# Model Order Reduction

## 1. Definition

Problem statement

Goal Replicate input-output behavior of large-scale system $\Sigma$ over a certain (restricted) range of

• forcing inputs and

• parameter inputs Problem statement Given large-scale system $\Sigma_\mathcal{N}$ of dimension $\mathcal{N}$, find a reduced order model $\Sigma_N$ of dimension $N << \mathcal{N}$ such that: The approximation error is small, i.e., there exists a global error bound such that

• $\|u(\mu) - u_N (\mu)\| \leq \varepsilon_{\mathrm{des}}$, and $|s(\mu) - s_N (\mu)| \leq \varepsilon^s_{\mathrm{des}} , \forall \mu \in D^{\mu}$.

• Stability (and passivity) is preserved.

• The procedure is computationally stable and efficient.

## 2. Motivation

Generalized Inverse Problem

• Given PDE($\mu$) constraints, find value(s) of parameter $\mu$ which:

• (OPT) minimizes (or maximizes) some functional;

• (EST) agrees with measurements;

• (CON) makes the system behave in a desired manner;

• or some combination of the above

• Full solution computationally very expensive due to a repeated evaluation for many different values of $\mu$

• Goal: or

## 3. Methodologies

Methodologies

• Reduced Basis Methods

• Proper Orthogonal Decomposition

• Balanced Truncation

• Krylov Subspace Methods

• Proper Generalized Decomposition

• Modal Decomposition

• Physical model reduction

• …​

Disclaimer Model Order Reduction Techniques

• replace your favorite discretization scheme (e.g. FE, FV, FD), but instead are build upon and supplement these schemes.

• useful if you are interested in a single high-fidelity solution of your high-dimensional problem, but instead if you are interested in the many-query or real-time context.

# Some Examples

## 1. Cooling of electronical components

Thermal Testcase Description

0.5

0.5

Overview

• Heat-Transfer with conduction and convection possibly coupled with Navier-Stokes

• Simple but complex enough to contain all difficulties to test the certified reduced basis

• non symmetric, non compliant

• physical and geometrical parameters

• coupled models

• Testcase can be easily complexified

## 2. Aerothermal flows

Airbus Use-Case Some Scientific Issues

• Turbulence

• Mixed forced and natural convection

• Boundary conditions coupled to an ECS (Environment Control System)

• Error prediction (Reduced Basis)

## 3. Modeling of high field magnets

`+`

High Field Magnet Modeling

Laboratoire National des Champs Magnétiques Intenses

Large scale user facility in France

• High magnetic field : from 24 T

• Grenoble : continuous magnetic field (36 T)

• Toulouse : pulsed magnetic field (90 T)

3.4cm

Application domains

• Magnetoscience

• Solide state physic

• Chemistry

• Biochemistry

2.4cm 3.9cm

Magnetic Field

• Earth : $5.8 \cdot 10^{-4} T$

• Supraconductors : $24 T$ *

• Pulsed field : $90 T$

Access

• Call for Magnet Time : $2 ~\times$ per year

• $\approx ~140$ projects per year

3.5cm

4cm 5cm 4.5cm Why use Reduced Basis Methods ?

Challenges

• Modeling : multi-physics non-linear models, complex geometries, genericity

• Account for uncertainties : uncertainty quantification, sensitivity analysis

• Optimization : shape of magnets, robustness of design

4.8cm

Objective 1 : Fast

• Complex geometries

• Large number of dofs

• Uncertainty quantification

• Large number of runs

4.4cm

Objective 2 : Reliable

• Field quality

• Design optimization

• Certified bounds

• Reach material limits

## 4. Summary

Summary Many problems in computational engineering require

many or real-time evaluations of PDE($\mu$)-induced
input-output relationships.

Model order reduction techniques enable

certified, real-time calculation
of outputs of PDE($\mu$)
for parameter estimation, optimization, and control.

# Reduced Basis Method

## 1. Problem Statement

The Reduced Basis Method

• <2→ Comparison to other model reduction techniques:

• Parametrized problems(material, constants, geometry,…​)

• A posteriori error estimation

• Offline-online decomposition

• Greedy algorithm (to construct reduced basis space)

• <3→ Motivation:

• Efficient solution of optimization and optimal control problems governed by parametrized PDEs.

Problem Statement

The Main Idea - Key Observation

General Problem Statement Given a system $\Sigma_\mathcal{N}$ of large dimension N, where

• $u(\mu, t) \in \mathbb{R}^{\mathcal{N}}$, the state

• $s(\mu, t)$, the outputs of interest

• $g(t)$, the forcing or control inputs

are functions of

• $\mu \in D$, the parameter inputs

• $t$, time

and the matrices $M$, $A$, $B$, and $L$ also depend on $\mu$ $\ldots$

General Problem Statement $\ldots$ construct a reduced order system $\Sigma_N$ of dimension $N << \mathcal{N}$, where $u_N(\mu) \in \mathbb{R}^N$ is the reduced state.

