At the heart of all quasi-Newton methods is an update rule that enables us to gradually improve the Hessian approximation using the already available gradient evaluations. Theoretical results show that the global performance of optimization algorithms can be improved with higher-order derivatives. This motivates an investigation of generalizations of quasi-Newton update rules to obtain for example third derivatives (which are tensors) from Hessian evaluations. Our generalization is based on the observation that quasi-Newton updates are least-change updates satisfying the secant equation, with different methods using different norms to measure the size of the change. We present a full characterization for least-change updates in weighted Frobenius norms (satisfying an analogue of the secant equation) for derivatives of arbitrary order. Moreover, we establish convergence of the approximations to the true derivative under standard assumptions and explore the quality of the generated approximations in numerical experiments.

Coupled differential equations generally present an important algebraic structure.
For example in the incompressible Navier-Stokes equations, the velocity is affected only by the selenoidal part of the applied force.
This structure can be translated naturally by the notion of complex.
One idea is then to exploit this complex structure at the discrete level in the creation of numerical methods.

The goal of the presentation is to expose the notion of complex by motivating its uses. 
We will present in more detail the creation of a scheme for the Navier-Stokes equations and study its properties.
 

The recent advancements in additive manufacturing have enabled the creation of lattice structures with intricate small-scale details. This has led to the need for new techniques in the field of topology optimization that can handle a vast number of design variables. Despite the efforts to develop multi-scale topology optimization techniques, their high computational cost has limited their application. To overcome this challenge, a new technique called dehomogenization has shown promising results in terms of performance and computational efficiency for optimizing compliance problems.

In this talk, we extend the application of the dehomogenization method to stress minimization problems, which are crucial in structural design. The method involves homogenizing the macroscopic response of a proposed family of microstructures. Next, the macroscopic structure is optimized using gradient-based methods while orienting the cells according to the principal stress components. The final step involves dehomogenization of the structure. The proposed methodology also considers singularities in the orientation field by incorporating singular functions in the dehomogenization process. The validity of the methodology is demonstrated through several numerical examples.

Sketch-and-precondition techniques are popular for solving large overdetermined least squares (LS) problems. This is when a right preconditioner is constructed from a randomized 'sketch' of the matrix. In this talk, we will see that the sketch-and-precondition technique is not numerically stable for ill-conditioned LS problems. We propose a modifciation: using an unpreconditioned iterative LS solver on the preconditioned matrix. Provided the condition number of A is smaller than the reciprocal of the unit round-off, we show that this modification ensures that the computed solution has a comparable backward error to the iterative LS solver applied to a well-conditioned matrix. Using smoothed analysis, we model floating-point rounding errors to provide a convincing argument that our modification is expected to compute a backward stable solution even for arbitrarily ill-conditioned LS problems.

We consider the problem of solving nonconvex quadratic optimization problems, potentially with additional integrality constraints on the variables.  Gurobi takes a branch-and-bound approach to solve such problems to global optimality, and in this talk we will review the three main algorithmic components that Gurobi uses:  Convex relaxations based on local linearization, globally valid cutting planes, and nonlinear local optimization heuristics.  We will explain how these parts play together, and discuss some of the implementation details.

The numerical solution of elliptic PDEs is often the most computationally intensive task in large-scale continuum mechanics simulations.  High-order finite element methods can efficiently exploit modern parallel hardware while offering very rapid convergence properties.  As the polynomial degree is increased, the efficient solution of such PDEs becomes difficult. In this talk we introduce preconditioners for high-order discretizations. We build upon the pioneering work of Pavarino, who proved in 1993 that the additive Schwarz method with vertex patches and a low-order coarse space gives a  solver for symmetric and coercive problems that is robust to the polynomial degree. 

However, for very high polynomial degrees it is not feasible to assemble or factorize the matrices for each vertex patch, as the patch matrices contain dense blocks, which couple together all degrees of freedom within a cell. The central novelty of the preconditioners we develop is that they have the same time and space complexity as sum-factorized operator application on unstructured meshes of tensor-product cells, i.e., we can solve $A x=b$ with the same complexity as evaluating $b-A x$. Our solver relies on new finite elements for the de Rham complex that enable the blocks in the stiffness matrix corresponding to the cell interiors to become diagonal for scalar PDEs or block diagonal for vector-valued PDEs.  With these new elements, the patch problems are as sparse as a low-order finite difference discretization, while having a sparser Cholesky factorization. In the non-separable case, themethod can be applied as a preconditioner by approximating the problem with a separable surrogate.  Through the careful use of incomplete factorizations and choice of space decomposition we achieve optimal fill-in in the patch factors, ultimately allowing for optimal-complexity storage and computational cost across the setup and solution stages.

We demonstrate the approach by solving the Riesz maps of $H^1$, $H(\mathrm{curl})$, and $H(\mathrm{div})$ in three dimensions at $p = 15$.

We investigate the theoretical properties of subsampling and hashing as tools for approximate Euclidean norm-preserving embeddings for vectors with (unknown) additive Gaussian noises. Such embeddings are called Johnson-Lindenstrauss embeddings due to their celebrated lemma. Previous work shows that as sparse embeddings, if a comparable embedding dimension to the Gaussian matrices is required, the success of subsampling and hashing closely depends on the $l_\infty$ to $l_2$ ratios of the vectors to be mapped. This paper shows that the presence of noise removes such constrain in high-dimensions; in other words, sparse embeddings such as subsampling and hashing with comparable embedding dimensions to dense embeddings have similar norm-preserving dimensionality-reduction properties, regardless of the $l_\infty$ to $l_2$ ratios of the vectors to be mapped. The key idea in our result is that the noise should be treated as information to be exploited, not simply a nuisance to be removed. Numerical illustrations show better performances of sparse embeddings in the presence of noise.