ ∥ L 2 ( ( 0 , T ) ; R d ) 2 = R 2 T − 3 ∥ r ^ ¨ ∥ L 2 ( ( 0 , 1 ) ; R d ) 2 \lVert\ddot{r}\rVert_{L^2((0,T);\mathbb{R}^d)}^2=R^2T^{-3}\lVert\ddot{\hat{r}}\rVert_{L^2((0,1);\mathbb{R}^d)}^2 ∥ r ¨ ∥ L 2 (( 0 , T ) ; R d ) 2 = R 2 T − 3 ∥ r ^ ¨ ∥ L 2 (( 0 , 1 ) ; R d ) 2 . Thus minimizers correspond bijectively. Hats denote normalization; we omit them when T = R = 1 T=R=1 T = R = 1 .
Remark 2.2 (Existence and nonuniqueness). The direct method [28 Theorem 2.3] guarantees a global minimizer of (1) . This solution may not be unique. In fact, uniqueness fails for opposite endpoint vectors, since symmetry gives equally costly optimal trajectories passing on opposite sides of the separation ball.
Optimal trajectory characterization
To characterize the structure of the optimal trajectories, we first derive optimality conditions (Proposition 2.3) and construct a global certificate (§ 2.2.2 ). Using the optimality conditions, we characterize the active set, namely the times at which the robots attain the minimum separation, as finitely many isolated contacts and boundary arcs (Proposition 2.4 ). We then prove that every optimal relative trajectory lies in a plane and establish side selection to complete the catalogue of eight contact families and determine their regions in endpoint space (§ 2.2.3 ). We conclude with a numerical procedure to recover the selected optimal trajectory and its cost (§ 2.2.4 ).
Contact laws
Write A = { t ∈ [ 0 , T ] : ∥ r ( t ) ∥ = R } \mathcal{A}=\{t\in[0,T]:\lVert r(t)\rVert=R\} A = { t ∈ [ 0 , T ] : ∥ r ( t )∥ = R } for the active set and g ( t ) = R 2 − ∥ r ( t ) ∥ 2 g(t)=R^2-\lVert r(t)\rVert^2 g ( t ) = R 2 − ∥ r ( t ) ∥ 2 for the clearance. The connected components of A \mathcal{A} A are the contacts. An isolated point is an isolated contact ; a contact of positive length is a boundary arc , necessarily a closed interval [ a , b ] [a,b] [ a , b ] with a < b a<b a < b . A measure represents both isolated contacts and contact maintained over an interval.
Proposition 2.3 (Karush–Kuhn–Tucker conditions for the relative problem). If ∥ r T ∥ > R \lVert r_T\rVert>R ∥ r T ∥ > R , every local minimizer r ∗ r^* r ∗ of (1) in H 2 ( ( 0 , T ) ; R d ) H^2((0,T);\mathbb{R}^d) H 2 (( 0 , T ) ; R d ) admits μ ∈ M + ( [ 0 , T ] ) \mu\in\mathcal{M}_+([0,T]) μ ∈ M + ([ 0 , T ]) such that ∫ 0 T ⟨ r ¨ ∗ ( t ) , η ˙ ( t ) ⟩ d t = ∫ 0 T ⟨ r ∗ ( t ) , η ( t ) ⟩ d μ ( t ) \int_0^T\langle\ddot{r}^*(t),\dot{\eta}(t)\rangle\,dt=\int_0^T\langle r^*(t),\eta(t)\rangle\,d\mu(t) ∫ 0 T ⟨ r ¨ ∗ ( t ) , η ˙ ( t )⟩ d t = ∫ 0 T ⟨ r ∗ ( t ) , η ( t )⟩ d μ ( t ) for every η ∈ H 2 ( ( 0 , T ) ; R d ) \eta\in H^2((0,T);\mathbb{R}^d) η ∈ H 2 (( 0 , T ) ; R d ) with zero endpoint values and first derivatives. Moreover, ∫ 0 T ( R 2 − ∥ r ∗ ( t ) ∥ 2 ) d μ ( t ) = 0 \int_0^T(R^2-\lVert r^*(t)\rVert^2)\,d\mu(t)=0 ∫ 0 T ( R 2 − ∥ r ∗ ( t ) ∥ 2 ) d μ ( t ) = 0 and supp ( μ ) ⊆ A \operatorname{supp}(\mu)\subseteq\mathcal{A} supp ( μ ) ⊆ A .
The multiplier μ \mu μ describes how the separation constraint changes the free cubic motion: at an isolated contact, positive mass produces a jump in jerk, whereas zero mass lets the same cubic continue through the contact.
If ∥ r T ∥ > R \lVert r_T\rVert> R ∥ r T ∥ > R , Proposition 2.3 gives every local minimizer a multiplier measure μ \mu μ satisfying
D t 4 r = r μ , μ ≥ 0 , supp ( μ ) ⊆ A , (2) D_t^4 r = r\,\mu,\qquad\mu\ge0,\qquad\operatorname{supp}(\mu) \subseteq\mathcal{A}, \tag*{(2)} D t 4 r = r μ , μ ≥ 0 , supp ( μ ) ⊆ A , (2)
in distributions on ( 0 , T ) (0,T) ( 0 , T ) . On boundary arcs, the density of μ \mu μ determines the rate of change of jerk. We call a feasible trajectory r r r stationary if some finite nonnegative Radon measure μ \mu μ satisfies (2) ; the conclusions below apply to every such pair ( r , μ ) (r,\mu) ( r , μ ) .
Proposition 2.4 (Contact structure). The active set has finitely many connected components, each an isolated contact or a boundary arc: A = { t i } i = 1 m ∪ ⋃ ℓ = 1 k [ a ℓ , b ℓ ] \mathcal{A} = \{t_i\}_{i=1}^{m} \cup\bigcup_{\ell=1}^{k}[a_\ell,b_\ell] A = { t i } i = 1 m ∪ ⋃ ℓ = 1 k [ a ℓ , b ℓ ] . The multiplier has no singular-continuous part. Its atoms lie at isolated contacts and arc endpoints; its remaining density is nonnegative, Lipschitz continuous on each arc interior, and zero almost everywhere outside the arcs.
A stationary trajectory is thus a finite chain of free cubics, isolated contacts, and boundary arcs.
Global optimality certificate
The conditions in § 2.2.1 characterize stationary trajectories, which may be only locally optimal. To certify global optimality, we compare a stationary trajectory r r r with every feasible trajectory q q q . We use a nonnegative multiplier measure μ \mu μ for the separation constraint along r r r , satisfying (2) , to derive an exact identity for their cost difference.
The free cubic and its correction. Let r ˉ \bar{r} r ˉ be the unique unconstrained minimizer of 1 2 ∥ r ¨ ∥ 2 \frac{1}{2}\lVert\ddot{r}\rVert^2 2 1 ∥ r ¨ ∥ 2 with the prescribed endpoint states. It is the cubic Hermite interpolant of the relative endpoint data. Every trajectory with these endpoint states has the form r = r ˉ + h r = \bar{r} + h r = r ˉ + h , where h h h belongs to the space H T H_T H T of functions in H 2 ( ( 0 , T ) ; R d ) H^2((0,T);\mathbb{R}^d) H 2 (( 0 , T ) ; R d ) with h ( 0 ) = h ( T ) = h ˙ ( 0 ) = h ˙ ( T ) = 0 h(0) = h(T) = \dot{h}(0) = \dot{h}(T) = 0 h ( 0 ) = h ( T ) = h ˙ ( 0 ) = h ˙ ( T ) = 0 . Equip this space with the inner product ⟨ h 1 , h 2 ⟩ H T : = ∫ 0 T ⟨ h ¨ 1 ( t ) , h ¨ 2 ( t ) ⟩ d t \langle h_1,h_2\rangle_{H_T} := \int_0^T \langle\ddot{h}_1(t),\ddot{h}_2(t)\rangle\,\mathrm{d}t ⟨ h 1 , h 2 ⟩ H T := ∫ 0 T ⟨ h ¨ 1 ( t ) , h ¨ 2 ( t )⟩ d t . The correction adds its own acceleration cost to that of the free cubic (Section A.2.3): 1 2 ∥ r ˉ ¨ + h ¨ ∥ 2 = 1 2 ∥ r ˉ ¨ ∥ 2 + 1 2 ∥ h ∥ H T 2 \frac{1}{2}\lVert\ddot{\bar{r}}+\ddot{h}\rVert^2 = \frac{1}{2}\lVert\ddot{\bar{r}}\rVert^2 + \frac{1}{2}\lVert h\rVert_{H_T}^2 2 1 ∥ r ˉ ¨ + h ¨ ∥ 2 = 2 1 ∥ r ˉ ¨ ∥ 2 + 2 1 ∥ h ∥ H T 2 .
From multipliers to trajectories. Restrict the measure μ \mu μ to ( 0 , T ) (0,T) ( 0 , T ) . The stationarity equation (2) gives D t 4 h = r μ D_t^4h = r\,\mu D t 4 h = r μ for the correction h = r − r ˉ h = r-\bar{r} h = r − r ˉ , with zero endpoint position and velocity. The scalar Green kernel G T ( t , s ) G_T(t,s) G T ( t , s ) describes the correction at time t t t produced by a unit jump in jerk at time s s s under these endpoint conditions. Rescaling the unit-interval formula [26 ] , Equation (3.3) by G T ( t , s ) = T 3 G 1 ( t / T , s / T ) G_T(t,s) = T^3G_1(t/T,s/T) G T ( t , s ) = T 3 G 1 ( t / T , s / T ) gives G T ( t , s ) = t 2 ( T − s ) 2 6 T 3 ( 3 T s − T t − 2 t s ) G_T(t,s) = \frac{t^2(T-s)^2}{6T^3}(3Ts-Tt-2ts) G T ( t , s ) = 6 T 3 t 2 ( T − s ) 2 ( 3 T s − T t − 2 t s ) (2.3) for $0 \le t \le s \le T$, with G T ( t , s ) = G T ( s , t ) G_T(t,s) = G_T(s,t) G T ( t , s ) = G T ( s , t ) for t > s t>s t > s . Adding the corrections from all contacts gives (Section A.2.3)
r ( t ) = r ˉ ( t ) + ∫ 0 + T − G T ( t , s ) r ( s ) d μ ( s ) r(t) = \bar{r}(t) + \int_{0+}^{T-} G_T(t,s)\,r(s)\,\mathrm{d}\mu(s) r ( t ) = r ˉ ( t ) + ∫ 0 + T − G T ( t , s ) r ( s ) d μ ( s ) (2.4) for t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] .
Comparing costs. Define Q μ [ h ] : = ∥ h ∥ H T 2 − ∫ 0 + T − ∥ h ( t ) ∥ 2 d μ ( t ) Q_{\mu}[h] := \lVert h\rVert_{\mathcal{H}_T}^{2} - \int_{0+}^{T-} \lVert h(t)\rVert^{2}\,\mathrm{d}\mu(t) Q μ [ h ] := ∥ h ∥ H T 2 − ∫ 0 + T − ∥ h ( t ) ∥ 2 d μ ( t ) for h ∈ H T h \in\mathcal{H}_T h ∈ H T . We can now compare r r r with any trajectory q = r + h q=r+h q = r + h with the same endpoint states through the exact identity
1 2 ( ∥ q ¨ ∥ 2 − ∥ r ¨ ∥ 2 ) = 1 2 Q μ [ h ] + 1 2 ∫ 0 + T − ( ∥ q ( t ) ∥ 2 − R 2 ) d μ ( t ) . (3) \frac{1}{2}\left(\lVert\ddot{q}\rVert^{2}-\lVert\ddot{r}\rVert^{2}\right)=\frac{1}{2}Q_{\mu}[h]+\frac{1}{2}\int_{0+}^{T-}\left(\lVert q(t)\rVert^{2}-R^{2}\right)\,\mathrm{d}\mu(t). \tag*{(3)} 2 1 ( ∥ q ¨ ∥ 2 − ∥ r ¨ ∥ 2 ) = 2 1 Q μ [ h ] + 2 1 ∫ 0 + T − ( ∥ q ( t ) ∥ 2 − R 2 ) d μ ( t ) . (3)
For every feasible q q q , the separation constraint makes the last integral nonnegative. Thus, Q μ ≥ 0 Q_{\mu}\geq0 Q μ ≥ 0 on H T \mathcal{H}_T H T certifies global optimality of r r r . The following theorem gives a sufficient condition for this nonnegativity in terms of projections of r r r and r ˉ \bar{r} r ˉ on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) .
Theorem 2.5 (A global optimality certificate). Let r r r be stationary with multiplier μ \mu μ , and let e ∈ R d e\in\mathbb{R}^{d} e ∈ R d be a unit vector and δ ≥ 0 \delta\geq0 δ ≥ 0 such that z ˉ ( t ) : = ⟨ e , r ˉ ( t ) ⟩ ≥ δ \bar{z}(t):=\langle e,\bar{r}(t)\rangle\geq\delta z ˉ ( t ) := ⟨ e , r ˉ ( t )⟩ ≥ δ and z ( t ) : = ⟨ e , r ( t ) ⟩ > 0 z(t):=\langle e,r(t)\rangle>0 z ( t ) := ⟨ e , r ( t )⟩ > 0 for t ∈ supp ( μ ) ∩ ( 0 , T ) t\in\operatorname{supp}(\mu)\cap(0,T) t ∈ supp ( μ ) ∩ ( 0 , T ) . Then, with κ c e r t : = min { 1 , δ / R } \kappa_{\mathrm{cert}}:=\min\{1,\delta/R\} κ cert := min { 1 , δ / R } , Q μ [ h ] ≥ κ c e r t ∥ h ∥ H T 2 Q_{\mu}[h]\geq\kappa_{\mathrm{cert}}\lVert h\rVert_{\mathcal{H}_T}^{2} Q μ [ h ] ≥ κ cert ∥ h ∥ H T 2 and 1 2 ( ∥ q ¨ ∥ 2 − ∥ r ¨ ∥ 2 ) ≥ κ c e r t 2 ∥ q − r ∥ H T 2 \frac{1}{2}\left(\lVert\ddot{q}\rVert^{2}-\lVert\ddot{r}\rVert^{2}\right)\geq\frac{\kappa_{\mathrm{cert}}}{2}\lVert q-r\rVert_{\mathcal{H}_T}^{2} 2 1 ( ∥ q ¨ ∥ 2 − ∥ r ¨ ∥ 2 ) ≥ 2 κ cert ∥ q − r ∥ H T 2 for every feasible q q q . If δ > 0 \delta>0 δ > 0 , r r r is the unique minimizer.
Geometrically, the contact positions at times in supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) lie in one open hemisphere and the free motion lies in the corresponding closed half-space at those times; this alignment certifies optimality even against trajectories passing around the opposite side of the ball.
The contact catalogue
With zero initial and final velocities the free cubic runs along the segment joining the endpoints, so its projection on the normal to that segment is constant: the endpoint data alone fix a direction for Theorem 2.5 . Scale time and separation to T = R = 1 T=R=1 T = R = 1 as in § 2.1 . The endpoint data satisfy ∥ r 0 ∥ > 1 \lVert r_0\rVert>1 ∥ r 0 ∥ > 1 and ∥ r T ∥ ≥ 1 \lVert r_T\rVert\geq1 ∥ r T ∥ ≥ 1 .
Normalized endpoint data. Choose orthonormal vectors e 1 , e 2 ∈ R d e_1,e_2\in\mathbb{R}^{d} e 1 , e 2 ∈ R d with r 0 = d 0 e 1 r_0=d_0e_1 r 0 = d 0 e 1 and r T = r T , 1 e 1 + r T , 2 e 2 r_T=r_{T,1}e_1+r_{T,2}e_2 r T = r T , 1 e 1 + r T , 2 e 2 , r T , 2 ≥ 0 r_{T,2}\geq0 r T , 2 ≥ 0 ; if r T r_T r T is parallel to r 0 r_0 r 0 , e 2 e_2 e 2 is any unit vector orthogonal to e 1 e_1 e 1 . Then the datum ω : = ( d 0 , r T , 1 , r T , 2 ) \omega:=(d_0,r_{T,1},r_{T,2}) ω := ( d 0 , r T , 1 , r T , 2 ) lies in D : = { ( d 0 , r T , 1 , r T , 2 ) ∈ R 3 : d 0 > 1 , r T , 1 2 + r T , 2 2 ≥ 1 , r T , 2 ≥ 0 } \mathcal{D}:=\{(d_0,r_{T,1},r_{T,2})\in\mathbb{R}^{3}:d_0>1,\ r_{T,1}^{2}+r_{T,2}^{2}\geq1,\ r_{T,2}\geq0\} D := {( d 0 , r T , 1 , r T , 2 ) ∈ R 3 : d 0 > 1 , r T , 1 2 + r T , 2 2 ≥ 1 , r T , 2 ≥ 0 } . The symmetry axis r T , 2 = 0 r_{T,2}=0 r T , 2 = 0 and the terminal wall ∥ r T ∥ = 1 \lVert r_T\rVert=1 ∥ r T ∥ = 1 are the faces of D \mathcal{D} D . Identify the chosen plane span { e 1 , e 2 } \operatorname{span}\{e_1,e_2\} span { e 1 , e 2 } with R 2 \mathbb{R}^{2} R 2 , and write a × b : = a 1 b 2 − a 2 b 1 a\times b:=a_1b_2-a_2b_1 a × b := a 1 b 2 − a 2 b 1 for planar vectors.
The free cubic r ˉ ( t ) = r 0 + ( r T − r 0 ) t 2 ( 3 − 2 t ) \bar{r}(t)=r_0+(r_T-r_0)t^{2}(3-2t) r ˉ ( t ) = r 0 + ( r T − r 0 ) t 2 ( 3 − 2 t ) , 0 ≤ t ≤ 1 0\leq t\leq1 0 ≤ t ≤ 1 , traverses the endpoint segment. Call the data blocked when that segment meets the open unit disk, that is, when the free cubic is infeasible. Unblocked data are settled by Theorem 2.5 at μ = 0 \mu=0 μ = 0 : the free cubic is their unique minimizer. Call data on the symmetry axis symmetric and all other data off-axis. For off-axis data the oriented unit normal to the line through the endpoints is e : = ( r T , 2 , d 0 − r T , 1 ) / ∥ r T − r 0 ∥ e:=(r_{T,2},d_0-r_{T,1})/\lVert r_T-r_0\rVert e := ( r T , 2 , d 0 − r T , 1 ) / ∥ r T − r 0 ∥ . With δ : = ⟨ e , r 0 ⟩ = ⟨ e , r T ⟩ = d 0 r T , 2 / ∥ r T − r 0 ∥ \delta:=\langle e,r_0\rangle=\langle e,r_T\rangle=d_0r_{T,2}/\lVert r_T-r_0\rVert δ := ⟨ e , r 0 ⟩ = ⟨ e , r T ⟩ = d 0 r T , 2 / ∥ r T − r 0 ∥ , we have ⟨ e , r ˉ ( t ) ⟩ = δ \langle e,\bar{r}(t)\rangle=\delta ⟨ e , r ˉ ( t )⟩ = δ throughout. For symmetric data set e = ( 0 , 1 ) e=(0,1) e = ( 0 , 1 ) and δ = 0 \delta=0 δ = 0 . With this e e e , the hypotheses of Theorem 2.5 reduce to ⟨ e , r ( t ) ⟩ > 0 \langle e,r(t)\rangle>0 ⟨ e , r ( t )⟩ > 0 on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) . The near side is the half-plane ⟨ e , ⋅ ⟩ ≥ δ \langle e,\cdot\rangle\geq\delta ⟨ e , ⋅ ⟩ ≥ δ in the chosen plane; for off-axis data, it is the side of the line through the endpoints away from the center of the disk. A trajectory in the chosen plane is upper if ⟨ e , r ( t ) ⟩ > 0 \langle e,r(t)\rangle>0 ⟨ e , r ( t )⟩ > 0 for 0 < t < 1 0<t<1 0 < t < 1 .
Proposition 2.6 (Planar reduction). For every r ∈ M ∗ r\in\mathcal{M}^{*} r ∈ M ∗ , the image r ( [ 0 , 1 ] ) r([0,1]) r ([ 0 , 1 ]) lies in a two-dimensional linear subspace of R d \mathbb{R}^{d} R d .
Intuitively, Proposition 2.6 shows that additional spatial directions cannot reduce the minimum effort.
Proposition 2.7 (Side selection). For off-axis data, M ∗ = { r ∗ } \mathcal{M}^{*}=\{r_{*}\} M ∗ = { r ∗ } and r ∗ ( [ 0 , 1 ] ) r_{*}([0,1]) r ∗ ([ 0 , 1 ]) is contained in the near side. For blocked symmetric data, M ∗ \mathcal{M}^{*} M ∗ is the orbit of its unique upper element under the orthogonal maps of R d \mathbb{R}^{d} R d that fix r 0 r_{0} r 0 and r T r_{T} r T .
For off-axis data with ∥ r T ∥ > 1 \lVert r_{T}\rVert>1 ∥ r T ∥ > 1 , the two ways around the disk each carry a local minimizer among trajectories in the chosen plane (Lemma A.7 ), and Proposition 2.7 puts the global one on the near side. The selected minimizer is thus well defined: the unique minimizer for off-axis or unblocked data, and the upper element of M ∗ \mathcal{M}^{*} M ∗ for blocked symmetric data.
The contact catalogue. With the side fixed, the laws of § 2.2.1 limit how the selected minimizer can touch the disk. Every global minimizer has a unique finite interior multiplier μ \mu μ satisfying (2) (Proposition A.6 ). The labels in Figure 1 (a–Figure 1 c) record its contact pattern. Write P for an isolated contact at t t t with positive mass μ ( { t } ) > 0 \mu(\{t\})>0 μ ({ t }) > 0 , A for a maximal boundary arc together with the atoms at its endpoints, and F for μ = 0 \mu=0 μ = 0 ; concatenation records temporal order, coincident atoms combine, and a zero-mass touch, like a prescribed endpoint contact, contributes no letter.
Theorem 2.8 (Canonical contact catalogue). For every datum in D \mathcal{D} D , the contact pattern of the selected minimizer is one of F, P, PP, PPP, A, PA, AP, and PAP.
The catalogue reduces the search for the optimum to eight finite-dimensional constructions. Figure 1 (d–Figure 1 g) shows representatives of four of the eight families.
Atomic families. For blocked data, choose m ∈ { 1 , 2 , 3 } m\in\{1,2,3\} m ∈ { 1 , 2 , 3 } , ordered times 0 < t 1 < ⋯ < t m < 1 0<t_{1}<\cdots<t_{m}<1 0 < t 1 < ⋯ < t m < 1 , and jerk jumps Δ i ∈ R 2 \Delta_{i}\in\mathbb{R}^{2} Δ i ∈ R 2 . With atoms at these times, the kernel response (2.4) takes the form r ( t ) = r ˉ ( t ) + ∑ i = 1 m G 1 ( t , t i ) Δ i r(t)=\bar{r}(t)+\sum_{i=1}^{m}G_{1}(t,t_{i})\Delta_{i} r ( t ) = r ˉ ( t ) + ∑ i = 1 m G 1 ( t , t i ) Δ i . With n i = r ( t i ) n_{i}=r(t_{i}) n i = r ( t i ) , c i = ( t i , Δ i ) c_{i}=(t_{i},\Delta_{i}) c i = ( t i , Δ i ) , and c = ( c 1 , … , c m ) c=(c_{1},\ldots,c_{m}) c = ( c 1 , … , c m ) , the contact conditions form the system
F ( c ) : = ( ∥ n i ∥ 2 − 1 , n i × Δ i , ⟨ n i , r ˙ ( t i ) ⟩ ) i = 1 m = 0. (4) F(c):=\left(\lVert n_{i}\rVert^{2}-1,\ n_{i}\times\Delta_{i},\ \langle n_{i},\dot{r}(t_{i})\rangle\right)_{i=1}^{m}=0. \tag*{(4)} F ( c ) := ( ∥ n i ∥ 2 − 1 , n i × Δ i , ⟨ n i , r ˙ ( t i )⟩ ) i = 1 m = 0. (4)
These are 3 m 3m 3 m equations in 3 m 3m 3 m unknowns for P, PP, and PPP; the atom masses must also satisfy m i : = ⟨ n i , Δ i ⟩ > 0 m_{i}:=\langle n_{i},\Delta_{i}\rangle>0 m i := ⟨ n i , Δ i ⟩ > 0 . Each real solution of (4) defines a candidate trajectory. It satisfies the separation constraint exactly when the polynomial ∥ r ( t ) ∥ 2 − 1 \lVert r(t)\rVert^{2}-1 ∥ r ( t ) ∥ 2 − 1 is nonnegative on every cubic piece; the tests are given in Section A.2.7.
Arc families. On a boundary arc, the robots remain at the minimum separation. Their relative position is described by a single angle, whose rate of change satisfies the differential equation in Proposition A.10. At either end a free cubic reaches the endpoint with zero velocity directly, giving A, or passes through one isolated contact first, giving PA, AP, or PAP according to the side. Section A.2.7 evaluates the arc system and derives both completions. A construction is admissible if its endpoint triple lies in D \mathcal{D} D , its atom masses are positive, every free piece clears the disk, ⟨ e , ⋅ ⟩ > 0 \langle e,\cdot\rangle>0 ⟨ e , ⋅ ⟩ > 0 on the support of its multiplier, and any arc has positive speed and nonnegative density.
Regime regions. Let L = { F , P , P P , P P P , A , P A , A P , P A P } \mathcal{L}=\{\mathrm{F},\mathrm{P},\mathrm{PP},\mathrm{PPP},\mathrm{A},\mathrm{PA},\mathrm{AP},\mathrm{PAP}\} L = { F , P , PP , PPP , A , PA , AP , PAP } . For σ ∈ L \sigma\in\mathcal{L} σ ∈ L , let P σ \mathcal{P}_{\sigma} P σ be the parameters of the admissible constructions of § 2.2.3 whose contact pattern has label σ \sigma σ , retaining all admitted roots, and let E σ ( p ) \mathcal{E}_{\sigma}(p) E σ ( p ) be the normalized endpoint triple of the assembled trajectory, using the map E c a n \mathcal{E}_{\mathrm{can}} E can in Section A.2.7 when rotation is needed. Set R σ : = E σ ( P σ ) ⊆ D \mathcal{R}_{\sigma}:=\mathcal{E}_{\sigma}(\mathcal{P}_{\sigma})\subseteq\mathcal{D} R σ := E σ ( P σ ) ⊆ D . Thus R F \mathcal{R}_{\mathrm{F}} R F is the set of unblocked data.
Theorem 2.9 (Regime map partition). The regions R σ \mathcal{R}_{\sigma} R σ , σ ∈ L \sigma\in\mathcal{L} σ ∈ L , are pairwise disjoint and cover D \mathcal{D} D . For ω ∈ R σ \omega\in\mathcal{R}_{\sigma} ω ∈ R σ , every admissible construction mapping to ω \omega ω gives the selected minimizer, with label σ \sigma σ , and every selected minimizer arises this way.
Figure 1(a–c) illustrates these regions numerically. The procedure in § 2.2.4 computes the selected optimal trajectory to a prescribed accuracy.
Numerical recovery
Algorithm 2.1 first checks the free cubic, then the families P \mathrm{P} P , P P \mathrm{PP} PP , and P P P \mathrm{PPP} PPP . If none applies, it searches A \mathrm{A} A , P A \mathrm{PA} PA , A P \mathrm{AP} AP , and P A P \mathrm{PAP} PAP by narrowing intervals for their contact parameters. The normalized endpoint data ω ∈ D \omega\in\mathcal{D} ω ∈ D (Remark 2.1) are algebraic, and the requested accuracy ε > 0 \varepsilon>0 ε > 0 is rational. Write J [ q ] = 1 2 ∥ q ¨ ∥ 2 J[q]=\frac{1}{2}\lVert\ddot{q}\rVert^2 J [ q ] = 2 1 ∥ q ¨ ∥ 2 for the cost. The output is a trial trajectory r ~ \tilde{r} r ~ , a bound ρ \rho ρ on its position, velocity, and acceleration errors relative to the selected minimizer at every time, and an interval I I I containing the optimal cost. Both ρ \rho ρ and the width of I I I must be at most ε \varepsilon ε .