Special case We start by considering $\dot{u} = 0$

Full Model

• <2→ For any new $\mu$, approximate $u$ by a linear combination of the $\xi_i$ $u(\mu) \approx \sum_{i=1}^N u_{N,i}(\mu) \xi_i = Z_N u_N(\mu)]$ determined by Galerkin projection, i.e., A posteriori error estimation • <1→ Assume well-posedness; $A(\mu)$ pos.def. with min eigenvalue $\alpha_a :=\lambda_1 >0$, where $A \xi=\lambda X \xi$ and $X$ corresponds to the $X$-inner product, $(v, v)_X = \|v\|_X^2$ • <2→ Let $\underbrace{e_N = u - Z_N\ u_N}_{\text{error}}$ , and $\underbrace{r = F - A\ Z_N\ u_N}_{\text{residual}}, \forall \mu \in D$, so that $\[A(\mu) e_N (\mu) = r(\mu)]$ • <3→ Then for any $\mu \in D$, $\[\|u(\mu)- Z_N u_N(\mu) \|_X \leq \frac{\|r(\mu)\|_{X'}}{\alpha_{LB}(\mu)} =: \Delta_N(\mu)]$ $\[|s(\mu)-s_N(\mu)| \leq \|L\|_{X'} \Delta_N(\mu) =: \Delta^s_N(\mu)]$ where $\alpha_{LB}(\mu)$ is a lower bound to $\alpha_a(\mu)$, and $\|r\|_{X'}=r^T X^{-1} r$. Offline-Online decomposition Greedy Algorithm ## 3. Summary Reduced Basis Opportunities Computational Opportunities • We restrict our attention to the typically smooth and low-dimensional manifold induced by the parametric dependence. $\Rightarrow$ Dimension reduction • We accept greatly increased offline cost in exchange for greatly decreased online cost. $\Rightarrow$ Real-time and/or many-query context Reduced Basis Relevance Real-Time Context (control,$\ldots$): \[\begin{align} \mu & \rightarrow & s_N(\mu), \Delta^s_N(\mu) & \\ t_0 (``input'') & & & t_0+\delta t_{\mathrm{comp}} (``response'') \end{align}] Many-Query Context (design,$\ldots$): \[\begin{align} \mu_j & \rightarrow & s_N(\mu_j), \Delta^s_N(\mu_j),\quad j=1\ldots J \\ t_0 & & t_0+\delta t_{\mathrm{comp}} J\quad (J \rightarrow \infty) \end{align}] $\Rightarrow$ (real-time) and/or (many-query) . Reduced Basis Challenges • A Posteriori error estimation • Rigorous error bounds for outputs of interest • Lower bounds to the stability ``constants'' • Offline-online computational procedures • Full decoupling of finite element and reduced basis spaces • A posteriori error estimation • Nonaffine and nonlinear problems • Effective sampling strategies • High parameter dimensions Reduced Basis Outline 1. Affine Elliptic Problems • (non)symmetric, (non)compliant, (non)coercive • (Convection)-diffusion, linear elasticity, Helmholtz 2. Affine Parabolic Problems • (Convection)-diffusion equation 3. Nonaffine and Nonlinear Problems • Nonaffine parameter dependence, nonpolynomial nonlinearities 4. Reduced Basis (RB) Method for Fluid Flow • Saddle-Point Problems (Stokes) • Navier-Stokes Equations 5. Applications • Parameter Optimization and Estimation (Inverse Problems) • Optimal Control === Linear Compliant Elliptic Problems # Notations, Definitions, Problem Statement, Example ## 1. Inner Product Spaces Definitions A space $Z$ is a linear or vector space if, for any $\alpha \in \mathbb{R}$ , $w,v \in Z$, $\alpha w+v \in Z$ Note: $\mathbb{R}$ denotes the real numbers, and $\mathbb{N}$ and $\mathbb{C}$ shall denote the natural and complex numbers, respectively. An inner product space (or Hilbert space) $Z$ is a linear space equipped with • an inner product $(w,v)_Z, \forall w,v \in Z$,and • induced norm $\|w\|_Z = (w,w)_Z, \forall w \in Z$. Inner Product An inner product $w,v \in Z \rightarrow (w,v)_Z \in \mathbb{R}$ has to satisfy • Bilinearity $(\alpha w+v,z)_Z =\alpha(w,z)_Z +(v,z)_Z \forall \alpha\in R,w,v,z\in Z$ `+` $(z,\alpha w+v)_Z =\alpha(z,w)_Z +(z,v)_Z, \forall \alpha\in R, w,v,z\in Z$ • Symmetry $(w,v)_Z = (v,w)_Z, \forall w,v \in Z$ • Positivity $(w,w)_Z >0, \forall w \in Z, w \neq 0$ `+` $(w,w)_Z =0, \text{ only if } w=0$ Cauchy-Schwarz inequality: $\[(w,v)_Z \leq \|w\|_Z\|v\|_Z,\forall w, v \in Z.]$ Norm A norm is a map $\| \cdot \| : Z \rightarrow \mathbb{R}$ such that • $\|w\|_Z > 0\quad \forall w\in Z,w\neq 0,$ • $\|\alpha w\|_Z = |\alpha |\|w\|_Z\quad \forall \alpha \in \mathbb{R},\ \forall w\in Z,$ • $\|w+v\|_Z \leq \|w\|_Z +\|v\|_Z\quad \forall w\in Z,\ \forall v\in Z.$ Equivalence of norms $\| \cdot \|_Z$ and $\| \cdot \|_Y$ : there exist positive constants $C_1$, $C_2$ such that $\[C_1\|v\|_Z \leq \|v\|_Y \leq C_2\|v\|_Z .]$ Cartesian Product Space Given two inner product spaces $Z_1$ and $Z_2$, we define $\[Z = Z_1 \times Z_2 \equiv \{(w_1,w_2)\ | \ w_1 \in Z_1,\ w_2 \in Z_2\}]$ and given $w = (w_1,w_2) \in Z, v = (v_1,v_2) \in Z$, we define $\[w + v \equiv (w_1 + v_1, w_2 + v_2).]$ We also equip $Z$ with the inner product $\[(w,v)_Z =(w_1,v_1)_{Z_1} +(w_2,v_2)_{Z_2}]$ and induced norm $\[\|w\|_Z = (w,w)_Z.]$ ## 2. Linear and Bilinear Forms Linear Forms A functional $g : Z \rightarrow \mathbb{R}$ is a linear functional if, for any $\alpha \in \mathbb{R}, w, v \in Z$ $\[g(\alpha w + v) = \alpha g(w) + g(v)]$ A linear form is bounded, or continuous, over $Z$ if $\[|g(v)| \leq C \|v\|_Z, \forall v \in Z,]$ for some finite real constant $C$. Dual Spaces Given $Z$, we define the dual space $Z'$ as the space of all bounded linear functionals over $Z$. We associate to $Z'$ the dual norm $\[\|g\|_{Z'} = \sup_{v \in Z} \frac{g(v)}{\|v\|_Z} , \forall g \in Z'.]$ For any $g \in Z'$, there exists a unique $w_g \in Z$ such that $\[(w_g, v)_Z =g(v), \forall v \in Z.]$ It directly follows that $\[\|g\|_{Z'} = \|w_g\|_Z.]$ Bilinear Forms A form $b:Z_1 \times Z_2 \rightarrow \mathbb{R}$ is bilinear if, for any $\alpha \in R$, • $b(\alpha w + v,z) = \alpha b(w,z) + b(v,z), \forall w,v \in Z_1, z \in Z_2$ • $b(z,\alpha w + v) =\alpha b(z,w) + b(z,v), \forall z \in Z_1, w,v \in Z_2$ The bilinear form $b : Z \times Z \rightarrow \mathbb{R}$ is • symmetric, if $\[b(w,v) = b(v,w),]$ • skew-symmetric, if $\[b(w,v) = -b(v,w),]$ • positive definite, if $\[b(v,v) \geq 0\text{ , with equality only for } v = 0.]$ Bilinear Forms The bilinear form $b : Z \times Z \rightarrow \mathbb{R}$ is positive semidefinite, if $\[b(v,v) \geq 0, \forall v \in Z.]$ We also define, for a general bilinear form $b : Z \times Z \rightarrow \mathbb{R}$, the • symmetric part as $\[b_S(w,v) = 1/2 (b(w,v) + b(v,w)), \forall w,v \in Z;]$ • the skew-symmetric part as $\[b_{SS}(w,v) = 1/2 (b(w,v) - b(v,w)), \forall w,v \in Z.]$ Bilinear Forms The bilinear form $b : Z \times Z \rightarrow \mathbb{R}$ is • over $Z$ if $\[\alpha \equiv \inf_{w\in Z} \frac{b(w,w)}{\|w\|^2_Z}]$ is positive; • over $Z$ if $\[\gamma \equiv \sup_{w\in Z} \sup_{v\in Z} \frac{b(w, v)}{\|w\|_Z \|v\|_Z}]$ is finite. Parametric Linear and Bilinear Forms We introduce • $D \in \mathbb{R}^P$ : closed bounded parameter domain; • $\mu = (\mu_1,\ldots,\mu_P) \in D$ : parameter vector. We shall say that • $g:Z\times D\rightarrow \mathbb{R}$ is a if, for all $\mu \in D, g( \cdot ; \mu) : Z \rightarrow \mathbb{R}$ is a linear form; • $b:Z\times Z\times D\rightarrow \mathbb{R}$ is a if,for all $\mu \in D, b( \cdot , \cdot ; \mu) : Z \times Z \rightarrow \mathbb{R}$ is a bilinear form. Concepts of symmetry,$\ldots$ directly extend to the parametric case. Parametric Linear and Bilinear Forms The parametric bilinear form $b : Z \times Z \times D \rightarrow \mathbb{R}$ is • coercive over Z if $\[\alpha(\mu) \equiv \inf_{w \in Z} \frac{b(w,w;\mu)}{\|w\|^2_Z}]$ is positive for all $\mu \in D$; • continuous over $Z$ if $\[\gamma(\mu)\equiv \sup_{w \in Z} \sup_{v \in Z} \frac{b(w, v; \mu)}{\|w\|_Z\|v\|_Z}]$ is finite for all $\mu \in D.$ We also define \[\begin{align} (0 <) \alpha _0 & \equiv \min_{\mu \in D} \alpha (\mu)\\ \gamma_0 & \equiv \max_{\mu \in D} \gamma (\mu) (< \infty ). \end{align}] Coercivity EigenProblem We have $\[\alpha (\mu) \equiv \inf_{w \in Z} \frac{b_S(w,w;\mu)}{\|w\|^2_Z}]$ Associated generalized eigenproblem: Given $\mu \in D$, find $(\chi^{co},\nu^{co})_i(\mu) \in Z \times \mathbb{R}, 1 \leq i \leq \dim(Z),$ such that $\[b_S(\chi_i^{co}(\mu), v; \mu) = \nu_i^{co}(\mu)(\chi_i^{co}(\mu), v)_Z]$ and $\[\|\chi_i^{co}(\mu)\|_Z=1]$ Let $\nu_1^{co}(\mu) \leq \nu_2^{co}(\mu) \leq \ldots \leq \nu_{\dim{Z}}^{co} (\mu)$ and b coercive, then $\[\alpha (\mu) = \nu_1^{co}(\mu) > 0.]