fixes and tweaks to docs

doc/docs/NLopt_Algorithms.md

@@ -305,11 +305,11 @@ My implementation of the globally-convergent method-of-moving-asymptotes (MMA) a
 
 This is an improved CCSA ("conservative convex separable approximation") variant of the original MMA algorithm published by Svanberg in 1987, which has become popular for topology optimization. (*Note:* "globally convergent" does *not* mean that this algorithm converges to the global optimum; it means that it is guaranteed to converge to *some* local minimum from any feasible starting point.)
 
-At each point **x**, MMA forms a local approximation using the gradient of *f* and the constraint functions, plus a quadratic "penalty" term to make the approximations "conservative" (upper bounds for the exact functions). The precise approximation MMA forms is difficult to describe in a few words, because it includes nonlinear terms consisting of a poles at some distance from *x* (outside of the current trust region), almost a kind of Pade approximant. The main point is that the approximation is both convex and separable, making it trivial to solve the approximate optimization by a dual method. Optimizing the approximation leads to a new candidate point **x**. The objective and constraints are evaluated at the candidate point. If the approximations were indeed conservative (upper bounds for the actual functions at the candidate point), then the process is restarted at the new **x**. Otherwise, the approximations are made more conservative (by increasing the penalty term) and re-optimized.
+At each point **x**, MMA forms a local approximation using the gradient of *f* and the constraint functions, plus a quadratic "penalty" term to make the approximations "conservative" (upper bounds for the exact functions). The precise approximation MMA forms is difficult to describe in a few words, because it includes nonlinear terms consisting of a poles at some distance from *x* (outside of the current trust region), almost a kind of Padé approximant. The main point is that the approximation is both convex and separable, making it trivial to solve the approximate optimization by a dual method. Optimizing the approximation leads to a new candidate point **x**. The objective and constraints are evaluated at the candidate point. If the approximations were indeed conservative (upper bounds for the actual functions at the candidate point), then the process is restarted at the new **x**. Otherwise, the approximations are made more conservative (by increasing the penalty term) and re-optimized.
 
 (If you contact [Professor Svanberg](http://researchprojects.kth.se/index.php/kb_7902/pb_2085/pb.html), he has been willing in the past to graciously provide you with his original code, albeit under restrictions on commercial use or redistribution. The MMA implementation in NLopt, however, is completely independent of Svanberg's, whose code we have not examined; any bugs are my own, of course.)
 
-I also implemented another CCSA algorithm from the same paper, `NLOPT_LD_CCSAQ`: instead of constructing local MMA approximations, it constructs simple quadratic approximations (or rather, affine approximations plus a quadratic penalty term to stay conservative). This is the ccsa_quadratic code. It seems to have similar convergence rates to MMA for most problems, which is not surprising as they are both essentially similar. However, for the quadratic variant I implemented the possibility of [preconditioning](NLopt_Reference#Preconditioning_with_approximate_Hessians.md): including a user-supplied Hessian approximation in the local model. It is easy to incorporate this into the proof in Svanberg's paper, and to show that global convergence is still guaranteed as long as the user's "Hessian" is positive semidefinite, and it practice it can greatly improve convergence if the preconditioner is a good approximation for the real Hessian (at least for the eigenvectors of the largest eigenvalues).
+I also implemented another CCSA algorithm from the same paper, `NLOPT_LD_CCSAQ`: instead of constructing local MMA approximations, it constructs simple quadratic approximations (or rather, affine approximations plus a quadratic penalty term to stay conservative). This is the ccsa_quadratic code. It seems to have similar convergence rates to MMA for most problems, which is not surprising as they are both essentially similar. However, for the quadratic variant I implemented the possibility of [preconditioning](NLopt_Reference.md#preconditioning-with-approximate-hessians): including a user-supplied Hessian approximation in the local model. It is easy to incorporate this into the proof in Svanberg's paper, and to show that global convergence is still guaranteed as long as the user's "Hessian" is positive semidefinite, and it practice it can greatly improve convergence if the preconditioner is a good approximation for the real Hessian (at least for the eigenvectors of the largest eigenvalues).
 
 ### SLSQP
 

doc/docs/NLopt_Introduction.md

@@ -4,6 +4,8 @@
 
 In this chapter of the manual, we begin by giving a general overview of the optimization problems that NLopt solves, the key distinctions between different types of optimization algorithms, and comment on ways to cast various problems in the form NLopt requires. We also describe the background and goals of NLopt.
 
+[TOC]
+
 Optimization problems
 ---------------------