Recovering isolated contacts. For m = 1 , 2 , 3 m=1,2,3 m = 1 , 2 , 3 , Algorithm 2.2 solves (4) for the contact times t i t_i t i and jerk jumps Δ i \Delta_i Δ i , retaining admissible solutions as defined in § 2.2.3 . To check clearance, minimize the polynomial ∥ r ( t ) ∥ 2 − 1 \lVert r(t)\rVert^2-1 ∥ r ( t ) ∥ 2 − 1 on each cubic piece using endpoint values and derivative roots, and reject any negative minimum. The Green representation (2.4) reconstructs the trajectory, and Theorem 2.5 certifies its optimality.
Constructing trials with an arc. For each of A \mathrm{A} A , P A \mathrm{PA} PA , A P \mathrm{AP} AP , and P A P \mathrm{PAP} PAP , the parameter vector p p p contains the ordered contact times, including arc entry and exit, the initial acceleration and jerk, and the nonnegative contact masses m i m_i m i . The bounds in Proposition A.13 give finite ranges for each parameter; P r e c \mathcal{P}_{\mathrm{rec}} P rec collects the choices for all four families, including coincident times and zero masses. Starting from the prescribed initial state, we construct a trial q p q_p q p by following cubic pieces, adding m i q p ( t i ) m_iq_p(t_i) m i q p ( t i ) to the jerk at each contact time, and integrating q p ( 4 ) ( t ) = Λ p ( t ) q p ( t ) q_p^{(4)}(t)=\Lambda_p(t)q_p(t) q p ( 4 ) ( t ) = Λ p ( t ) q p ( t ) along the arc, where Λ p ( t ) = − 3 ∥ q ¨ p ( t ) ∥ 2 − 4 ⟨ q ˙ p ( t ) , q p ( 3 ) ( t ) ⟩ \Lambda_p(t)=-3\lVert\ddot{q}_p(t)\rVert^2-4\langle\dot{q}_p(t),q_p^{(3)}(t)\rangle Λ p ( t ) = − 3 ∥ q ¨ p ( t ) ∥ 2 − 4 ⟨ q ˙ p ( t ) , q p ( 3 ) ( t )⟩ . When evaluating Λ p q p \Lambda_pq_p Λ p q p for a trial, we limit each component of position, velocity, acceleration, and jerk to a fixed range containing every valid arc; exceeding this range counts as a violation.
Searching parameter intervals. A parameter box assigns one closed interval with rational endpoints to each coordinate of p p p and represents all allowed parameter choices within those intervals. Algorithm 2.3 covers the search ranges by such boxes. For each trial, R ( p ) \mathcal{R}(p) R ( p ) adds the squared endpoint and contact errors to the largest violation of each remaining requirement: clearance, the near-side condition, staying on the circle along the arc, nonnegative multiplier, and the integration bounds. Each term is nonnegative, so R ( p ) = 0 R(p) = 0 R ( p ) = 0 exactly when all requirements hold. We bound R R R over each box, including numerical integration errors, and discard a box only if its lower bound is positive. Thus a discarded box contains no valid solution.
The parameter η \eta η bounds box diameters and numerical evaluation errors. Among the retained boxes, we choose one trial r ~ \tilde{r} r ~ and bound its difference from every retained trial in position, velocity, and acceleration by ρ \rho ρ ; I I I encloses all their costs. These bounds contain the selected minimizer and its cost, even though r ~ \tilde{r} r ~ itself need not be feasible. Algorithm 2.1 halves η \eta η until both error bounds meet the requested accuracy.
Algorithm 2.1
Input: ω \omega ω , ϵ \epsilon ϵ
Output: r ~ \tilde{r} r ~ , ρ \rho ρ , I I I with ρ \rho ρ , width( I ) ≤ ϵ (I) \le\epsilon ( I ) ≤ ϵ
if ω \omega ω is unblocked then
return r ~ , 0 , [ J [ r ~ ] , J [ r ~ ] ] \tilde{r}, 0, [J[\tilde{r}], J[\tilde{r}]] r ~ , 0 , [ J [ r ~ ] , J [ r ~ ]]
end if
for m = 1 , 2 , 3 m = 1, 2, 3 m = 1 , 2 , 3 do
r ← r \leftarrow r ← Algorithm 2.2( ω , m ) (\omega,m) ( ω , m )
if r ≠ ∅ r \ne\varnothing r = ∅ then
return r , 0 , [ J [ r ] , J [ r ] ] r, 0, [J[r], J[r]] r , 0 , [ J [ r ] , J [ r ]]
end if
end for
for n = 1 , 2 , … n = 1, 2, \ldots n = 1 , 2 , … do
( r ~ , ρ , I ) ← (\tilde{r}, \rho, I) \leftarrow ( r ~ , ρ , I ) ← Algorithm 2.3( ω , 2 − n ) (\omega,2^{-n}) ( ω , 2 − n )
if max { ρ , width ( I ) } ≤ ϵ \max\{\rho,\operatorname{width}(I)\} \le\epsilon max { ρ , width ( I )} ≤ ϵ then
return r ~ , ρ , I \tilde{r}, \rho, I r ~ , ρ , I
end if
end for
Algorithm 2.2
Input: ω \omega ω , m m m
Output: Minimizer r r r , or ∅ \varnothing ∅
Solve (4) for admissible c c c
return ∅ \varnothing ∅ if none exists
Reconstruct r r r by (2.4)
return r r r
Algorithm 2.3
Input: ω \omega ω , rational η > 0 \eta> 0 η > 0
Output: r ~ \tilde{r} r ~ , ρ \rho ρ , I I I
Cover P r e c \mathcal{P}_{\mathrm{rec}} P rec by boxes of diameter ≤ η \le\eta ≤ η
for all parameter boxes C C C do
Bound R R R on C C C to error η \eta η
Discard C C C if its lower bound is positive
end for
Choose a retained trial r ~ \tilde{r} r ~
Bound distances to r ~ \tilde{r} r ~ by ρ \rho ρ
Enclose retained costs in I I I
return r ~ , ρ , I \tilde{r}, \rho, I r ~ , ρ , I
Discussion
We derive a complete analytical characterization of minimum-effort collision-free motion for two robots with double-integrator dynamics and zero initial and final velocities. The optimal relative trajectory combines cubic pieces with isolated contacts or a boundary arc. We use the fact that every optimal relative trajectory lies in a plane and the contact laws to reduce the trajectory optimization problem to eight contact families, whose admissible constructions partition the space of endpoint configurations. We certify the global optimality of these constructions through an exact comparison of trajectory costs.
These results provide structured candidate trajectories for use within fleet-level planners. The companion work [20 ] incorporates acceleration constraints and demonstrates collision-free planning for one million robots in minutes and for thousands in a fraction of a second.
We identify three promising directions for future work. First, extending the contact classification and global optimality certificate to pointwise bounds on position, velocity, and acceleration would account for workspace restrictions and robot operating limits. Second, studying different dynamics would clarify which features of the optimal contact patterns depend on the double-integrator model. Finally, we believe that a complete characterization for nonzero initial and final velocities would be particularly useful for online replanning, where robots must update their trajectories while moving. Altogether, this paper provides an analytical foundation for understanding optimal two-robot motion and using its structure to design efficient planners for larger robot fleets.
Proofs
Proof roadmap. The characterization follows from four steps: (i) Exhaustion. Every optimal relative trajectory lies in a plane. Together with side selection, the contact bounds allow at most three isolated contacts, or one boundary arc with at most one isolated contact on each side, yielding the eight families of Theorem 2.8 . (ii) Identification. The construction rules capture every selected minimizer. Conversely, admissibility and the global certificate (Theorem 2.5 ) identify every admitted construction with the selected minimizer. (iii) Partition. Existence and exhaustion ensure coverage of D \mathcal{D} D . Uniqueness of the selected trajectory and its multiplier prevents different family labels at the same endpoint datum, establishing Theorem 2.9 . (iv) Recovery. Exact algebraic tests handle isolated contacts. For the remaining families, compactness and uniqueness force the certified trajectory and cost enclosures to shrink, guaranteeing termination of Algorithm 2.1 (Section A.2.8).
Additional notation
We write C 2 , 1 ( [ 0 , T ] ; R d ) C^{2,1}([0,T];\mathbb{R}^{d}) C 2 , 1 ([ 0 , T ] ; R d ) for twice continuously differentiable functions with a Lipschitz second derivative, and BV ( ( 0 , T ) ; R d ) \operatorname{BV}((0,T);\mathbb{R}^{d}) BV (( 0 , T ) ; R d ) for functions of bounded variation. We use distributional derivatives and bounded-variation calculus as in [16 Sections 3.5 and 9.1] . A finite nonnegative measure on [ 0 , T ] [0,T] [ 0 , T ] has the decomposition [16 Section 3.5] μ = Λ ( t ) d t + ∑ k m k δ t k + μ s c \mu=\Lambda(t)\,\mathrm{d}t+\sum_k m_k\delta_{t_k}+\mu_{\mathrm{sc}} μ = Λ ( t ) d t + ∑ k m k δ t k + μ sc , where Λ ≥ 0 \Lambda\ge0 Λ ≥ 0 is the density, m k = μ ( { t k } ) m_k=\mu(\{t_k\}) m k = μ ({ t k }) are the atomic masses, and μ s c \mu_{\mathrm{sc}} μ sc is the singular-continuous part.
Proofs for Section 2.2
Proof of Theorem 2.3
Lemma A.1 (Interior stationarity). Every local minimizer admits a locally finite nonnegative Radon measure μ \mu μ on ( 0 , T ) (0,T) ( 0 , T ) satisfying (2) .
Proof of Lemma A.1. Let r r r be a local minimizer. For a nonnegative smooth test ϕ \phi ϕ compactly supported in ( 0 , T ) (0,T) ( 0 , T ) , the radial variation ( 1 + ϵ ϕ ) r (1+\epsilon\phi)r ( 1 + ϵ ϕ ) r is feasible for ϵ ≥ 0 \epsilon\ge0 ϵ ≥ 0 . The distribution L [ ϕ ] : = R − 2 ∫ 0 T ⟨ r ¨ ( t ) , ( ϕ r ) ¨ ( t ) ⟩ d t L[\phi]:=R^{-2}\int_{0}^{T}\langle\ddot{r}(t),\ddot{(\phi r)}(t)\rangle\,\mathrm{d}t L [ ϕ ] := R − 2 ∫ 0 T ⟨ r ¨ ( t ) , ( ϕ r ) ¨ ( t )⟩ d t is positive, hence is a locally finite nonnegative Radon measure μ \mu μ . Indeed, a smooth cutoff χ ≥ 1 \chi\ge1 χ ≥ 1 on a fixed compact support gives ∣ L [ ϕ ] ∣ ≤ L [ χ ] max ∣ ϕ ∣ |L[\phi]|\le L[\chi]\max|\phi| ∣ L [ ϕ ] ∣ ≤ L [ χ ] max ∣ ϕ ∣ , so the Riesz representation theorem applies locally. Both signs of the radial variation are allowed away from contact, giving supp ( μ ) ⊆ A ∩ ( 0 , T ) \operatorname{supp}(\mu)\subseteq A\cap(0,T) supp ( μ ) ⊆ A ∩ ( 0 , T ) . Every smooth compactly supported vector test η \eta η decomposes as η = ϕ r + η ⊥ \eta=\phi r+\eta_{\perp} η = ϕ r + η ⊥ , with ϕ ( t ) = ⟨ η ( t ) , r ( t ) ⟩ / ∥ r ( t ) ∥ 2 \phi(t)=\langle\eta(t),r(t)\rangle/\lVert r(t)\rVert^{2} ϕ ( t ) = ⟨ η ( t ) , r ( t )⟩ / ∥ r ( t ) ∥ 2 and ⟨ η ⊥ ( t ) , r ( t ) ⟩ = 0 \langle\eta_{\perp}(t),r(t)\rangle=0 ⟨ η ⊥ ( t ) , r ( t )⟩ = 0 . These components belong to H 2 ( ( 0 , T ) ; R ) H^2((0,T);\mathbb{R}) H 2 (( 0 , T ) ; R ) and H 2 ( ( 0 , T ) ; R d ) H^2((0,T);\mathbb{R}^d) H 2 (( 0 , T ) ; R d ) and have compact support, since r ∈ H 2 ( ( 0 , T ) ; R d ) r \in H^2((0,T);\mathbb{R}^d) r ∈ H 2 (( 0 , T ) ; R d ) and ∥ r ( t ) ∥ ≥ R > 0 \lVert r(t)\rVert\ge R > 0 ∥ r ( t )∥ ≥ R > 0 . Since ∥ r ( t ) + ε η ⊥ ( t ) ∥ 2 = ∥ r ( t ) ∥ 2 + ε 2 ∥ η ⊥ ( t ) ∥ 2 \lVert r(t) + \varepsilon\eta_{\perp}(t)\rVert^2 = \lVert r(t)\rVert^2 + \varepsilon^2\lVert\eta_{\perp}(t)\rVert^2 ∥ r ( t ) + ε η ⊥ ( t ) ∥ 2 = ∥ r ( t ) ∥ 2 + ε 2 ∥ η ⊥ ( t ) ∥ 2 , both signs are feasible, and local minimality gives ∫ 0 T ⟨ r ¨ ( t ) , η ¨ ⊥ ( t ) ⟩ d t = 0 \int_0^T \langle\ddot{r}(t),\ddot{\eta}_{\perp}(t)\rangle\,\mathrm{d}t = 0 ∫ 0 T ⟨ r ¨ ( t ) , η ¨ ⊥ ( t )⟩ d t = 0 . Smooth approximation on a fixed compact subinterval extends the radial identity to ϕ \phi ϕ : multiplication is continuous in H 2 ( ( 0 , T ) ; R ) H^2((0,T);\mathbb{R}) H 2 (( 0 , T ) ; R ) , and convergence is uniform for integration against the measure. Since ∥ r ( t ) ∥ = R \lVert r(t)\rVert= R ∥ r ( t )∥ = R on its support,
∫ 0 T ⟨ r ¨ ( t ) , η ¨ ( t ) ⟩ d t = R 2 ∫ 0 + T − ϕ ( t ) d μ ( t ) = ∫ 0 + T − ⟨ r ( t ) , η ( t ) ⟩ d μ ( t ) . \int_0^T \langle\ddot{r}(t),\ddot{\eta}(t)\rangle\,\mathrm{d}t = R^2\int_{0+}^{T-} \phi(t)\,\mathrm{d}\mu(t) = \int_{0+}^{T-} \langle r(t),\eta(t)\rangle\,\mathrm{d}\mu(t). ∫ 0 T ⟨ r ¨ ( t ) , η ¨ ( t )⟩ d t = R 2 ∫ 0 + T − ϕ ( t ) d μ ( t ) = ∫ 0 + T − ⟨ r ( t ) , η ( t )⟩ d μ ( t ) .
Proof of Proposition 2.3. Strict endpoint separation makes A \mathcal{A} A a compact subset of ( 0 , T ) (0,T) ( 0 , T ) , so the measure of Lemma A.1 is finite. Extend it by zero to [ 0 , T ] [0,T] [ 0 , T ] . Density of smooth compactly supported tests in H T \mathcal{H}_T H T gives the stationarity identity, and the support condition gives complementarity.
Proof of Theorem 2.4
We first establish the regularity and clearance identity needed to exclude accumulating contacts, using the stationary pair ( r , μ ) (r,\mu) ( r , μ ) of § 2.2.1 .
Proposition A.2 (Trajectory regularity). Let r r r be a feasible trajectory and μ \mu μ a finite nonnegative Radon measure satisfying (2.2). The distributional derivative D t 3 r D_t^3r D t 3 r is represented by a right-continuous function r ( 3 ) ∈ BV ( ( 0 , T ) ; R d ) r^{(3)} \in\operatorname{BV}((0,T);\mathbb{R}^d) r ( 3 ) ∈ BV (( 0 , T ) ; R d ) , which we call the jerk. The weak acceleration D t 2 r D_t^2r D t 2 r has a Lipschitz continuous representative r ¨ \ddot{r} r ¨ on [ 0 , T ] [0,T] [ 0 , T ] , with ∥ r ¨ ( t ) − r ¨ ( s ) ∥ ≤ ∥ r ( 3 ) ∥ L ∞ ( ( 0 , T ) ; R d ) ∣ t − s ∣ \lVert\ddot{r}(t)-\ddot{r}(s)\rVert\le\lVert r^{(3)}\rVert_{L^\infty((0,T);\mathbb{R}^d)}\lvert t-s\rvert ∥ r ¨ ( t ) − r ¨ ( s )∥ ≤ ∥ r ( 3 ) ∥ L ∞ (( 0 , T ) ; R d ) ∣ t − s ∣ for s , t ∈ [ 0 , T ] s,t \in[0,T] s , t ∈ [ 0 , T ] . Consequently, r ∈ C 2 , 1 ( [ 0 , T ] ; R d ) r \in C^{2,1}([0,T];\mathbb{R}^d) r ∈ C 2 , 1 ([ 0 , T ] ; R d ) and g ∈ C 2 , 1 ( [ 0 , T ] ; R ) g \in C^{2,1}([0,T];\mathbb{R}) g ∈ C 2 , 1 ([ 0 , T ] ; R ) .
Proof of Proposition A.2. By the Sobolev embedding H 2 ( ( 0 , T ) ; R d ) ↪ C 1 ( [ 0 , T ] ; R d ) H^2((0,T);\mathbb{R}^d) \hookrightarrow C^1([0,T];\mathbb{R}^d) H 2 (( 0 , T ) ; R d ) ↪ C 1 ([ 0 , T ] ; R d ) [1 Theorem 4.12, Part II] , r r r is bounded and continuous. The primitive F μ ( t ) : = ∫ 0 + t r ( s ) d μ ( s ) F_\mu(t) := \int_{0+}^{t} r(s)\,\mathrm{d}\mu(s) F μ ( t ) := ∫ 0 + t r ( s ) d μ ( s ) is right-continuous with bounded variation and satisfies D t 1 F μ = r μ D_t^1F_\mu= r\mu D t 1 F μ = r μ [16 Sections 3.5 and 9.1] . Thus (2.2) gives r ( 3 ) ( t ) = r 0 ( 3 ) + F μ ( t ) r^{(3)}(t) = r_0^{(3)} + F_\mu(t) r ( 3 ) ( t ) = r 0 ( 3 ) + F μ ( t ) for a constant vector r 0 ( 3 ) ∈ R d r_0^{(3)} \in\mathbb{R}^d r 0 ( 3 ) ∈ R d . In particular, r ( 3 ) ∈ BV ( ( 0 , T ) ; R d ) ∩ L ∞ ( ( 0 , T ) ; R d ) r^{(3)} \in\operatorname{BV}((0,T);\mathbb{R}^d) \cap L^\infty((0,T);\mathbb{R}^d) r ( 3 ) ∈ BV (( 0 , T ) ; R d ) ∩ L ∞ (( 0 , T ) ; R d ) . Integrating once more gives r ¨ ( t ) = r ¨ 0 + ∫ 0 t r ( 3 ) ( s ) d s \ddot{r}(t) = \ddot{r}_0 + \int_0^t r^{(3)}(s)\,\mathrm{d}s r ¨ ( t ) = r ¨ 0 + ∫ 0 t r ( 3 ) ( s ) d s , which proves the stated Lipschitz bound and r ∈ C 2 , 1 ( [ 0 , T ] ; R d ) r \in C^{2,1}([0,T];\mathbb{R}^d) r ∈ C 2 , 1 ([ 0 , T ] ; R d ) . Finally, g ¨ ( t ) = − 2 ∥ r ˙ ( t ) ∥ 2 − 2 ⟨ r ( t ) , r ¨ ( t ) ⟩ \ddot{g}(t) = -2\lVert\dot{r}(t)\rVert^2 - 2\langle r(t),\ddot{r}(t)\rangle g ¨ ( t ) = − 2 ∥ r ˙ ( t ) ∥ 2 − 2 ⟨ r ( t ) , r ¨ ( t )⟩ is Lipschitz, so g ∈ C 2 , 1 ( [ 0 , T ] ; R ) g \in C^{2,1}([0,T];\mathbb{R}) g ∈ C 2 , 1 ([ 0 , T ] ; R ) .
At every interior contact, g g g attains a local maximum of zero, which gives tangency: g ˙ ( t ) = − 2 ⟨ r ( t ) , r ˙ ( t ) ⟩ = 0 \dot{g}(t) = -2\langle r(t),\dot{r}(t)\rangle= 0 g ˙ ( t ) = − 2 ⟨ r ( t ) , r ˙ ( t )⟩ = 0 for t ∈ A ∩ ( 0 , T ) t \in\mathcal{A} \cap(0,T) t ∈ A ∩ ( 0 , T ) .
Let t ∗ ∈ ( 0 , T ) t^* \in(0,T) t ∗ ∈ ( 0 , T ) be an isolated contact or an endpoint of a boundary arc. The acceleration is continuous by Proposition A.2, so [ r ¨ ] t ∗ = 0 [\ddot{r}]_{t^*} = 0 [ r ¨ ] t ∗ = 0 . The atom of D t 1 r ( 3 ) = r μ D_t^1r^{(3)} = r\mu D t 1 r ( 3 ) = r μ gives the jerk jump [ r ( 3 ) ] t ∗ = μ ( { t ∗ } ) r ( t ∗ ) [r^{(3)}]_{t^*} = \mu(\{t^*\})r(t^*) [ r ( 3 ) ] t ∗ = μ ({ t ∗ }) r ( t ∗ ) [16 Section 3.5] .
On every open interval disjoint from supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) , D t 4 r = 0 D_t^4r = 0 D t 4 r = 0 , so the trajectory is cubic. A free segment between two interior contacts is a free bridge. For the clearance g g g of § 2.2.1 , define the clearance forcing f g ( t ) : = 6 ∥ r ¨ ( t ) ∥ 2 + 8 ⟨ r ˙ ( t ) , r ( 3 ) ( t ) ⟩ f_g(t) := 6\lVert\ddot{r}(t)\rVert^2 + 8\langle\dot{r}(t),r^{(3)}(t)\rangle f g ( t ) := 6 ∥ r ¨ ( t ) ∥ 2 + 8 ⟨ r ˙ ( t ) , r ( 3 ) ( t )⟩ .
Lemma A.3 (Clearance identity). The forcing f g f_g f g has a Lipschitz continuous representative on [ 0 , T ] [0,T] [ 0 , T ] , and D t 4 g = − f g ( t ) d t − 2 R 2 d μ D_t^4g = -f_g(t)\,\mathrm{d}t - 2R^2\,\mathrm{d}\mu D t 4 g = − f g ( t ) d t − 2 R 2 d μ (A.1) in distributions on ( 0 , T ) (0,T) ( 0 , T ) . On every boundary arc [ a , b ] [a,b] [ a , b ] , μ ∣ ( a , b ) = − f g ( t ) / ( 2 R 2 ) d t \mu|_{(a,b)} = -f_g(t)/(2R^2)\,\mathrm{d}t μ ∣ ( a , b ) = − f g ( t ) / ( 2 R 2 ) d t , with f g ≤ 0 f_g \le0 f g ≤ 0 on [ a , b ] [a,b] [ a , b ] .
Proof of Lemma A.3. By Proposition A.2 and the Lebesgue–Stieltjes product rule [16 Theorem 3.36] , D t 3 g D_t^3g D t 3 g has the right-continuous representative g ( 3 ) ( t ) = − 6 ⟨ r ˙ ( t ) , r ¨ ( t ) ⟩ − 2 ⟨ r ( t ) , r ( 3 ) ( t ) ⟩ g^{(3)}(t) = -6\langle\dot{r}(t),\ddot{r}(t)\rangle- 2\langle r(t),r^{(3)}(t)\rangle g ( 3 ) ( t ) = − 6 ⟨ r ˙ ( t ) , r ¨ ( t )⟩ − 2 ⟨ r ( t ) , r ( 3 ) ( t )⟩ in B V ( ( 0 , T ) ; R ) \mathrm{BV}((0,T);\mathbb{R}) BV (( 0 , T ) ; R ) . Differentiating once more and using D t 1 r ( 3 ) = r μ D_t^1r^{(3)}=r\,\mu D t 1 r ( 3 ) = r μ gives − D t 1 g ( 3 ) = ( 6 ∥ r ¨ ( t ) ∥ 2 + 8 ⟨ r ˙ ( t ) , r ( 3 ) ( t ) ⟩ ) d t + 2 ∥ r ( t ) ∥ 2 d μ ( t ) -D_t^1g^{(3)}=(6\lVert\ddot r(t)\rVert^2+8\langle\dot r(t),r^{(3)}(t)\rangle)\,\mathrm{d}t+2\lVert r(t)\rVert^2\,\mathrm{d}\mu(t) − D t 1 g ( 3 ) = ( 6 ∥ r ¨ ( t ) ∥ 2 + 8 ⟨ r ˙ ( t ) , r ( 3 ) ( t )⟩) d t + 2 ∥ r ( t ) ∥ 2 d μ ( t ) . Since ∥ r ( t ) ∥ = R \lVert r(t)\rVert=R ∥ r ( t )∥ = R on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) , this is (A.1). The same product rule gives D t 1 f g = 20 ⟨ r ¨ ( t ) , r ( 3 ) ( t ) ⟩ d t + 8 ⟨ r ˙ ( t ) , r ( t ) ⟩ d μ ( t ) D_t^1f_g=20\langle\ddot r(t),r^{(3)}(t)\rangle\,\mathrm{d}t+8\langle\dot r(t),r(t)\rangle\,\mathrm{d}\mu(t) D t 1 f g = 20 ⟨ r ¨ ( t ) , r ( 3 ) ( t )⟩ d t + 8 ⟨ r ˙ ( t ) , r ( t )⟩ d μ ( t ) . Tangency on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) removes the measure term. The remaining density is bounded, so f g f_g f g has a Lipschitz continuous representative. On a boundary arc ( a , b ) (a,b) ( a , b ) , g = 0 g=0 g = 0 , so (A.1) gives μ = − f g ( t ) / ( 2 R 2 ) d t \mu=-f_g(t)/(2R^2)\,\mathrm{d}t μ = − f g ( t ) / ( 2 R 2 ) d t there. Nonnegativity of μ \mu μ and continuity of f g f_g f g yield f g ≤ 0 f_g\leq0 f g ≤ 0 on [ a , b ] [a,b] [ a , b ] .
On a boundary arc [ a , b ] [a,b] [ a , b ] , the continuous density gives r ( 4 ) ( t ) = Λ ( t ) r ( t ) r^{(4)}(t)=\Lambda(t)r(t) r ( 4 ) ( t ) = Λ ( t ) r ( t ) classically for t ∈ ( a , b ) t\in(a,b) t ∈ ( a , b ) . Its arc-side endpoint traces are Λ ( a + ) \Lambda(a^+) Λ ( a + ) and Λ ( b − ) \Lambda(b^-) Λ ( b − ) .
Proof of Proposition 2.4 . Use Proposition A.2 and Lemma A.3.
Exclusion of contact accumulation. Suppose a sequence of contacts ( t n ) n ∈ N (t_n)_{n\in\mathbb{N}} ( t n ) n ∈ N , with t n ↑ t a c c t_n\uparrow t_{\mathrm{acc}} t n ↑ t acc , accumulates without forming a boundary interval to the left of t a c c t_{\mathrm{acc}} t acc . Tangency gives g ( t n ) = g ˙ ( t n ) = 0 g(t_n)=\dot g(t_n)=0 g ( t n ) = g ˙ ( t n ) = 0 . Rolle’s theorem [34 Theorem 2.3.8] gives a sequence ( s n ) n ∈ N (s_n)_{n\in\mathbb{N}} ( s n ) n ∈ N with s n ∈ ( t n , t n + 1 ) s_n\in(t_n,t_{n+1}) s n ∈ ( t n , t n + 1 ) and g ¨ ( s n ) = 0 \ddot g(s_n)=0 g ¨ ( s n ) = 0 . As s n ↑ t a c c s_n\uparrow t_{\mathrm{acc}} s n ↑ t acc , continuity gives g ( t a c c ) = g ˙ ( t a c c ) = g ¨ ( t a c c ) = 0 g(t_{\mathrm{acc}})=\dot g(t_{\mathrm{acc}})=\ddot g(t_{\mathrm{acc}})=0 g ( t acc ) = g ˙ ( t acc ) = g ¨ ( t acc ) = 0 . Finally, g ( 3 ) ( t a c c − ) = 0 g^{(3)}(t_{\mathrm{acc}}^-)=0 g ( 3 ) ( t acc − ) = 0 ; otherwise g ( 3 ) g^{(3)} g ( 3 ) would have a fixed sign near t a c c t_{\mathrm{acc}} t acc , making g ¨ \ddot g g ¨ strictly monotone there and precluding the zeros s n → t a c c s_n\to t_{\mathrm{acc}} s n → t acc .