$ Parameter affine Dependence We assume $\[g(v;\mu)= \sum_{q=1}^{Q_g} \theta^q_g(\mu)g^q(v), \forall v \in Z,]$ where, for $1 \leq q \leq Q_g$ (finite), • functions $\theta^q_g : D \rightarrow \mathbb{R}$, • forms $g^q : Z \rightarrow \mathbb{R};$ and $\[b(w,v;\mu)= \sum_{q=1}^{Q_b} \theta^q_b(\mu) b^q(w,v),\quad \forall w,v \in Z,]$ where, for $1 \leq q \leq Q_b$ (finite), • functions $\theta^q_b : D \rightarrow \mathbb{R}$, • forms $b^q : Z \times Z \rightarrow \mathbb{R}$. Parametric Coercivity The coercive bilinear form $b : Z \times Z \times D \rightarrow \mathbb{R}$ $\[b(w,v;\mu)= \sum_{q=1}^{Q_b} \theta^q_b(\mu) b^q(w,v),\quad \forall w,v \in Z,]$ is if $c\equiv b_S$ is affine $\[c(w,v;\mu)= \sum_{q=1}^{Q_c} \theta^q_c(\mu) c^q(w,v),\quad \forall w,v \in Z,]$ and satisfies and • $\theta^q_c(\mu)>0, \forall \mu \in D, 1\leq q\leq Q_c,$ • $c^q(v,v)\geq 0,\forall v \in Z, 1\leq q\leq Q_c.$ ## 3. Classes of Functions Scalar and Vector Fields We consider (real) • scalar-valued field variables (e.g., temperature, pressure) $w : \Omega \rightarrow \mathbb{R}^{d=1}$ • vector-valued field variables (e.g., displacement, velocity) $\mathbf{w} : \Omega \rightarrow \mathbb{R}^d$ , where $\mathbf{w}(x) = (w_1(x), \ldots , w_d (x));$ and • $\Omega \in \mathbb{R}^d, d=1, 2, \text{or} 3$ is an open bounded domain • $x = (x_1,...,x_d) \in \Omega$; • $\Omega$ has Lipschitz continuous boundary $\partial \Omega$ ; and • we define the canonical basis vectors as $e_i, 1 \leq i \leq d.$ Multi-Index Derivative Given a scalar (or one component of a vector) • field $w : \Omega \rightarrow \mathbb{R}$SPATIAL DERIVATIVE $\[(D^\sigma w)(x) = \frac{\partial^\sigma w}{\partial x_1^{\sigma_1} ...\partial x_d^{\sigma d}}]$ • field $w : \Omega \times D \rightarrow \mathbb{R}$ SENSITIVITY DERIVATIVE $\[(D_\sigma w)(x) = \frac{\partial^\sigma w}{\partial \mu_1^{\sigma_1} ...\partial \mu_d^{\sigma d}}]$ where • $\sigma = (\sigma_1,\ldots,\sigma_d)$, $\sigma_i, 1 \leq i \leq d$, non-negative integers; • $|\sigma| = \sum_{j=1}^{d} \sigma_j$ is the order of the derivative; and • $I^{d,n}$ is set of all index vectors $\sigma \in N^d_0$ such that $|\sigma | \leq n.$ Function Spaces Let $m \in N_0$, the space $C^m(\Omega )$ is defined as $\[C^m(\Omega )\equiv \{w | D^\sigma w \in C^0(\Omega ), \forall \sigma \in I^{d,m}\},]$ and $C^0(\Omega )$ is the space of continuous functions over $\Omega \in \mathbb{R}^d$. We denote by $C^\infty (\Omega )$ the space of functions $w$ for which $D^\sigma$ exists and is continuous for any order $|\sigma |.$ Lebesgue Spaces We define, for $1 \leq p < \infty$ , the Lebesgue space $L^p(\Omega )$ as $\[L^p(\Omega )\equiv \{ w \text{ measurable } |\quad \|w\|_{L^p(\Omega )} < \infty\}]$ where • $\|w\|_{L^p(\Omega )} \equiv \left( \int_\Omega |w|^pdx\right)^{1/p} , 1\leq p<\infty,$ • $\|w\|_{L^\infty (\Omega )} \equiv \mathrm{ess} \sup_{x\in\Omega} |w(x)|, p = \infty .$ Hilbert Space Let $m \in \mathbb{N}_0$, the space $H^m(\Omega )$ is then defined as $\[H^m(\Omega )\equiv \{w |\quad D^\sigma w \in L^2(\Omega ), \forall \sigma \in I^{d,m}\},]$ with associated inner product $\[(w,v)_{H^m(\Omega )}\equiv \sum_{\sigma \in I^{d,m}}\int_\Omega D^\sigma w D^\sigma v dx,]$ and induced norm $\[\|w\|_{H^m(\Omega )} \equiv \sqrt{(w, w)_{H^m(\Omega )}}.]$ Special (most important) cases Since we only consider , we require mostly • $L^2(\Omega ) = H^0(\Omega )$: Lebesgue Space $p = 2$ $\[(w,v)_{L^2(\Omega)} = \int_\Omega w v \quad \forall w, v \in L^2(\Omega )]$ $\[\|w\|_{L^2(\Omega)} = \sqrt{(w,w)_{L^2(\Omega)}} \forall w \in L^2(\Omega ),]$ $\Rightarrow$ Space of all functions $w : \Omega \rightarrow \mathbb{R}$ square-integrable over $\Omega$ . Special (most important) cases Since we only consider , we require mostly • $H^1(\Omega)$ $\[H^1(\Omega ) \equiv \{w \in L^2(\Omega )| \frac{\partial w}{ \partial xi} \in L^2(\Omega ), 1\leq i\leq d\}]$ with inner product and induced norm $\[(w,v)_{H^1(\Omega )} \equiv \int_\Omega \nabla w \cdot \nabla v + wv\quad \forall w,v \in H^1(\Omega ),]$, $\[\|w\|_{H^1(\Omega )} \equiv \sqrt{(w,w)_{H^1(\Omega)}}\quad \forall w \in H^1(\Omega ),]$ and seminorm $\[|w|_{H^1(\Omega )} \equiv \int_\Omega \nabla w \cdot \nabla w,\quad \forall w \in H^1(\Omega ).]$ Special (most important) cases Since we only consider , we require mostly • the space $H_0^1(\Omega )$ $\[H^1_0(\Omega) \equiv \{v \in H^1(\Omega )|v_{|\partial \Omega}=0 \}]$ where $v = 0$ on the boundary $\partial \Omega .$ Note that, for any $v \in H_0^1(\Omega )$, we have $\[C_{PF} \|v\|_{H^1(\Omega )} \leq |v|_{H^1(\Omega )} \leq \|v\|_{H^1(\Omega )},]$ and thus $\[\|v\|_{H^1(\Omega)} = 0 \, \Rightarrow v = 0]$ $\Rightarrow |v|_{H^1(\Omega )}$ constitutes a norm for $v \in H_0^1(\Omega ).$ Projection Given Hilbert Spaces $Y$ and $Z \subset Y$ , the projection, $\Pi : Y \rightarrow Z$, of $y \in Y$ onto $Z$ is defined as $\[(\Pi y,v)_Y = (y,v)_Y , \forall v \in Z]$ Properties: • Orthogonality: $(y - \Pi y, v)_Y = 0$ • Idempotence: $\Pi (\Pi y) = \Pi y$ • Best Approximation $\|y-\Pi y\|^2_Y = \inf_{v \in Z} \|y-v\|^2_Y, \,$ Given an orthonormal basis $\{ \varphi i\}_{i=1, N = \dim(Z)}$, then $\[\Pi y= \sum_{i=1}^{\dim(Z)} ( \varphi i,y)_Y \varphi_i, \forall y \in Y]$ ## 4. Notations Notations and Definitions Notations • $(\cdot)^\mathcal{N}$ finite element approximation • $(\cdot)_N$ reduced basis approximation • $\mu$ input parameter (physical, geometrical,…​) • $s(t;\mu) \approx s^\mathcal{N}(t;\mu)\approx s_N(t;\mu)$ output approximations • $\mu \rightarrow s(t;\mu)$ input-output relationship Definitions • $\Omega \subset \mathbb{R}^d$ spatial domain • $\mu$ $P$-uplet • $\mathcal{D}^\mu \subset \mathbb{R}^P$ parameter space • $s$ output, $\ell, f$ functionals • $u$ field variable • $X$ function space $H^1_0(\Omega)^\nu \subset X \subset H^1(\Omega)^\nu$ ($\nu=1$ for simplicity) $(\cdot,\cdot)_X$ scalar product and $\|\cdot\|_X$ norm associated to $X$ ## 5. Problem Statement Problem Statement The formal problem statement reads: Given $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$, evaluate $\[s(\mu) = \ell(u(\mu);\mu)]$ where $u(x;\mu) \in X$ satisfies $\[a(u(\mu), v; \mu ) = f(v; \mu), \quad \forall v \in X]$ [rem:problem-statement] We consider first the case of linear affine compliant elliptic problem and then complexify Hypothesis: Reference Geometry In these notes $\Omega$ is considered • To apply the reduced basis methodology exposed later, we need to setup a reference spatial domain $\Omega_{\mathrm{ref}}$ • We introduce an affine mapping $\matcal{T}(\cdot;\mu) : \Omega (\equiv \Omega_{\mathrm{ref}} = \Omega_o(\bar{\mu})) \rightarrow \Omega_o(\mu)$ such that $\[a(u,v;\mu) = a_o(u_o \circ \mathcal{T}_\mu,v_o \circ \mathcal{T}_\mu;\mu)]$ Hypothesis: Continuity, stability, compliance We consider the following $\mu-$PDE rl $a(\cdot,\cdot;\mu)$ & bilinear & symmetric & continuous & coercive ($\forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$) $f(\cdot;\mu), \ell(\cdot;\mu)$ & linear & bounded ($\forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$) and in particular, to start, the compliant case • $a$ symmetric • $f(\cdot;\mu) = \ell(\cdot;\mu)\quad \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ Hypothesis: Affine dependence in the parameter We require for the RB methodology $\[a(u,v;\mu) = \sum_{q=1}^{Q_a} \Theta^q_a(\mu)\ a^q( u, v ),]$ where for $q=1,...,Q_a$ $\[\begin{array}[c${rll} \Theta^q_a :& {\ensuremath{\mathcal{D}^\mu}\xspace}\rightarrow \mathbb{R} & \mu-\text{\alert{dependent} functions}\\ a^q :& X \times X \rightarrow \mathbb{R} & \mu-\text{\alert{independent} bilinear forms} \end{array}]

