From PDEs to Certified Reduced Models: A Bird's-Eye View
A summary of notes from the lecture about model order reduction in the summer semester 2026.
Many similar PDEs:
A huge range of physical models, such as heat spreading, ink dispersing in a current, waves, even the Schrödinger equation are all special cases of one template, the general linear second-order operator:
\[B_\mu u(x) = \underbrace{\nabla \cdot \big(\boldsymbol{\alpha}_\mu(x)\, \nabla u(x)\big)}_{\text{diffusion}} \;+\; \underbrace{\boldsymbol{\beta}_\mu(x) \cdot \nabla u(x)}_{\text{advection}} \;+\; \underbrace{\gamma_\mu(x)\, u(x)}_{\text{reaction}}.\]The subscript $\mu$ says the operator depends on some parameter, like a material property, a geometry or a physical constant. The right hand side of the equation all represent different physical aspects. The first term represents the diffusion or how much the disturbance spreads out. The advection term carries it along rigidly without spreading it and the reaction grows or shrinks it in place without moving or spreading it at all. A concrete way to picture the combination is a drop of ink released into a flowing, chemically active river. Click on the visualization below to see what happens.
Click or tap the water to add a drop of ink
How fast the ink spreads on its own, in every direction.
Speed and direction of the flow carrying the ink downstream.
Negative fades the ink away over time, positive lets it bloom.
u(x,t) is the ink concentration in the tank. The diffusion term spreads it evenly outward, the advection term carries it along with the current, and the reaction term lets it decay or grow locally. All three run at once, on the same field, every frame — exactly the operator Bμu described above.
In practice you rarely want to solve $B_\mu u = f_\mu$ for just one $\mu$; you want it for thousands of them (design sweeps, optimization, real-time control). Doing each one from scratch is wasteful, so the whole field is built around one question: can we do the expensive work once, and reuse it cheaply for every new parameter? That’s only possible when the parameter dependence is affine,
\[A_\mu = \sum_{q=1}^{Q} \vartheta_q(\mu)\, A_q,\]a sum of fixed, $\mu$-independent operators, each scaled by a scalar function of $\mu$. The simplest example is a thermal block: a domain split into pieces with a different, constant diffusivity on each one.
Ω is split into P = 6 disjoint blocks Ωp. Inside each block the diffusivity is a single
constant μp, so αμ(x) = ∑p μp χΩp(x)
— no smooth variation, just a jump at every block boundary.
Step one: make sure a solution exists at all
Classical PDE theory demands very smooth solutions. Unfortunately most real problems such as kinked materials, sharp corners and point sources do not fullfill this criteria. Therefore we loosen the requirement and do the following: multiply by a test function $v$, integrate by parts, and search for a solution in a larger, “weaker” space instead, requiring only one derivative of $u$ and $v$ rather than two. This turns the PDE into a variational problem: find $u \in \mathcal U$ such that
\[b_\mu(u,v) = f_\mu(v) \qquad \text{for all } v \in \mathcal V.\]Whether this is well-posed comes down to two properties of the bilinear form $b_\mu$:
Coercivity \(b_\mu(u,u) \;\ge\; \alpha_\mu \, \|u\|_{\mathcal U}^2 \qquad \text{for all } u \in \mathcal U, \quad \alpha_\mu > 0.\)
Inf-sup condition \(\inf_{u \in \mathcal U}\; \sup_{v \in \mathcal V} \; \frac{b_\mu(u,v)}{\|u\|_{\mathcal U}\|v\|_{\mathcal V}} \;\ge\; \beta_\mu > 0.\)
Coercivity is the special case $\mathcal U = \mathcal V$; it’s easy to check (it follows directly from ellipticity of the diffusion coefficient plus a Poincaré–Friedrichs inequality) and it implies inf-sup for free, with $\beta_\mu = \alpha_\mu$. When it holds, Lax–Milgram hands you existence, uniqueness, and stability outright. When trial and test spaces genuinely differ (advection-dominated problems, the wave equation, ultraweak formulations), the more general Banach–Nečas theorem takes over, resting entirely on the inf-sup condition instead.
The constant $\beta_\mu$ is not merely a theoretical footnote: it appears directly in the denominator of every error bound discussed below, so a small $\beta_\mu$ is an early indicator that a model built later will be unreliable. Geometrically, $\beta_\mu$ and the continuity constant $C_\mu$ correspond to the shortest and longest semi-axes of the ellipsoid obtained by stretching the unit ball of $\mathcal U$ through the supremizing operator, the map sending each $u$ to its ideal test partner:
Drag right to make the problem less stable: the shortest axis of the ellipse (βμ) shrinks toward zero while the longest axis (Cμ) grows.
Step two: shrink the problem, without losing the guarantee
Once the problem is discretized (using finite elements, for instance), the same argument repeats one level down. The discrete spaces require their own LBB (discrete inf-sup) condition with constant $\beta_\mu^\circ$. Given this, the quasi-optimality result known as Céa’s lemma (or its non-coercive generalization, the Xu–Zikatanov theorem) states that the computed solution is essentially as good as the best possible approximation in that space:
\[\|u_\mu - u_\mu^\delta\|_{\mathcal U} \;\le\; \frac{C_\mu}{\beta_\mu^\circ} \inf_{w^\delta \in \mathcal{U}^\delta} \|u_\mu - w^\delta\|_{\mathcal U}.\]The reduced basis method repeats this shrinking process once more, reducing a truth space with possibly hundreds of thousands of degrees of freedom down to a handful of actual solutions (“snapshots”) computed at well-chosen parameter values. For coercive problems, reusing the snapshots themselves as test functions preserves stability at no additional cost. For non-coercive problems this does not hold: the ideal test partner for a given snapshot (its supremizer) is generally not expressible as a combination of the snapshots, and $\beta_N$ collapses toward zero. The remedy is to compute these supremizers explicitly and build a separate test space from them, a Petrov–Galerkin setup, which restores $\beta_N \ge \beta_\mu^\circ > 0$ by construction.
Step three: certify the shortcut
Speed alone is not sufficient: one also needs to know how far off the cheap, reduced solution might be, without ever computing the expensive true one. The certified bound has the same structure as Céa’s lemma, applied one level further down:
\[\|u_\mu^\delta - u_\mu^N\|_{\mathcal U} \;\le\; \Delta_N(\mu) = \frac{\|r_N^\delta(\mu)\|_{\mathcal V'}}{\beta_\mu^\circ},\]where the residual $r_N^\delta(\mu) = f_\mu - B_\mu u_\mu^N$ is obtained essentially at no extra cost from the reduced solution, and $\beta_\mu^\circ$ is bounded cheaply online (for instance via the Successive Constraint Method) rather than by solving the full problem again. With a fast reduced solve and an equally fast, reliable error bound available, training the reduced space becomes a search problem: identify the parameter at which the current model performs worst according to $\Delta_N(\mu)$ (the true error is never actually available), add a snapshot there, and repeat. This is the weak greedy algorithm.
The recurring pattern
Everything above is, in essence, one idea applied at three different scales: continuous to truth, truth to reduced, and estimated to true error.
\[\text{error} \;\le\; \text{best possible approximation} \;\times\; \text{a condition-number-like stability factor}.\]Coercivity, inf-sup, supremizers, and affine decompositions all serve the same purpose: keeping the stability factor under control.
The above content is based on my own notes from the Model Reduction lecture by Prof. Urban, Summer Semester 2026.