Let r ∗ ( 3 ) : = r ( 3 ) ( t a c c − ) r_*^{(3)}:=r^{(3)}(t_{\mathrm{acc}}^-) r ∗ ( 3 ) := r ( 3 ) ( t acc − ) and let r c u b r_{\mathrm{cub}} r cub be the free cubic with position r ( t a c c ) r(t_{\mathrm{acc}}) r ( t acc ) , velocity r ˙ ( t a c c ) \dot r(t_{\mathrm{acc}}) r ˙ ( t acc ) , acceleration r ¨ ( t a c c ) \ddot r(t_{\mathrm{acc}}) r ¨ ( t acc ) , and jerk r ∗ ( 3 ) r_*^{(3)} r ∗ ( 3 ) at t a c c t_{\mathrm{acc}} t acc . Write f ˙ ∗ : = 20 ⟨ r ¨ ( t a c c ) , r ∗ ( 3 ) ⟩ \dot f_*:=20\langle\ddot r(t_{\mathrm{acc}}),r_*^{(3)}\rangle f ˙ ∗ := 20 ⟨ r ¨ ( t acc ) , r ∗ ( 3 ) ⟩ for the left limit of the density of D t 1 f g D_t^1f_g D t 1 f g . The clearance g c u b ( s ) : = R 2 − ∥ r c u b ( t a c c + s ) ∥ 2 g_{\mathrm{cub}}(s):=R^2-\lVert r_{\mathrm{cub}}(t_{\mathrm{acc}}+s)\rVert^2 g cub ( s ) := R 2 − ∥ r cub ( t acc + s ) ∥ 2 has vanishing derivatives through order three at zero, so (A.1) gives
g c u b ( s ) = − f g ( t a c c ) 24 s 4 − f ˙ ∗ 120 s 5 − ∥ r ∗ ( 3 ) ∥ 2 36 s 6 . g_{\mathrm{cub}}(s)=-\frac{f_g(t_{\mathrm{acc}})}{24}s^4-\frac{\dot f_*}{120}s^5-\frac{\lVert r_*^{(3)}\rVert^2}{36}s^6. g cub ( s ) = − 24 f g ( t acc ) s 4 − 120 f ˙ ∗ s 5 − 36 ∥ r ∗ ( 3 ) ∥ 2 s 6 .
Thus g c u b g_{\mathrm{cub}} g cub is positive or nonpositive throughout some interval ( − ε , 0 ) (-\varepsilon,0) ( − ε , 0 ) .
Suppose first that g c u b > 0 g_{\mathrm{cub}}>0 g cub > 0 on ( − ε , 0 ) (-\varepsilon,0) ( − ε , 0 ) . The expansion then gives f g ( t a c c ) < 0 f_g(t_{\mathrm{acc}})<0 f g ( t acc ) < 0 , or f g ( t a c c ) = 0 f_g(t_{\mathrm{acc}})=0 f g ( t acc ) = 0 and f ˙ ∗ > 0 \dot f_*>0 f ˙ ∗ > 0 . In the second case the identity for D t 1 f g D_t^1f_g D t 1 f g gives f g ( t ) = − ∫ t t a c c 20 ⟨ r ¨ ( u ) , r ( 3 ) ( u ) ⟩ d u < 0 f_g(t)=-\int_t^{t_{\mathrm{acc}}}20\langle\ddot r(u),r^{(3)}(u)\rangle\,\mathrm{d}u<0 f g ( t ) = − ∫ t t acc 20 ⟨ r ¨ ( u ) , r ( 3 ) ( u )⟩ d u < 0 for t < t a c c t<t_{\mathrm{acc}} t < t acc close to t a c c t_{\mathrm{acc}} t acc . Either way, f g < 0 f_g<0 f g < 0 on a left neighborhood of t a c c t_{\mathrm{acc}} t acc . Fix a contact t n 0 t_{n_0} t n 0 there. Since g = 0 g=0 g = 0 on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) and g = g ˙ = 0 g=\dot g=0 g = g ˙ = 0 at the contact endpoints, two integrations by parts in (A.1) give, for every later contact t n t_n t n ,
∫ t n 0 t n ( − g ( t ) ) f g ( t ) d t = ∫ t n 0 t n ( g ¨ ( t ) ) 2 d t . \int_{t_{n_0}}^{t_n}(-g(t))f_g(t)\,\mathrm{d}t=\int_{t_{n_0}}^{t_n}(\ddot g(t))^2\,\mathrm{d}t. ∫ t n 0 t n ( − g ( t )) f g ( t ) d t = ∫ t n 0 t n ( g ¨ ( t ) ) 2 d t .
The left side is nonpositive and the right side nonnegative, so both vanish. Hence g = 0 g=0 g = 0 on [ t n 0 , t n ] [t_{n_0},t_n] [ t n 0 , t n ] ; letting t n ↑ t a c c t_n\uparrow t_{\mathrm{acc}} t n ↑ t acc gives the excluded boundary interval.
Otherwise g c u b ≤ 0 g_{\mathrm{cub}}\leq0 g cub ≤ 0 on ( − ε , 0 ) (-\varepsilon,0) ( − ε , 0 ) : the limiting free cubic is feasible near t a c c t_{\mathrm{acc}} t acc . Integrating (2) backwards from t a c c t_{\mathrm{acc}} t acc gives r ( t ) = r c u b ( t ) + ∫ t + t a c c ( s − t ) 3 6 r ( s ) d μ ( s ) r(t)=r_{\mathrm{cub}}(t)+\int_{t+}^{t_{\mathrm{acc}}}\frac{(s-t)^3}{6}r(s)\,\mathrm{d}\mu(s) r ( t ) = r cub ( t ) + ∫ t + t acc 6 ( s − t ) 3 r ( s ) d μ ( s ) , t < t a c c t<t_{\mathrm{acc}} t < t acc . For t t t sufficiently close to t a c c t_{\mathrm{acc}} t acc , the kernel is positive and ⟨ r c u b ( t ) , r ( s ) ⟩ > 0 \langle r_{\mathrm{cub}}(t),r(s)\rangle>0 ⟨ r cub ( t ) , r ( s )⟩ > 0 . Hence any mass of μ \mu μ in ( t , t a c c ) (t,t_{\mathrm{acc}}) ( t , t acc ) pushes r ( t ) r(t) r ( t ) strictly farther from the origin than r c u b ( t ) r_{\mathrm{cub}}(t) r cub ( t ) : if μ ( ( t , t a c c ) ) > 0 \mu((t,t_{\mathrm{acc}}))>0 μ (( t , t acc )) > 0 , then ∥ r ( t ) ∥ > ∥ r c u b ( t ) ∥ ≥ R \lVert r(t)\rVert>\lVert r_{\mathrm{cub}}(t)\rVert\geq R ∥ r ( t )∥ > ∥ r cub ( t )∥ ≥ R . A contact sufficiently close to t a c c t_{\mathrm{acc}} t acc therefore cannot have mass of μ \mu μ between itself and t a c c t_{\mathrm{acc}} t acc . Along the accumulating contacts this yields t 0 < t a c c t_0<t_{\mathrm{acc}} t 0 < t acc such that μ ( ( t 0 , t a c c ) ) = 0 \mu((t_0,t_{\mathrm{acc}}))=0 μ (( t 0 , t acc )) = 0 . Hence r = r c u b r=r_{\mathrm{cub}} r = r cub and g ( t ) = g c u b ( t − t a c c ) g(t)=g_{\mathrm{cub}}(t-t_{\mathrm{acc}}) g ( t ) = g cub ( t − t acc ) on ( t 0 , t a c c ) (t_0,t_{\mathrm{acc}}) ( t 0 , t acc ) . The accumulating contacts are zeros of this polynomial, so g = 0 g=0 g = 0 on ( t 0 , t a c c ) (t_0,t_{\mathrm{acc}}) ( t 0 , t acc ) , again a boundary interval contrary to the assumption. The same argument applies from the right and at the endpoints of the time horizon. Thus contacts can accumulate only into boundary arcs.
Finitely many contact components. If A \mathcal{A} A had infinitely many components, compactness would give a monotone sequence of points from distinct components converging to t a c c t_{\mathrm{acc}} t acc . The preceding argument places its tail in one boundary interval, a contradiction. Thus A \mathcal{A} A has finitely many components, each an isolated point or a closed interval by closedness.
Multiplier decomposition. By Lemma A.3, μ \mu μ has nonnegative Lipschitz density − f g / ( 2 R 2 ) -f_g/(2R^2) − f g / ( 2 R 2 ) on each arc interior. Complementarity confines the remaining support to finitely many isolated contacts and arc endpoints, so the remaining part is atomic and μ s c = 0 \mu_{\mathrm{sc}}=0 μ sc = 0 .
Remark on related arguments. For a scalar minimum-acceleration problem, Dontchev and Kolmanovsky [11 proof of Theorem 2.1] show that two contacts with the same linear bound force contact throughout the intervening interval. A linear bound has f g = 0 f_g=0 f g = 0 , whereas disk clearance allows either sign; their comparison therefore leaves contact accumulation open here. Bonnans and Hermant [8 Proposition 3.6 and Equation (7.17)] also decompose multipliers into arc densities and atoms at contact times. They assume finitely many entries, exits, and isolated contacts in (A4)(i); here we prove finiteness.
Proof of Theorem 2.5
Recall the free cubic r ˉ \bar r r ˉ and correction space H T \mathcal{H}_T H T of § 2.2.2 .
Lemma A.4 (Green kernel identities). The Green kernel (2.3) is strictly positive on ( 0 , T ) 2 (0,T)^2 ( 0 , T ) 2 . For s ∈ ( 0 , T ) s\in(0,T) s ∈ ( 0 , T ) , it satisfies D t 4 G T ( ⋅ , s ) = δ s D_t^4G_T(\cdot,s)=\delta_s D t 4 G T ( ⋅ , s ) = δ s and has zero value and first derivative at both endpoints. Its evaluation identity, acting componentwise, is h ( s ) = ∫ 0 T ∂ 2 G T ∂ t 2 ( t , s ) h ¨ ( t ) d t h(s)=\int_0^T\frac{\partial^2G_T}{\partial t^2}(t,s)\,\ddot h(t)\,\mathrm{d}t h ( s ) = ∫ 0 T ∂ t 2 ∂ 2 G T ( t , s ) h ¨ ( t ) d t for h ∈ H T h\in\mathcal{H}_T h ∈ H T .
Proof of Lemma A.4. Positivity follows from 3 T s − T t − 2 t s ≥ 2 s ( T − s ) > 0 3Ts-Tt-2ts\ge2s(T-s)>0 3 T s − T t − 2 t s ≥ 2 s ( T − s ) > 0 for 0 < t ≤ s < T 0<t\le s<T 0 < t ≤ s < T . Direct differentiation of (2.3) on either side of t = s t=s t = s gives cubic pieces, matching values through the second derivative, and a unit jump in the third derivative. Integrating by parts separately on ( 0 , s ) (0,s) ( 0 , s ) and ( s , T ) (s,T) ( s , T ) eliminates the clamped endpoint terms, and the unit jump gives the evaluation identity.
Proof of Theorem 2.5. Twice integrating by parts, using the clamped endpoint conditions and r ˉ ( 4 ) = 0 \bar r^{(4)}=0 r ˉ ( 4 ) = 0 , gives ∫ 0 T ⟨ r ˉ ¨ ( t ) , h ¨ ( t ) ⟩ d t = 0 \int_0^T\langle\ddot{\bar r}(t),\ddot h(t)\rangle\,\mathrm{d}t=0 ∫ 0 T ⟨ r ˉ ¨ ( t ) , h ¨ ( t )⟩ d t = 0 for h ∈ H T h\in\mathcal{H}_T h ∈ H T , hence the cost decomposition in § 2.2.2 . The evaluation identity in Lemma A.4 and finiteness of μ \mu μ give the Green reconstruction (2.4). The integral belongs to H T \mathcal{H}_T H T . Indeed, each scalar kernel section has norm G T ( s , s ) \sqrt{G_T(s,s)} G T ( s , s ) , uniformly bounded for s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] , and r r r is bounded. Four distributional derivatives give r μ r\,\mu r μ ; subtracting it from r − r ˉ r-\bar r r − r ˉ leaves a clamped cubic, which is zero. Equivalently, by the reproducing identity and orthogonality of the free cubic, ∫ 0 T ⟨ r ¨ ( t ) , h ¨ ( t ) ⟩ d t = ∫ 0 + T − ⟨ r ( t ) , h ( t ) ⟩ d μ ( t ) \int_0^T\langle\ddot r(t),\ddot h(t)\rangle\,\mathrm{d}t=\int_{0+}^{T-}\langle r(t),h(t)\rangle\,\mathrm{d}\mu(t) ∫ 0 T ⟨ r ¨ ( t ) , h ¨ ( t )⟩ d t = ∫ 0 + T − ⟨ r ( t ) , h ( t )⟩ d μ ( t ) , h ∈ H T h\in\mathcal{H}_T h ∈ H T . Expanding 1 2 ∥ r ¨ + h ¨ ∥ 2 \frac{1}{2}\|\ddot r+\ddot h\|^2 2 1 ∥ r ¨ + h ¨ ∥ 2 , using 2 ⟨ r ( t ) , h ( t ) ⟩ = ∥ r ( t ) + h ( t ) ∥ 2 − ∥ r ( t ) ∥ 2 − ∥ h ( t ) ∥ 2 2\langle r(t),h(t)\rangle=\|r(t)+h(t)\|^2-\|r(t)\|^2-\|h(t)\|^2 2 ⟨ r ( t ) , h ( t )⟩ = ∥ r ( t ) + h ( t ) ∥ 2 − ∥ r ( t ) ∥ 2 − ∥ h ( t ) ∥ 2 , and using ∥ r ( t ) ∥ = R \|r(t)\|=R ∥ r ( t ) ∥ = R on the support proves the cost identity (3) . Its last integral is nonnegative for a feasible q q q .
For the projection test, the bounded trace map B : H T → L 2 ( ( 0 , T ) , μ ; R d ) B:\mathcal{H}_T\to L^2((0,T),\mu;\mathbb{R}^d) B : H T → L 2 (( 0 , T ) , μ ; R d ) , B h = h ∣ supp ( μ ) Bh=h|_{\operatorname{supp}(\mu)} B h = h ∣ supp ( μ ) , has adjoint B ∗ f = ∫ 0 + T − G T ( ⋅ , s ) f ( s ) d μ ( s ) B^*f=\int_{0+}^{T-}G_T(\cdot,s)f(s)\,\mathrm{d}\mu(s) B ∗ f = ∫ 0 + T − G T ( ⋅ , s ) f ( s ) d μ ( s ) . Thus S = B B ∗ S=BB^* S = B B ∗ acts componentwise by the scalar kernel G T G_T G T . On supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) , projection of (2.4) gives S z = z − z ˉ Sz=z-\bar z S z = z − z ˉ and 0 < z ≤ R 0<z\le R 0 < z ≤ R . If μ ≠ 0 \mu\ne0 μ = 0 , strict positivity of G T G_T G T gives S z > 0 Sz>0 S z > 0 there, so δ < R \delta<R δ < R and S z ≤ ( 1 − δ / R ) z Sz\le(1-\delta/R)z S z ≤ ( 1 − δ / R ) z . For a scalar f ∈ L 2 ( ( 0 , T ) , μ ; R ) f\in L^2((0,T),\mu;\mathbb{R}) f ∈ L 2 (( 0 , T ) , μ ; R ) , the inequality 2 ∣ f ( t ) f ( s ) ∣ ≤ ∣ f ( t ) ∣ 2 z ( s ) z ( t ) + ∣ f ( s ) ∣ 2 z ( t ) z ( s ) 2|f(t)f(s)|\le|f(t)|^2\frac{z(s)}{z(t)}+|f(s)|^2\frac{z(t)}{z(s)} 2∣ f ( t ) f ( s ) ∣ ≤ ∣ f ( t ) ∣ 2 z ( t ) z ( s ) + ∣ f ( s ) ∣ 2 z ( s ) z ( t ) and symmetry of the positive kernel imply ⟨ f , S f ⟩ L 2 ( ( 0 , T ) , μ ; R ) ≤ ∫ 0 + T − ∣ f ( t ) ∣ 2 ( S z ) ( t ) z ( t ) d μ ( t ) ≤ ( 1 − κ c e r t ) ∥ f ∥ L 2 ( ( 0 , T ) , μ ; R ) 2 \left\langle f,Sf\right\rangle_{L^{2}((0,T),\mu;\mathbb{R})}\leq\int_{0+}^{T-} |f(t)|^{2}\frac{(Sz)(t)}{z(t)}\,d\mu(t)\leq(1-\kappa_{\mathrm{cert}})\|f\|^{2}_{L^{2}((0,T),\mu;\mathbb{R})} ⟨ f , S f ⟩ L 2 (( 0 , T ) , μ ; R ) ≤ ∫ 0 + T − ∣ f ( t ) ∣ 2 z ( t ) ( S z ) ( t ) d μ ( t ) ≤ ( 1 − κ cert ) ∥ f ∥ L 2 (( 0 , T ) , μ ; R ) 2 . Tonelli’s theorem justifies this weighted estimate even if z z z has no positive uniform lower bound on the support. Applying it componentwise gives ∥ S ∥ ≤ 1 − κ c e r t \|S\|\leq1-\kappa_{\mathrm{cert}} ∥ S ∥ ≤ 1 − κ cert ; the Hilbert-space identity ∥ B ∥ 2 = ∥ B B ∗ ∥ \|B\|^{2}=\|BB^{*}\| ∥ B ∥ 2 = ∥ B B ∗ ∥ therefore yields ∫ 0 + T − ∥ h ( t ) ∥ 2 d μ ( t ) ≤ ( 1 − κ c e r t ) ∥ h ∥ H T 2 \int_{0+}^{T-}\|h(t)\|^{2}\,d\mu(t)\leq(1-\kappa_{\mathrm{cert}})\|h\|^{2}_{\mathcal{H}_{T}} ∫ 0 + T − ∥ h ( t ) ∥ 2 d μ ( t ) ≤ ( 1 − κ cert ) ∥ h ∥ H T 2 . This proves the bounds in Theorem 2.5 . When μ = 0 \mu=0 μ = 0 , (2.4) gives r = r ˉ r=\bar r r = r ˉ and Q μ [ h ] = ∥ h ∥ H T 2 Q_{\mu}[h]=\|h\|^{2}_{\mathcal{H}_{T}} Q μ [ h ] = ∥ h ∥ H T 2 directly. Finally, zero acceleration norm for a clamped difference forces that difference to vanish, proving uniqueness whenever κ c e r t > 0 \kappa_{\mathrm{cert}}>0 κ cert > 0 and also in the free case.
From here on, use the normalized endpoint data of Section § 2.2.3 . Before proving that every optimal relative trajectory lies in a plane, we show that the interior multiplier remains finite at a terminal contact with zero velocity. The argument uses the following conserved quantities.
Proposition A.5 (Noether invariants). Let r r r be feasible and let μ \mu μ be a finite nonnegative Radon measure satisfying (2) . With r ( 3 ) r^{(3)} r ( 3 ) denoting the right-continuous jerk, the quantities Γ : = r ˙ ( t ) r ¨ ( t ) T − r ¨ ( t ) r ˙ ( t ) T − r ( t ) r ( 3 ) ( t ) T + r ( 3 ) ( t ) r ( t ) T \Gamma:=\dot r(t)\ddot r(t)^{\mathsf T}-\ddot r(t)\dot r(t)^{\mathsf T}-r(t)r^{(3)}(t)^{\mathsf T}+r^{(3)}(t)r(t)^{\mathsf T} Γ := r ˙ ( t ) r ¨ ( t ) T − r ¨ ( t ) r ˙ ( t ) T − r ( t ) r ( 3 ) ( t ) T + r ( 3 ) ( t ) r ( t ) T and E : = 1 2 ∥ r ¨ ( t ) ∥ 2 − ⟨ r ˙ ( t ) , r ( 3 ) ( t ) ⟩ E:=\frac{1}{2}\|\ddot r(t)\|^{2}-\langle\dot r(t),r^{(3)}(t)\rangle E := 2 1 ∥ r ¨ ( t ) ∥ 2 − ⟨ r ˙ ( t ) , r ( 3 ) ( t )⟩ are constant on ( 0 , T ) (0,T) ( 0 , T ) . In particular, their values agree across isolated contacts and boundary-arc endpoints.
Proof of Proposition A.5. By Proposition A.2, the bounded-variation product rule [16, Theorem 3.36]} applies entrywise and gives D t 1 Γ = ( D t 4 r ) r T − r ( D t 4 r ) T = ( r r T − r r T ) μ = 0 D^{1}_{t}\Gamma=(D^{4}_{t}r)r^{\mathsf T}-r(D^{4}_{t}r)^{\mathsf T}=(rr^{\mathsf T}-rr^{\mathsf T})\mu=0 D t 1 Γ = ( D t 4 r ) r T − r ( D t 4 r ) T = ( r r T − r r T ) μ = 0 and D t 1 E = − ⟨ r ˙ , D t 4 r ⟩ = − ⟨ r ˙ ( t ) , r ( t ) ⟩ d μ ( t ) = 0 D^{1}_{t}E=-\langle\dot r,D^{4}_{t}r\rangle=-\langle\dot r(t),r(t)\rangle\,d\mu(t)=0 D t 1 E = − ⟨ r ˙ , D t 4 r ⟩ = − ⟨ r ˙ ( t ) , r ( t )⟩ d μ ( t ) = 0 . The last equality follows from tangency on A ∩ ( 0 , T ) A\cap(0,T) A ∩ ( 0 , T ) and the support condition for μ \mu μ . Both right-continuous representatives are therefore constant on ( 0 , T ) (0,T) ( 0 , T ) .
Proposition A.6 (Interior multiplier). Every global minimizer has initial and terminal free pieces of positive duration and a unique finite nonnegative interior multiplier measure satisfying (2) , compactly supported in ( 0 , 1 ) (0,1) ( 0 , 1 ) , including at a terminal contact.
Proof of Proposition A.6. Let r r r be a global minimizer. Lemma A.1 gives its locally finite interior multiplier μ \mu μ .
Terminal contact with zero velocity. On compact interior intervals, the proofs of Propositions A.2 and A.5 apply to the locally finite measure. Thus r ¨ \ddot r r ¨ is locally Lipschitz, the right-continuous jerk r ( 3 ) r^{(3)} r ( 3 ) has locally bounded variation, and E E E and Γ \Gamma Γ are constants. At any interior contact write r ( t ) = n r(t)=n r ( t ) = n and r ¨ ( t ) = A n + B \ddot r(t)=An+B r ¨ ( t ) = A n + B with ⟨ B , n ⟩ = 0 \langle B,n\rangle=0 ⟨ B , n ⟩ = 0 ; tangency gives ⟨ r ˙ ( t ) , n ⟩ = 0 \langle\dot r(t),n\rangle=0 ⟨ r ˙ ( t ) , n ⟩ = 0 . The component of r ( 3 ) ( t ) r^{(3)}(t) r ( 3 ) ( t ) orthogonal to n n n is Γ n − A r ˙ ( t ) \Gamma n-A\dot r(t) Γ n − A r ˙ ( t ) , and eliminating it from E E E gives ( A + ∥ r ˙ ( t ) ∥ 2 ) 2 + ∥ B ∥ 2 = ∥ r ˙ ( t ) ∥ 4 + 2 E + 2 ⟨ r ˙ ( t ) , Γ n ⟩ (A+\|\dot r(t)\|^{2})^{2}+\|B\|^{2}=\|\dot r(t)\|^{4}+2E+2\langle\dot r(t),\Gamma n\rangle ( A + ∥ r ˙ ( t ) ∥ 2 ) 2 + ∥ B ∥ 2 = ∥ r ˙ ( t ) ∥ 4 + 2 E + 2 ⟨ r ˙ ( t ) , Γ n ⟩ . The velocity is bounded on [ 0 , 1 ] [0,1] [ 0 , 1 ] , so this identity bounds r ¨ \ddot r r ¨ at all interior contacts. On a free component r ¨ \ddot r r ¨ is affine; its contact-end values give the same bound throughout. A component meeting an endpoint is a single polynomial there. Initial strict separation supplies an initial free interval. These observations bound r ¨ \ddot r r ¨ on the entire open interval without assuming finitely many contact components.
If ∥ r T ∥ > 1 \|r_{T}\|>1 ∥ r T ∥ > 1 , strict separation already gives a final free interval. Suppose instead Y = r T Y=r_{T} Y = r T has unit norm. Choose s < 1 s<1 s < 1 so that y ( t ) : = ⟨ Y , r ( t ) ⟩ ≥ 1 / 2 y(t):=\langle Y,r(t)\rangle\geq1/2 y ( t ) := ⟨ Y , r ( t )⟩ ≥ 1/2 for s < t < 1 s<t<1 s < t < 1 . Use the right-continuous representative y ( 3 ) ( t ) = ⟨ Y , r ( 3 ) ( t ) ⟩ y^{(3)}(t)=\langle Y,r^{(3)}(t)\rangle y ( 3 ) ( t ) = ⟨ Y , r ( 3 ) ( t )⟩ . Then D t 1 y ( 3 ) = y μ ≥ 1 2 μ D^{1}_{t}y^{(3)}=y\,\mu\ge\frac{1}{2}\mu D t 1 y ( 3 ) = y μ ≥ 2 1 μ . The bounded function y ¨ \ddot y y ¨ is convex and y ( 3 ) y^{(3)} y ( 3 ) is nondecreasing. Hence y ( 3 ) ∈ L 1 ( ( s , 1 ) ; R ) y^{(3)}\in L^{1}((s,1);\mathbb{R}) y ( 3 ) ∈ L 1 (( s , 1 ) ; R ) : its negative part is bounded and its integral is a difference of bounded values of y ¨ \ddot y y ¨ . For s < t < 1 s<t<1 s < t < 1 , increasing s s s if necessary to fix its representative,
∥ r ( 3 ) ( t ) − r ( 3 ) ( s ) ∥ ≤ ∫ s + t + ∥ r ( u ) ∥ d μ ( u ) = μ ( ( s , t ] ) ≤ 2 ( y ( 3 ) ( t ) − y ( 3 ) ( s ) ) . \|r^{(3)}(t)-r^{(3)}(s)\|\le\int_{s+}^{t+}\|r(u)\|\,\mathrm{d}\mu(u)=\mu((s,t])\le2(y^{(3)}(t)-y^{(3)}(s)). ∥ r ( 3 ) ( t ) − r ( 3 ) ( s ) ∥ ≤ ∫ s + t + ∥ r ( u ) ∥ d μ ( u ) = μ (( s , t ]) ≤ 2 ( y ( 3 ) ( t ) − y ( 3 ) ( s )) .