• similar decomposition is required for $\ell(v;\mu)$ and $f(v;\mu)$, and denote $Q_\ell$ and $Q_f$ the corresponding number of terms

• applicable to a large class of problems including geometric variations

• can be relaxed (see non affine/non linear problems)

Inner Products and Norms

• and associated norm () \begin{aligned} (((w,v)))_\mu &= a(w,v;\mu) &\ \forall u,v \in X\\ |||v|||_\mu &= \sqrt{a(v,v;\mu)} &\ \forall v \in X \end{aligned}] • $X$-inner product and associated norm () \[\begin{aligned} (w,v)_X &= (((w,v)))_{\bar{\mu}} \ (\equiv a(w,v;\bar{\mu})) &\ \forall u,v \in X\\ ||v||_X &= |||v|||_{\bar{\mu}} \ (\equiv \sqrt{a(v,v;\bar{\mu})}) & \ \forall v \in X \end{aligned}] Coercivity and Continuity Constants We assume $a$ and Recall that • constant $\[(0 < ) \alpha(\mu) \equiv \inf_{v\in X}\frac{a(v,v;\mu)}{||v||^2_X}]$ • constant $\[\gamma(\mu) \equiv \sup_{w\in X} \sup_{v\in X}\frac{a(w,v;\mu)}{\|w\|_X \|v\|_X} ( < \infty)]$ ## 6. Example ### 6.1. Example Thermal Block: Heat Transfer 6 (0,0) rectangle (3,3); (0,0) grid (3,3); (0,0) rectangle (3,3); (1.5,-0.5)node[right]$\Gamma_0$ (Heat Flux) to[out=180,in=90] (1.5,0); (1.5,3.5)node[right]${\Gamma_{\mathrm{top}}}$ (Zero Dirichlet) to[out=180,in=90] (1.5,3); (3.5,1.5) node[right]${\Gamma_{\mathrm{sides}}}$ (Insulated) to[out=180,in=90] (0,1.5) ; (3.5,1.5) to[out=180,in=-90] (3,1.5); in 5mm,15mm,25mm in 5mm,15mm,25mm at (+0.45,+0.3) $\mu_\theindex$; 3 Example Thermal Block: Problem statement Given $\mu \in (\mu_1,...\mu_P) \in {\ensuremath{\mathcal{D}^\mu}\xspace}\equiv [\mu^{\text{min}},\mu^{\text{max}}$^P], evaluate (recall that $\ell = f$) $\[s(\mu) = f(u(\mu))]$ where $u(\mu) \in X \equiv \{ v \in H^1(\Omega), v|_{\Gamma_{\text{top}}} = 0\}$ satisfies $\[a(u(\mu), v; \mu) = f(v;\mu) \quad \forall v \in X]$ we have $P = 8$ and given $1 < \mu_r < \infty$ we set $\[\mu^{\mathrm{min}} = 1/\sqrt{\mu_r},\quad \mu^{\mathrm{max}} = \sqrt{\mu_r}]$ such that $\mu^{\mathrm{max}}/\mu^{\mathrm{min}}=\mu_r.$ Example Thermal Block Recall we are in the compliant case $\ell = f$, we have $\[f(v) = \int_{\Gamma_{0}} v\quad \forall v \in X]$ and $\[a(u,v;\mu) = \sum_{i=1}^{P} \mu_i \int_{\Omega_i} \nabla u \cdot \nabla v + 1 \int_{\Omega_{P+1}} \nabla u \cdot \nabla v \quad\forall u,\ v\ \in X]$ where $\Omega = \cup_{i=1}^{P+1} \Omega_i$. Example Thermal Block The inner product is defined as follows $\[(u,v)_X = \sum_{i=1}^P \bar{\mu}_i \int_{\Omega_i}\nabla u \cdot \nabla v + 1 \int_{\Omega_{P+1}} \nabla u \cdot \nabla v]$ where $\bar{\mu}_i$ is a . We have readily that $a$ is * * $\[0 < \frac{1}{\sqrt{\mu_r}} \leq \mathrm{min}(\mu_1/\bar{\mu}_1, \ldots, \mu_P/\bar{\mu}_P,1) \leq \alpha(\mu)]$ * and $\[\gamma(\mu) \leq \mathrm{max}(\mu_1/\bar{\mu}_1, \ldots, \mu_P/\bar{\mu}_P,1) \leq \sqrt{\mu_r} < \infty]$ and the linear form $f$ is . Example Thermal Block: Affine decomposition We $\[a(u,v;\mu) = \sum_{q=1}^{P+1} \Theta^q(\mu) a^q(u,v)]$ with \[\begin{aligned} \Theta^1(\mu) = \mu_1 & & a^1(u,v) = \int_{\Omega_1} \nabla u \cdot \nabla v\\ & \vdots & \\ \Theta^P(\mu) = \mu_P & & a^P(u,v) = \int_{\Omega_P} \nabla u \cdot \nabla v\\ \Theta^{P+1}(\mu) = 1 & & a^{P+1}(u,v) = \int_{\Omega_{P+1}} \nabla u \cdot \nabla v \end{aligned}] Example Thermal Block • 0.5 ** 0.5 ** • ``Truth'' FEM Approximation Let $\mu \in \mathcal{D}^{\mu}$, evaluate $\[\displaystyle s^{\mathcal{N}} (\mu) = \ell (u^{\mathcal{N}} (\mu)) \ ,]$ where $u^{\mathcal{N}} (\mu) \in X^{\mathcal{N}}$ satisfies $\[a (u^{\mathcal{N}} (\mu), v; \mu ) = f (v), \quad \forall \: v \in X^{\mathcal{N}} \ .]$ Here $X^{\mathcal{N}} \subset X$ is a finite element approximation of dimension equiped with an inner product $(\cdot,\cdot)_X$ and induced norm $||\cdot||_X$. Denote also $X'$ and associated norm $\[\ell \in X',\qquad\displaystyle ||\ell||_{X'} \equiv \operatorname{sup}_{v\in X}\frac{\ell(v)}{||v||_X}]$. Purpose • $u(\mu)$ and $u_{\mathcal{N}}(\mu)$ in the sense that $\[||u(\mu)-u_{\mathcal{N}}(\mu)||_X \leq \mathrm{tol}\quad\forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}]$ • the reduced basis approximation using the FEM approximation • the error associated with the reduced basis approximation relative to the FEM approximation $\Rightarrow u^{\mathcal{N}} (\mu)$ is a calculable surrogate for $u(\mu).$ $\[\|u(\mu)-u^\mathcal{N}(\mu)\|_{X} \leq \underbrace{\|u(\mu)-u^\mathcal{N}(\mu)\|_{X}}_{\leq \varepsilon^\mathcal{N}} + \underbrace{\|u^\mathcal{N}(\mu)-u^N(\mu)\|_X}_{\varepsilon_{\mathrm{tol,min}}}]$ with $\varepsilon^\mathcal{N} << \varepsilon_{\mathrm{tol,min}}$ # Reduced Basis Approximation Reduced Basis Objectives For given accuracy $\epsilon$, evaluate $\[\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}\rightarrow s_N(\mu) (\approx s^\mathcal{N}(\mu)) \text{ and } \Delta^s_N(\mu)]$ that achieves the desired accuracy Reliability $\[|s^\mathcal{N}(\mu)-s_N(\mu)| \leq \Delta^s_N(\mu) \leq \epsilon]$ for a $t_{\textsc{comp}}$ Efficiency $\[\text{\alert{Independent} of } \mathcal{N} \text{ as } \mathcal{N} \rightarrow \infty]$ where $t_{\textsc{comp}}$ is the time to perform the input-output relationship $\[\hfill\mu \rightarrow (s_N(\mu),\Delta^s_N(\mu))]$ Reduced Basis Objective : Rapid Convergence Build a rapidly convergent approximation of $\[s_N(\mu) \in \mathbb{R} \text{ and } u_N(\mu) \in X^N \subset X^{\mathcal{N}} \subset X]$ such that for all $\mu$, we have $\[s_N(\mu) \rightarrow s^{\mathcal{N}}(\mu) \text{ and } u_N(\mu) \rightarrow u^{\mathcal{N}}(\mu)]$ rapidly as $N = {\ensuremath{{\operatorname{dim}}}\xspace}{X_N} \rightarrow \infty (= 10-200)$ (and of $\mathcal{N}$) Reduced Basis Objective : Reliability and Sharpness Provide error bound $\Delta_N(\mu)$ and $\Delta^s_N(\mu)$ : $\[1 (\text{rigor}) \leq \frac{\Delta_N(\mu)}{\|u^{\mathcal{N}}(\mu) - u_N(\mu)\|_X} \leq \ E (\text{sharpness})]$ and $\[1 (\text{rigor}) \leq \frac{\Delta^s_N(\mu)}{|s^{\mathcal{N}}(\mu) - s_N(\mu)|} \leq \ E (\text{sharpness})]$ for all $N = 1 \ldots N_{\textsc{max}}$ and $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$. Reduced Basis Objective : Efficiency Develop a two stage strategy : Offline/Online Offline very expensive pre-processing, we have typically that for a given $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ $\[t^{\textsc{offline}}_{\textsc{comp}} >> t^{\mu\rightarrow s^{\mathcal{N}}(\mu)}_{\textsc{comp}}]$ Online very rapid convergent certified reduced basis input-output relationship $\[t^{\textsc{online}}_{\textsc{comp}} \text{ independent of } \mathcal{N}]$ [rem:rbobjectives-efficiency] $\mathcal{N}$ may/should be chosen Parametric Manifold $\mathcal{M}^\mathcal{N}$ We assume • the form $a$ is continuous and coercive (or inf-sup stable); and • affine -dependence; and • the $\theta^q(\mu), 1 \leq q \leq Q$, are smooth (i.e., $\theta^q \in C^\infty(\mathcal{D})$ ; then $\[\mathcal{M}^\mathcal{N} = \{ u^\mathcal{N}(\mu),\, \mu \in \mathcal{D}\}]$ is a smooth $P$-dimensional manifold in $X^\mathcal{N}$, since $\[\| D_\sigma y^\mathcal{N}(\mu) \| \leq C_\sigma \forall \mu \in \mathcal{D}, \text{ for any order } |\sigma| \in \mathbb{N}_{+0}]$ <1-3>Approximation opportunities: Low-Dimension Manifold 5 (0,0,0) – (1,0,0); (0,0,0) – (0,1,0); (0,0,0) – (0,0,1) node[right]$Y \equiv H^1(\Omega \subset \mathbb{R}^d)$; 5 Spaces & Bases We define the RB approximation space $\[X_N =\operatorname*{span}\{\xi^n, 1 \leq n \leq N \},\, 1 \leq N \leq N_{max}]$ with linearly independent basis functions $\[\xi^n \in X,\, 1 \leq n \leq N_{max}]$ We thus obtain $\[X_N \subset X, \, {\ensuremath{{\operatorname{dim}}}\xspace}(X_N) = N,\, 1 \leq N \leq N_{max}]$ and $\[X_1 \subset X_2 \subset \ldots X_{N_{max}} (\subset X)]$ We denote non-hierarchical RB spaces as $X^{nh}_N, 1 \leq N \leq Nmax,$ $\[X^{nh}_N \subset X, \, {\ensuremath{{\operatorname{dim}}}\xspace}(X^{nh}_N) = N,\, 1 \leq N \leq N_{max}]$ Spaces & Bases - Lagrangian Parameter Samples: $\[\mbox{\alert{Sample}}: \ \ S_N = \{ \mu_1 \in \mathcal{D}^{\mu}, \ldots, \mu_N \in \mathcal{D}^{\mu} \}\quad 1 \leq N \leq N_{\mathrm{max}},]$ with $\[S_1 \subset S_2 \ldots S_{N_\mathrm{max}-1} \subset S_{N_\mathrm{max}} \subset {\ensuremath{\mathcal{D}^\mu}\xspace}]$ Lagrangian Hierarchical Space $\[W_N = {\rm span} \: \{ \xi^n \equiv \underbrace{ u (\mu^n)}_{u^{\mathcal{N}} (\mu^n)}, n = 1, \ldots, N \}.]$ with $\[W_1 \subset W_2 \ldots \subset W_{N_\mathrm{max}} \subset X^{\mathcal{N}} \subset{X}]$ Sampling strategies? • Equidistributed points in $\mathcal{D}^\mu$(curse of dimensionality) • Log-random distributed points in $\mathcal{D}^\mu$ • See later for more efficient, adaptive strategies Space & Bases - Taylor & Hermite • Taylor reduced basis spaces: $\[W^{Taylor}_N = \operatorname*{span}\{D_\sigma u(\mu), \forall \sigma \in I^{P,N-1} \}, 1 \leq N \leq N_{max},]$ field variable sensitivity derivatives • Hermite reduced basis spaces: $\[W^{Hermite}_N ``='' W^{Lagrangian}_N \cup W^{Taylor}_N]$ field variable sensitivity derivatives Note: We will exclusively use Lagrangian RB spaces in this course. Space & Bases - Orthogonal Basis Given $\xi^n = u(\mu^n), 1 \leq n \leq N_{max}$ (Lagrange case) we construct the basis set $\{\zeta^n, 1 \leq n \leq Nmax\}$, from ^1 = 1/1_X; for n = 2 : Nmax z^n =^n- _m=1^n-1 (n,m )_X ^m; ^n = zn/zn_X; end. Note: $(\zeta^n,\zeta^m)_X = \delta_{nm}, 1 \leq n,m \leq Nmax$ Space & Bases - Orthogonal Basis Given reduced basis space $\[X_N = {\rm span} \: \{ \zeta^n, n = 1, \ldots, N \}, 1 \leq N \leq N_{max}]$ we can express any $w_N \in X_N$ as $\[w_N = \sum_{k=1}^N {w_N}_n \zeta^n]$ for unique ${w_N}_n \in \mathbb{R}, 1 \leq n \leq N$ Reduced basis ``matrices'' $Z_N \in \mathbb{R}^{\mathcal{N}\times N} , 1 \leq N \leq N_{max}:$ $\[Z_N=[\zeta^1,\zeta^2,...,\zeta^N$, 1 \leq N \leq N_{max}] where, from orthogonality, $Z^T_{N_{max}} X Z^T_{N_{max}} = I_{N_{max}},$ and $I_M$ is the Identity matrix in $\mathbb{R}^{M\times M}$.

