533DAnimationPhysics-UBCComputerScience.ppt

cs542g-term1-2007 cs533d-winter-2005 Notes Extra class this Friday 1-2pm Assignment 2 is out Error in last lecture: quasi-Newton methods based on the secant condition: Really just Taylor series applied to the gradient function g: Nonlinear Least Squares One particular nonlinear problem of interest: Can we exploit the structure of this objective? Gradient: Hessian: Gauss-Newton Assuming g is close to linear, just keep the JTJ term Automatically positive (semi-)definite:in fact, modified Newton solve is now a linear least-squares problem! Convergence may not be quite as fast as Newton, but more robust. Time Integration A core problem across many applications Physics, robotics, finance, control, … Typical statement:system of ordinary differential equations,first orderSubject to initial conditions: Higher order equations can be reduced to a system - though not always a good thing to do Forward Euler Derive from Taylor series: Truncation error is second-order (Global) accuracy is first-order Heuristic: to get to a fixed time T, need T/?t time steps Analyzing Forward Euler Euler proved convergence as ?t goes to 0 In practice, need to pick ?t 0- how close to zero do you need to go? The first and foremost tool is using a simple model problem that we can solve exactly Solving Linear ODE’s First: what is the exact solution? For scalars: This is in fact true for systems as well, though needs the “matrix exponential” More elementary approach:change basis to diagonalize A(or reduce to Jordan canonical form) The Test Equation We can thus simplify our model down to a single scalar equation, the test equation Relate this to a full nonlinear system by taking the eigenvalues of the Jacobian of f: But obviously nonlinearities may cause unexpected things to happen: full nonlinear analysis is generally problem-specific, and hard or impossible Solution of the Test Equation Split scalar into real and imaginary parts: For initial conditions y0=1 The mag