updates according to pull request comments

src/gradient_descent.jl

@@ -56,15 +56,15 @@ function optimize{T}(d::DifferentiableFunction,
 
     # The current search direction
     s = similar(x)
-    
+
     # Buffers for use in line search
     x_ls = similar(x)
     gr_ls = similar(x)
-    
+
     # Store f(x) in f_x
     f_x_previous, f_x = NaN, d.fg!(x, gr)
     f_calls, g_calls = f_calls + 1, g_calls + 1
-    
+
     # Keep track of step-sizes
     alpha = alphainit(one(T), x, gr, f_x)
 

src/l_bfgs.jl

@@ -38,9 +38,8 @@ function twoloop!(s::Vector,
     end
 
     # Copy q into s for forward pass
-    # UNPRECONDITIONED: copy!(s, q)
-    # this is equivalent to choosing Id as the preconditioner / initial guess
-    # PRECONDITIONING: (preconditioner update was done outside!)
+    # apply preconditioner if precon != nothing
+    # (Note: preconditioner update was done outside of this function)
     precondfwd!(s, precon, q)
 
     # Forward pass

src/precon.jl

@@ -31,7 +31,7 @@ precondprep!(P::Void, x) = nothing
 #####################################################
 #  [2] Diagonal preconditioner
 
-function precondfwd!{T}(out::Array, p::Vector, A::Array{T})
+function precondfwd!(out::Array, p::Vector, A::Array)
     @simd for i in 1:length(A)
         @inbounds out[i] = p[i] * A[i]
     end
@@ -56,11 +56,9 @@ end
 
 #####################################################
 #  [3] Matrix Preconditioner
-#  this assumption here is that P is given by its inverse, which
-#  is more typical
+#  the assumption here is that P is given by its inverse, which is typical
 
-PMAT = Union{Matrix,SparseMatrixCSC}
-precondfwd!(out::Vector, P::PMAT, A::Vector) = copy!(out, P \ A)
-precondfwddot(A::Vector, P::PMAT, B::Vector) = dot(A, P \ B)
-precondinvdot(A::Vector, P::PMAT, B::Vector) = dot(A, P * B)
+precondfwd!(out::Vector, P::AbstractMatrix, A::Vector) = copy!(out, P \ A)
+precondfwddot(A::Vector, P::AbstractMatrix, B::Vector) = dot(A, P \ B)
+precondinvdot(A::Vector, P::AbstractMatrix, B::Vector) = dot(A, P * B)