Formulation (Linear Compliant Case): a Galerkin method

Galerkin Projection Given $\mu \in \mathcal{D}^{\mu}$ evaluate

$\label{eq:1} s_N (\mu) = f(u_N (\mu);\mu)]$ where $u_N (\mu) \in X_N$ satisfies $\[a (u_N (\mu), v; \mu) = f(v;\mu), \ \forall \: v \in X_N \ .]$ Formulation (Linear Compliant Case): Optimality For any $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$, we have the following optimality results (thanks to Galerkin) \[\begin{aligned} |||u(\mu) - u_N(\mu)|||_{\mu} &= \inf_{v_N \in X_N} |||u(\mu) - v_N(\mu)|||_\mu,\\ ||u(\mu) - u_N(\mu)||_X &\leq \sqrt{\frac{\gamma(\mu)}{\alpha(\mu)}} \inf_{v_N \in X_N} ||u(\mu) - v_N(\mu)||_X,\\ \end{aligned}] and \[\begin{aligned} s(\mu)-s_N(\mu) &= |||u(\mu) - u_N(\mu)|||^{\alert{2}}_\mu,\\ &= \inf_{v_N \in X_N} |||u(\mu) - v_N(\mu)|||^{\alert{2}}_\mu, \end{aligned}] and finally $\[0 \leq s(\mu)-s_N(\mu) \leq \gamma(\mu)\inf_{v_N \in X_N} ||u(\mu) - v_N(\mu)||^{\alert{2}}_X]$ Formulation (Linear Compliant Case): offline-online decomposition our RB approximations: \[\begin{aligned} u_N(\mu)\ =&\ \sum_{j=1}^N\ {u_N}_j(\mu)\ \zeta_j\label{eq:4} \end{aligned}] $s_N(\mu)$ $\[\label{eq:5} \displaystyle s_N(\mu) = \displaystyle\sum_{j=1}^N {u_N}_j(\mu)\ \left\{ \sum_{q=1}^{Q_f}\ \Theta^q_f(\mu)\ f^q(\zeta_j)\right\}]$ where ${u_N}_i(\mu), 1 \leq i \leq N$ satisfies \[\begin{aligned} \sum_{j=1}^N \left\{ \sum_{q=1}^{Q_a}\ \Theta^q_a(\mu)\ a^q( \zeta_i, \zeta_{j}) \right\} {u_N}_j(\mu) =& \sum_{q=1}^{Q_f}\ \Theta^q_f(\mu)\ f^q(\zeta_i),\notag \\ & 1 \leq i \leq N \label{eq:6}\\ \end{aligned}] Formulation (Linear Compliant Case): matrix form Solve $\[\label{eq:10} \underline{A}_N (\mu) \: \underline{u}_N (\mu) = \underline{F}_N]$ where \[\begin{aligned} (A_N)_{i \: j} (\mu) &= \sum_{q=1}^{Q_a}\ \Theta^q_a(\mu)\ a^q( \zeta_i, \zeta_{j}) , \\ & \\ F_{N \: i} &= \sum_{q=1}^{Q_f}\ \Theta^q_f(\mu) f^q (\zeta_i) \ . \\[.5ex & 1 \leq i,j \leq N, \quad 1 \leq i \leq N \end{aligned}]

Formulation (Linear Compliant Case): complexity analysis

Offline: independent of $\mu$

• Solve: $N$ FEM system depending on $\mathcal{N}$

• Form and store: $f^q (\zeta_i)$

• Form and store: $a^q( \zeta_i, \zeta_{j})$

Online: independent of $\mathcal{N}$

• Given a new $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$

• Form and solve $A_N(\mu)$ : $O(Q N^2)$ and $O(N^3)$

• Compute $s_N(\mu)$

Online: $N << \mathcal{N}$ Online we realize often orders of magnitude computational economies relative to FEM in the context of

Formulation (Linear Compliant Case): Condition number

[prop:1] Thanks to the orthonormalization of the basis function, we have that the condition number of $A_N(\mu)$ is bounded by the ratio $\gamma(\mu)/\alpha(\mu)$.

• Write the Rayleigh Quotient $\frac{v_N^T A_N(\mu) v_N}{v_N^T v_N}, \quad \forall v_N \in \mathbb{R}^N]$ • Express $\[v_N = \sum_{n=1}^N v_{N_n} \zeta^n]$ • Use coercivity, continuity and orthonormality. # A Posteriori Error Estimation ## 1. Motivations & Preliminaries ``Truth'' Problem statement Let $\mu \in \mathcal{D}^{\mu}$, evaluate $\[\displaystyle s (\mu) = \ell (u (\mu)) \ ,]$ where $u (\mu) \in X$ satisfies $\[a (u (\mu), v; \mu ) = f (v), \quad \forall \: v \in X \ .]$ Assumptions • linearity, coercivity, continuity • affine parameter dependence; and • compliance: $\ell=f$, $a$ symmetric Reduced Basis Sample and Space Parameter Samples: $\[\mbox{\alert{Sample}}: \ \ S_N = \{ \mu_1 \in \mathcal{D}^{\mu}, \ldots, \mu_N \in \mathcal{D}^{\mu} \}\quad 1 \leq N \leq N_{\mathrm{max}},]$ with $\[S_1 \subset S_2 \ldots S_{N_\mathrm{max}-1} \subset S_{N_\mathrm{max}} \subset {\ensuremath{\mathcal{D}^\mu}\xspace}]$ Lagrangian Hierarchical Space $\[W_N = {\rm span} \: \{ \xi^n \equiv \underbrace{ u (\mu^n)}_{u^{\mathcal{N}} (\mu^n)}, n = 1, \ldots, N \}.]$ with $\[W_1 \subset W_2 \ldots \subset W_{N_\mathrm{max}} \subset X^{\mathcal{N}} \subset{X}]$ Reduced basis approximation Given $\mu \in \mathcal{D}^{\mu}$ evaluate $\[\label{eq:1} s_N (\mu) = f(u_N (\mu);\mu)]$ where $u_N (\mu) \in X_N$ satisfies $\[a (u_N (\mu), v; \mu) = f(v;\mu), \ \forall \: v \in X_N \ .]$ Recall: • RB Space: $X_N=``\text{Gram-Schmidt}''(W_N)$ • $u_N(\mu)$ unique (coercivity, continuity, linear dependence) Coercivity and Continuity Constants We assume $a$ and Recall that • constant $\[(0 < ) \alpha(\mu) \equiv \inf_{v\in X}\frac{a(v,v;\mu)}{||v||^2_X}]$ • constant $\[\gamma(\mu) \equiv \sup_{w\in X} \sup_{v\in X}\frac{a(w,v;\mu)}{\|w\|_X \|v\|_X} ( < \infty)]$ Affine dependence and parametric coercivity We assume that $a: X\times X \times \mathcal{D} \rightarrow \mathbb{R}$ is • $\[a(u,v;\mu) = \sum_{q=1}^{Q_a} \Theta^q_a(\mu)\ a^q( u, v ),\, \forall u,v \in X]$ • and $\[\Theta^q_a(\mu) > 0\quad \forall \mu \in \mathcal{D}, \, 1 \leq q \leq Q_a]$ and $\[a^q(u,v) \geq 0\quad \forall u,v \in X, \, 1 \leq q \leq Q_a]$ Inner Products and Norms • and associated norm () \[\begin{aligned} (((w,v)))_\mu &= a(w,v;\mu) &\ \forall u,v \in X\\ |||v|||_\mu &= \sqrt{a(v,v;\mu)} &\ \forall v \in X \end{aligned}] • $X$-inner product and associated norm () \[\begin{aligned} (w,v)_X &= (((w,v)))_{\bar{\mu}} \ (\equiv a(w,v;\bar{\mu})) &\ \forall u,v \in X\\ ||v||_X &= |||v|||_{\bar{\mu}} \ (\equiv \sqrt{a(v,v;\bar{\mu})}) & \ \forall v \in X \end{aligned}] ## 2. Bound theorems Questions • What is the accuracy of $u_N(\mu)$ and $s_N(\mu)$ ? Online \[\begin{aligned} \|u(\mu)-u_N(\mu)\|_X &\leq \epsilon_{\mathrm{tol}}, \quad \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}\\ |s(\mu)-s_N(\mu)\|_X &\leq \epsilon^s_{\mathrm{tol}}, \quad \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}\\ \end{aligned}] • What is the best value for $N$ ? Offline/Online • $N$ too large $\Rightarrow$ computational inefficiency • $N$ too small $\Rightarrow$ unacceptable error • How should we build $S_N$ ? is there an optimal construction ? Offline • Good approximation of the manifold $\mathcal{M}$ through the RB space, but • need for well conditioned RB matrices A Posteriori Error Estimation: Requirements We shall develop the following error bounds ${\ensuremath{\Delta_N(\mu)}\xspace}$ and $\Delta^s_N(\mu)$ with the following properties • $1 \leq N \leq N_{\mathrm{max}}$ \[\begin{aligned} \|u(\mu)-u_N(\mu)\|_X &\leq \Delta_N(\mu), \quad \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}\\ |s(\mu)-s_N(\mu)| &\leq \Delta^s_N(\mu), \quad \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}\end{aligned}] • $1 \leq N \leq N_{\mathrm{max}}$ $\[\begin{gathered} \frac{\Delta_N(\mu)}{\|u(\mu)-u_N(\mu)\|_X} \leq C, \frac{\Delta^s_N(\mu)}{|s(\mu)-s_N(\mu)|} \leq C,\\C\approx 1 \end{gathered}]$ • Online cost depend only on $Q$ and $N$ $u_N(\mu)$ : Error equation and residual dual norm Given our RB approximation $u_N(\mu)$, we have $\[\label{eq:20} e(\mu) \equiv u(\mu) - u_N(\mu)]$ that satisfies $\[\label{eq:21} a( e(\mu), v; \mu ) \ = \ r( u_N(\mu), v; \mu ), \forall v \in X]$ where $r( u_N(\mu), v; \mu ) = f(v) - a( u_N(\mu), v; \mu )$ is the . We have then from coercivity and the definitions above that $\[\label{eq:22} ||e(\mu)||_{X} \ \leq\ \frac{||r( u_N(\mu), v; \mu )||_{X'}}{\alpha(\mu)}\ =\ \frac{\varepsilon_N(\mu)}{\alpha(\mu)}]$ A Posteriori error estimation: Dual norm of the residual [prop:1] Given $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$, the dual norm of $r(u_N(\mu),\cdot;\mu)$ is defined as follows \[\begin{aligned} ||r(u_N(\mu),\cdot;\mu)||_{X'} & \equiv \sup_{v\in X} \frac{r(u_N(\mu),v;\mu)}{||v||_X}\\ & = ||{\ensuremath{\Hat{e}(\mu)}\xspace}||_X \end{aligned}] where ${\ensuremath{\Hat{e}(\mu)}\xspace}\in X$ satisfies \[\begin{aligned} ({\ensuremath{\Hat{e}(\mu)}\xspace},v)_X = r(u_N(\mu),v;\mu) \end{aligned}] The error residual equation can then be rewritten $\[a( e(\mu), v; \mu ) \ = ({\ensuremath{\Hat{e}(\mu)}\xspace},v)_X, \quad \forall v \in X]$ $u_N(\mu)$ : Definitions of energy error bounds and effectivity Given ${\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}$ a nonnegative lower bound of ${\ensuremath{\alpha(\mu)}\xspace}$: $\[\label{eq:23} {\ensuremath{\alpha(\mu)}\xspace}\geq {\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}\geq \epsilon_{\alpha} {\ensuremath{\alpha(\mu)}\xspace},\ \epsilon_{\alpha} \ \in\$0,1[,\, \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}] Denote $\varepsilon_N(\mu) = \|{\ensuremath{\Hat{e}(\mu)}\xspace}\|_X = \|r(u_N(\mu),v;\mu\|_{X'}$