Thus r ( 3 ) ∈ L 1 ( ( s , 1 ) ; R d ) r^{(3)}\in L^{1}((s,1);\mathbb{R}^{d}) r ( 3 ) ∈ L 1 (( s , 1 ) ; R d ) and r ¨ \ddot r r ¨ has a continuous, absolutely continuous extension to 1 1 1 . Monotonicity and integrability also give ( 1 − t ) ∣ y ( 3 ) ( t ) ∣ → 0 (1-t)|y^{(3)}(t)|\to0 ( 1 − t ) ∣ y ( 3 ) ( t ) ∣ → 0 , and the preceding bound gives ( 1 − t ) ∥ r ( 3 ) ( t ) ∥ → 0 (1-t)\|r^{(3)}(t)\|\to0 ( 1 − t ) ∥ r ( 3 ) ( t ) ∥ → 0 . Since r ˙ ( 1 ) = 0 \dot r(1)=0 r ˙ ( 1 ) = 0 and r ¨ \ddot r r ¨ is bounded, ∥ r ˙ ( t ) ∥ = O ( 1 − t ) \|\dot r(t)\|=O(1-t) ∥ r ˙ ( t ) ∥ = O ( 1 − t ) , whence ⟨ r ˙ ( t ) , r ( 3 ) ( t ) ⟩ → 0 \langle\dot r(t),r^{(3)}(t)\rangle\to0 ⟨ r ˙ ( t ) , r ( 3 ) ( t )⟩ → 0 . Integrating the energy invariant and using the zero initial and final velocities now yields E = 3 2 ∫ 0 1 ∥ r ¨ ( t ) ∥ 2 d t > 0 E=\frac{3}{2}\int_{0}^{1}\|\ddot r(t)\|^{2}\,\mathrm{d}t>0 E = 2 3 ∫ 0 1 ∥ r ¨ ( t ) ∥ 2 d t > 0 , ∥ r ¨ ( 1 ) ∥ 2 = 2 E \|\ddot r(1)\|^{2}=2E ∥ r ¨ ( 1 ) ∥ 2 = 2 E . The strict inequality holds here because r 0 ≠ Y r_{0}\ne Y r 0 = Y .
Use the positive gap b ( t ) = ∥ r ( t ) ∥ 2 − 1 = − g ( t ) b(t)=\|r(t)\|^{2}-1=-g(t) b ( t ) = ∥ r ( t ) ∥ 2 − 1 = − g ( t ) . Locally, the same product rule as in (A.1) gives D t 4 b = f g d t + 2 μ D^{4}_{t}b=f_{g}\,\mathrm{d}t+2\mu D t 4 b = f g d t + 2 μ , f g ( t ) = 10 ∥ r ¨ ( t ) ∥ 2 − 8 E → 12 E > 0 f_{g}(t)=10\|\ddot r(t)\|^{2}-8E\to12E>0 f g ( t ) = 10∥ r ¨ ( t ) ∥ 2 − 8 E → 12 E > 0 as t ↑ 1 t\uparrow1 t ↑ 1 . If interior contacts accumulated at 1 1 1 , their tangent zeros of b b b would, by Rolle’s theorem, give infinitely many zeros of the continuous function b ¨ \ddot b b ¨ . But the preceding measure identity makes this function strictly convex near 1 1 1 , where it can have at most two zeros. A final boundary interval is excluded by the same identity. There is therefore a final free interval of positive duration. The multiplier is supported in a compact subset of ( 0 , 1 ) (0,1) ( 0 , 1 ) and is finite. Its uniqueness follows from D t 4 r = r μ D^{4}_{t}r=r\,\mu D t 4 r = r μ and ∥ r ( t ) ∥ = 1 \|r(t)\|=1 ∥ r ( t ) ∥ = 1 on its support. Possible endpoint atoms do not enter the interior equation or the clamped Green reconstruction, and are omitted.
Proof of Propositions 2.6 and 2.7
For off-axis data, recall the normal e : = ( r T , 2 , d 0 − r T , 1 ) / ∥ r T − r 0 ∥ e:=(r_{T,2},d_{0}-r_{T,1})/\|r_{T}-r_{0}\| e := ( r T , 2 , d 0 − r T , 1 ) /∥ r T − r 0 ∥ and offset δ : = ⟨ e , r 0 ⟩ = ⟨ e , r T ⟩ = d 0 r T , 2 / ∥ r T − r 0 ∥ \delta:=\langle e,r_{0}\rangle=\langle e,r_{T}\rangle=d_{0}r_{T,2}/\|r_{T}-r_{0}\| δ := ⟨ e , r 0 ⟩ = ⟨ e , r T ⟩ = d 0 r T , 2 /∥ r T − r 0 ∥ (A.2).
Proof of Propositions 2.6 and 2.7. If the free cubic is feasible, Theorem 2.5 at μ = 0 \mu=0 μ = 0 makes it the unique global minimizer. Its image lies in the span of the endpoints and, for off-axis data, in the near side. Hence assume blocked data, so that r 0 ≠ r T r_{0}\ne r_{T} r 0 = r T , and put a : = ( r 0 − r T ) / ∥ r 0 − r T ∥ a:=(r_{0}-r_{T})/\|r_{0}-r_{T}\| a := ( r 0 − r T ) /∥ r 0 − r T ∥ . For off-axis data use the normal e e e and offset δ \delta δ of (A.2); for symmetric data let e ⊥ a e\perp a e ⊥ a be any unit vector and δ = 0 \delta=0 δ = 0 . In both cases r 0 − ⟨ a , r 0 ⟩ a = r T − ⟨ a , r T ⟩ a = δ e r_{0}-\langle a,r_{0}\rangle a=r_{T}-\langle a,r_{T}\rangle a=\delta e r 0 − ⟨ a , r 0 ⟩ a = r T − ⟨ a , r T ⟩ a = δ e .
Transverse comparison. Let r ∈ M ∗ r\in\mathcal{M}^{*} r ∈ M ∗ and write r = X a + Y r=Xa+Y r = X a + Y , with X = ⟨ a , r ⟩ X=\langle a,r\rangle X = ⟨ a , r ⟩ and ⟨ a , Y ( t ) ⟩ = 0 \langle a,Y(t)\rangle=0 ⟨ a , Y ( t )⟩ = 0 . The endpoint states give Y − δ e ∈ H 1 Y-\delta e\in\mathcal{H}_{1} Y − δ e ∈ H 1 and ∥ r ¨ ∥ 2 = ∥ X ¨ ∥ 2 + ∥ Y − δ e ∥ H 1 2 \|\ddot r\|^{2}=\|\ddot X\|^{2}+\|Y-\delta e\|_{\mathcal{H}_{1}}^{2} ∥ r ¨ ∥ 2 = ∥ X ¨ ∥ 2 + ∥ Y − δ e ∥ H 1 2 . Let V = { w ∈ H 2 ( ( 0 , 1 ) ; R ) : w ( 0 ) = w ( 1 ) = w ˙ ( 0 ) = w ˙ ( 1 ) = 0 } \mathcal{V}=\{w\in H^{2}((0,1);\mathbb{R}):w(0)=w(1)=\dot w(0)=\dot w(1)=0\} V = { w ∈ H 2 (( 0 , 1 ) ; R ) : w ( 0 ) = w ( 1 ) = w ˙ ( 0 ) = w ˙ ( 1 ) = 0 } , ∥ w ∥ V 2 = ∫ 0 1 ∣ w ¨ ( t ) ∣ 2 d t \|w\|_{\mathcal{V}}^{2}=\int_{0}^{1}|\ddot w(t)|^{2}\,\mathrm{d}t ∥ w ∥ V 2 = ∫ 0 1 ∣ w ¨ ( t ) ∣ 2 d t . Put I = { t ∈ ( 0 , 1 ) : ∣ X ( t ) ∣ < 1 } \mathcal{I}=\{t\in(0,1):|X(t)|<1\} I = { t ∈ ( 0 , 1 ) : ∣ X ( t ) ∣ < 1 } , and write ξ ( t ) = 1 − X ( t ) 2 \xi(t)=\sqrt{1-X(t)^{2}} ξ ( t ) = 1 − X ( t ) 2 on its closure. Feasibility requires ∥ Y ∥ ≥ ξ \|Y\|\ge\xi ∥ Y ∥ ≥ ξ on I \mathcal{I} I . For n ≥ 3 n\ge3 n ≥ 3 , let I n = I ∩ [ 1 / n , 1 − 1 / n ] \mathcal{I}_{n}=\mathcal{I}\cap[1/n,1-1/n] I n = I ∩ [ 1/ n , 1 − 1/ n ] and minimize ∥ w ∥ V 2 \|w\|_{\mathcal{V}}^{2} ∥ w ∥ V 2 subject to w ≥ ξ − δ w\ge\xi-\delta w ≥ ξ − δ on I n \mathcal{I}_{n} I n . These compact constraints admit a strictly feasible smooth cutoff. Their unique minimizer u n u_{n} u n therefore has a nonnegative finite multiplier ν n \nu_{n} ν n supported on I n \mathcal{I}_{n} I n , with u n = ∫ 0 + 1 − G 1 ( ⋅ , s ) d ν n ( s ) u_{n}=\int_{0+}^{1-}G_{1}(\cdot,s)\,\mathrm{d}\nu_{n}(s) u n = ∫ 0 + 1 − G 1 ( ⋅ , s ) d ν n ( s ) , ∥ u n ∥ V 2 = ∫ 0 + 1 − ( ξ ( s ) − δ ) d ν n ( s ) \|u_{n}\|_{\mathcal{V}}^{2}=\int_{0+}^{1-}(\xi(s)-\delta)\,\mathrm{d}\nu_{n}(s) ∥ u n ∥ V 2 = ∫ 0 + 1 − ( ξ ( s ) − δ ) d ν n ( s ) . The strict Slater condition gives the multiplier [10 Theorem 1.1] . The second identity follows from complementarity and the reproducing property. If I n \mathcal{I}_{n} I n is empty, take u n = 0 u_{n}=0 u n = 0 and ν n = 0 \nu_{n}=0 ν n = 0 . Positivity of G 1 G_{1} G 1 gives u n ≥ 0 u_{n}\ge0 u n ≥ 0 .
Let Y ^ ( s ) : = Y ( s ) / ∥ Y ( s ) ∥ \widehat{Y}(s):=Y(s)/\|Y(s)\| Y ( s ) := Y ( s ) /∥ Y ( s ) ∥ where Y ( s ) ≠ 0 Y(s)\ne0 Y ( s ) = 0 , and Y ^ ( s ) : = e \widehat{Y}(s):=e Y ( s ) := e otherwise, and set W n = ∫ 0 + 1 − G 1 ( ⋅ , s ) Y ^ ( s ) d ν n ( s ) W_{n}=\int_{0+}^{1-}G_{1}(\cdot,s)\widehat{Y}(s)\,\mathrm{d}\nu_{n}(s) W n = ∫ 0 + 1 − G 1 ( ⋅ , s ) Y ( s ) d ν n ( s ) . Since ⟨ Y − δ e , Y ^ ⟩ ≥ ξ − δ ⟨ e , Y ^ ⟩ \langle Y-\delta e,\widehat{Y}\rangle\ge\xi-\delta\langle e,\widehat{Y}\rangle ⟨ Y − δ e , Y ⟩ ≥ ξ − δ ⟨ e , Y ⟩ on I n \mathcal{I}_{n} I n , the reproducing property and complementarity give ⟨ Y − δ e , W n ⟩ H 1 ≥ ∥ u n ∥ V 2 + δ ∫ 0 + 1 − ( 1 − ⟨ e , Y ^ ( s ) ⟩ ) d ν n ( s ) \langle Y-\delta e,W_n\rangle_{\mathcal{H}_1}\geq\|u_n\|_{\mathcal{V}}^2+\delta\int_{0+}^{1-}(1-\langle e,\widehat{Y}(s)\rangle)\,d\nu_n(s) ⟨ Y − δ e , W n ⟩ H 1 ≥ ∥ u n ∥ V 2 + δ ∫ 0 + 1 − ( 1 − ⟨ e , Y ( s )⟩) d ν n ( s ) . Expanding ∥ Y − δ e − W n ∥ H 1 2 \|Y-\delta e-W_n\|_{\mathcal{H}_1}^2 ∥ Y − δ e − W n ∥ H 1 2 yields the transverse comparison ∥ Y − δ e ∥ H 1 2 − ∥ u n ∥ V 2 ≥ ∥ Y − δ e − W n ∥ H 1 2 + ∥ u n ∥ V 2 − ∥ W n ∥ H 1 2 + 2 δ ∫ 0 + 1 − ( 1 − ⟨ e , Y ^ ( s ) ⟩ ) d ν n ( s ) \|Y-\delta e\|_{\mathcal{H}_1}^2-\|u_n\|_{\mathcal{V}}^2\geq\|Y-\delta e-W_n\|_{\mathcal{H}_1}^2+\|u_n\|_{\mathcal{V}}^2-\|W_n\|_{\mathcal{H}_1}^2+2\delta\int_{0+}^{1-}(1-\langle e,\widehat{Y}(s)\rangle)\,d\nu_n(s) ∥ Y − δ e ∥ H 1 2 − ∥ u n ∥ V 2 ≥ ∥ Y − δ e − W n ∥ H 1 2 + ∥ u n ∥ V 2 − ∥ W n ∥ H 1 2 + 2 δ ∫ 0 + 1 − ( 1 − ⟨ e , Y ( s )⟩) d ν n ( s ) . The middle difference equals 1 2 ∫ 0 + 1 − ∫ 0 + 1 − G 1 ( s , t ) ∥ Y ^ ( s ) − Y ^ ( t ) ∥ 2 d ν n ( s ) d ν n ( t ) ≥ 0 \frac{1}{2}\int_{0+}^{1-}\int_{0+}^{1-}G_1(s,t)\|\widehat{Y}(s)-\widehat{Y}(t)\|^2\,d\nu_n(s)\,d\nu_n(t)\geq0 2 1 ∫ 0 + 1 − ∫ 0 + 1 − G 1 ( s , t ) ∥ Y ( s ) − Y ( t ) ∥ 2 d ν n ( s ) d ν n ( t ) ≥ 0 . Thus ∥ u n ∥ V ≤ ∥ Y − δ e ∥ H 1 \|u_n\|_{\mathcal{V}}\leq\|Y-\delta e\|_{\mathcal{H}_1} ∥ u n ∥ V ≤ ∥ Y − δ e ∥ H 1 . The constraint sets are nested, and the minimum-norm variational inequality gives, for m ≥ n m\geq n m ≥ n , ∥ u m − u n ∥ V 2 ≤ ∥ u m ∥ V 2 − ∥ u n ∥ V 2 \|u_m-u_n\|_{\mathcal{V}}^2\leq\|u_m\|_{\mathcal{V}}^2-\|u_n\|_{\mathcal{V}}^2 ∥ u m − u n ∥ V 2 ≤ ∥ u m ∥ V 2 − ∥ u n ∥ V 2 . Hence u n u_n u n converges strongly in V \mathcal{V} V to a nonnegative u + u_+ u + satisfying every upper constraint. The trajectory X a + ( δ + u + ) e Xa+(\delta+u_+)e X a + ( δ + u + ) e is feasible and has the endpoint states of r r r . Minimality of r r r gives ∥ Y − δ e ∥ H 1 = ∥ u + ∥ V \|Y-\delta e\|_{\mathcal{H}_1}=\|u_+\|_{\mathcal{V}} ∥ Y − δ e ∥ H 1 = ∥ u + ∥ V , so every nonnegative term on the right of the transverse comparison tends to zero.
If δ > 0 \delta>0 δ > 0 , then ∫ 0 + 1 − ∥ Y ^ ( s ) − e ∥ 2 d ν n ( s ) = 2 ∫ 0 + 1 − ( 1 − ⟨ e , Y ^ ( s ) ⟩ ) d ν n ( s ) → 0 \int_{0+}^{1-}\|\widehat{Y}(s)-e\|^2\,d\nu_n(s)=2\int_{0+}^{1-}(1-\langle e,\widehat{Y}(s)\rangle)\,d\nu_n(s)\to0 ∫ 0 + 1 − ∥ Y ( s ) − e ∥ 2 d ν n ( s ) = 2 ∫ 0 + 1 − ( 1 − ⟨ e , Y ( s )⟩) d ν n ( s ) → 0 . With M = sup s ∈ ( 0 , 1 ) G 1 ( s , s ) < ∞ M=\sup_{s\in(0,1)}\sqrt{G_1(s,s)}<\infty M = sup s ∈ ( 0 , 1 ) G 1 ( s , s ) < ∞ , the reproducing property gives u n ≤ M ∥ u n ∥ V u_n\leq M\|u_n\|_{\mathcal{V}} u n ≤ M ∥ u n ∥ V , and positivity and symmetry of the kernel give ∥ W n − u n e ∥ H 1 2 ≤ ∫ 0 + 1 − ∥ Y ^ ( s ) − e ∥ 2 u n ( s ) d ν n ( s ) → 0 \|W_n-u_ne\|_{\mathcal{H}_1}^2\leq\int_{0+}^{1-}\|\widehat{Y}(s)-e\|^2u_n(s)\,d\nu_n(s)\to0 ∥ W n − u n e ∥ H 1 2 ≤ ∫ 0 + 1 − ∥ Y ( s ) − e ∥ 2 u n ( s ) d ν n ( s ) → 0 . Together with ∥ Y − δ e − W n ∥ H 1 → 0 \|Y-\delta e-W_n\|_{\mathcal{H}_1}\to0 ∥ Y − δ e − W n ∥ H 1 → 0 , this proves Y = ( δ + u + ) e Y=(\delta+u_+)e Y = ( δ + u + ) e . Thus r r r lies in span { a , e } = span { r 0 , r T } \operatorname{span}\{a,e\}=\operatorname{span}\{r_0,r_T\} span { a , e } = span { r 0 , r T } , and ⟨ e , r ( t ) ⟩ ≥ δ > 0 \langle e,r(t)\rangle\geq\delta>0 ⟨ e , r ( t )⟩ ≥ δ > 0 throughout. With the multiplier of Proposition A.6 , Theorem 2.5 therefore gives uniqueness.
For blocked symmetric data, I ‾ \overline{I} I is compactly contained in ( 0 , 1 ) (0,1) ( 0 , 1 ) . This follows immediately at strict endpoints. At a terminal contact r T = − a r_T=-a r T = − a , the final free cubic of Proposition A.6 gives X ( t ) = − 1 + 1 2 X ¨ ( 1 ) ( 1 − t ) 2 − 1 6 X ( 3 ) ( 1 − ) ( 1 − t ) 3 X(t)=-1+\frac{1}{2}\ddot{X}(1)(1-t)^2-\frac{1}{6}X^{(3)}(1-)(1-t)^3 X ( t ) = − 1 + 2 1 X ¨ ( 1 ) ( 1 − t ) 2 − 6 1 X ( 3 ) ( 1 − ) ( 1 − t ) 3 and Y ( t ) = O ( ( 1 − t ) 2 ) Y(t)=O((1-t)^2) Y ( t ) = O (( 1 − t ) 2 ) . Separation forces X ¨ ( 1 ) ≤ 0 \ddot{X}(1)\leq0 X ¨ ( 1 ) ≤ 0 and, if it vanishes, X ( 3 ) ( 1 − ) ≥ 0 X^{(3)}(1-)\geq0 X ( 3 ) ( 1 − ) ≥ 0 . Hence X ≤ − 1 X\leq-1 X ≤ − 1 near the terminal endpoint. For all sufficiently large n n n , the set I n I_n I n is therefore the fixed compact set I ‾ \overline{I} I , and u n = u + u_n=u_+ u n = u + with a fixed multiplier ν \nu ν . The middle difference in the transverse comparison vanishes, and G 1 G_1 G 1 is bounded below by a positive constant on I ‾ × I ‾ \overline{I}\times\overline{I} I × I . Hence Y ^ = b 0 \widehat{Y}=b_0 Y = b 0 ν \nu ν -almost everywhere for a unit vector b 0 ⊥ a b_0\perp a b 0 ⊥ a , so W n = u + b 0 W_n=u_+b_0 W n = u + b 0 and Y = u + b 0 Y=u_+b_0 Y = u + b 0 . The multiplier is nonzero because X X X crosses zero, where u + ≥ 1 u_+\geq1 u + ≥ 1 , and strict kernel positivity gives u + ( t ) > 0 u_+(t)>0 u + ( t ) > 0 for every t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) . Thus r r r lies in span { a , b 0 } \operatorname{span}\{a,b_0\} span { a , b 0 } , which contains r 0 r_0 r 0 and r T r_T r T .
Equality for symmetric data. Write this minimizer as r = X a + z b 0 r=Xa+zb_0 r = X a + z b 0 , with z = u + > 0 z=u_+>0 z = u + > 0 in ( 0 , 1 ) (0,1) ( 0 , 1 ) . Its multiplier μ \mu μ from Proposition A.6 is nonzero, Theorem 2.5 applies with unit vector b 0 b_0 b 0 and δ = 0 \delta=0 δ = 0 , and projection of (2.4) gives B z = S [ B z ] Bz=S[Bz] B z = S [ B z ] . Here B : V → L 2 ( ( 0 , 1 ) , μ ; R ) B:\mathcal{V}\to L^2((0,1),\mu;\mathbb{R}) B : V → L 2 (( 0 , 1 ) , μ ; R ) is the scalar trace map from the certificate and S = B B ∗ S=BB^* S = B B ∗ has kernel G 1 G_1 G 1 . Compactness of supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) makes z z z bounded away from zero there. For f ∈ L 2 ( ( 0 , 1 ) , μ ; R ) f\in L^2((0,1),\mu;\mathbb{R}) f ∈ L 2 (( 0 , 1 ) , μ ; R ) , symmetry gives ⟨ f , ( I − S ) f ⟩ L 2 ( ( 0 , 1 ) , μ ; R ) = 1 2 ∫ 0 + 1 − ∫ 0 + 1 − G 1 ( t , s ) z ( t ) z ( s ) ( f ( t ) z ( t ) − f ( s ) z ( s ) ) 2 d μ ( s ) d μ ( t ) \langle f,(I-S)f\rangle_{L^2((0,1),\mu;\mathbb{R})}=\frac{1}{2}\int_{0+}^{1-}\int_{0+}^{1-}G_1(t,s)z(t)z(s)\left(\frac{f(t)}{z(t)}-\frac{f(s)}{z(s)}\right)^2\,d\mu(s)\,d\mu(t) ⟨ f , ( I − S ) f ⟩ L 2 (( 0 , 1 ) , μ ; R ) = 2 1 ∫ 0 + 1 − ∫ 0 + 1 − G 1 ( t , s ) z ( t ) z ( s ) ( z ( t ) f ( t ) − z ( s ) f ( s ) ) 2 d μ ( s ) d μ ( t ) . Strict kernel positivity makes the nullspace of I − S I-S I − S the span of the trace of z z z . If a scalar h ∈ V h\in\mathcal{V} h ∈ V has Q μ [ h ] = 0 Q_\mu[h]=0 Q μ [ h ] = 0 , positivity of I − B ∗ B I-B^*B I − B ∗ B gives h = B ∗ B h h=B^*Bh h = B ∗ B h , and therefore B h = S [ B h ] Bh=S[Bh] B h = S [ B h ] . Since B ∗ [ B z ] = z B^*[Bz]=z B ∗ [ B z ] = z , every such h h h is a scalar multiple of z z z . Any other global minimizer q q q has equality in the cost identity (3) . Componentwise application of this scalar equality case yields q = r + ζ ζ q=r+\zeta\zeta q = r + ζ ζ for a constant ζ ∈ R d \zeta\in\mathbb{R}^d ζ ∈ R d . The remaining term in that identity forces ∥ q ( t ) ∥ = 1 \|q(t)\|=1 ∥ q ( t ) ∥ = 1 on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) . With ζ 1 = ⟨ a , ζ ⟩ \zeta_1=\langle a,\zeta\rangle ζ 1 = ⟨ a , ζ ⟩ and ζ 2 = ⟨ b 0 , ζ ⟩ \zeta_2=\langle b_0,\zeta\rangle ζ 2 = ⟨ b 0 , ζ ⟩ , 2 ζ 1 X ( t ) z ( t ) + 2 ζ 2 + ∥ ζ ∥ 2 = 0 2\zeta_1\frac{X(t)}{z(t)}+2\zeta_2+\|\zeta\|^2=0 2 ζ 1 z ( t ) X ( t ) + 2 ζ 2 + ∥ ζ ∥ 2 = 0 on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) . If the support contains two distinct contact normals, their ratios X / z X/z X / z differ. Thus ζ 1 = 0 \zeta_1=0 ζ 1 = 0 and ∥ b 0 + ζ ∥ = 1 \|b_0+\zeta\|=1 ∥ b 0 + ζ ∥ = 1 , so q = X a + z b q=Xa+zb q = X a + z b for the fixed unit vector b = b 0 + ζ ⊥ a b=b_0+\zeta\perp a b = b 0 + ζ ⊥ a .
If the contact normal on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) is one fixed n = n 1 a + n 2 b 0 n=n_1a+n_2b_0 n = n 1 a + n 2 b 0 with n 2 > 0 n_2>0 n 2 > 0 , Green reconstruction gives r = r ‾ + n A r=\overline{r}+nA r = r + n A , where A ( t ) = ∫ 0 + 1 − G 1 ( t , s ) d μ ( s ) A(t)=\int_{0+}^{1-}G_1(t,s)\,d\mu(s) A ( t ) = ∫ 0 + 1 − G 1 ( t , s ) d μ ( s ) . At a support time the transverse coordinate @page 17/30 starts_mid=1 ends_mid=1 and complementarity give ⟨ Y − δ e , W n ⟩ H 1 ≥ ∥ u n ∥ V 2 + δ ∫ 0 + 1 − ( 1 − ⟨ e , Y ^ ( s ) ⟩ ) d ν n ( s ) \langle Y-\delta e,W_n\rangle_{\mathcal{H}_1}\geq\|u_n\|_{\mathcal{V}}^2+\delta\int_{0+}^{1-}(1-\langle e,\widehat{Y}(s)\rangle)\,d\nu_n(s) ⟨ Y − δ e , W n ⟩ H 1 ≥ ∥ u n ∥ V 2 + δ ∫ 0 + 1 − ( 1 − ⟨ e , Y ( s )⟩) d ν n ( s ) . Expanding ∥ Y − δ e − W n ∥ H 1 2 \|Y-\delta e-W_n\|_{\mathcal{H}_1}^2 ∥ Y − δ e − W n ∥ H 1 2 yields the transverse comparison ∥ Y − δ e ∥ H 1 2 − ∥ u n ∥ V 2 ≥ ∥ Y − δ e − W n ∥ H 1 2 + ∥ u n ∥ V 2 − ∥ W n ∥ H 1 2 + 2 δ ∫ 0 + 1 − ( 1 − ⟨ e , Y ^ ( s ) ⟩ ) d ν n ( s ) \|Y-\delta e\|_{\mathcal{H}_1}^2-\|u_n\|_{\mathcal{V}}^2\geq\|Y-\delta e-W_n\|_{\mathcal{H}_1}^2+\|u_n\|_{\mathcal{V}}^2-\|W_n\|_{\mathcal{H}_1}^2+2\delta\int_{0+}^{1-}(1-\langle e,\widehat{Y}(s)\rangle)\,d\nu_n(s) ∥ Y − δ e ∥ H 1 2 − ∥ u n ∥ V 2 ≥ ∥ Y − δ e − W n ∥ H 1 2 + ∥ u n ∥ V 2 − ∥ W n ∥ H 1 2 + 2 δ ∫ 0 + 1 − ( 1 − ⟨ e , Y ( s )⟩) d ν n ( s ) . The middle difference equals 1 2 ∫ 0 + 1 − ∫ 0 + 1 − G 1 ( s , t ) ∥ Y ^ ( s ) − Y ^ ( t ) ∥ 2 d ν n ( s ) d ν n ( t ) ≥ 0 \frac{1}{2}\int_{0+}^{1-}\int_{0+}^{1-}G_1(s,t)\|\widehat{Y}(s)-\widehat{Y}(t)\|^2\,d\nu_n(s)\,d\nu_n(t)\geq0 2 1 ∫ 0 + 1 − ∫ 0 + 1 − G 1 ( s , t ) ∥ Y ( s ) − Y ( t ) ∥ 2 d ν n ( s ) d ν n ( t ) ≥ 0 . Thus ∥ u n ∥ V ≤ ∥ Y − δ e ∥ H 1 \|u_n\|_{\mathcal{V}}\leq\|Y-\delta e\|_{\mathcal{H}_1} ∥ u n ∥ V ≤ ∥ Y − δ e ∥ H 1 . The constraint sets are nested, and the minimum-norm variational inequality gives, for m ≥ n m\geq n m ≥ n , ∥ u m − u n ∥ V 2 ≤ ∥ u m ∥ V 2 − ∥ u n ∥ V 2 \|u_m-u_n\|_{\mathcal{V}}^2\leq\|u_m\|_{\mathcal{V}}^2-\|u_n\|_{\mathcal{V}}^2 ∥ u m − u n ∥ V 2 ≤ ∥ u m ∥ V 2 − ∥ u n ∥ V 2 . Hence u n u_n u n converges strongly in V \mathcal{V} V to a nonnegative u + u_+ u + satisfying every upper constraint. The trajectory X a + ( δ + u + ) e Xa+(\delta+u_+)e X a + ( δ + u + ) e is feasible and has the endpoint states of r r r . Minimality of r r r gives ∥ Y − δ e ∥ H 1 = ∥ u + ∥ V \|Y-\delta e\|_{\mathcal{H}_1}=\|u_+\|_{\mathcal{V}} ∥ Y − δ e ∥ H 1 = ∥ u + ∥ V , so every nonnegative term on the right of the transverse comparison tends to zero.
