Rest-to-Rest Flexible Momentum Updates
Implementation & benchmark of arXiv:2608.11360 — Time-Optimal Control of One-Mode Flexible Structures: Analytical Solution and the Cost of Flexibility
Source paper: Time-Optimal Control of One-Mode Flexible Structures: Analytical Solution and the Cost of Flexibility arXiv:2608.11360 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace an unconstrained momentum update by a bounded-acceleration, four-arc bang-bang maneuver in an augmented state containing parameter position, velocity, and an oscillator coordinate. Each micro-maneuver targets a gradient-derived displacement while ending with zero velocity and zero oscillator amplitude, so flexible or momentum-like modes do not carry ringing into the next update.
Formulas
Mathematical statement
For each update direction q in R^d, with norm q equal to 1, introduce scalar states x_1, x_2, x_3, x_4, where x_1 is displacement along q, x_2 is its velocity, and (x_3,x_4) is an oscillator with unit dimensionless frequency. The controlled dynamics are dot{x}_1=x_2, dot{x}_2=u, dot{x}_3=x_4, and dot{x}_4=-x_3+u, with absolute value of u at most 1. For a rest-to-rest displacement L>0, the paper gives T implicitly by (T/2)^2 - 2 arccos^2(cos^2(T/4)) = L. Define T_s = 2 arccos(cos^2(T/4)); then T^2 = T_r^2 + 2T_s^2, with T_r=2 sqrt(L) and 0 <= T_s <= pi. The optimal input is an alternating bang-bang signal with signs -1,+1,-1,+1 and symmetric switching times. The resulting maneuver satisfies 2 sqrt(L) <= T <= 2 sqrt(L+pi^2/2), has T asymptotic to 384^(1/4) L^(1/4) for small L, and has T asymptotic to 2 sqrt(L) for large L.
Implementation notes
1. Integration point: place this controller between gradient computation and parameter assignment. Given parameters theta, compute a preconditioned descent vector g, set q=-g/(norm(g)+epsilon), and choose a desired parameter displacement D=eta_0(norm(g)+epsilon), where eta_0 is a baseline step scale. Choose acceleration bound A and oscillator frequency omega. Convert to dimensionless displacement L=omega^2 D/A, execute the scalar maneuver, and update theta along q. The physical time is t=tau/omega, and the physical acceleration is A u. 2. Pseudocode: compute L; solve the scalar equation F(T)=(T/2)^2-2 arccos^2(cos^2(T/4))-L=0 by bisection; set T_s=2 arccos(cos^2(T/4)); initialize x=(0,0,0,0); integrate the state equations with RK4 or small-step Euler using the paper's symmetric alternating bang-bang control. During the maneuver, hold q fixed and apply parameter acceleration A u q. 3. Computed from the paper: the state equations, bang-bang structure, scalar root equation, synchronization time, and time bounds. Estimated empirically: A, omega, the mapping from gradient magnitude to D, and whether minibatch-gradient variation is slow enough to justify freezing q during one maneuver. Log terminal velocity norm and oscillator residual sqrt(x_3^2+x_4^2). 4. First cheap experiment: train a two-layer MLP on MNIST or Fashion-MNIST and compare SGD with momentum, Adam, and this controller at matched parameter-update displacement. Before full training, test it on a diagonal quadratic f(theta)=0.5 theta^T H theta with condition numbers 1, 10, and 100. The falsifiable signatures are that measured maneuver time agrees with the implicit formula within 5%, terminal oscillator residual is numerical-error limited, and the transition from T proportional to L^(1/4) to T approximately 2 sqrt(L) appears as L varies. At natural-motion points where T_s=0, the flexible controller should coincide with rigid double-integrator timing; at anti-resonant points it should approach the upper bound. If minibatch noise destroys the benefit, oscillator residual and loss variance should increase sharply when the gradient direction changes substantially within one maneuver.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.