$\label{eq:25} \Delta^{\mathrm{en}}_N(\mu) \ \equiv \ \frac{\varepsilon_N(\mu)}{\sqrt{{\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}}}]$ $\[\label{eq:25} \eta^{\mathrm{en}}_N(\mu) \ \equiv \ \frac{\Delta^{\mathrm{en}}_N(\mu)}{|||e(\mu)|||_\mu}]$ $u_N(\mu)$ : energy error bounds $\[\label{eq:26} 1 \ \leq\ \eta^{\mathrm{en}}_N(\mu) \ \leq \sqrt{\frac{{\ensuremath{{\gamma_{{\mathrm{UB}}}}(\mu)}\xspace}}{{\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}}}, \quad 1 \leq N \leq N_{\max}, \quad \forall \mu\ \in \ {\ensuremath{\mathcal{D}^\mu}\xspace}]$ Remarks • : Left inequality ensures rigorous upper bound measured in $||\cdot||_{X}$ , i.e. $||e(\mu)||_{X} \leq {\ensuremath{\Delta_N(\mu)}\xspace},\ \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ • : Right inequality states that $\Delta_N(\mu)$overestimates the ``true'' error by at most $\gamma(\mu) / {\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}$ • for $a$ and symmetric $\[\theta^{\bar{\mu}} \equiv \frac{\Theta^{\max,\bar{\mu}}_a(\mu)}{\Theta^{\min,\bar{\mu}}_a(\mu)} = \frac{\gamma_{\mathrm{ub}}(\mu)}{\alpha_{\mathrm{lb}}(\mu)}]$ $s_N(\mu)$ : output error bounds $\[1 \ \leq\ \eta^s_N(\mu) \ \leq \frac{{\ensuremath{{\gamma_{{\mathrm{UB}}}}(\mu)}\xspace}}{{\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}}, \quad 1 \leq N \leq N_{\max}, \quad \forall \mu\ \in \ {\ensuremath{\mathcal{D}^\mu}\xspace}]$ where $\[\label{eq:30} \Delta^s_N (\mu) = {\Delta_N^{\mathrm{en}}}(\mu)^2]$ and $\[\eta^s_N(\mu)\equiv \frac{\Delta^s_N(\mu)}{s(\mu)-s_N(\mu)}]$ Rapid convergence of the error in the output Note that the error in the output vanishes quadratically Relative output error bounds We define • the $\[\Delta^{s,rel}_N (\mu) \equiv \frac{\|\hat{e}(\mu)\|^2_X}{ \alpha_\mathrm{lb}(\mu) s_N(\mu)}= \frac{\Delta_N^{\mathrm{en}}(\mu)^2}{s_N(\mu)}]$ • the $\[\eta^{s,rel}_N(\mu)\equiv \frac{\Delta^{s,rel}_N(\mu)}{s(\mu)-s_N(\mu)/s(\mu)}]$ $\[1 \ \leq\ \eta^{s,rel}_N(\mu) \ \leq 2 \frac{{\ensuremath{{\gamma_{{\mathrm{UB}}}}(\mu)}\xspace}}{{\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}}, \quad 1 \leq N \leq N_{\max}, \quad \forall \mu\ \in \ {\ensuremath{\mathcal{D}^\mu}\xspace}]$ for $\Delta^{s,rel}_N \leq 1$ $X$-norm error bounds We define • the $\[\Delta_N (\mu) \equiv \frac{\|\hat{e}(\mu)\|_X}{\alpha_\mathrm{lb}(\mu)}]$ • the $\[\eta_N(\mu)\equiv \frac{\Delta_N(\mu)}{\|e(\mu)\|_X}]$ $\[1 \ \leq\ \eta_N(\mu) \ \leq \frac{{\ensuremath{{\gamma_{{\mathrm{UB}}}}(\mu)}\xspace}}{{\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}}, \quad 1 \leq N \leq N_{\max}, \quad \forall \mu\ \in \ {\ensuremath{\mathcal{D}^\mu}\xspace}]$ Remarks on error bounds Remarks: • The error bounds are rigorous upper bounds for the reduced basis error for any $N = 1,\ldots,N_{max}$ and for all $\mu \in \mathcal{D}$. • The upper bounds for the effectivities are • independent of $N$ , and • independent of $\mathcal{N}$ if $\alpha_{\mathrm{lb}}(\mu)$ only depends on $\mu$, and are thus stable with respect to RB and FEM refinement. • Results for energy norm (and $X$-norm) bound directly extend to noncompliant (& nonsymmetric) problems • if we choose an appropriate definition for the energy (and $X$) norm ## 3. Offline-Online decomposition Offline-Online decomposition Denote ${\ensuremath{\Hat{e}(\mu)}\xspace}\in X$ $\[\label{eq:34} ||{\ensuremath{\Hat{e}(\mu)}\xspace}||_X = \varepsilon_N(\mu) = ||r(u_N(\mu),\cdot;\mu)||_X]$ such that $\[\label{eq:36} ({\ensuremath{\Hat{e}(\mu)}\xspace},v)_X = -r(u_N(\mu),v;\mu), \quad \forall v \in X]$ And recall that $\[\label{eq:35} -r(u_N(\mu),v;\mu) = f(v) - \sum_{q=1}^Q \sum_{n=1}^N\ \Theta^q(\mu)\ {u_N}_n(\mu)\ a^q( \zeta_n,v), \quad \forall v\ \in\ X]$ Offline-Online decomposition • It follows next that ${\ensuremath{\Hat{e}(\mu)}\xspace}\in X$ satisfies $\[({\ensuremath{\Hat{e}(\mu)}\xspace},v)_X \ = \ f(v) - \sum_{q=1}^Q \sum_{n=1}^N\ \Theta^q(\mu)\ {u_N}_n(\mu)\ a^q( \zeta_n,v), \quad \forall v\ \in\ X]$ • Observe then that the rhs is the sum of products of parameter dependent functions and parameter independent linear functionals, thus invoking $\[{\ensuremath{\Hat{e}(\mu)}\xspace}\ = \ \mathcal{C} - \sum_{q=1}^Q \sum_{n=1}^N\ \Theta^q(\mu)\ {u_N}_n(\mu)\ \mathcal{L}^q_n]$ where • $\mathcal{C} \in X$ satisfies $\[(\mathcal{C},v) = f(v), \forall v \in X]$ • $\mathcal{L} \in X$ satisfies $\[(\mathcal{L}^q_n,v)_X = -a^q(\zeta_n,v), \forall v \in X, \, 1 \leq n \leq N, 1 \leq q \leq Q]$ which are parameter independent problems Offline-Online decomposition: Error bounds From ([eq:12]) we get \[\begin{aligned} ||{\ensuremath{\Hat{e}(\mu)}\xspace}||_X^2\ =\ & (\mathcal{C},\mathcal{C})_X\ +\ \sum_{q=1}^Q \sum_{n=1}^N\ \Theta^q(\mu)\ {u_N}_n(\mu)\ \displaystyle \Bigg\{ 2 ( \mathcal{C}, \mathcal{L}^q_n)_X \notag\\ & + \sum_{q'=1}^{Q'} \sum_{n'=1}^{N'}\ \Theta^{q'}(\mu)\ {u_N}_{n'}(\mu)\ ( \mathcal{L}^{q}_{n}, \mathcal{L}^{q'}_{n'})_X \Bigg\} \label{eq:rbellipticlinear_error:37} \end{aligned}] Remark In ([eq:rbellipticlinear_error:37]), $||{\ensuremath{\Hat{e}(\mu)}\xspace}||_X^2$ is the sum of products of • and • , the offline-online for the error bounds is now clear. Offline-Online decomposition: steps and complexity Offline: • Solve for $\mathcal{C}$ and $\mathcal{L}^q_n,\ 1 \leq n \leq N,\ 1 \leq q \leq Q$ • Form and save $(\mathcal{C},\mathcal{C})_X$, $( \mathcal{C}, \mathcal{L}^q_n)_X$ and $( \mathcal{L}^{q}_{n}, \mathcal{L}^{q'}_{n'})_X$, $1 \leq n,n' \leq N,\ 1 \leq q, q' \leq Q$ Online • Given a new $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ • Evaluate the sum $||{\ensuremath{\Hat{e}(\mu)}\xspace}||_X^2$ ([eq:rbellipticlinear_error:37]) in terms of $\Theta^q(\mu)$ and ${u_N}_n(\mu)$ • Complexity in $O(Q^2 N^2)$ independent of $\mathcal{N}$ ## 4. Sampling strategy: a Greedy algorithm Offline-Online Scenarii Offline Given a tolerance $\tau$, build $S_N$ and $W_N$ s.t. $\[\forall \ \mu\ \in \mathcal{P} \equiv \mathcal{D}^{\mu} \ , \ \Delta_N(\mu) < \tau]$ Online Given $\mu$ and a tolerance $\tau$, find $N^*$ and thus $s_{N^*}(\mu)$ s.t. $\[N^* = \operatorname{arg\ max}_N\ \left( \Delta_{N}(\mu) < \tau \right)]$ or given $\mu$ and a max execution time $T$, find $N^*$ and thus $s_{N^*}(\mu)$ s.t. $\[N^* = \operatorname{arg\ min}_N\ \left( \Delta_{N}(\mu) \mbox{ and execution time } < T \right)]$ $S_N$ and $W_N$ Generation Strategies Offline Generation • Given a tolerance $\epsilon$, set $N = 0$ and $S_0 = \emptyset$ • While ${\ensuremath{\Delta_N^{\mathrm{max}}}\xspace}> \epsilon$ • $N = N+1$ • If N == 1; then Pick ((log-)randomly) $\mu_1 \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ • Build ${\ensuremath{S_N}\xspace}:= \{ \mu_N \} \cup S_{N-1}$ • Build ${\ensuremath{W_N}\xspace}:= \{ \xi = u(\mu_N) \} \cup W_{N-1}$ • Compute ${\ensuremath{\Delta_N^{\mathrm{max}}}\xspace}:= \mathrm{max}_{\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}}\, \Delta_N(\mu)$ * • End While Condition number recall that the $\zeta_n$ are , this ensures that the condition number will stay bounded by $\gamma(\mu)/\alpha(\mu)$ Online Algorithm I $\mu$ adaptive online • Given $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$, compute $({\ensuremath{s_{N^{*}}}\space}(\mu), {\ensuremath{\Delta_{N^{*}}}\space}(\mu))$ such that ${\ensuremath{\Delta_{N^{*}}}\space}(\mu) < \tau.$ • $N = 2$ • While ${\ensuremath{\Delta_N}\space}(\mu) > \tau$ • Compute $({\ensuremath{s_N}\space}(\mu), {\ensuremath{\Delta_N}\space}(\mu)) \mbox{ using } ({\ensuremath{S_N}\xspace},{\ensuremath{W_N}\xspace})$ • $N = N * 2\qquad$ use the (very) fast convergence properties of RB • End While Online Algorithm II Offline • While $i <= \mathrm{Imax} >> 1$ • Pick log-randomly $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ • Store in table $\mathcal{T}, {\ensuremath{\Delta_N}\space}(\mu)$ if for $N=1,..., {\ensuremath{{N^{\mathrm{max}}}}\xspace}$ • $i = i + 1$; End While Online Algorithm II – $\mu$ adaptive online – worst case • Given $\mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$, compute $({\ensuremath{s_{N^{*}}}\space}(\mu), {\ensuremath{\Delta_{N^{*}}}\space}(\mu))$ such that ${\ensuremath{\Delta_{N^{*}}}\space}(\mu) < \tau.$ • $N^{*} := \mathrm{arg} \mathrm{max}_{\mathcal{T}}\, {{\ensuremath{\Delta_N}\space}(\mu) \, < \, \tau}$ • Use ${\ensuremath{W_{N^{*}}}\xspace}$ to compute $({\ensuremath{s_{N^{*}}}\space}(\mu),{\ensuremath{\Delta_{N^{*}}}\space}(\mu))$ # Inf-sup lower bound Lower bound for coercivity constant We require a ${\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}$ for ${\ensuremath{\alpha(\mu)}\xspace}= \alpha_c(\mu),\ \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ Two strategies are available: • ``Min $Theta$'' approach if $a$ is parametrically coercive (i.e. the coercivity constant depends solely on $\mu$) • and more generally the Successive Constraint Method(SCM) which can also be applied in case of ``Inf-Sup'' stable problems (Stokes, Helmholtz,…​) ## 1. “Min $\Theta$“ Approach ``Min $Theta$'' approach: Lower bound for $\alpha(\mu)$ For a parametrically coercive bilinear form • $\Theta^q(\mu) > 0,\ \forall \mu \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ and • $a^q(v,v) \geq 0,\ \forall v \in X,\ 1 \leq q \leq Q$ We have $\[\label{eq:38} \Theta^{\mathrm{min},\Bar{\mu}}_a(\mu) = \alpha(\Bar{\mu})\min_{q=1...Q}\frac{\Theta_a^q(\mu)}{\Theta_a^q(\Bar{\mu})} \leq \alpha(\mu)]$ where $\Bar{\mu} \in {\ensuremath{\mathcal{D}^\mu}\xspace}$ which was used to define the $X$-inner product and induced norm Recall that \[\begin{aligned} (u,v)_X &= a(u,v;\alert{\Bar{\mu}}), \quad \forall u,v \in X\\ \|v\|_X &= \sqrt{(u,v)_X}, \quad \forall v \in X \end{aligned}] ``Min $Theta$'' approach: Upper bound for $\gamma(\mu)$ Similarly we develop an upper bound ${\ensuremath{{\gamma_{{\mathrm{UB}}}}(\mu)}\xspace}$ for $\gamma(\mu)$. We define $\[\label{eq:38} \infty > \Theta^{\mathrm{max},\Bar{\mu}}_a(\mu) = \gamma(\Bar{\mu}) \max_{q=1...Q}\frac{\Theta_q^q(\mu)}{\Theta_a^q(\Bar{\mu})}\geq \gamma(\mu)]$ [rem:mintheta-rem] ${\ensuremath{{\gamma_{{\mathrm{UB}}}}(\mu)}\xspace}$ is actually not required in practice but relevant in the theory. ``Min $Theta$'' approach: Summary if $a$ is parametrically coercive we then choose • the coercivity constant lower bound to be $\[{\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}\equiv \Theta^{\mathrm{min},\Bar{\mu}}_a(\mu)]$ • and the continuity constant upper bound to be ($a$ symmetric) $\[{\ensuremath{{\gamma_{{\mathrm{UB}}}}(\mu)}\xspace}\equiv \Theta^{\mathrm{max},\Bar{\mu}}_a(\mu)]$ • Online cost to evaluate ${\ensuremath{{\alpha_{{\mathrm{LB}}}}(\mu)}\xspace}$ : $O(Q_a)$ • Choice of inner product important $(u,v)_X &= a(u,v;\alert{\Bar{\mu}})$ (see multiple inner products approach) • Extends to non-symmetric problems by considering the symmetric part $\[a_s(u,v;\mu) = \frac{1}{2}( a(u,v;\mu)+a(v,u;\mu) )]$ ## 2. Successive constraint method (SCM) ### 2.1. Successive Constraint method: Stability estimates We wish to compute $\alpha_{\mathrm{LB}}: \mathcal{D} \rightarrow \mathbb{R}$ such that $\[\label{eq:38} 0 < \alpha_{\mathrm{LB}}(\mu) \leq \alpha^{\mathcal{N}}(\mu), \quad \mu \in \mathcal{D}]$ and it computation is rapid $O(1)$ where $\[\label{eq:39} \alpha^{\mathcal{N}}(\mu)= \mathrm{inf}_{w \in X^{\mathcal{N}}} \frac{a(w,w;\mu)}{\|w\|_X^2}]$ Computation of $\alpha^{\mathcal{N}}(\mu)$ $\alpha^{\mathcal{N}}(\mu)$ is the minimum eigenvalue of the following generalized eigenvalue problem $\[\label{eq:40} a(w,v;\mu) = \lambda(\mu)\ m(w,v;\mu), \quad (A w = \lambda B w)]$ where $m(\cdot,\cdot)$ is the bi-linear form associated with $\|\cdot\|_X$ and $B$ is the associated matrix. ### 2.2. Successive Constraint method:Reformulation The problem as a minimization one First recall $\[\label{eq:41} a(w,v;\mu) = \sum_{q=1}^{Q_{a}}\ \theta_q(\mu)\ a_q(w,v)]$ Hence we have $\[\label{eq:42} \alpha^{\mathcal{N}}(\mu)= \mathrm{inf}_{w \in X^{\mathcal{N}}} \sum_{q=1}{^Q_a}\ \theta_q(\mu) \frac{a_q(w,w)}{\|w\|_X^2}]$ and we note $\[\label{eq:43} \mathcal{J}^{\mathrm{obj}}(w;\mu) = \sum_{q=1}^{Q_a}\ \theta_q(\mu) \frac{a_q(w,w)}{\|w\|_X^2}]$ Reformulation We have the following optimisation problem $\[\label{eq:44} \alpha^{\mathcal{N}}(\mu)= \mathrm{inf}_{y \in \mathcal{Y}} \mathcal{J}^{\mathrm{obj}}(\mu; y)]$ where $\[\label{eq:46} \mathcal{J}^{\mathrm{obj}}(\mu; y) \equiv \sum_{q=1}^{Q_a}\ \theta_q(\mu) y_q]$ and $\[\label{eq:45} \mathcal{Y} = \Big\{ y \in \mathbb{R}^{Q_a} |\ \exists w \in X^{\mathcal{N}}\ \mathrm{s.t.}\ y_q = \frac{a_q(w,w)}{\|w\|_{X^{\mathcal{N}}}^2}, 1 \leq q \leq Q_a \Big\}]$ We now need to characterize $\mathcal{Y}$, to do this we construct two sets $\mathcal{Y}_{\mathrm{LB}}$ and $\mathcal{Y}_{\mathrm{UB}}$ such that $\mathcal{Y}_{\mathrm{UB}} \subset \mathcal{Y} \subset \mathcal{Y}_{\mathrm{LB}}$ over which finding $\alpha^{\mathcal{N}}(\mu)$ is feasible. Successive Constraint method: Ingredients First we set the design space for the minimisation problem . We introduce $\[\label{eq:21} \mathcal{B} = \prod_{q=1}^{Q_a} \Big[ \mathrm{inf}_{w\in X^{\mathcal{N}}} \frac{a_q(w,w)}{\|w\|_X^2}; \mathrm{sup}_{w\in X^{\mathcal{N}}} \frac{a_q(w,w)}{\|w\|_X^2} \Big$] $\[\label{eq:22} \Xi = \Big\{ \mu_i \in \mathcal{D}; i=1,...,J \Big\}]$ and $\[\label{eq:23} C_K = \Big\{ \mu_i \in \Xi; i=1,...,K \Big\} \subset \Xi]$