If δ > 0 \delta>0 δ > 0 , then ∫ 0 + 1 − ∥ Y ^ ( s ) − e ∥ 2 d ν n ( s ) = 2 ∫ 0 + 1 − ( 1 − ⟨ e , Y ^ ( s ) ⟩ ) d ν n ( s ) → 0 \int_{0+}^{1-}\|\widehat{Y}(s)-e\|^2\,d\nu_n(s)=2\int_{0+}^{1-}(1-\langle e,\widehat{Y}(s)\rangle)\,d\nu_n(s)\to0 ∫ 0 + 1 − ∥ Y ( s ) − e ∥ 2 d ν n ( s ) = 2 ∫ 0 + 1 − ( 1 − ⟨ e , Y ( s )⟩) d ν n ( s ) → 0 . With M = sup s ∈ ( 0 , 1 ) G 1 ( s , s ) < ∞ M=\sup_{s\in(0,1)}\sqrt{G_1(s,s)}<\infty M = sup s ∈ ( 0 , 1 ) G 1 ( s , s ) < ∞ , the reproducing property gives u n ≤ M ∥ u n ∥ V u_n\leq M\|u_n\|_{\mathcal{V}} u n ≤ M ∥ u n ∥ V , and positivity and symmetry of the kernel give ∥ W n − u n e ∥ H 1 2 ≤ ∫ 0 + 1 − ∥ Y ^ ( s ) − e ∥ 2 u n ( s ) d ν n ( s ) → 0 \|W_n-u_ne\|_{\mathcal{H}_1}^2\leq\int_{0+}^{1-}\|\widehat{Y}(s)-e\|^2u_n(s)\,d\nu_n(s)\to0 ∥ W n − u n e ∥ H 1 2 ≤ ∫ 0 + 1 − ∥ Y ( s ) − e ∥ 2 u n ( s ) d ν n ( s ) → 0 . Together with ∥ Y − δ e − W n ∥ H 1 → 0 \|Y-\delta e-W_n\|_{\mathcal{H}_1}\to0 ∥ Y − δ e − W n ∥ H 1 → 0 , this proves Y = ( δ + u + ) e Y=(\delta+u_+)e Y = ( δ + u + ) e . Thus r r r lies in span { a , e } = span { r 0 , r T } \operatorname{span}\{a,e\}=\operatorname{span}\{r_0,r_T\} span { a , e } = span { r 0 , r T } , and ⟨ e , r ( t ) ⟩ ≥ δ > 0 \langle e,r(t)\rangle\geq\delta>0 ⟨ e , r ( t )⟩ ≥ δ > 0 throughout. With the multiplier of Proposition A.6 , Theorem 2.5 therefore gives uniqueness.
For blocked symmetric data, I ‾ \overline{I} I is compactly contained in ( 0 , 1 ) (0,1) ( 0 , 1 ) . This follows immediately at strict endpoints. At a terminal contact r T = − a r_T=-a r T = − a , the final free cubic of Proposition A.6 gives X ( t ) = − 1 + 1 2 X ¨ ( 1 ) ( 1 − t ) 2 − 1 6 X ( 3 ) ( 1 − ) ( 1 − t ) 3 X(t)=-1+\frac{1}{2}\ddot{X}(1)(1-t)^2-\frac{1}{6}X^{(3)}(1-)(1-t)^3 X ( t ) = − 1 + 2 1 X ¨ ( 1 ) ( 1 − t ) 2 − 6 1 X ( 3 ) ( 1 − ) ( 1 − t ) 3 and Y ( t ) = O ( ( 1 − t ) 2 ) Y(t)=O((1-t)^2) Y ( t ) = O (( 1 − t ) 2 ) . Separation forces X ¨ ( 1 ) ≤ 0 \ddot{X}(1)\leq0 X ¨ ( 1 ) ≤ 0 and, if it vanishes, X ( 3 ) ( 1 − ) ≥ 0 X^{(3)}(1-)\geq0 X ( 3 ) ( 1 − ) ≥ 0 . Hence X ≤ − 1 X\leq-1 X ≤ − 1 near the terminal endpoint. For all sufficiently large n n n , the set I n I_n I n is therefore the fixed compact set I ‾ \overline{I} I , and u n = u + u_n=u_+ u n = u + with a fixed multiplier ν \nu ν . The middle difference in the transverse comparison vanishes, and G 1 G_1 G 1 is bounded below by a positive constant on I ‾ × I ‾ \overline{I}\times\overline{I} I × I . Hence Y ^ = b 0 \widehat{Y}=b_0 Y = b 0 ν \nu ν -almost everywhere for a unit vector b 0 ⊥ a b_0\perp a b 0 ⊥ a , so W n = u + b 0 W_n=u_+b_0 W n = u + b 0 and Y = u + b 0 Y=u_+b_0 Y = u + b 0 . The multiplier is nonzero because X X X crosses zero, where u + ≥ 1 u_+\geq1 u + ≥ 1 , and strict kernel positivity gives u + ( t ) > 0 u_+(t)>0 u + ( t ) > 0 for every t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) . Thus r r r lies in span { a , b 0 } \operatorname{span}\{a,b_0\} span { a , b 0 } , which contains r 0 r_0 r 0 and r T r_T r T .
Equality for symmetric data. Write this minimizer as r = X a + z b 0 r=Xa+zb_0 r = X a + z b 0 , with z = u + > 0 z=u_+>0 z = u + > 0 in ( 0 , 1 ) (0,1) ( 0 , 1 ) . Its multiplier μ \mu μ from Proposition A.6 is nonzero, Theorem 2.5 applies with unit vector b 0 b_0 b 0 and δ = 0 \delta=0 δ = 0 , and projection of (2.4) gives B z = S [ B z ] Bz=S[Bz] B z = S [ B z ] . Here B : V → L 2 ( ( 0 , 1 ) , μ ; R ) B:\mathcal{V}\to L^2((0,1),\mu;\mathbb{R}) B : V → L 2 (( 0 , 1 ) , μ ; R ) is the scalar trace map from the certificate and S = B B ∗ S=BB^* S = B B ∗ has kernel G 1 G_1 G 1 . Compactness of supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) makes z z z bounded away from zero there. For f ∈ L 2 ( ( 0 , 1 ) , μ ; R ) f\in L^2((0,1),\mu;\mathbb{R}) f ∈ L 2 (( 0 , 1 ) , μ ; R ) , symmetry gives ⟨ f , ( I − S ) f ⟩ L 2 ( ( 0 , 1 ) , μ ; R ) = 1 2 ∫ 0 + 1 − ∫ 0 + 1 − G 1 ( t , s ) z ( t ) z ( s ) ( f ( t ) z ( t ) − f ( s ) z ( s ) ) 2 d μ ( s ) d μ ( t ) \langle f,(I-S)f\rangle_{L^2((0,1),\mu;\mathbb{R})}=\frac{1}{2}\int_{0+}^{1-}\int_{0+}^{1-}G_1(t,s)z(t)z(s)\left(\frac{f(t)}{z(t)}-\frac{f(s)}{z(s)}\right)^2\,d\mu(s)\,d\mu(t) ⟨ f , ( I − S ) f ⟩ L 2 (( 0 , 1 ) , μ ; R ) = 2 1 ∫ 0 + 1 − ∫ 0 + 1 − G 1 ( t , s ) z ( t ) z ( s ) ( z ( t ) f ( t ) − z ( s ) f ( s ) ) 2 d μ ( s ) d μ ( t ) . Strict kernel positivity makes the nullspace of I − S I-S I − S the span of the trace of z z z . If a scalar h ∈ V h\in\mathcal{V} h ∈ V has Q μ [ h ] = 0 Q_\mu[h]=0 Q μ [ h ] = 0 , positivity of I − B ∗ B I-B^*B I − B ∗ B gives h = B ∗ B h h=B^*Bh h = B ∗ B h , and therefore B h = S [ B h ] Bh=S[Bh] B h = S [ B h ] . Since B ∗ [ B z ] = z B^*[Bz]=z B ∗ [ B z ] = z , every such h h h is a scalar multiple of z z z . Any other global minimizer q q q has equality in the cost identity (3) . Componentwise application of this scalar equality case yields q = r + ζ ζ q=r+\zeta\zeta q = r + ζ ζ for a constant ζ ∈ R d \zeta\in\mathbb{R}^d ζ ∈ R d . The remaining term in that identity forces ∥ q ( t ) ∥ = 1 \|q(t)\|=1 ∥ q ( t ) ∥ = 1 on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) . With ζ 1 = ⟨ a , ζ ⟩ \zeta_1=\langle a,\zeta\rangle ζ 1 = ⟨ a , ζ ⟩ and ζ 2 = ⟨ b 0 , ζ ⟩ \zeta_2=\langle b_0,\zeta\rangle ζ 2 = ⟨ b 0 , ζ ⟩ , 2 ζ 1 X ( t ) z ( t ) + 2 ζ 2 + ∥ ζ ∥ 2 = 0 2\zeta_1\frac{X(t)}{z(t)}+2\zeta_2+\|\zeta\|^2=0 2 ζ 1 z ( t ) X ( t ) + 2 ζ 2 + ∥ ζ ∥ 2 = 0 on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) . If the support contains two distinct contact normals, their ratios X / z X/z X / z differ. Thus ζ 1 = 0 \zeta_1=0 ζ 1 = 0 and ∥ b 0 + ζ ∥ = 1 \|b_0+\zeta\|=1 ∥ b 0 + ζ ∥ = 1 , so q = X a + z b q=Xa+zb q = X a + z b for the fixed unit vector b = b 0 + ζ ⊥ a b=b_0+\zeta\perp a b = b 0 + ζ ⊥ a .
If the contact normal on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) is one fixed n = n 1 a + n 2 b 0 n=n_1a+n_2b_0 n = n 1 a + n 2 b 0 with n 2 > 0 n_2>0 n 2 > 0 , Green reconstruction gives r = r ‾ + n A r=\overline{r}+nA r = r + n A , where A ( t ) = ∫ 0 + 1 − G 1 ( t , s ) d μ ( s ) A(t)=\int_{0+}^{1-}G_1(t,s)\,d\mu(s) A ( t ) = ∫ 0 + 1 − G 1 ( t , s ) d μ ( s ) . At a support time the transverse coordinate forces A = 1 A=1 A = 1 , hence r ˉ = 0 \bar{r}=0 r ˉ = 0 . The free longitudinal coordinate is strictly decreasing in ( 0 , 1 ) (0,1) ( 0 , 1 ) , so it has a single zero t ∗ t_* t ∗ . The multiplier is one atom and r ( t ) = r ˉ ( t ) + G 1 ( t , t ∗ ) G 1 ( t ∗ , t ∗ ) n r(t)=\bar{r}(t)+\frac{G_1(t,t_*)}{G_1(t_*,t_*)}n r ( t ) = r ˉ ( t ) + G 1 ( t ∗ , t ∗ ) G 1 ( t , t ∗ ) n . Since z = n 2 A z=n_2A z = n 2 A , every feasible trajectory with the same cost has the same representation with a unit vector n ~ = n + n 2 ζ \tilde{n}=n+n_2\zeta n ~ = n + n 2 ζ in place of n n n . Feasibility gives tangency at t ∗ t_* t ∗ : ⟨ r ˉ ˙ ( t ∗ ) , n ~ ⟩ + ∂ 1 G 1 ( t ∗ , t ∗ ) G 1 ( t ∗ , t ∗ ) = 0 \langle\dot{\bar{r}}(t_*),\tilde{n}\rangle+\frac{\partial_1G_1(t_*,t_*)}{G_1(t_*,t_*)}=0 ⟨ r ˉ ˙ ( t ∗ ) , n ~ ⟩ + G 1 ( t ∗ , t ∗ ) ∂ 1 G 1 ( t ∗ , t ∗ ) = 0 . The nonzero free longitudinal derivative fixes ⟨ a , n ~ ⟩ = n 1 \langle a,\tilde{n}\rangle=n_1 ⟨ a , n ~ ⟩ = n 1 . Its unit norm then gives n ~ = n 1 a + n 2 b \tilde{n}=n_1a+n_2b n ~ = n 1 a + n 2 b with b ⊥ a b\perp a b ⊥ a unit, again yielding q = X a + z b q=Xa+zb q = X a + z b .
Conversely, every such choice of b b b preserves the endpoint states, clearance, and cost. Thus the complete family of global minimizers is r b ( t ) = X ( t ) a + z ( t ) b r_b(t)=X(t)a+z(t)b r b ( t ) = X ( t ) a + z ( t ) b for b ∈ a ⊥ b\in a^\perp b ∈ a ⊥ with ∥ b ∥ = 1 \lVert b\rVert=1 ∥ b ∥ = 1 . Each lies in the fixed plane span { a , b } \operatorname{span}\{a,b\} span { a , b } . For d = 2 d=2 d = 2 there are two choices; for d > 2 d>2 d > 2 they are parametrized by S d − 2 S^{d-2} S d − 2 . The orthogonal maps fixing the endpoint axis act transitively on these choices. In the chosen plane b = ± e b=\pm e b = ± e , and only b = e b=e b = e gives the upper trajectory since z > 0 z>0 z > 0 .
Local minimizers on either side. The local minimizers mentioned in § 2.2.3 follow from the next lemma.
Let d = 2 d=2 d = 2 . A feasible trajectory r r r has a continuous angle θ \theta θ with r ( t ) = ∥ r ( t ) ∥ ( cos θ ( t ) , sin θ ( t ) ) r(t)=\lVert r(t)\rVert(\cos\theta(t),\sin\theta(t)) r ( t ) = ∥ r ( t )∥ ( cos θ ( t ) , sin θ ( t )) , unique up to an additive multiple of 2 π 2\pi 2 π , so its total angle θ ( T ) − θ ( 0 ) \theta(T)-\theta(0) θ ( T ) − θ ( 0 ) is well defined. Feasible trajectories with a common total angle form a homotopy class ; the two ways around the disk are two such classes, whose total angles differ by 2 π 2\pi 2 π .
Lemma A.7 (Minimizers in a homotopy class). If ∥ r T ∥ > R \lVert r_T\rVert>R ∥ r T ∥ > R , every homotopy class of feasible trajectories contains a minimizer of 1 2 ∥ r ¨ ∥ 2 \frac{1}{2}\lVert\ddot{r}\rVert^2 2 1 ∥ r ¨ ∥ 2 over that class. Each such minimizer is a local minimizer of (1) and satisfies (2) for a finite nonnegative Radon measure μ \mu μ .
Proof of Lemma A.7. Take a feasible q q q (§ 2.1 ) and choose β ∈ C ∞ ( [ 0 , T ] ; R ) \beta\in C^\infty([0,T];\mathbb{R}) β ∈ C ∞ ([ 0 , T ] ; R ) equal to zero near 0 0 0 and one near T T T . If Rot ( ϕ ) \operatorname{Rot}(\phi) Rot ( ϕ ) denotes planar rotation by ϕ \phi ϕ , then, for k ∈ Z k\in\mathbb{Z} k ∈ Z , q k ( t ) = Rot ( 2 π k β ( t ) ) q ( t ) q_k(t)=\operatorname{Rot}(2\pi k\beta(t))q(t) q k ( t ) = Rot ( 2 π k β ( t )) q ( t ) is feasible, agrees with q q q near both endpoints, and adds 2 π k 2\pi k 2 π k to its total angle. Thus every class is nonempty.
If feasible q , r q,r q , r satisfy sup t ∈ [ 0 , T ] ∥ q ( t ) − r ( t ) ∥ < R \sup_{t\in[0,T]}\lVert q(t)-r(t)\rVert<R sup t ∈ [ 0 , T ] ∥ q ( t ) − r ( t )∥ < R , then ⟨ q ( t ) , r ( t ) ⟩ > 0 \langle q(t),r(t)\rangle>0 ⟨ q ( t ) , r ( t )⟩ > 0 . Their relative angle lies in ( − π / 2 , π / 2 ) (-\pi/2,\pi/2) ( − π /2 , π /2 ) and vanishes at both endpoints, so they belong to the same class. A minimizing sequence in a class is bounded in H 2 ( ( 0 , T ) ; R 2 ) H^2((0,T);\mathbb{R}^2) H 2 (( 0 , T ) ; R 2 ) by its acceleration cost and fixed initial data. A subsequence converges weakly in H 2 ( ( 0 , T ) ; R 2 ) H^2((0,T);\mathbb{R}^2) H 2 (( 0 , T ) ; R 2 ) and strongly in C 1 ( [ 0 , T ] ; R 2 ) C^1([0,T];\mathbb{R}^2) C 1 ([ 0 , T ] ; R 2 ) ; its limit is feasible and remains in that class. Weak lower semicontinuity gives a minimizer. Class stability and the Sobolev embedding make it a local minimizer, and Proposition 2.3 supplies its multiplier measure.
Proof of Theorem 2.8
The counting argument concerns geometric contacts, including zero-mass touches; the contact labels are assigned afterward.
Lemma A.8 (Canonical contact order). For blocked data the interior geometric contacts of the selected minimizer occur on the near side, strictly ordered counterclockwise, with w ( t ) : = r ( t ) × r ˙ ( t ) > 0 w(t) := r(t) \times\dot{r}(t) > 0 w ( t ) := r ( t ) × r ˙ ( t ) > 0 and τ ( t ) : = r ( t ) × r ( 3 ) ( t ) < 0 \tau(t) := r(t) \times r^{(3)}(t) < 0 τ ( t ) := r ( t ) × r ( 3 ) ( t ) < 0 at each of them and τ \tau τ continuous across an atom. Proof. Write z ( t ) = ⟨ e , r ( t ) ⟩ z(t) = \langle e,r(t)\rangle z ( t ) = ⟨ e , r ( t )⟩ , so that every interior contact normal n n n has ⟨ e , n ⟩ > 0 \langle e,n\rangle> 0 ⟨ e , n ⟩ > 0 . By Propositions A.6 and 2.7 , the interior multiplier is finite, nonzero and compactly supported, and z > 0 z > 0 z > 0 in ( 0 , 1 ) (0,1) ( 0 , 1 ) . Since the free projection is the constant δ \delta δ , the Green reconstruction (2.4) gives, for 0 < t < 1 0 < t < 1 0 < t < 1 , u ( t ) : = z ( t ) − δ = ∫ 0 + 1 − G 1 ( t , s ) z ( s ) d μ ( s ) > 0 u(t) := z(t)-\delta= \int_{0+}^{1-} G_1(t,s)z(s)\,d\mu(s) > 0 u ( t ) := z ( t ) − δ = ∫ 0 + 1 − G 1 ( t , s ) z ( s ) d μ ( s ) > 0 , D t 4 u = z μ ≥ 0 D_t^4u = z\mu\ge0 D t 4 u = z μ ≥ 0 . Thus u ¨ \ddot{u} u ¨ is convex. Differentiating the kernel at the endpoints gives u ¨ ( 0 ) = ∫ 0 + 1 − s ( 1 − s ) 2 z ( s ) d μ ( s ) > 0 \ddot{u}(0) = \int_{0+}^{1-} s(1-s)^2z(s)\,d\mu(s) > 0 u ¨ ( 0 ) = ∫ 0 + 1 − s ( 1 − s ) 2 z ( s ) d μ ( s ) > 0 , u ¨ ( 1 ) = ∫ 0 + 1 − s 2 ( 1 − s ) z ( s ) d μ ( s ) > 0 \ddot{u}(1) = \int_{0+}^{1-} s^2(1-s)z(s)\,d\mu(s) > 0 u ¨ ( 1 ) = ∫ 0 + 1 − s 2 ( 1 − s ) z ( s ) d μ ( s ) > 0 . Its integral is u ˙ ( 1 ) − u ˙ ( 0 ) = 0 \dot{u}(1)-\dot{u}(0) = 0 u ˙ ( 1 ) − u ˙ ( 0 ) = 0 , so u ¨ \ddot{u} u ¨ is negative between exactly two zeros and positive outside them. Integrating once, z ˙ \dot{z} z ˙ starts positive, decreases strictly across this negative interval, and ends negative before reaching z ˙ ( 1 ) = 0 \dot{z}(1) = 0 z ˙ ( 1 ) = 0 . It therefore has exactly one interior zero t ∗ t_* t ∗ , with z ¨ ( t ∗ ) < 0 \ddot{z}(t_*) < 0 z ¨ ( t ∗ ) < 0 .
Put f = ( d 0 − r T , 1 , − r T , 2 ) / ∥ r T − r 0 ∥ f = (d_0-r_{T,1},-r_{T,2})/\lVert r_T-r_0\rVert f = ( d 0 − r T , 1 , − r T , 2 ) / ∥ r T − r 0 ∥ and X ( t ) = ⟨ f , r ( t ) ⟩ X(t) = \langle f,r(t)\rangle X ( t ) = ⟨ f , r ( t )⟩ . The frame ( f , e ) (f,e) ( f , e ) is positively oriented. Blockage of the endpoint segment implies δ < 1 \delta< 1 δ < 1 , X ( 0 ) > 0 X(0) > 0 X ( 0 ) > 0 and X ( 1 ) < 0 X(1) < 0 X ( 1 ) < 0 . At any crossing X = 0 X = 0 X = 0 , feasibility and z > 0 z > 0 z > 0 require z ≥ 1 z \ge1 z ≥ 1 . An interior contact before t ∗ t_* t ∗ must have X > 0 X > 0 X > 0 : if X < 0 X < 0 X < 0 , an earlier crossing would have height at least one, contradicting strict increase of z z z up to a contact with z ≤ 1 z \le1 z ≤ 1 ; if X = 0 X = 0 X = 0 , tangency would give z ˙ = 0 \dot{z} = 0 z ˙ = 0 . Similarly every contact after t ∗ t_* t ∗ has X < 0 X < 0 X < 0 . At these contacts tangency gives z ˙ = w X \dot{z} = wX z ˙ = w X , hence w > 0 w > 0 w > 0 .
If t ∗ t_* t ∗ itself is a contact, the existence of a crossing and the maximality of z ( t ∗ ) z(t_*) z ( t ∗ ) imply z ( t ∗ ) = 1 z(t_*) = 1 z ( t ∗ ) = 1 and X ( t ∗ ) = 0 X(t_*) = 0 X ( t ∗ ) = 0 . It is the only crossing, from positive to negative X X X . For b ( t ) = ∥ r ( t ) ∥ 2 − 1 b(t) = \lVert r(t)\rVert^2 - 1 b ( t ) = ∥ r ( t ) ∥ 2 − 1 , feasibility gives 0 ≤ b ¨ ( t ∗ ) = 2 ( w ( t ∗ ) 2 + z ¨ ( t ∗ ) ) 0 \le\ddot{b}(t_*) = 2(w(t_*)^2+\ddot{z}(t_*)) 0 ≤ b ¨ ( t ∗ ) = 2 ( w ( t ∗ ) 2 + z ¨ ( t ∗ )) , so w ( t ∗ ) ≠ 0 w(t_*) \ne0 w ( t ∗ ) = 0 . The crossing direction gives X ˙ ( t ∗ ) = − w ( t ∗ ) < 0 \dot{X}(t_*) = -w(t_*) < 0 X ˙ ( t ∗ ) = − w ( t ∗ ) < 0 . The strict increase of z z z on the right half of the circle and its strict decrease on the left half now strictly order all interior contact normals in the open semicircle z > 0 z > 0 z > 0 .
Fix any interior geometric contact t t t , write n = r ( t ) n = r(t) n = r ( t ) , e t = ( − n 2 , n 1 ) e_t = (-n_2,n_1) e t = ( − n 2 , n 1 ) and τ = ⟨ e t , r ( 3 ) ( t ± ) ⟩ \tau= \langle e_t,r^{(3)}(t^\pm)\rangle τ = ⟨ e t , r ( 3 ) ( t ± )⟩ . The last value is independent of the side because jerk jumps are radial. Integrating D t 1 r ( 3 ) = r μ D_t^1r^{(3)} = r\mu D t 1 r ( 3 ) = r μ to both endpoints with zero velocity gives 0 = w ( t ) − t ⟨ e t , r ¨ ( t ) ⟩ + 1 2 t 2 τ + P L 0 = w(t)-t\langle e_t,\ddot{r}(t)\rangle+\frac{1}{2}t^2\tau+P_L 0 = w ( t ) − t ⟨ e t , r ¨ ( t )⟩ + 2 1 t 2 τ + P L and 0 = w ( t ) + ( 1 − t ) ⟨ e t , r ¨ ( t ) ⟩ + 1 2 ( 1 − t ) 2 τ + P R 0 = w(t)+(1-t)\langle e_t,\ddot{r}(t)\rangle+\frac{1}{2}(1-t)^2\tau+P_R 0 = w ( t ) + ( 1 − t ) ⟨ e t , r ¨ ( t )⟩ + 2 1 ( 1 − t ) 2 τ + P R , where P L = − ∫ 0 + t − 1 2 s 2 ⟨ e t , r ( s ) ⟩ d μ ( s ) ≥ 0 P_L = -\int_{0+}^{t-}\frac{1}{2}s^2\langle e_t,r(s)\rangle\,d\mu(s) \ge0 P L = − ∫ 0 + t − 2 1 s 2 ⟨ e t , r ( s )⟩ d μ ( s ) ≥ 0 and P R = ∫ t + 1 − 1 2 ( 1 − s ) 2 ⟨ e t , r ( s ) ⟩ d μ ( s ) ≥ 0 P_R = \int_{t+}^{1-}\frac{1}{2}(1-s)^2\langle e_t,r(s)\rangle\,d\mu(s) \ge0 P R = ∫ t + 1 − 2 1 ( 1 − s ) 2 ⟨ e t , r ( s )⟩ d μ ( s ) ≥ 0 . The signs follow from the strict angular order on supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) , whose angular span is below π \pi π . If τ ≥ 0 \tau\ge0 τ ≥ 0 , the first equality forces ⟨ e t , r ¨ ( t ) ⟩ > 0 \langle e_t,\ddot{r}(t)\rangle> 0 ⟨ e t , r ¨ ( t )⟩ > 0 , while the second forces ⟨ e t , r ¨ ( t ) ⟩ < 0 \langle e_t,\ddot{r}(t)\rangle< 0 ⟨ e t , r ¨ ( t )⟩ < 0 . Hence τ < 0 \tau< 0 τ < 0 .
Three-contact curvature. The following local inequality uses the ordered normals, contact curvatures, and tangential-jerk signs supplied by Lemma A.8 .
