[Reference] Constant-length energy function - Hessian

December 23, 2013

Recall energy function

E=12xDDx+12μDDμxDη

and its gradient,

E=Ex=xDD+μDDJμηDxDJη

The derivative w.r.t. t_i of E' (i.e. the i-th row of the Hessian) is given by:

Eti=(DD):i+(JμDDJμ):i+JμtiDDμ(DJη):i(JηD):i{xDJηti}

Of the six terms, the third and sixth are particularly problematic when generalizing to the full hessian, because they involve the Jacobian of a Jacobian, which is a 3D tensor.

To keep running time to \(O(n3)\), we'll use diagonal approximations for those terms.

2μ(ti)2=ti(Jμ):i=ti[diag3x1(Δ3x3z)+(Δ1x3)repmat(z3,N/3,1)+KJz]:i=ti(00δi3x3z00)+δi,3x1repmat(z(3)i,N/3,1)+K(Jz):i=(00Ci3x3z00)+(00δi3x3zi00)+Ci,3x1repmat(z(3)i,N/3,1)+δi,3x1repmat(z(3)i,N/3,1)+K(Jz):i+K(Jz):i

Started deriving Jz, but was stymied by the complexity. For now, we'll resort to ignoring these terms and see how the optimization goes.

Posted by Kyle Simek