$\Xi$ is constructed using a $\frac{1}{2^p}$ division of $\mathcal{D}$: in 1D, $0, 1; \frac{1}{2}; \frac{1}{4}, \frac{3}{4};...$. $C_K$ will be constructed using a greedy algorithm.

Finally we shall denote $P_M(\mu;E)$ the set of $M$ points closest to $\mu$ in the set $E$. We shall need this type of set to construct the lower bounds.

### 2.3. Successive Constraint method: Lower bounds $\mathcal{Y}_{\mathrm{LB}}$

Given $M_\alpha, M_+ \in \mathbb{N}$ we are now ready to define $\mathcal{Y}_{\mathrm{LB}}$ $\[\begin{gathered} \label{eq:24} \mathcal{Y}_{\mathrm{LB}}(\mu; C_K) = \Big\{ y \in \mathbb{R}^{Q_a}\ |\ y \in \mathcal{B}, \\ \ \sum_{q=1}^{Q_a} \theta_q(\mu') y_q \geq \alpha^{\mathcal{N}}(\mu'),\ \forall \mu' \in P_{M_\alpha}(\mu;C_K) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\\ \sum_{q=1}^{Q_a} \theta_q(\mu') y_q \geq \alpha_{\mathrm{LB}}(\mu';C_{K-1}),\ \forall \mu' \in P_{M_+}(\mu;\Xi\backslash C_K) \Big\}~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \end{gathered}]$ We now set $\[\label{eq:25} \alpha_{\mathrm{LB}}(\mu;C_K) = \mathrm{inf}_{y \in \mathcal{Y}_{\mathrm{LB}(\mu;C_K)}}\ \mathcal{J}^{\mathrm{obj}}(\mu;y)]$

Computing $\alpha_{\mathrm{LB}}(\mu;C_K)$ is in fact a linear program with $Q_a$ design variables, $y_q$, and $2 Q_a+M_\alpha+M_+$ constraints online. It requires the construction of $C_K$ offline.

