Reduced Bases and Proper Orthogonal Decomposition

A reduced basis replaces a high-dimensional relativistic solution with its coordinates in a small, problem-adapted subspace. The compression is scientifically useful only when the inner product, preprocessing, parameter domain, training catalog, and held-out errors are explicit. Rapid singular-value decay shows that the sampled family is compact in the chosen norm; it does not prove that an unsampled corner, new physical regime, or altered alignment convention lies near the same subspace.

Required background. Probability and Inference for Gravity supplies sampling designs, held-out tests, and uncertainty summaries.

Helpful background. Formulations of General Relativity supplies the physical variables and gauge content being reduced; Differential Equations for Relativity supplies the underlying solution family; Computational Relativity supplies convergence, provenance, and reproducibility controls.

A solution manifold in a declared norm

Let $h(\lambda)$ denote a waveform or field generated at parameter point $\lambda\in\Lambda$ and represented in a Hilbert space $\mathcal H$ with inner product $\langle\cdot,\cdot\rangle$. The solution manifold is

$$\mathcal S=\{h(\lambda):\lambda\in\Lambda\}.$$

Choose $e_1,\ldots,e_n$ orthonormal in this inner product. The $n$-dimensional reduced space $V_n=\operatorname{span}\{e_1,\ldots,e_n\}$ then gives the orthogonal projection

$$P_nh=\sum_{j=1}^{n}\langle e_j,h\rangle e_j, \qquad \epsilon_n(\lambda)= \frac{\|h(\lambda)-P_nh(\lambda)\|}{\|h(\lambda)\|}.$$

Every symbol carries a choice. A time-domain $L^2$ norm weights inspiral and merger differently from a detector-noise-weighted norm. Maximizing over time and phase removes degrees of freedom that a downstream application may need. Truncating all waveforms to their common duration can make the family look artificially compact. State the mode content, time or frequency interval, measure, quadrature, alignment, taper, normalization, and whether complex strain means $h_+-ih_\times$ under the inherited $e^{-i\omega t}$ convention.

The mathematical quantity that asks for the best possible $n$-plane is the Kolmogorov width

$$d_n(\mathcal S)= \inf_{\dim V=n}\sup_{h\in\mathcal S} \inf_{v\in V}\|h-v\|.$$

A finite training catalog estimates this continuum object but cannot equal it unless coverage has been justified. Field and collaborators connect greedy reduced bases, catalog error, and the Kolmogorov-width idea for gravitational waveforms in Field et al. 2011, “Reduced Basis Method,” Eqs. (1)–(3) and Figs. 1–3.

Proper orthogonal decomposition

Let the discrete state inner product be

$$\langle x,y\rangle_M=x^*My,$$

where the Gram or mass matrix $M$ is Hermitian positive definite; $M=I$ is the Euclidean special case. Given quadrature-weighted snapshot columns

$$S=\begin{bmatrix} \sqrt{w_1}h(\lambda_1)&\cdots&\sqrt{w_m}h(\lambda_m) \end{bmatrix},$$

form an ordinary SVD only after whitening the state metric,

$$\widetilde S=M^{1/2}S=\widetilde U\Sigma V^*, \qquad \Sigma=\operatorname{diag}(\sigma_1,\sigma_2,\ldots), \qquad \sigma_1\ge\sigma_2\ge\cdots.$$

The physical POD modes are the columns of $U=M^{-1/2}\widetilde U$, so $U^*MU=I$ and $P_n=U_nU_n^*M$. The first $n$ modes minimize the weighted mean-square projection error of these snapshots in the $M$ norm, and the discarded training error is

$$\sum_{i=1}^{m}w_i\|h_i-P_nh_i\|_M^2 =\sum_{j>n}\sigma_j^2.$$

Sirovich’s method of snapshots obtains the same nonzero modes from the smaller correlation matrix $S^*MS$ when the state dimension greatly exceeds the number of samples; see Sirovich 1987, pp. 561–571. The optimality is precise but narrow: it uses the empirical weights and average error. A rare boundary waveform can have large error while contributing little to the objective.

A greedy alternative begins with one normalized waveform and repeatedly adds the training point with the largest current projection error. It directly controls the maximum training error and identifies informative parameters. POD and greedy bases optimize different objectives, so they need not reach the same dimension at a stated tolerance even with correct implementations. Inspect average and worst-case errors in the same norm before attributing a difference to conditioning or coverage.