Lemma A.9 (Three-contact curvature inequality). Let three distinct unit-circle normals n i = r ( t i ) n_i = r(t_i) n i = r ( t i ) , t 1 < t 2 < t 3 t_1 < t_2 < t_3 t 1 < t 2 < t 3 , occur in counterclockwise order with total angular span below π \pi π . Suppose r r r is cubic on each intervening interval, has continuous position, velocity and acceleration at t 2 t_2 t 2 , and has tangent velocities r ˙ ( t i ) = w i ( − n i , 2 , n i , 1 ) \dot{r}(t_i) = w_i(-n_{i,2},n_{i,1}) r ˙ ( t i ) = w i ( − n i , 2 , n i , 1 ) with w i > 0 w_i > 0 w i > 0 . Assume γ ( t i ) : = w i 2 + ⟨ n i , r ¨ ( t i ) ⟩ ≥ 0 \gamma(t_i) := w_i^2+\langle n_i,\ddot{r}(t_i)\rangle\ge0 γ ( t i ) := w i 2 + ⟨ n i , r ¨ ( t i )⟩ ≥ 0 , [ r ( 3 ) ] t 2 = m n 2 [r^{(3)}]_{t_2} = mn_2 [ r ( 3 ) ] t 2 = m n 2 , m > 0 m > 0 m > 0 , and negative outside tangential jerks n 1 × r ( 3 ) ( t 1 + ) < 0 n_1 \times r^{(3)}(t_1^+) < 0 n 1 × r ( 3 ) ( t 1 + ) < 0 , n 3 × r ( 3 ) ( t 3 − ) < 0 n_3 \times r^{(3)}(t_3^-) < 0 n 3 × r ( 3 ) ( t 3 − ) < 0 . Then γ ( t 2 ) < min { γ ( t 1 ) , γ ( t 3 ) } \gamma(t_2) < \min\{\gamma(t_1),\gamma(t_3)\} γ ( t 2 ) < min { γ ( t 1 ) , γ ( t 3 )} . Proof. Rescale time to ( t 1 , t 2 , t 3 ) = ( − 1 , 0 , h ) (t_1,t_2,t_3) = (-1,0,h) ( t 1 , t 2 , t 3 ) = ( − 1 , 0 , h ) , h > 0 h > 0 h > 0 , and rotate n 2 n_2 n 2 to ( 1 , 0 ) (1,0) ( 1 , 0 ) . Write n 1 = ( C , − S ) n_1 = (C,-S) n 1 = ( C , − S ) , n 3 = ( D , V ) n_3 = (D,V) n 3 = ( D , V ) , with S , V > 0 S,V > 0 S , V > 0 , and α n = r ¨ 1 ( 0 ) \alpha_n = \ddot{r}_1(0) α n = r ¨ 1 ( 0 ) . Let r L , r R r_L,r_R r L , r R be the polynomial extensions of the left
and right cubics, and set b L ( t ) = ∥ r L ( t ) ∥ 2 − 1 b_L(t)=\lVert r_L(t)\rVert^2-1 b L ( t ) = ∥ r L ( t ) ∥ 2 − 1 and b R ( t ) = ∥ r R ( t ) ∥ 2 − 1 b_R(t)=\lVert r_R(t)\rVert^2-1 b R ( t ) = ∥ r R ( t ) ∥ 2 − 1 . The double zeros of b L b_L b L at − 1 , 0 -1,0 − 1 , 0 , its endpoint second derivatives and leading coefficient give b L ( t ) = t 2 ( t + 1 ) 2 [ a 6 t 2 + ( a 6 + γ ( 0 ) − γ ( − 1 ) ) t + γ ( 0 ) ] b_L(t)=t^2(t+1)^2\left[a_6t^2+\left(a_6+\gamma(0)-\gamma(-1)\right)t+\gamma(0)\right] b L ( t ) = t 2 ( t + 1 ) 2 [ a 6 t 2 + ( a 6 + γ ( 0 ) − γ ( − 1 ) ) t + γ ( 0 ) ] , a 6 = ∥ r L ( 3 ) ( 0 ) ∥ 2 / 36 a_6=\lVert r_L^{(3)}(0)\rVert^2/36 a 6 = ∥ r L ( 3 ) ( 0 ) ∥ 2 /36 . Suppose γ ( 0 ) ≥ γ ( − 1 ) \gamma(0)\ge\gamma(-1) γ ( 0 ) ≥ γ ( − 1 ) . The bracket has nonnegative coefficients, not all zero: otherwise the polynomial curve would lie on the circle identically and hence be constant, contrary to the distinct normals. Consequently b L ( t ) > 0 b_L(t)>0 b L ( t ) > 0 and d d t ( b L ( t ) / t 6 ) < 0 \frac{\mathrm{d}}{\mathrm{d}t}(b_L(t)/t^6)<0 d t d ( b L ( t ) / t 6 ) < 0 for t > 0 t>0 t > 0 . For the latter assertion, write the quotient as ( 1 + t − 1 ) 2 [ a 6 + ( a 6 + γ ( 0 ) − γ ( − 1 ) ) t − 1 + γ ( 0 ) t − 2 ] (1+t^{-1})^2[a_6+(a_6+\gamma(0)-\gamma(-1))t^{-1}+\gamma(0)t^{-2}] ( 1 + t − 1 ) 2 [ a 6 + ( a 6 + γ ( 0 ) − γ ( − 1 )) t − 1 + γ ( 0 ) t − 2 ] . If m = 0 m=0 m = 0 , this already precludes the third contact. For m > 0 m>0 m > 0 , use r R = r L + m t 3 ( 1 , 0 ) / 6 r_R=r_L+mt^3(1,0)/6 r R = r L + m t 3 ( 1 , 0 ) /6 to obtain d d t b R − b L t 6 = m 3 ( − 3 t 4 − α n 2 t 2 ) \frac{\mathrm{d}}{\mathrm{d}t}\frac{b_R-b_L}{t^6}=\frac{m}{3}\left(-\frac{3}{t^4}-\frac{\alpha_n}{2t^2}\right) d t d t 6 b R − b L = 3 m ( − t 4 3 − 2 t 2 α n ) . The tangent contact at h h h makes d d t ( b R ( t ) / t 6 ) ∣ t = h = 0 \left.\frac{\mathrm{d}}{\mathrm{d}t}(b_R(t)/t^6)\right|_{t=h}=0 d t d ( b R ( t ) / t 6 ) t = h = 0 ; therefore − α n h 2 > 6 -\alpha_nh^2>6 − α n h 2 > 6 . Hermite interpolation at the right contact gives α n h 2 = 6 ( D − 1 ) + 2 h V w 3 \alpha_nh^2=6(D-1)+2hVw_3 α n h 2 = 6 ( D − 1 ) + 2 hV w 3 , hence D < 0 D<0 D < 0 . Thus − 1 < D < 0 -1<D<0 − 1 < D < 0 and D 2 + V 2 = 1 D^2+V^2=1 D 2 + V 2 = 1 . The total angular span gives 0 < C < 1 0<C<1 0 < C < 1 ; write A = 1 − C > 0 A=1-C>0 A = 1 − C > 0 .
Negative middle tangential acceleration. Set u = h w 2 u=hw_2 u = h w 2 and v = h w 3 v=hw_3 v = h w 3 , both positive. Since γ ( 0 ) = w 2 2 + α n ≥ 0 \gamma(0)=w_2^2+\alpha_n\ge0 γ ( 0 ) = w 2 2 + α n ≥ 0 , − α n h 2 > 6 -\alpha_nh^2>6 − α n h 2 > 6 implies u 2 > 6 u^2>6 u 2 > 6 . The right Hermite cubic also gives r ¨ 2 ( 0 ) h 2 = 6 V − 4 u − 2 D v \ddot r_2(0)h^2=6V-4u-2Dv r ¨ 2 ( 0 ) h 2 = 6 V − 4 u − 2 D v , h 3 6 ( n 3 × r R ( 3 ) ) = v + D u − 2 V < 0 \frac{h^3}{6}(n_3\times r_R^{(3)})=v+Du-2V<0 6 h 3 ( n 3 × r R ( 3 ) ) = v + D u − 2 V < 0 . If r ¨ 2 ( 0 ) ≥ 0 \ddot r_2(0)\ge0 r ¨ 2 ( 0 ) ≥ 0 , these two relations imply ( 2 − D 2 ) u < V ( 3 − 2 D ) (2-D^2)u<V(3-2D) ( 2 − D 2 ) u < V ( 3 − 2 D ) . But 2 − D 2 = 1 + V 2 2-D^2=1+V^2 2 − D 2 = 1 + V 2 and − 2 D V ≤ D 2 + V 2 = 1 -2DV\le D^2+V^2=1 − 2 D V ≤ D 2 + V 2 = 1 , so u < 3 V + 1 1 + V 2 < 9 4 < 6 u<\frac{3V+1}{1+V^2}<\frac{9}{4}<\sqrt{6} u < 1 + V 2 3 V + 1 < 4 9 < 6 , where the middle strict inequality follows from 9 ( 1 + V 2 ) − 4 ( 3 V + 1 ) = ( 3 V − 2 ) 2 + 1 > 0 9(1+V^2)-4(3V+1)=(3V-2)^2+1>0 9 ( 1 + V 2 ) − 4 ( 3 V + 1 ) = ( 3 V − 2 ) 2 + 1 > 0 . This contradicts u 2 > 6 u^2>6 u 2 > 6 , proving r ¨ 2 ( 0 ) < 0 \ddot r_2(0)<0 r ¨ 2 ( 0 ) < 0 .
Ordering the first two speeds. The left Hermite cubic gives r ¨ 2 ( 0 ) = − 6 S + 2 C w 1 + 4 w 2 \ddot r_2(0)=-6S+2Cw_1+4w_2 r ¨ 2 ( 0 ) = − 6 S + 2 C w 1 + 4 w 2 and γ ( − 1 ) = w 1 2 + 2 S w 2 − 6 A \gamma(-1)=w_1^2+2Sw_2-6A γ ( − 1 ) = w 1 2 + 2 S w 2 − 6 A . Suppose w 2 ≥ w 1 w_2\ge w_1 w 2 ≥ w 1 . Then r ¨ 2 ( 0 ) < 0 \ddot r_2(0)<0 r ¨ 2 ( 0 ) < 0 implies 0 < w 1 < W : = 3 S / ( C + 2 ) 0<w_1<W:=3S/(C+2) 0 < w 1 < W := 3 S / ( C + 2 ) . Define Q ( z ) = z 2 − C S z − 3 A 2 Q(z)=z^2-CSz-3A^2 Q ( z ) = z 2 − C S z − 3 A 2 . The identity γ ( − 1 ) = Q ( w 1 ) + S r ¨ 2 ( 0 ) / 2 \gamma(-1)=Q(w_1)+S\ddot r_2(0)/2 γ ( − 1 ) = Q ( w 1 ) + S r ¨ 2 ( 0 ) /2 gives γ ( − 1 ) < Q ( w 1 ) \gamma(-1)<Q(w_1) γ ( − 1 ) < Q ( w 1 ) . This convex quadratic is negative throughout [ 0 , W ] [0,W] [ 0 , W ] , since Q ( 0 ) = − 3 A 2 < 0 Q(0)=-3A^2<0 Q ( 0 ) = − 3 A 2 < 0 and Q ( W ) = − 3 A 2 ( C + 2 ) 2 < 0 Q(W)=-\frac{3A^2}{(C+2)^2}<0 Q ( W ) = − ( C + 2 ) 2 3 A 2 < 0 . This contradicts γ ( − 1 ) ≥ 0 \gamma(-1)\ge0 γ ( − 1 ) ≥ 0 , so w 2 < w 1 w_2<w_1 w 2 < w 1 .
The common transverse cubic. We have γ ( 0 ) − γ ( − 1 ) = ( w 2 − w 1 ) ( w 1 + w 2 − 2 S ) \gamma(0)-\gamma(-1)=(w_2-w_1)(w_1+w_2-2S) γ ( 0 ) − γ ( − 1 ) = ( w 2 − w 1 ) ( w 1 + w 2 − 2 S ) . Our assumed γ ( 0 ) ≥ γ ( − 1 ) \gamma(0)\ge\gamma(-1) γ ( 0 ) ≥ γ ( − 1 ) therefore gives w 1 + w 2 ≤ 2 S w_1+w_2\le2S w 1 + w 2 ≤ 2 S and w 2 < S w_2<S w 2 < S . The radial jump leaves the transverse jerk unchanged; write r L ( 3 ) ( 0 ) = r R ( 3 ) ( 0 ) = 6 ( C w 1 + w 2 − 2 S ) r_L^{(3)}(0)=r_R^{(3)}(0)=6(Cw_1+w_2-2S) r L ( 3 ) ( 0 ) = r R ( 3 ) ( 0 ) = 6 ( C w 1 + w 2 − 2 S ) . Using α n = 2 S w 1 − 6 A \alpha_n=2Sw_1-6A α n = 2 S w 1 − 6 A , we obtain r L ( 3 ) ( 0 ) − α n w 2 = 6 C ( w 1 + w 2 − 2 S ) + 12 A ( w 2 − S ) − 2 S w 1 w 2 < 0 r_L^{(3)}(0)-\alpha_nw_2=6C(w_1+w_2-2S)+12A(w_2-S)-2Sw_1w_2<0 r L ( 3 ) ( 0 ) − α n w 2 = 6 C ( w 1 + w 2 − 2 S ) + 12 A ( w 2 − S ) − 2 S w 1 w 2 < 0 . Together with r ¨ 2 ( 0 ) < 0 \ddot r_2(0)<0 r ¨ 2 ( 0 ) < 0 and − α n h 2 > 6 -\alpha_nh^2>6 − α n h 2 > 6 , this forces r 2 ( h ) h = w 2 + r ¨ 2 ( 0 ) h 2 + r L , 2 ( 3 ) ( 0 ) h 2 6 < w 2 ( 1 + α n h 2 6 ) < 0 \frac{r_2(h)}{h}=w_2+\frac{\ddot r_2(0)h}{2}+\frac{r_{L,2}^{(3)}(0)h^2}{6}<w_2\left(1+\frac{\alpha_nh^2}{6}\right)<0 h r 2 ( h ) = w 2 + 2 r ¨ 2 ( 0 ) h + 6 r L , 2 ( 3 ) ( 0 ) h 2 < w 2 ( 1 + 6 α n h 2 ) < 0 , contrary to r 2 ( h ) = V > 0 r_2(h)=V>0 r 2 ( h ) = V > 0 . Thus γ ( 0 ) < γ ( − 1 ) \gamma(0)<\gamma(-1) γ ( 0 ) < γ ( − 1 ) .
Time reversal followed by a reflection F F F restores counterclockwise order. The transformed jets are − F r ˙ -F\dot r − F r ˙ , F r ¨ F\ddot r F r ¨ , and − F r ( 3 ) -Fr^{(3)} − F r ( 3 ) . Curvatures and positive oriented speeds are preserved, the middle jump is m F n 2 mFn_2 m F n 2 , and ( F n ) × ( − F r ( 3 ) ) = n × r ( 3 ) (Fn)\times(-Fr^{(3)})=n\times r^{(3)} ( F n ) × ( − F r ( 3 ) ) = n × r ( 3 ) . The same argument therefore proves γ ( 0 ) < γ ( h ) \gamma(0)<\gamma(h) γ ( 0 ) < γ ( h ) .
Angular dynamics on an arc. For motion in an oriented plane, Γ : = r ˙ × r ¨ − r × r ( 3 ) \Gamma:=\dot r\times\ddot r-r\times r^{(3)} Γ := r ˙ × r ¨ − r × r ( 3 ) is the corresponding component of Γ \Gamma Γ in Proposition A.5 . On an arc write r = ( cos θ , sin θ ) r=(\cos\theta,\sin\theta) r = ( cos θ , sin θ ) and w = θ ˙ w=\dot\theta w = θ ˙ .
Proposition A.10 (Angular dynamics on a boundary arc). On a boundary arc [ a , b ] [a,b] [ a , b ] , write r = ( cos θ , sin θ ) r=(\cos\theta,\sin\theta) r = ( cos θ , sin θ ) with continuous θ \theta θ and w = θ ˙ w=\dot\theta w = θ ˙ . Then w ¨ = 2 w 3 − Γ \ddot w=2w^3-\Gamma w ¨ = 2 w 3 − Γ on ( a , b ) (a,b) ( a , b ) , with first integral ( w ˙ ) 2 = w 4 − 2 Γ w + 2 E (\dot w)^2=w^4-2\Gamma w+2E ( w ˙ ) 2 = w 4 − 2Γ w + 2 E .
Proof of Proposition A.10. The continuous arc density in Lemma A.3 gives continuous fourth derivatives of r r r and its angle in the arc interior. With n = r n=r n = r and e θ = ( − sin θ , cos θ ) e_\theta=(-\sin\theta,\cos\theta) e θ = ( − sin θ , cos θ ) , differentiation gives r ˙ = w e θ \dot r=we_\theta r ˙ = w e θ , r ¨ = w ˙ e θ − w 2 n \ddot r=\dot we_\theta-w^2n r ¨ = w ˙ e θ − w 2 n , and r ( 3 ) = ( w ¨ − w 3 ) e θ − 3 w w ˙ n r^{(3)}=(\ddot w-w^3)e_\theta-3w\dot wn r ( 3 ) = ( w ¨ − w 3 ) e θ − 3 w w ˙ n . Substitution into the invariants of Proposition A.5 gives Γ = 2 w 3 − w ¨ \Gamma= 2w^3-\ddot{w} Γ = 2 w 3 − w ¨ and E = 1 2 ( ( w ˙ ) 2 − w 4 ) + Γ w E = \frac{1}{2}\left((\dot{w})^2-w^4\right)+\Gamma w E = 2 1 ( ( w ˙ ) 2 − w 4 ) + Γ w , proving both identities.
With α = w ˙ \alpha=\dot{w} α = w ˙ , Lemma A.3 and Proposition A.10 give Λ = w 4 − 3 α 2 − 4 w w ¨ = 4 Γ w − 7 w 4 − 3 α 2 = 10 Γ w − 10 w 4 − 6 E \Lambda=w^4-3\alpha^2-4w\ddot{w}=4\Gamma w-7w^4-3\alpha^2=10\Gamma w-10w^4-6E Λ = w 4 − 3 α 2 − 4 w w ¨ = 4Γ w − 7 w 4 − 3 α 2 = 10Γ w − 10 w 4 − 6 E (A.3).
Proposition A.11 (Geometric contact bounds). For the selected minimizer, either A ∩ ( 0 , 1 ) A\cap(0,1) A ∩ ( 0 , 1 ) has at most three elements, or there exist 0 < a < b < 1 0<a<b<1 0 < a < b < 1 and sets S − ⊆ ( 0 , a ) S_-\subseteq(0,a) S − ⊆ ( 0 , a ) , S + ⊆ ( b , 1 ) S_+\subseteq(b,1) S + ⊆ ( b , 1 ) , each with at most one element, such that A ∩ ( 0 , 1 ) = S − ∪ [ a , b ] ∪ S + A\cap(0,1)=S_-\cup[a,b]\cup S_+ A ∩ ( 0 , 1 ) = S − ∪ [ a , b ] ∪ S + .
Proof. For unblocked data the free cubic has at most one interior geometric contact. Otherwise Propositions A.6 and 2.4 give finitely many contact components between free endpoint pieces, ordered by Lemma A.8 . Write b ( t ) = ∥ r ( t ) ∥ 2 − 1 = − g ( t ) b(t)=\lVert r(t)\rVert^2-1=-g(t) b ( t ) = ∥ r ( t ) ∥ 2 − 1 = − g ( t ) . At geometric contacts, γ ( t ) = b ¨ ( t ) / 2 ≥ 0 \gamma(t)=\ddot{b}(t)/2\geq0 γ ( t ) = b ¨ ( t ) /2 ≥ 0 ; at arc endpoints, γ ( t ) = 0 \gamma(t)=0 γ ( t ) = 0 .
At most one arc. For a free bridge [ a 0 , b 0 ] [a_0,b_0] [ a 0 , b 0 ] of duration ℓ \ell ℓ between tangent unit contacts, set A 0 = 2 γ ( a 0 ) A_0=2\gamma(a_0) A 0 = 2 γ ( a 0 ) , A 1 = ℓ 2 γ ( b 0 ) A_1=\ell^2\gamma(b_0) A 1 = ℓ 2 γ ( b 0 ) , and κ = ℓ 6 ∥ r ( 3 ) ( a 0 + ) ∥ 2 / 36 \kappa=\ell^6\lVert r^{(3)}(a_0^+)\rVert^2/36 κ = ℓ 6 ∥ r ( 3 ) ( a 0 + ) ∥ 2 /36 . The double endpoint zeros factor its gap as
b ( a 0 + ℓ u ) = u 2 ( 1 − u ) 2 [ A 0 ( 1 − u ) + A 1 u − κ u ( 1 − u ) ] (5) b(a_0+\ell u)=u^2(1-u)^2\left[A_0(1-u)+A_1u-\kappa u(1-u)\right] \tag*{(5)} b ( a 0 + ℓ u ) = u 2 ( 1 − u ) 2 [ A 0 ( 1 − u ) + A 1 u − κ u ( 1 − u ) ] (5)
Differentiating gives b ( 3 ) ( b 0 − ) − b ( 3 ) ( a 0 + ) = 12 ( γ ( a 0 ) + γ ( b 0 ) ) / ℓ + ℓ 3 ∥ r ( 3 ) ( a 0 + ) ∥ 2 / 3 > 0 b^{(3)}(b_0^-)-b^{(3)}(a_0^+)=12(\gamma(a_0)+\gamma(b_0))/\ell+\ell^3\lVert r^{(3)}(a_0^+)\rVert^2/3>0 b ( 3 ) ( b 0 − ) − b ( 3 ) ( a 0 + ) = 12 ( γ ( a 0 ) + γ ( b 0 )) / ℓ + ℓ 3 ∥ r ( 3 ) ( a 0 + ) ∥ 2 /3 > 0 . If the right side vanished, the preceding factorization would force the polynomial curve to lie on the circle and hence be constant, contrary to the distinct ordered normals. At every intervening atom, b ( 3 ) b^{(3)} b ( 3 ) jumps by 2 m i ≥ 0 2m_i\geq0 2 m i ≥ 0 , whereas it vanishes on an arc. Thus no finite chain of free bridges and isolated contacts connects two arcs.
Counting geometric contacts. By Lemma A.9 , four consecutive isolated interior contacts at t 1 < t 2 < t 3 < t 4 t_1<t_2<t_3<t_4 t 1 < t 2 < t 3 < t 4 would give both γ ( t 2 ) < γ ( t 3 ) \gamma(t_2)<\gamma(t_3) γ ( t 2 ) < γ ( t 3 ) and γ ( t 3 ) < γ ( t 2 ) \gamma(t_3)<\gamma(t_2) γ ( t 3 ) < γ ( t 2 ) . Two isolated contacts at t 1 < t 2 t_1<t_2 t 1 < t 2 immediately before an arc with entry a a a would give γ ( t 2 ) < γ ( a ) = 0 \gamma(t_2)<\gamma(a)=0 γ ( t 2 ) < γ ( a ) = 0 ; time reversal excludes two after an arc. The contact laws and Lemma A.8 supply the lemma’s hypotheses, also at zero-mass contacts.
Proposition A.12 (Arc signs). In the arc case of Proposition A.11 , the angular speed w : = r × r ˙ w:=r\times\dot{r} w := r × r ˙ satisfies w ( t ) > 0 w(t)>0 w ( t ) > 0 for t ∈ [ a , b ] t\in[a,b] t ∈ [ a , b ] . Moreover, Λ ( t ) > 0 \Lambda(t)>0 Λ ( t ) > 0 for t ∈ ( a , b ) ∗ a n d ∗ t\in(a,b) *and* t ∈ ( a , b ) ∗ an d ∗ \mu({t})>0$ for every t ∈ S − ∪ { a , b } ∪ S + t\in S_-\cup\{a,b\}\cup S_+ t ∈ S − ∪ { a , b } ∪ S + .
Proof. Write the arc as [ t e n t r y , t e x i t ] [t_{\mathrm{entry}},t_{\mathrm{exit}}] [ t entry , t exit ] and set b ( t ) = ∥ r ( t ) ∥ 2 − 1 b(t)=\lVert r(t)\rVert^2-1 b ( t ) = ∥ r ( t ) ∥ 2 − 1 .
Arc speed and density. Integrating the energy invariant by parts with zero initial and final velocities gives E = 3 2 ∫ 0 1 ∥ r ¨ ( t ) ∥ 2 d t > 0 E=\frac{3}{2}\int_0^1\lVert\ddot{r}(t)\rVert^2\,\mathrm{d}t>0 E = 2 3 ∫ 0 1 ∥ r ¨ ( t ) ∥ 2 d t > 0 . By Lemma A.8 , w > 0 w>0 w > 0 on the closed arc. Use α = w ˙ \alpha=\dot{w} α = w ˙ , Proposition A.10, and (A.3), with their continuous arc-side endpoint limits.
The arc equations are smooth, so an interior zero of Λ ≥ 0 \Lambda\geq0 Λ ≥ 0 would satisfy Λ ˙ = 10 ( Γ − 4 w 3 ) α = 0 \dot{\Lambda}=10(\Gamma-4w^3)\alpha=0 Λ ˙ = 10 ( Γ − 4 w 3 ) α = 0 and Λ ¨ ≥ 0 \ddot{\Lambda}\geq0 Λ ¨ ≥ 0 . If α = 0 \alpha=0 α = 0 , the zero-density equation gives Γ = 7 w 3 / 4 \Gamma=7w^3/4 Γ = 7 w 3 /4 and then Λ ¨ = − 45 w 6 / 8 < 0 \ddot{\Lambda}=-45w^6/8<0 Λ ¨ = − 45 w 6 /8 < 0 . Otherwise Γ = 4 w 3 \Gamma=4w^3 Γ = 4 w 3 , α 2 = 3 w 4 \alpha^2=3w^4 α 2 = 3 w 4 and Λ ¨ = − 360 w 6 < 0 \ddot{\Lambda}=-360w^6<0 Λ ¨ = − 360 w 6 < 0 . Thus the density is strictly positive in the arc interior. Its arc-side endpoint traces are nonnegative and need not be strictly positive.
Positive atoms at arc endpoints. Consider first an arc exit at time t e x i t t_{\mathrm{exit}} t exit followed by a geometric contact after a free bridge of duration ℓ \ell ℓ . Let m ≥ 0 m\geq0 m ≥ 0 be its exit atom mass and set b ^ ( u ) = b ( t e x i t + ℓ u ) \hat{b}(u)=b(t_{\mathrm{exit}}+\ell u) b ^ ( u ) = b ( t exit + ℓ u ) for 0 ≤ u ≤ 1 0 \le u \le1 0 ≤ u ≤ 1 . Since b = b ˙ = b ¨ = 0 b=\dot{b}=\ddot{b}=0 b = b ˙ = b ¨ = 0 on the arc and b ( 3 ) ( t e x i t + ) = 2 m b^{(3)}(t_{\mathrm{exit}}^+)=2m b ( 3 ) ( t exit + ) = 2 m , we have d 3 b ^ d u 3 ( 0 + ) = 2 ℓ 3 m \frac{\mathrm{d}^3\hat{b}}{\mathrm{d}u^3}(0^+)=2\ell^3m d u 3 d 3 b ^ ( 0 + ) = 2 ℓ 3 m .
The factorization (A.4) gives b ^ ( u ) = u 3 ( 1 − u ) 2 [ ℓ 3 m / 3 + u ℓ 6 ∥ r ( 3 ) ( t e x i t + ) ∥ 2 / 36 ] \hat{b}(u)=u^3(1-u)^2\left[\ell^3m/3+u\ell^6\lVert r^{(3)}(t_{\mathrm{exit}}^+)\rVert^2/36\right] b ^ ( u ) = u 3 ( 1 − u ) 2 [ ℓ 3 m /3 + u ℓ 6 ∥ r ( 3 ) ( t exit + ) ∥ 2 /36 ] (A.5). If m = 0 m=0 m = 0 , the fourth derivative in physical time is b ( 4 ) ( t e x i t + ) = 2 ℓ 2 ∥ r ( 3 ) ( t e x i t + ) ∥ 2 / 3 > 0 b^{(4)}(t_{\mathrm{exit}}^+)=2\ell^2\lVert r^{(3)}(t_{\mathrm{exit}}^+)\rVert^2/3>0 b ( 4 ) ( t exit + ) = 2 ℓ 2 ∥ r ( 3 ) ( t exit + ) ∥ 2 /3 > 0 . Here r ( 3 ) ( t e x i t + ) ≠ 0 r^{(3)}(t_{\mathrm{exit}}^+)\ne0 r ( 3 ) ( t exit + ) = 0 , since otherwise the factorization would again force a constant circle-valued polynomial. But the continuous clearance forcing in (A.1) gives this fourth derivative as f g ( t e x i t + ) = − 2 Λ ( t e x i t + ) ≤ 0 f_g(t_{\mathrm{exit}}^+)=-2\Lambda(t_{\mathrm{exit}}^+)\le0 f g ( t exit + ) = − 2Λ ( t exit + ) ≤ 0 , a contradiction. Time reversal and reflection prove the same assertion at entry.