### Successive Constraint method: Upper bounds $\mathcal{Y}_{\mathrm{UB}}$

Let $\[\label{eq:26} \mathcal{Y}_{\mathrm{UB}}( C_K ) = \Big\{ y^*(\mu_k), 1 \leq k \leq K \Big\}]$ with $\[\label{eq:27} y^*(\mu) = \mathrm{arg}\mathrm{min}_{y \in \mathcal{Y}}\ \mathcal{J}^{\mathrm{obj}}( \mu; y )]$ We set $\[\label{eq:28} \alpha_{\mathrm{UB}}( \mu; C_K) = \mathrm{inf}_{y \in \mathcal{Y}_{\mathrm{UB}}(C_K)}\ \mathcal{J}^{\mathrm{obj}}(\mu;y)]$

$\mathcal{Y}_{\mathrm{UB}}$ requires $K$ eigensolves to compute the eigenmode $\eta_k$ associated with $w_k, k=1,...,K$ and $K Q \mathcal{N}$ inner products to compute the $y^*_q(w_k)=\frac{a_q(\eta_k,\eta_k;\mu)}{\|\eta_k\|_{X^{\mathcal{N}}}^2}, k=1,...,K$ offline . Then computing $\alpha_{\mathrm{UB}}( \mu; C_K)$ is a simple enumeration online.

Successive Constraint method: $C_K$

$[C_{K_\mathrm{max}}$ = \mathrm{Greedy}(\Xi, \epsilon]) Given $\Xi$ and $\epsilon \in [0;1$]

• While $\mathrm{max}_{\mu \in \Xi}\ \frac{\alpha_{\mathrm{UB}}( \mu; C_K) - \alpha_{\mathrm{LB}}( \mu; C_K)}{\alpha_{\mathrm{UB}}( \mu; C_K)} > \epsilon$

• $\mu_{K+1} = \mathrm{arg} \mathrm{max}_{\mu \in \Xi}\ \frac{\alpha_{\mathrm{UB}}( \mu; C_K) - \alpha_{\mathrm{LB}}( \mu; C_K)}{\alpha_{\mathrm{UB}}( \mu; C_K)}$

• $C_{K+1} = C_K \cup \{ \mu_{K+1} \}$

• $K \leftarrow K+1$

• set $K_{\mathrm{max}} = K$

Successive Constraint method:Offline-Online

$\mathrm{Offline}$

• $2Q_a+M_\alpha+M_+$ eigensolves $\alpha^{\mathcal{N}}(\mu), y^*(\mu_k)$

• $n_\Xi K_{\mathrm{max}} LP(Q,M_\alpha,M_+)$ to build $C_{K_{\mathrm{max}}}$

• $K_{\mathrm{max}} Q$ inner products over $X^{\mathcal{N}} \Rightarrow \mathcal{Y}_{\mathrm{UB}}$

$[\alpha_{\mathrm{LB}}(\mu)$ = \mathrm{Online}( \mu, C_{K_{\mathrm{max}}}, M_\alpha, M_+ )] Given $\mu \in \mathcal{D}$

• sort over $C_{K_{\mathrm{max}}} \Rightarrow P_{M_\alpha}(\mu;C_{K_{\mathrm{max}}})$ and $P_{M_+}(\mu;\Xi\backslash C_{K_{\mathrm{max}}})$

• $(M_\alpha+M_++2) Q_a$ evaluation of $\theta_q(\mu')$

• $M_\alpha$ lookups to get $\mu' \rightarrow \alpha^{\mathcal{N}}(\mu')$

• $LP(Q_a,M_\alpha,M+)$ to get $\alpha_{\mathrm{LB}} (\mu)$

# Numerical Experiments

## 1. Problem Statement

### 1.1. Example Thermal Block: Heat Transfer

6

(0,0) rectangle (3,3); (0,0) grid (3,3); (0,0) rectangle (3,3);

(1.5,-0.5)node[right]$\Gamma_0$ (Heat Flux) to[out=180,in=90] (1.5,0); (1.5,3.5)node[right]${\Gamma_{\mathrm{top}}}$ (Zero Dirichlet) to[out=180,in=90] (1.5,3); (3.5,1.5) node[right]${\Gamma_{\mathrm{sides}}}$ (Insulated) to[out=180,in=90] (0,1.5) ; (3.5,1.5) to[out=180,in=-90] (3,1.5);

in 5mm,15mm,25mm

in 5mm,15mm,25mm

at (+0.45,+0.3) $\mu_\theindex$;

3

Example Thermal Block: Problem statement Given $\mu \in (\mu_1,...\mu_P) \in {\ensuremath{\mathcal{D}^\mu}\xspace}\equiv [\mu^{\text{min}},\mu^{\text{max}}$^P], evaluate (recall that $\ell = f$)
$\[s(\mu) = f(u(\mu))]$ where $u(\mu) \in X \equiv \{ v \in H^1(\Omega), v|_{\Gamma_{\text{top}}} = 0\}$ satisfies $\[a(u(\mu), v; \mu) = f(v;\mu) \quad \forall v \in X]$ we have $P = 8$ and given $1 < \mu_r < \infty$ we set $\[\mu^{\mathrm{min}} = 1/\sqrt{\mu_r},\quad \mu^{\mathrm{max}} = \sqrt{\mu_r}]$ such that $\mu^{\mathrm{max}}/\mu^{\mathrm{min}}=\mu_r.$

Example Thermal Block Recall we are in the compliant case $\ell = f$, we have $\[f(v) = \int_{\Gamma_{0}} v\quad \forall v \in X]$ and $\[a(u,v;\mu) = \sum_{i=1}^{P} \mu_i \int_{\Omega_i} \nabla u \cdot \nabla v + 1 \int_{\Omega_{P+1}} \nabla u \cdot \nabla v \quad\forall u,\ v\ \in X]$ where $\Omega = \cup_{i=1}^{P+1} \Omega_i$.

Example Thermal Block The inner product is defined as follows $\[(u,v)_X = \sum_{i=1}^P \bar{\mu}_i \int_{\Omega_i}\nabla u \cdot \nabla v + 1 \int_{\Omega_{P+1}} \nabla u \cdot \nabla v]$ where $\bar{\mu}_i$ is a . We have readily that $a$ is

* * $\[0 < \frac{1}{\sqrt{\mu_r}} \leq \mathrm{min}(\mu_1/\bar{\mu}_1, \ldots, \mu_P/\bar{\mu}_P,1) \leq \alpha(\mu)]$ * and $\[\gamma(\mu) \leq \mathrm{max}(\mu_1/\bar{\mu}_1, \ldots, \mu_P/\bar{\mu}_P,1) \leq \sqrt{\mu_r} < \infty]$

and the linear form $f$ is .

Example Thermal Block: Affine decomposition We $\[a(u,v;\mu) = \sum_{q=1}^{P+1} \Theta^q(\mu) a^q(u,v)]$ with \[\begin{aligned} \Theta^1(\mu) = \mu_1 & & a^1(u,v) = \int_{\Omega_1} \nabla u \cdot \nabla v\\ & \vdots & \\ \Theta^P(\mu) = \mu_P & & a^P(u,v) = \int_{\Omega_P} \nabla u \cdot \nabla v\\ \Theta^{P+1}(\mu) = 1 & & a^{P+1}(u,v) = \int_{\Omega_{P+1}} \nabla u \cdot \nabla v \end{aligned}]

Example Thermal Block

• 0.5 ** 0.5 ** • ## 2. Thermal Block $P=1$

ExampleThermal Block $P=1$ We assume $\[1/\mu^{\min}_1=\mu^{\max}_1=\sqrt{\mu_r}=10]$ and choose $\mathcal{N}=256$

we set $\bar{\mu}=1$ and have $\[\Theta^1_a(\mu)=\mu_1,\, \Theta^2_a(\mu) = 1]$ Thus, $\[\Theta_a^{\min,\bar{\mu}}(\mu_1)=\min(\mu_1,1)\\ \Theta_a^{\max,\bar{\mu}}(\mu_1)=\max(\mu_1,1)]$

hence

$\[\theta^{\bar{\mu}}(\mu_1) = \max(\frac{1}{\mu_1},\mu_1)]$

and $\[\theta^{\bar{\mu}}(\mu_1) \leq\sqrt{\mu_r} \forall \mu_1 \in \mathcal{D}]$

Example: Thermal Block $P=1$ in [RHP2008]

Table 1. Convergence results for $P=1$
$N$ $\Delta^s_{N,\mathrm{max}}(\mu)$ $\eta^s_{N,\mathrm{ave}}$ $\eta^s_{N,\mathrm{max}}$

1

7.2084E+00

2.3417

3.3305

2

4.5371E–01

2.4858

3.6850

3

6.9652E–04

6.2195

9.8551

4

1.3744E–07

3.3219

7.2632

5

3.1140E–11

6.0789

7.0453

Note that: $\eta^s_{N,\mathrm{max}}(\mu_1) \leq \eta^s_{\mathrm{max,UB}} \equiv \sqrt{\mu_r} = 10$

• Maximum output error bound: $\Delta^s_{N,\mathrm{max}} = \max_{\mu \in \Xi_{\mathrm{train}}} \Delta^s_N(\mu)$

• Average output effectivity: $\eta^s_{N,\mathrm{ave}} = \frac{1}{\Xi_\mathrm{train}}\sum_{\mu \in \Xi_{\mathrm{train}}} \eta^s_N(\mu)$

• Maximum output effectivity: $\eta^s_{N,\mathrm{max}} = \max_{\mu \in \Xi_{\mathrm{train}}} \eta^s_N(\mu)$

## 3. Example Thermal Block $P=8$

Thermal Block $P=8$

• Configuration :

• 47 600 dofs;

• Preconditionner : LU – Solver : MUMPS

• $\Xi :$ parameter sampling of dimension 1 000.

• Plot $\ds{ \max_{\mu \in \Xi} \frac{ |s^{\mathcal{N}}(\mu)-s_N(\mu)|}{s^{\mathcal{N}}(\mu)} }$

table[x=NbBasis,y=Max]; table[x=NbBasis,y=Max]; ;

Thermal Block $P=8$

• More parameters there are, more rich the problem is;

• Notations :

• $e^i(\mu)$ is the relative error on the output when $i$ parameters vary.

table[x=NbBasis,y=Max]; table[x=NbBasis,y=Max]; table[x=NbBasis,y=Max]; table[x=NbBasis,y=Max]; table[x=NbBasis,y=Max]; table[x=NbBasis,y=Max]; table[x=NbBasis,y=Max]; table[x=NbBasis,y=Max]; ;