Fluid stability problem from IFISS package, Mark Embree, Virginia Tech   
                                                                         
The matrix comes from a shift-invert eigenvalue computation that arises  
during linear stability analysis of a 2D backward facing step.  The      
discretizations were created with IFISS, which we also use to compute the
steady-state flow whose stability we are assessing.                      
                                                                         
Created by the IFISS package, by David Silvester (School of Mathematics, 
Univ. of Manchester), Howard Elman (Computer Science, Univ. of Maryland),
and Alison Ramage (Dept. of Mathematics and Statistics, Univ. of         
Strathclyde).  http://www.maths.manchester.ac.uk/~djs/ifiss/             
                                                                         
This matrix requires one step of iterative refinement after LU           
factorization.  x=A\b in MATLAB (using UMFPACK) does iterative refinement
by default but with just lu, no iterative refinement is done:            
                                                                         
    In MATLAB R2014a:                                                    
    x = randn(96307,1);                                                  
    b = A*x;                                                             
    x1 = A\b;                                                            
    norm(x-x1) is 1.1045984e-12                                          
    norm(b-A*x1) is 9.3538018e-15                                        
                                                                         
    [L,U,P,Q,R] = lu(A);                                                 
    x2 = Q*(U\(L\(P*(R\b))));                                            
    norm(x-x2) is 4.9150874e-05                                          
    norm(b-A*x2) is 4.0055911e-07                                        
                                                                         
The matrix is well conditioned, with singular values ranging from        
1.003 to 2e-5.  The singluar values themselves are in                    
Problem.aux.singular_values.  The need for iterative refinement          
comes from the threshold partial pivoting in UMFPACK, which tries        
to balance reduction in fill with finding good numerical pivots.         
Thus UMFPACK uses iterative refinement with sparse backward error using  
the method described in Arioli, Demmel, and Duff, "Sovling sparse linear 
systems with sparse backward error", SIAM J. Matrix Analysis & Applic,   
vol 10, no 2, pp 165-190, Apr 1989).                                     
                                                                         
The matrix itself is described in this paper:                            
                                                                         
https://arxiv.org/abs/1601.00044                                         
                                                                         
Pseudospectra of Matrix Pencils for Transient Analysis of                
Differential-Algebraic Equations                                         
                                                                         
Mark Embree, Blake Keeler                                                
(Submitted on 1 Jan 2016 (v1),                                           
last revised 27 Jun 2017 (this version, v3))                             
                                                                         
To understand the solution of a linear, time-invariant                   
differential-algebraic equation, one must analyze a matrix pencil (A,E)  
with singular E. Even when this pencil is stable (all its finite         
eigenvalues fall in the left-half plane), the solution can exhibit       
transient growth before its inevitable decay. When the equation results  
from the linearization of a nonlinear system, this transient growth gives
a mechanism that can promote nonlinear instability. One might hope to    
enrich the conventional large-scale eigenvalue calculation used for      
linear stability analysis to signal the potential for such transient     
growth. Toward this end, we introduce a new definition of the            
pseudospectrum of a matrix pencil, use it to bound transient growth,     
explain how to incorporate a physically-relevant norm, and derive        
approximate pseudospectra using the invariant subspace computed in       
conventional linear stability analysis. We apply these tools to several  
canonical test problems in fluid mechanics, an important source of       
differential-algebraic equations.