It remains to treat a direct cubic completion with zero endpoint velocity, including an endpoint on the circle. Orient this cubic from that endpoint to arc entry at time t e n t r y t_{\mathrm{entry}} t entry , put its duration a 0 > 0 a_0>0 a 0 > 0 , and use the contact frame at entry n = ( 1 , 0 ) \boldsymbol{n}=(1,0) n = ( 1 , 0 ) , r ˙ = ( 0 , w ) \dot{\boldsymbol{r}}=(0,w) r ˙ = ( 0 , w ) , r ¨ = ( − w 2 , α ) \ddot{\boldsymbol{r}}=(-w^2,\alpha) r ¨ = ( − w 2 , α ) , with w > 0 w>0 w > 0 . Set a ^ 0 = a 0 w \hat{a}_0=a_0w a ^ 0 = a 0 w and β t a i l = a 0 α / w \beta_{\mathrm{tail}}=a_0\alpha/w β tail = a 0 α / w . Integrating the cubic back to zero velocity and matching tangential jerk gives Z 0 = ( 1 − a ^ 0 2 6 , a ^ 0 ( β t a i l − 4 ) 6 ) Z_0=\left(1-\frac{\hat{a}_0^2}{6},\frac{\hat{a}_0(\beta_{\mathrm{tail}}-4)}{6}\right) Z 0 = ( 1 − 6 a ^ 0 2 , 6 a ^ 0 ( β tail − 4 ) ) , m = w 2 a 0 2 ( 2 − 3 β t a i l ) m=\frac{w^2}{a_0^2}(2-3\beta_{\mathrm{tail}}) m = a 0 2 w 2 ( 2 − 3 β tail ) , Γ − w 3 = 2 w ( 1 − β t a i l ) a 0 2 \Gamma-w^3=\frac{2w(1-\beta_{\mathrm{tail}})}{a_0^2} Γ − w 3 = a 0 2 2 w ( 1 − β tail ) . For 0 ≤ u ≤ 1 0\le u\le1 0 ≤ u ≤ 1 , define the normalized gap b ^ 0 ( u ) = b ( t e n t r y − a 0 + a 0 u ) \hat{b}_0(u)=b(t_{\mathrm{entry}}-a_0+a_0u) b ^ 0 ( u ) = b ( t entry − a 0 + a 0 u ) . Then b ^ 0 ( u ) = a ^ 0 2 36 ( 1 − u ) 3 Q 0 ( u ) \hat{b}_0(u)=\frac{\hat{a}_0^2}{36}(1-u)^3Q_0(u) b ^ 0 ( u ) = 36 a ^ 0 2 ( 1 − u ) 3 Q 0 ( u ) . In the standard cubic Bernstein basis ( 3 i ) u i ( 1 − u ) 3 − i \binom{3}{i}u^i(1-u)^{3-i} ( i 3 ) u i ( 1 − u ) 3 − i , i = 0 , 1 , 2 , 3 i=0,1,2,3 i = 0 , 1 , 2 , 3 , the coefficients of Q 0 Q_0 Q 0 are B 0 B_0 B 0 , 2 B 0 2B_0 2 B 0 , 3 B 0 + 4 ( 1 − β t a i l ) 3B_0+4(1-\beta_{\mathrm{tail}}) 3 B 0 + 4 ( 1 − β tail ) , 12 ( 2 − 3 β t a i l ) 12(2-3\beta_{\mathrm{tail}}) 12 ( 2 − 3 β tail ) , B 0 = a ^ 0 2 + β t a i l 2 − 8 β t a i l + 4 = 36 ( ∥ Z 0 ∥ 2 − 1 ) a ^ 0 2 B_0=\hat{a}_0^2+\beta_{\mathrm{tail}}^2-8\beta_{\mathrm{tail}}+4=\frac{36(\lVert Z_0\rVert^2-1)}{\hat{a}_0^2} B 0 = a ^ 0 2 + β tail 2 − 8 β tail + 4 = a ^ 0 2 36 (∥ Z 0 ∥ 2 − 1 ) . The exterior endpoint and m ≥ 0 m\ge0 m ≥ 0 give B 0 ≥ 0 B_0\ge0 B 0 ≥ 0 and β t a i l ≤ 2 / 3 \beta_{\mathrm{tail}}\le2/3 β tail ≤ 2/3 . Hence b ^ 0 ( u ) > 0 \hat{b}_0(u)>0 b ^ 0 ( u ) > 0 for 0 < u < 1 0<u<1 0 < u < 1 , even when ∥ Z 0 ∥ = 1 \lVert Z_0\rVert=1 ∥ Z 0 ∥ = 1 . If m = 0 m=0 m = 0 , then β t a i l = 2 / 3 \beta_{\mathrm{tail}}=2/3 β tail = 2/3 and B 0 ≥ 0 B_0\ge0 B 0 ≥ 0 requires a ^ 0 2 ≥ 8 / 9 \hat{a}_0^2\ge8/9 a ^ 0 2 ≥ 8/9 . On the other hand, substituting the tangential matching relation in the arc density gives 0 ≤ Λ ( t e n t r y + ) = w 2 a 0 2 ( 4 3 − 3 a ^ 0 2 ) 0\le\Lambda(t_{\mathrm{entry}}^+)=\frac{w^2}{a_0^2}\left(\frac{4}{3}-3\hat{a}_0^2\right) 0 ≤ Λ ( t entry + ) = a 0 2 w 2 ( 3 4 − 3 a ^ 0 2 ) , so a ^ 0 2 ≤ 4 / 9 \hat{a}_0^2\le4/9 a ^ 0 2 ≤ 4/9 , a contradiction. Thus both arc endpoints carry positive atoms.
Isolated contacts beside the arc. There is at most one isolated geometric contact on each side of the arc. If its mass were zero, its neighboring cubics would have identical jets through jerk and would join into a single cubic from the endpoint with zero velocity to the arc. That cubic would have an interior touch, contrary to the strict clearance of b ^ 0 \hat{b}_0 b ^ 0 . Thus any such isolated contact has positive mass and there are no incidental zero-mass touches beside a persisting arc.
Proof of Theorem 2.8 . The geometric configurations of Proposition A.11 give F, P, PP, or PPP without an arc, and A, PA, AP, or PAP with one, using the contact-label convention of § 2.2.3 . Proposition A.12 excludes zero-mass contacts beside an arc. Multiplier uniqueness (Proposition A.6 ) fixes the label.
Proof of Theorem 2.9
Atomic construction and clearance. The free trajectory is feasible exactly when the endpoint segment avoids the open disk: for D = r T − r 0 ≠ 0 D=r_{\mathrm{T}}-r_0\ne0 D = r T − r 0 = 0 , test ∥ r 0 + s ∗ D ∥ ≥ 1 \lVert r_0+s_*D\rVert\ge1 ∥ r 0 + s ∗ D ∥ ≥ 1 , where s ∗ = min { 1 , max { 0 , − ⟨ r 0 , D ⟩ / ∥ D ∥ 2 } } s_*=\min\{1,\max\{0,-\langle r_0,D\rangle/\lVert D\rVert^2\}\} s ∗ = min { 1 , max { 0 , − ⟨ r 0 , D ⟩ / ∥ D ∥ 2 }} . For D = 0 D=0 D = 0 the test succeeds.
Atomic representation. For blocked data write G = G 1 G=G_1 G = G 1 . The atomic contact equations in § 2.2.3 imply Δ i = m i ν i \Delta_i=m_i\nu_i Δ i = m i ν i , so four distributional derivatives of the Green representation give D t 4 r = r ∑ i m i δ t i D_t^4r=r\sum_i m_i\delta_{t_i} D t 4 r = r ∑ i m i δ t i . The kernel enforces the endpoints and continuity through acceleration. Admissibility supplies the hypotheses of Theorem 2.5 , and Proposition 2.7 identifies the selected minimizer. Conversely, (2.4) represents every purely atomic selected minimizer, and Proposition A.6 determines its positive jumps and times uniquely.
On each time segment the Green branches are polynomials. Clearance can be imposed without introducing a root selector as ∀ u ( t i ≤ u ≤ t i + 1 ⇒ ∥ r i ( u ) ∥ 2 ≥ 1 ) \forall u\ \left(t_i \le u \le t_{i+1}\Rightarrow\lVert r_i(u)\rVert^2\ge1\right) ∀ u ( t i ≤ u ≤ t i + 1 ⇒ ∥ r i ( u ) ∥ 2 ≥ 1 ) , t 0 = 0 t_0=0 t 0 = 0 , t m + 1 = 1 t_{m+1}=1 t m + 1 = 1 , where r i r_i r i is the corresponding cubic polynomial. Impose ⟨ e , n i ⟩ > 0 \langle e,n_i\rangle>0 ⟨ e , n i ⟩ > 0 at every atom. Multiply this projection inequality by the positive ∥ r T − r 0 ∥ \lVert r_T-r_0\rVert ∥ r T − r 0 ∥ to use the polynomial normal ( r T , 2 , d 0 − r T , 1 ) (r_{T,2},d_0-r_{T,1}) ( r T , 2 , d 0 − r T , 1 ) . On the ordered time simplex, the equations and inequalities thus provide a real-algebraic specification in the 3 m 3m 3 m unknowns. Retain all real components, with distinct times and positive masses as required by the contact-label convention in § 2.2.3 .
Polynomial clearance tests. For a contact-to-contact bridge, test the quadratic factor in (A.4). Its minimum occurs at u ∗ = min { 1 , max { 0 , ( A 0 + κ − A 1 ) / ( 2 κ ) } } u_*=\min\{1,\max\{0,(A_0+\kappa-A_1)/(2\kappa)\}\} u ∗ = min { 1 , max { 0 , ( A 0 + κ − A 1 ) / ( 2 κ )}} when κ > 0 \kappa>0 κ > 0 ; for κ = 0 \kappa=0 κ = 0 , evaluate both endpoints. Outer pieces require only the endpoints and the real roots of a cubic derivative after their forced double zero is removed. A constant quotient is checked at one point. At a terminal contact with zero velocity a further double factor is available only on that endpoint stratum. These are exact tests of nonnegativity, retaining zero-mass touches.
Arc system and completion. In unit-duration time, (A.1) gives the Cartesian arc equation r ( 4 ) ( t ) = λ ( r ( t ) , r ˙ ( t ) , r ¨ ( t ) , r ( 3 ) ( t ) ) r ( t ) r^{(4)}(t)=\lambda(r(t),\dot{r}(t),\ddot{r}(t),r^{(3)}(t))r(t) r ( 4 ) ( t ) = λ ( r ( t ) , r ˙ ( t ) , r ¨ ( t ) , r ( 3 ) ( t )) r ( t ) , where λ ( z 0 , z 1 , z 2 , z 3 ) : = − 3 ∥ z 2 ∥ 2 − 4 ⟨ z 1 , z 3 ⟩ \lambda(z_0,z_1,z_2,z_3):=-3\lVert z_2\rVert^2-4\langle z_1,z_3\rangle λ ( z 0 , z 1 , z 2 , z 3 ) := − 3 ∥ z 2 ∥ 2 − 4 ⟨ z 1 , z 3 ⟩ . Here Λ ( t ) = λ ( r ( t ) , r ˙ ( t ) , r ¨ ( t ) , r ( 3 ) ( t ) ) \Lambda(t)=\lambda(r(t),\dot{r}(t),\ddot{r}(t),r^{(3)}(t)) Λ ( t ) = λ ( r ( t ) , r ˙ ( t ) , r ¨ ( t ) , r ( 3 ) ( t )) on the arc.
On an arc [ a , b ] [a,b] [ a , b ] , the entry speed w a = θ ˙ ( a ) > 0 w_a=\dot{\theta}(a)>0 w a = θ ˙ ( a ) > 0 (Proposition A.12) defines s : = w a ( t − a ) s:=w_a(t-a) s := w a ( t − a ) and θ ^ ( s ) : = θ ( a + s / w a ) − θ ( a ) \hat{\theta}(s):=\theta(a+s/w_a)-\theta(a) θ ^ ( s ) := θ ( a + s / w a ) − θ ( a ) . Dots on hatted quantities denote derivatives with respect to s s s . Set w ^ : = θ ^ ˙ \hat{w}:=\dot{\hat{\theta}} w ^ := θ ^ ˙ , α ^ : = w ^ ˙ \hat{\alpha}:=\dot{\hat{w}} α ^ := w ^ ˙ , Γ ^ : = Γ / w a 3 \hat{\Gamma}:=\Gamma/w_a^3 Γ ^ := Γ/ w a 3 , E ^ : = E / w a 4 \hat{E}:=E/w_a^4 E ^ := E / w a 4 , and T ^ A : = w a ( b − a ) \hat{T}_A:=w_a(b-a) T ^ A := w a ( b − a ) . The normalized equations are θ ^ ˙ = w ^ \dot{\hat{\theta}}=\hat{w} θ ^ ˙ = w ^ , w ^ ˙ = α ^ \dot{\hat{w}}=\hat{\alpha} w ^ ˙ = α ^ , and α ^ ˙ = 2 w ^ 3 − Γ ^ \dot{\hat{\alpha}}=2\hat{w}^3-\hat{\Gamma} α ^ ˙ = 2 w ^ 3 − Γ ^ on [ 0 , T ^ A ] [0,\hat{T}_A] [ 0 , T ^ A ] . For a choice of ( α ^ 0 , Γ ^ , T ^ A ) (\hat{\alpha}_0,\hat{\Gamma},\hat{T}_A) ( α ^ 0 , Γ ^ , T ^ A ) with T ^ A > 0 \hat{T}_A>0 T ^ A > 0 , impose θ ^ ( 0 ) = 0 \hat{\theta}(0)=0 θ ^ ( 0 ) = 0 , w ^ ( 0 ) = 1 \hat{w}(0)=1 w ^ ( 0 ) = 1 , and α ^ ( 0 ) = α ^ 0 \hat{\alpha}(0)=\hat{\alpha}_0 α ^ ( 0 ) = α ^ 0 ; then E ^ = 1 2 ( α ^ 0 2 − 1 + 2 Γ ^ ) \hat{E}=\frac{1}{2}(\hat{\alpha}_0^2-1+2\hat{\Gamma}) E ^ = 2 1 ( α ^ 0 2 − 1 + 2 Γ ^ ) and the arc is r ^ ( s ) = ( cos θ ^ ( s ) , sin θ ^ ( s ) ) \hat{r}(s)=(\cos\hat{\theta}(s),\sin\hat{\theta}(s)) r ^ ( s ) = ( cos θ ^ ( s ) , sin θ ^ ( s )) .
On intervals where the angular speed is strictly monotone, the identity in Proposition A.10 gives elapsed time by integrating with respect to angular speed. The raw density formula (A.3) applies to Λ ^ \hat{\Lambda} Λ ^ with w ^ = w ^ \hat{w}=\hat{w} w ^ = w ^ , α = α ^ \alpha=\hat{\alpha} α = α ^ , Γ = Γ ^ \Gamma=\hat{\Gamma} Γ = Γ ^ , and E ^ = E ^ \hat{E}=\hat{E} E ^ = E ^ . Require w ^ > 0 \hat{w}>0 w ^ > 0 and Λ ^ ≥ 0 \hat{\Lambda}\ge0 Λ ^ ≥ 0 throughout the arc.
Completion relation.
By Proposition A.11, each end of an arc reaches its endpoint with zero velocity either directly or through one isolated contact. Both completions are determined by the jet at the arc end, up to finitely many roots. For an arc [ a , b ] [a,b] [ a , b ] , use local outward time s s s , with s = 0 s=0 s = 0 at the chosen endpoint and s < 0 s<0 s < 0 on the arc. Write the local trajectory as r ( s ) r(s) r ( s ) , with dots denoting derivatives in s s s . The density traces Λ ( a + ) \Lambda(a^+) Λ ( a + ) and Λ ( b − ) \Lambda(b^-) Λ ( b − ) retain the original arc coordinate. Choose the frame so that r ( 0 ) = n = ( 1 , 0 ) r(0)=n=(1,0) r ( 0 ) = n = ( 1 , 0 ) , r ˙ ( 0 ) = ( 0 , w ) \dot{r}(0)=(0,w) r ˙ ( 0 ) = ( 0 , w ) , r ¨ ( 0 ) = ( − w 2 , α ) \ddot{r}(0)=(-w^2,\alpha) r ¨ ( 0 ) = ( − w 2 , α ) , r A ( 3 ) = ( − 3 w α , τ A ) r_A^{(3)}=(-3w\alpha,\tau_A) r A ( 3 ) = ( − 3 w α , τ A ) , τ A : = w 3 − Γ < 0 \tau_A:=w^3-\Gamma<0 τ A := w 3 − Γ < 0 . Here r A ( 3 ) = r ( 3 ) ( 0 − ) r_A^{(3)}=r^{(3)}(0^-) r A ( 3 ) = r ( 3 ) ( 0 − ) is the arc-side jerk. At entry, Λ ( a + ) = − 4 τ A w − 3 w 4 − 3 α 2 \Lambda(a^+)=-4\tau_A w-3w^4-3\alpha^2 Λ ( a + ) = − 4 τ A w − 3 w 4 − 3 α 2 ; at exit, the same expression equals Λ ( b − ) \Lambda(b^-) Λ ( b − ) . Nonnegative arc-side density and w > 0 w>0 w > 0 imply τ A < 0 \tau_A<0 τ A < 0 . For the normalized arc, set Γ = Γ ^ \Gamma=\hat{\Gamma} Γ = Γ ^ . At entry use w = 1 w=1 w = 1 , α = − α ^ 0 \alpha=-\hat{\alpha}_0 α = − α ^ 0 and reflect the resulting positions in the local horizontal axis; at exit use w = w ^ ( T ^ A ) w=\hat{w}(\hat{T}_A) w = w ^ ( T ^ A ) , α = α ^ ( T ^ A ) \alpha=\hat{\alpha}(\hat{T}_A) α = α ^ ( T ^ A ) and rotate by θ ^ ( T ^ A ) \hat{\theta}(\hat{T}_A) θ ^ ( T ^ A ) . The invariant Γ \Gamma Γ is unchanged by the combined time reversal and reflection.
The stopping relation. Every completion ends with a free cubic from its last contact to the endpoint with zero velocity. At the last contact, whose local time is t c t_c t c , write p = r ( t c ) p=r(t_c) p = r ( t c ) , r ˙ c = r ˙ ( t c ) \dot{r}_c=\dot{r}(t_c) r ˙ c = r ˙ ( t c ) , r ¨ c = r ¨ ( t c ) \ddot{r}_c=\ddot{r}(t_c) r ¨ c = r ¨ ( t c ) , and r − ( 3 ) = r ( 3 ) ( t c − ) r_-^{(3)}=r^{(3)}(t_c^-) r − ( 3 ) = r ( 3 ) ( t c − ) , r + ( 3 ) = r ( 3 ) ( t c + ) r_+^{(3)}=r^{(3)}(t_c^+) r + ( 3 ) = r ( 3 ) ( t c + ) . Set w c = p × r ˙ c > 0 w_c=p\times\dot{r}_c>0 w c = p × r ˙ c > 0 , α c = p × r ¨ c \alpha_c=p\times\ddot{r}_c α c = p × r ¨ c , and τ c = p × r − ( 3 ) < 0 \tau_c=p\times r_-^{(3)}<0 τ c = p × r − ( 3 ) < 0 . The final cubic’s duration, endpoint, and outgoing jerk are T t a i l = 2 w c / α c 2 − 2 τ c w c − α c T_{\mathrm{tail}}=2w_c/\sqrt{\alpha_c^2-2\tau_cw_c-\alpha_c} T tail = 2 w c / α c 2 − 2 τ c w c − α c , Z = p + 2 T t a i l 3 r ˙ c + T t a i l 2 6 r ¨ c Z=p+\frac{2T_{\mathrm{tail}}}{3}\dot{r}_c+\frac{T_{\mathrm{tail}}^2}{6}\ddot{r}_c Z = p + 3 2 T tail r ˙ c + 6 T tail 2 r ¨ c , and r + ( 3 ) = − 2 r ˙ c / T t a i l − 2 r ˙ c / T t a i l 2 r_+^{(3)}=-2\dot{r}_c/T_{\mathrm{tail}}-2\dot{r}_c/T_{\mathrm{tail}}^2 r + ( 3 ) = − 2 r ˙ c / T tail − 2 r ˙ c / T tail 2 (A.6). The outgoing cubic reaches zero velocity after time T t a i l T_{\mathrm{tail}} T tail precisely when its jerk has this value. Tangential jerk continuity gives τ c T t a i l 2 + 2 α c T t a i l + 2 w c = 0 \tau_c T_{\mathrm{tail}}^2 + 2\alpha_c T_{\mathrm{tail}} + 2w_c = 0 τ c T tail 2 + 2 α c T tail + 2 w c = 0 , whose unique positive root is (A.6). Integration gives Z Z Z . The jump [ r ( 3 ) ] t c [\mathbf{r}^{(3)}]_{t_c} [ r ( 3 ) ] t c is radial, with positive atom mass ⟨ p , [ r ( 3 ) ] t c ⟩ \langle p,[\mathbf{r}^{(3)}]_{t_c}\rangle ⟨ p , [ r ( 3 ) ] t c ⟩ . For direct completion from the arc to the endpoint with zero velocity, p = n p=n p = n and r ( 3 ) = r A ( 3 ) = r ( 3 ) \mathbf{r}^{(3)}=\mathbf{r}^{(3)}_A=\mathbf{r}^{(3)} r ( 3 ) = r A ( 3 ) = r ( 3 ) , giving m = ⟨ n , [ r ( 3 ) ] t c ⟩ = 2 w 2 / T t a i l + 3 w α m=\langle n,[\mathbf{r}^{(3)}]_{t_c}\rangle=2w^2/T_{\mathrm{tail}}+3w\alpha m = ⟨ n , [ r ( 3 ) ] t c ⟩ = 2 w 2 / T tail + 3 w α .
An optional isolated contact. Before it stops, the trajectory may touch the disk once more. Launch r ( s ) = n + s r ˙ ( 0 ) + s 2 r ¨ ( 0 ) / 2 + s 3 r B ( 3 ) / 6 \mathbf{r}(s)=n+s\dot{\mathbf{r}}(0)+s^2\ddot{\mathbf{r}}(0)/2+s^3\mathbf{r}^{(3)}_B/6 r ( s ) = n + s r ˙ ( 0 ) + s 2 r ¨ ( 0 ) /2 + s 3 r B ( 3 ) /6 , where r B ( 3 ) = r ( 3 ) ( 0 + ) = r A ( 3 ) + m n \mathbf{r}^{(3)}_B=\mathbf{r}^{(3)}(0^+)=\mathbf{r}^{(3)}_A+mn r B ( 3 ) = r ( 3 ) ( 0 + ) = r A ( 3 ) + mn is the bridge jerk. For a contact at s = ℓ > 0 s=\ell>0 s = ℓ > 0 , set p 2 = w ℓ + α ℓ 2 / 2 + τ A ℓ 3 / 6 p_2=w\ell+\alpha\ell^2/2+\tau_A\ell^3/6 p 2 = w ℓ + α ℓ 2 /2 + τ A ℓ 3 /6 , A ℓ = 3 − w ℓ 2 / 2 A_\ell=3-w\ell^2/2 A ℓ = 3 − w ℓ 2 /2 , and B ℓ = 3 − ( 2 w ℓ + α ℓ 2 / 2 ) p 2 B_\ell=3-(2w\ell+\alpha\ell^2/2)p_2 B ℓ = 3 − ( 2 w ℓ + α ℓ 2 /2 ) p 2 , and define P c o n ( ℓ ) = 144 [ B ℓ 2 + A ℓ 2 ( p 2 2 − 1 ) ] / ℓ 4 P_{\mathrm{con}}(\ell)=144[B_\ell^2+A_\ell^2(p_2^2-1)]/\ell^4 P con ( ℓ ) = 144 [ B ℓ 2 + A ℓ 2 ( p 2 2 − 1 )] / ℓ 4 . The quotient extends to a sextic polynomial with leading coefficient τ A 2 ( w 4 + α 2 ) > 0 \tau_A^2(w^4+\alpha^2)>0 τ A 2 ( w 4 + α 2 ) > 0 . Indeed, if p = r ( ℓ ) = ( p 1 , p 2 ) p=\mathbf{r}(\ell)=(p_1,p_2) p = r ( ℓ ) = ( p 1 , p 2 ) , direct substitution gives ℓ ⟨ p , r ˙ ( ℓ ) ⟩ − 3 ( ∥ p ∥ 2 − 1 ) = B ℓ − A ℓ p 1 \ell\langle p,\dot{\mathbf{r}}(\ell)\rangle-3(\lVert p\rVert^2-1)=B_\ell-A_\ell p_1 ℓ ⟨ p , r ˙ ( ℓ )⟩ − 3 (∥ p ∥ 2 − 1 ) = B ℓ − A ℓ p 1 . Consequently unit position and tangency are equivalent to A ℓ p 1 = B ℓ A_\ell p_1=B_\ell A ℓ p 1 = B ℓ and p 1 2 + p 2 2 = 1 p_1^2+p_2^2=1 p 1 2 + p 2 2 = 1 . Ordered normals require p 2 > 0 p_2>0 p 2 > 0 , so retain every root of P c o n P_{\mathrm{con}} P con in the explicit bracket 0 < ℓ < 3 α + 9 α 2 − 24 τ A w − 2 τ A 0<\ell<\frac{3\alpha+\sqrt{9\alpha^2-24\tau_Aw}}{-2\tau_A} 0 < ℓ < − 2 τ A 3 α + 9 α 2 − 24 τ A w . For A ℓ ≠ 0 A_\ell\ne0 A ℓ = 0 set p 1 = B ℓ / A ℓ p_1=B_\ell/A_\ell p 1 = B ℓ / A ℓ . For A ℓ = 0 A_\ell=0 A ℓ = 0 require B ℓ = 0 B_\ell=0 B ℓ = 0 and 0 < p 2 ≤ 1 0<p_2\le1 0 < p 2 ≤ 1 , and enumerate both real values p 1 = ± 1 − p 2 2 p_1=\pm\sqrt{1-p_2^2} p 1 = ± 1 − p 2 2 , deduplicating zero. In either case set m = 6 ( p 1 − 1 + w 2 ℓ 2 / 2 ) ℓ 3 + 3 w α m=\frac{6(p_1-1+w^2\ell^2/2)}{\ell^3}+3w\alpha m = ℓ 3 6 ( p 1 − 1 + w 2 ℓ 2 /2 ) + 3 w α and restore the original unit-position and tangency equations. There are at most six distinct durations. Since A ℓ = 0 A_\ell=0 A ℓ = 0 has only one positive solution, there are at most seven position-duration candidates, including that exceptional case. Keep m > 0 m>0 m > 0 , positive tangent speed and negative tangent jerk at p p p , and attach its unique final cubic with zero endpoint velocity using (A.6). Retain the admissible completions, including all admitted multiple roots.
Clearance. Equation (A.5) and the factorization for a cubic completion in the proof of Proposition A.12 give strict clearance inside admissible bridges and direct cubic completions. An outer tail following P still requires the polynomial test above.
Representation completeness. For fixed ( α ^ 0 , Γ ^ , T ^ A ) (\hat{\alpha}_0,\hat{\Gamma},\hat{T}_A) ( α ^ 0 , Γ ^ , T ^ A ) there are at most eight raw completion choices per side, including the direct choice. After joining and filtering, let s 0 s_0 s 0 be the initial temporary time and L L L the total duration, so the assembled trajectory r t m p \mathbf{r}_{\mathrm{tmp}} r tmp is defined on [ s 0 , s 0 + L ] [s_0,s_0+L] [ s 0 , s 0 + L ] . Then r u n i t ( t ) = r t m p ( s 0 + L t ) \mathbf{r}_{\mathrm{unit}}(t)=\mathbf{r}_{\mathrm{tmp}}(s_0+Lt) r unit ( t ) = r tmp ( s 0 + L t ) , t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , normalizes the full trajectory to unit duration. Velocities, accelerations, jerk and atom masses acquire factors L L L , L 2 L^2 L 2 , L 3 L^3 L 3 , L 3 L^3 L 3 , respectively; the density acquires L 4 L^4 L 4 . Endpoints are unchanged. For endpoints X , Y X,Y X , Y in the construction frame, first require ∥ X ∥ ≥ 1 \lVert X\rVert\ge1 ∥ X ∥ ≥ 1 , ∥ Y ∥ ≥ 1 \lVert Y\rVert\ge1 ∥ Y ∥ ≥ 1 , and X × Y ≥ 0 X\times Y\ge0 X × Y ≥ 0 . Rotation of X X X onto the positive axis then gives the datum E c a n ( X , Y ) = ( ∥ X ∥ , ∥ X ∥ ∥ X ∥ , X × Y ∥ X ∥ ) \mathcal{E}_{\mathrm{can}}(X,Y)=\left(\lVert X\rVert,\frac{\lVert X\rVert}{\lVert X\rVert},\frac{X\times Y}{\lVert X\rVert}\right) E can ( X , Y ) = ( ∥ X ∥ , ∥ X ∥ ∥ X ∥ , ∥ X ∥ X × Y ) . Retain only admissible constructions as defined in § 2.2.3 , matching the datum when endpoints are prescribed. The matched jets give D t 4 r = r μ D_t^4\mathbf{r}=\mathbf{r}\,\mu D t 4 r = r μ with finite nonnegative multiplier compactly supported in the time interval. Theorem 2.5 and Proposition 2.7 identify the selected optimum.
Conversely, take a selected optimum with arc [ a , b ] [a,b] [ a , b ] , angular speed w ( t ) w(t) w ( t ) , and entry speed w a = w ( a ) > 0 w_a=w(a)>0 w a = w ( a ) > 0 . Rotate its entry normal to ( 1 , 0 ) (1,0) ( 1 , 0 ) and use s = w a ( t − a ) s=w_a(t-a) s = w a ( t − a ) . Its normalized initial conditions are those of Section A.2.7, with α ^ 0 = w ˙ ( a ) / w a 2 \hat{\alpha}_0=\dot{w}(a)/w_a^2 α ^ 0 = w ˙ ( a ) / w a 2 , Γ ^ = Γ / w a 3 \hat{\Gamma}=\Gamma/w_a^3 Γ ^ = Γ/ w a 3 , and T ^ A = w a ( b − a ) > 0 \hat{T}_A=w_a(b-a)>0 T ^ A = w a ( b − a ) > 0 . Lemma A.8 and Propositions A.11 and A.12 supply the order, signs, and allowed contact counts. The final cubic durations are the unique positive roots in (A.6). Any intervening isolated contact satisfies P c o n ( ℓ ) = 0 P_{\mathrm{con}}(\ell)=0 P con ( ℓ ) = 0 and its bracket, so the completion rules retain it, including exceptional and multiple roots. The initial temporary time is s 0 = − w a a s_0=-w_a a s 0 = − w a a and the total duration is L = w a L=w_a L = w a , so the time normalization and E c a n \mathcal{E}_{\mathrm{can}} E can recover the original datum and trajectory.
The plane basis of § 2.2.3 gives U = ( e 1 e 2 ) U=(e_1\ e_2) U = ( e 1 e 2 ) ; Remark 2.1 restores physical trajectories by r p h y s ( t ) = R U r c a n ( t / T ) r_{\mathrm{phys}}(t)=RU r_{\mathrm{can}}(t/T) r phys ( t ) = R U r can ( t / T ) . Varying the permitted basis recovers every minimizer by Proposition 2.7 .
Proof of Theorem 2.9. Existence and Theorem 2.8 , together with the representation and certification proved above, give coverage and identify every admitted construction with the selected minimizer. If two constructions have the same datum, their trajectories therefore agree. By Proposition A.6 , their interior multipliers agree as well, so their minimal labels coincide. Hence the regions are pairwise disjoint.
Bounds and correctness of Algorithm 2.1
Represent the algebraic inputs by polynomials and isolating intervals.
Explicit bounds.
Proposition A.13 (Explicit bounds for canonical arc recovery). For blocked data, set D = r T − r 0 D=r_T-r_0 D = r T − r 0 , U = 162 [ ( 1 + r T , 2 ) 2 + ( r T , 1 − d 0 ) 2 + 1 ] U=162\left[(1+r_{T,2})^2+(r_{T,1}-d_0)^2+1\right] U = 162 [ ( 1 + r T , 2 ) 2 + ( r T , 1 − d 0 ) 2 + 1 ] , t f r e e = min { 1 2 , ( 3 ( d 0 − 1 ) 2 8 U ) 1 / 3 } t_{\mathrm{free}}=\min\left\{\frac{1}{2},\left(\frac{3(d_0-1)^2}{8U}\right)^{1/3}\right\} t free = min { 2 1 , ( 8 U 3 ( d 0 − 1 ) 2 ) 1/3 } , J 0 = 24 U / t f r e e 3 J_0=\sqrt{24U/t_{\mathrm{free}}^3} J 0 = 24 U / t free 3 , B = d 0 J 0 B=d_0J_0 B = d 0 J 0 , w min = 54 ∥ D ∥ 2 5 B w_{\min}=\frac{54\lVert D\rVert^2}{5B} w m i n = 5 B 54 ∥ D ∥ 2 , V = 2 U V=\sqrt{2U} V = 2 U , A = V 2 + V 4 + 6 U A=V^2+\sqrt{V^4+6U} A = V 2 + V 4 + 6 U , C τ = B + V A C_\tau=B+VA C τ = B + V A , δ e n d = min { 1 / 4 , w min 2 / ( 4 U ) } \delta_{\mathrm{end}}=\min\{1/4,w_{\min}^2/(4U)\} δ end = min { 1/4 , w m i n 2 / ( 4 U )} , and
C μ = 2 A 5 t f r e e 4 δ e n d 4 + 2 ( A + C τ / 2 ) 2 + 2 A δ e n d . C_\mu=\frac{2A}{5t_{\mathrm{free}}^4\delta_{\mathrm{end}}^4}+2(A+C_\tau/2)^2+\frac{2A}{\delta_{\mathrm{end}}}. C μ = 5 t free 4 δ end 4 2 A + 2 ( A + C τ /2 ) 2 + δ end 2 A . If the selected optimum contains an arc, then for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] it satisfies 1 2 ∥ r ¨ ∥ 2 ≤ U \frac{1}{2}\lVert\ddot r\rVert^2\le U 2 1 ∥ r ¨ ∥ 2 ≤ U , 0 < Γ ≤ B 0<\Gamma\le B 0 < Γ ≤ B , ∥ r ˙ ( t ) ∥ ≤ V \lVert\dot r(t)\rVert\le V ∥ r ˙ ( t )∥ ≤ V , ∥ r ¨ ( t ) ∥ ≤ A \lVert\ddot r(t)\rVert\le A ∥ r ¨ ( t )∥ ≤ A , μ ( ( 0 , 1 ) ) ≤ C μ \mu((0,1))\le C_\mu μ (( 0 , 1 )) ≤ C μ , ∥ r ( 3 ) ∥ L ∞ ( ( 0 , 1 ) ; R 2 ) ≤ J 0 + C μ \lVert r^{(3)}\rVert_{L^\infty((0,1);\mathbb{R}^2)}\le J_0+C_\mu ∥ r ( 3 ) ∥ L ∞ (( 0 , 1 ) ; R 2 ) ≤ J 0 + C μ . The initial free interval contains [ 0 , t f r e e ] [0,t_{\mathrm{free}}] [ 0 , t free ] , and ∥ r ¨ ( 0 ) ∥ ≤ 6 U \lVert\ddot r(0)\rVert\le\sqrt{6U} ∥ r ¨ ( 0 )∥ ≤ 6 U , ∥ r ( 3 ) ( 0 ) ∥ ≤ J 0 \lVert r^{(3)}(0)\rVert\le J_0 ∥ r ( 3 ) ( 0 )∥ ≤ J 0 . On the arc, w ( t ) ≥ w min w(t)\ge w_{\min} w ( t ) ≥ w m i n .
Proof. Join r 0 r_0 r 0 , ( d 0 , 1 + r T , 2 ) (d_0,1+r_{T,2}) ( d 0 , 1 + r T , 2 ) , ( r T , 1 , 1 + r T , 2 ) (r_{T,1},1+r_{T,2}) ( r T , 1 , 1 + r T , 2 ) , and r T r_T r T by three straight segments, each traversed in time 1 / 3 1/3 1/3 with cubic timing and zero initial and final velocities. The first segment has first coordinate d 0 > 1 d_0>1 d 0 > 1 and the second has height at least one. On the third, either ∣ r T , 1 ∣ ≥ 1 \lvert r_{T,1}\rvert\ge1 ∣ r T , 1 ∣ ≥ 1 or the height is at least r T , 2 ≥ 1 − r T , 1 2 r_{T,2}\ge\sqrt{1-r_{T,1}^2} r T , 2 ≥ 1 − r T , 1 2 . This is a feasible comparison trajectory with cost U U U , since a cubic segment with displacement z z z , duration ℓ \ell ℓ , and zero initial and final velocities costs 6 ∥ z ∥ 2 / ℓ 3 6\lVert z\rVert^2/\ell^3 6 ∥ z ∥ 2 / ℓ 3 . The free lower bound and the energy identity for zero initial and final velocities give 18 ∥ D ∥ 2 ≤ E = 3 2 ∥ r ˙ ∥ 2 ≤ 3 U 18\lVert D\rVert^2\le E=\frac{3}{2}\lVert\dot r\rVert^2\le3U 18 ∥ D ∥ 2 ≤ E = 2 3 ∥ r ˙ ∥ 2 ≤ 3 U . Blocked data have D ≠ 0 D\ne0 D = 0 .
Cauchy–Schwarz and r ˙ ( 0 ) = 0 \dot r(0)=0 r ˙ ( 0 ) = 0 give ∥ r ( t ) − r 0 ∥ ≤ 2 U / 3 t 3 / 2 \lVert r(t)-r_0\rVert\le\sqrt{2U/3}\,t^{3/2} ∥ r ( t ) − r 0 ∥ ≤ 2 U /3 t 3/2 and ∥ r ˙ ( t ) ∥ ≤ V \lVert\dot r(t)\rVert\le V ∥ r ˙ ( t )∥ ≤ V . Thus no contact occurs on [ 0 , t f r e e ] [0,t_{\mathrm{free}}] [ 0 , t free ] . There r ¨ ( t ) = r ¨ ( 0 ) + t r ( 3 ) ( 0 ) \ddot r(t)=\ddot r(0)+t r^{(3)}(0) r ¨ ( t ) = r ¨ ( 0 ) + t r ( 3 ) ( 0 ) , and completing the square gives
2 U ≥ ∫ 0 t f r e e ∥ r ¨ ( 0 ) + t r ( 3 ) ( 0 ) ∥ 2 d t ≥ t f r e e 3 12 ∥ r ( 3 ) ( 0 ) ∥ 2 . 2U\ge\int_0^{t_{\mathrm{free}}}\lVert\ddot r(0)+t r^{(3)}(0)\rVert^2\,\mathrm{d}t\ge\frac{t_{\mathrm{free}}^3}{12}\lVert r^{(3)}(0)\rVert^2. 2 U ≥ ∫ 0 t free ∥ r ¨ ( 0 ) + t r ( 3 ) ( 0 ) ∥ 2 d t ≥ 12 t free 3 ∥ r ( 3 ) ( 0 ) ∥ 2 .
Since r ˙ ( 0 ) = 0 \dot r(0)=0 r ˙ ( 0 ) = 0 , Γ = − r 0 × r ( 3 ) ( 0 ) \Gamma=-r_0\times r^{(3)}(0) Γ = − r 0 × r ( 3 ) ( 0 ) , hence ∣ Γ ∣ ≤ B \lvert\Gamma\rvert\le B ∣ Γ ∣ ≤ B . On an arc, positive speed and nonnegative density imply Γ w ≥ w 4 + 3 E / 5 > 0 \Gamma w\ge w^4+3E/5>0 Γ w ≥ w 4 + 3 E /5 > 0 , hence Γ > 0 \Gamma>0 Γ > 0 and w ≥ 3 E / ( 5 Γ ) ≥ w min w\ge3E/(5\Gamma)\ge w_{\min} w ≥ 3 E / ( 5Γ ) ≥ w m i n .
At a contact t t t , write n = r ( t ) n=r(t) n = r ( t ) , r ˙ ( t ) = w ( − n 2 , n 1 ) \dot r(t)=w(-n_2,n_1) r ˙ ( t ) = w ( − n 2 , n 1 ) . The energy and momentum invariants give
∥ r ¨ + w 2 n ∥ 2 = w 4 − 2 Γ w + 2 E , τ n : = n × r ( 3 ) ( t ) = − Γ − w ( n , r ˙ ) . \lVert\ddot r+w^2n\rVert^2=w^4-2\Gamma w+2E,\qquad\tau_n:=n\times r^{(3)}(t)=-\Gamma-w(n,\dot r). ∥ r ¨ + w 2 n ∥ 2 = w 4 − 2Γ w + 2 E , τ n := n × r ( 3 ) ( t ) = − Γ − w ( n , r ˙ ) .
Contact order gives w > 0 w>0 w > 0 , so ∥ r ¨ ( t ) ∥ ≤ A \lVert\ddot r(t)\rVert\le A ∥ r ¨ ( t )∥ ≤ A and ∣ τ n ∣ ≤ C τ \lvert\tau_n\rvert\le C_\tau ∣ τ n ∣ ≤ C τ . Since the initial and final velocities are zero, both endpoint accelerations have norm 2 E ≤ 6 U ≤ A \sqrt{2E}\le\sqrt{6U}\le A 2 E ≤ 6 U ≤ A . Acceleration is affine on every free piece, proving the global acceleration bound.
At the arc exit b b b , w min ≤ ∥ r ˙ ( b ) ∥ ≤ 2 U ( 1 − b ) w_{\min}\le\lVert\dot r(b)\rVert\le\sqrt{2U(1-b)} w m i n ≤ ∥ r ˙ ( b )∥ ≤ 2 U ( 1 − b ) , so b ≤ 1 − 2 δ e n d b\le1-2\delta_{\mathrm{end}} b ≤ 1 − 2 δ end . For ϕ ( t ) = t 2 ( 1 − t ) 2 \phi(t)=t^2(1-t)^2 ϕ ( t ) = t 2 ( 1 − t ) 2 , apply Theorem 2.5 to h = ϕ e 1 h=\phi e_1 h = ϕ e 1 , using the projection established in Section A.2.5. This gives ∫ 0 + 1 − ϕ ( t ) 2 d μ ( t ) ≤ ∫ 0 1 ∣ ϕ ¨ ( t ) ∣ 2 d t = 4 5 \int_{0^+}^{1^-}\phi(t)^2\,\mathrm{d}\mu(t)\le\int_0^1\lvert\ddot{\phi}(t)\rvert^2\,\mathrm{d}t=\frac45 ∫ 0 + 1 − ϕ ( t ) 2 d μ ( t ) ≤ ∫ 0 1 ∣ ϕ ¨ ( t ) ∣ 2 d t = 5 4 . Since ϕ ≥ t f r e e 2 δ e n d 2 \phi\ge t_{\mathrm{free}}^2\delta_{\mathrm{end}}^2 ϕ ≥ t free 2 δ end 2 on [ t f r e e , 1 − δ e n d ] [t_{\mathrm{free}},1-\delta_{\mathrm{end}}] [ t free , 1 − δ end ] , the mass of μ \mu μ there is at most 4 / ( 5 t f r e e 2 δ e n d 2 ) 4/(5t_{\mathrm{free}}^2\delta_{\mathrm{end}}^2) 4/ ( 5 t free 2 δ end 2 ) . By Theorem 2.8 , any mass of μ \mu μ outside this interval lies in a single isolated atom at t P > 1 − δ e n d t_P>1-\delta_{\mathrm{end}} t P > 1 − δ end . The preceding free bridge starts at b ≤ 1 − 2 δ e n d b\le1-2\delta_{\mathrm{end}} b ≤ 1 − 2 δ end , so its constant jerk satisfies ∥ r L ( 3 ) ∥ ≤ 2 A / δ e n d \lVert r_L^{(3)}\rVert\le2A/\delta_{\mathrm{end}} ∥ r L ( 3 ) ∥ ≤ 2 A / δ end .
Set T t a i l = 1 − t P T_{\mathrm{tail}}=1-t_P T tail = 1 − t P , n = r ( t P ) n=r(t_P) n = r ( t P ) and w = n × r ˙ ( t P ) w=n\times\dot{r}(t_P) w = n × r ˙ ( t P ) ; then T t a i l < δ e n d ≤ 1 / 4 T_{\mathrm{tail}}<\delta_{\mathrm{end}}\le1/4 T tail < δ end ≤ 1/4 . The zero final tangential velocity gives w = − T t a i l ( n × r ¨ ( t P ) ) − T t a i l 2 τ n / 2 w=-T_{\mathrm{tail}}(n\times\ddot{r}(t_P))-T_{\mathrm{tail}}^2\tau n/2 w = − T tail ( n × r ¨ ( t P )) − T tail 2 τ n /2 , hence w ≤ T t a i l ( A + C τ / 2 ) w\le T_{\mathrm{tail}}(A+C_\tau/2) w ≤ T tail ( A + C τ /2 ) . The contact curvature gives ⟨ n , r ¨ ( t P ) ⟩ ≥ − w 2 \langle n,\ddot{r}(t_P)\rangle\ge-w^2 ⟨ n , r ¨ ( t P )⟩ ≥ − w 2 . Therefore its mass obeys m P = − 2 ⟨ n , r ¨ ( t P ) ⟩ T t a i l − ⟨ n , r L ( 3 ) ⟩ ≤ 2 w 2 T t a i l + 2 A δ e n d ≤ 2 ( A + C τ / 2 ) 2 + 2 A δ e n d m_P=-\frac{2\langle n,\ddot{r}(t_P)\rangle}{T_{\mathrm{tail}}}-\langle n,r_L^{(3)}\rangle\le\frac{2w^2}{T_{\mathrm{tail}}}+\frac{2A}{\delta_{\mathrm{end}}}\le2(A+C_\tau/2)^2+\frac{2A}{\delta_{\mathrm{end}}} m P = − T tail 2 ⟨ n , r ¨ ( t P )⟩ − ⟨ n , r L ( 3 ) ⟩ ≤ T tail 2 w 2 + δ end 2 A ≤ 2 ( A + C τ /2 ) 2 + δ end 2 A . Summing proves the bound C μ C_\mu C μ . Finally D t 1 r ( 3 ) = r μ D_t^1r^{(3)}=r\,\mu D t 1 r ( 3 ) = r μ and ∥ r ( t ) ∥ = 1 \lVert r(t)\rVert=1 ∥ r ( t )∥ = 1 on its support yield ∥ r ( 3 ) ( t ) ∥ ≤ ∥ r ( 3 ) ( 0 ) ∥ + μ ( ( 0 , 1 ) ) ≤ J 0 + C μ \lVert r^{(3)}(t)\rVert\le\lVert r^{(3)}(0)\rVert+\mu((0,1))\le J_0+C_\mu ∥ r ( 3 ) ( t )∥ ≤ ∥ r ( 3 ) ( 0 )∥ + μ (( 0 , 1 )) ≤ J 0 + C μ .
Correctness and termination. The free test and Algorithm 2.2 terminate by real quantifier elimination and sample-point recovery [3 Section 2] : the coefficients and clearance conditions in Section A.2.7 are algebraic. If these tests fail, Theorem 2.8 leaves the four arc families.
Bounded parameters. For each arc family, use m = 2 m=2 m = 2 , 3 3 3 , or 4 4 4 event times 0 ≤ t 1 ≤ ⋯ ≤ t m ≤ 1 0\le t_1\le\cdots\le t_m\le1 0 ≤ t 1 ≤ ⋯ ≤ t m ≤ 1 , with consecutive arc endpoints, initial acceleration and jerk r ¨ ( 0 ) , r ( 3 ) ( 0 ) ∈ R 2 \ddot{r}(0),r^{(3)}(0)\in\mathbb{R}^2 r ¨ ( 0 ) , r ( 3 ) ( 0 ) ∈ R 2 , and nonnegative atom masses m i m_i m i . From ( r 0 , 0 ) (r_0,0) ( r 0 , 0 ) , propagate cubic pieces with jumps r ( 3 ) ↦ r ( 3 ) + m i r ( t i ) r^{(3)}\mapsto r^{(3)}+m_i r(t_i) r ( 3 ) ↦ r ( 3 ) + m i r ( t i ) . Propagate the arc between its entry and exit jumps. Use rational upper bounds from Proposition A.13 for the initial jets and masses. The four closed parameter domains, including zero durations and masses, form a compact disjoint union P r e c \mathcal{P}_{\mathrm{rec}} P rec with computable finite rational covers.
In the arc system of Section A.2.7, choose rational H > max { 1 , V , A , J 0 + C μ } H>\max\{1,V,A,J_0+C_\mu\} H > max { 1 , V , A , J 0 + C μ } and clip each state coordinate to [ − H , H ] [-H,H] [ − H , H ] only when evaluating the last component λ r \lambda r λ r . This leaves true arc states unchanged and gives a globally Lipschitz vector field, with constant L H = max { 1 , 42 H 2 } L_H=\max\{1,42H^2\} L H = max { 1 , 42 H 2 } in the maximum coordinate norm.
For p ∈ P r e c p\in\mathcal{P}_{\mathrm{rec}} p ∈ P rec , write q p q_p q p for the assembled trial and z p z_p z p for its one-sided arc state. Euler error estimates for the clipped field, which vanishes at zero, and the polynomial cubic and jump maps give computable state bounds and uniform parameter and time moduli. Moving knots preserve continuity of position, velocity, and acceleration, whose one-sided derivatives remain bounded. Thus p ↦ q p ∈ C 2 ( [ 0 , 1 ] ; R 2 ) p\mapsto q_p\in C^2([0,1];\mathbb{R}^2) p ↦ q p ∈ C 2 ([ 0 , 1 ] ; R 2 ) is computable with a computable uniform modulus for ∥ q ∥ C 2 ( [ 0 , 1 ] ; R 2 ) : = max 0 ≤ ℓ ≤ 2 sup 0 ≤ t ≤ 1 ∥ q ( ℓ ) ( t ) ∥ \lVert q\rVert_{C^2([0,1];\mathbb{R}^2)}:=\max_{0\le\ell\le2}\sup_{0\le t\le1}\lVert q^{(\ell)}(t)\rVert ∥ q ∥ C 2 ([ 0 , 1 ] ; R 2 ) := max 0 ≤ ℓ ≤ 2 sup 0 ≤ t ≤ 1 ∥ q ( ℓ ) ( t )∥ . Integrating the squared acceleration gives the same properties for p ↦ 1 2 ∥ q ˙ p ∥ 2 p\mapsto\frac12\lVert\dot{q}_p\rVert^2 p ↦ 2 1 ∥ q ˙ p ∥ 2 .
Admissible parameters. Impose q p ( 1 ) = r T q_p(1)=r_T q p ( 1 ) = r T , q ˙ p ( 1 ) = 0 \dot{q}_p(1)=0 q ˙ p ( 1 ) = 0 , and ∥ q p ( t i ) ∥ = 1 \lVert q_p(t_i)\rVert=1 ∥ q p ( t i )∥ = 1 ; require ∥ q p ( t ) ∥ ≥ 1 \lVert q_p(t)\rVert\ge1 ∥ q p ( t )∥ ≥ 1 and ⟨ e , q p ( t ) ⟩ ≥ δ \langle e,q_p(t)\rangle\ge\delta ⟨ e , q p ( t )⟩ ≥ δ throughout [ 0 , 1 ] [0,1] [ 0 , 1 ] ; and on the closed arc require ∥ q p ( t ) ∥ = 1 \lVert q_p(t)\rVert=1 ∥ q p ( t )∥ = 1 , λ ( z p ( t ) ) ≥ 0 \lambda(z_p(t))\ge0 λ ( z p ( t )) ≥ 0 , and z p ( t ) ∈ [ − H , H ] 8 z_p(t)\in[-H,H]^8 z p ( t ) ∈ [ − H , H ] 8 . Use (A.2), with e = ( 0 , 1 ) e=(0,1) e = ( 0 , 1 ) , δ = 0 \delta=0 δ = 0 for symmetric data. At coincident events apply jumps in template order; arc tests use the state after entry and before exit, including the single state of a zero-duration arc.
These conditions define Z = R − 1 ( 0 ) \mathcal{Z}=\mathcal{R}^{-1}(0) Z = R − 1 ( 0 ) , where R ≥ 0 \mathcal{R}\ge0 R ≥ 0 sums squared endpoint and event errors, suprema of positive inequality violations, and the supremum of the absolute arc circle error. After parametrizing variable intervals by u ∈ [ 0 , 1 ] u\in[0,1] u ∈ [ 0 , 1 ] , finite meshes and the uniform state and time moduli compute R \mathcal{R} R and a uniform modulus.
Representation completeness and Proposition A.13 give Z ≠ ∅ \mathcal{Z}\ne\varnothing Z = ∅ . Every p ∈ Z p\in\mathcal{Z} p ∈ Z is feasible and satisfies D t 4 q p = q p μ p D_t^4q_p=q_p\mu_p D t 4 q p = q p μ p , where μ p = ∑ 0 < t i < 1 m i δ t i + λ ( z p ( t ) ) d t \mu_p=\sum_{0<t_i<1}m_i\delta_{t_i}+\lambda(z_p(t))\,\mathrm{d}t μ p = ∑ 0 < t i < 1 m i δ t i + λ ( z p ( t )) d t has its density restricted to the arc and is nonnegative and supported on the circle. Omit endpoint masses and combine coincident ones. For δ > 0 \delta>0 δ > 0 , Theorem 2.5 applies directly. For symmetric data, z ( t ) = ⟨ e , q p ( t ) ⟩ ≥ 0 z(t)=\langle e,q_p(t)\rangle\geq0 z ( t ) = ⟨ e , q p ( t )⟩ ≥ 0 satisfies
z ( t ) = ∫ 0 + 1 − G 1 ( t , s ) z ( s ) d μ p ( s ) . z(t)=\int_{0^+}^{1^-}G_1(t,s)z(s)\,\mathrm{d}\mu_p(s). z ( t ) = ∫ 0 + 1 − G 1 ( t , s ) z ( s ) d μ p ( s ) .
If z μ p = 0 z\mu_p=0 z μ p = 0 , then z = 0 z=0 z = 0 , which cannot clear the disk between blocked symmetric endpoints. Otherwise positivity of G 1 G_1 G 1 gives z > 0 z>0 z > 0 in ( 0 , 1 ) (0,1) ( 0 , 1 ) , so the certificate identifies the selected upper optimum. Hence { q p : p ∈ Z } = { r ∗ } \{q_p:p\in\mathcal{Z}\}=\{r_*\} { q p : p ∈ Z } = { r ∗ } . A zero-duration arc would give at most three combined interior atoms, contradicting the exact atomic rejection.
Search and termination. Rational sample evaluations and the uniform moduli give the bounds on parameter boxes in Algorithm 2.3. Every zero remains covered, and R \mathcal{R} R tends uniformly to zero on retained boxes as resolution and evaluation errors vanish. Time meshes, derivative bounds, and the same moduli give trajectory and cost enclosures with errors tending to zero. Compactness places every limit of retained parameters in Z \mathcal{Z} Z . Its trajectory image is { r ∗ } \{r_*\} { r ∗ } , so continuity forces the enclosure diameters to zero. Thus Algorithm 2.1 terminates with bounds containing r ∗ r_* r ∗ and its cost; the center trial need not be feasible. This is effective compactness with a unique image [12 Fact 3.25 and the proof of Corollary 4.11] .