A weighted two-snapshot calculation

Take $h_1=(1,0)^{\mathsf T}$, $h_2=(0,1)^{\mathsf T}$, equal snapshot weights $w_1=w_2=1/2$, and state metric $M=\operatorname{diag}(4,1)$. Then

$$\widetilde S= \begin{pmatrix}\sqrt2&0\\0&1/\sqrt2\end{pmatrix}, \qquad \sigma_1^2=2,\quad \sigma_2^2=\frac12.$$

The first physical mode is $e_1=(1/2,0)^{\mathsf T}$, which indeed has $e_1^*Me_1=1$. Keeping it reproduces $h_1$ exactly and maps $h_2$ to zero. The weighted mean-square error is therefore $1/2$, equal to the discarded singular-value square. Its fraction of total snapshot energy is $(1/2)/(2+1/2)=1/5$.

This makes the norm choice tangible: the first component counts four times as strongly as the second. It also shows why “80% of training energy retained” does not imply a 20% error on every waveform—the second snapshot has 100% relative projection error.

Compact-binary coverage benchmark

Freeze a versioned family of high-fidelity complex $(\ell,m)=(2,2)$ compact-binary waveforms on a declared dimensionless time interval. Record the generating model or simulation revisions, mass ratio and spin coordinates, extraction and extrapolation choices, sampling cadence, units, alignment, and every preprocessing checksum. Split by physical parameter point before basis construction:

• an interpolation training set over the declared domain;
• a scattered held-out set not used for alignment choices or stopping;
• all parameter-domain faces and corners as boundary tests;
• one contiguous interior region omitted from training as a coverage-hole test; and
• points beyond the declared domain used only to test failure detection, never to claim extrapolation accuracy.

For each basis size $n$, report the singular values, mean and maximum training projection errors, the same held-out errors, and the location of the worst case. Also report detector-weighted mismatch after a separately declared time-and-phase maximization, because a small raw $L^2$ error and a small application mismatch are related but not identical claims. Predeclare a representative acceptance target such as maximum in-domain held-out mismatch below $10^{-4}$ and relative projection error below $10^{-3}$; the actual measured discrepancies must be emitted rather than replaced by pass labels.

Increase training density and basis size separately. If the error plateaus with $n$ but falls when the catalog is densified, coverage rather than rank is limiting. If training error falls while boundary or hole error does not, the compactness claim has not generalized. Repeat the SVD in higher precision or with a rank-revealing decomposition when $\sigma_n/\sigma_1$ approaches the arithmetic floor, and verify $\|U_n^*MU_n-I\|$ independently.

The environment record includes dataset revision and checksums, split manifest, preprocessing source, inner-product weights, linear-algebra backend, precision, SVD tolerances, basis orientation convention, operating system or container, compiler or interpreter, accelerator, process and thread counts, and random seeds used to design any split. A second quadrature and direct high-fidelity overlap calculation provide independent checks.

Detecting structure the catalog missed

Random train/test splits can be misleading when nearby parameter points are highly correlated. A held-out waveform may be almost duplicated by a neighbor in training. The contiguous-hole test asks the harder interpolation question: can the basis represent a region whose local structure was not shown to it? Moving the hole and repeating the construction distinguishes a genuinely low-dimensional manifold from fortunate sampling.

Inspect residual fields, not only scalar norms. Concentrated phase error near merger, a missing higher-mode feature, or a discontinuity introduced by alignment can be scientifically important even when an integrated norm is small. Compare symmetry-related parameters when the underlying model supplies an exact transformation. Add a newly computed high-fidelity waveform at the current worst point and check whether the claimed error estimate predicted it.

The main failure modes are data leakage through preprocessing, a basis tuned to one norm and advertised for another, near-duplicate random holdouts, ill-conditioned modes below the precision floor, omitted domain boundaries, phase alignment that absorbs physical variation, and extrapolation presented as interpolation. A compact basis for one waveform model also compresses that model’s approximation error; it does not validate the model against general relativity or observation.

Clean reconstruction sequence

A fully specified rerun verifies the catalog and split checksums; applies the frozen preprocessing; constructs POD and greedy bases over an increasing training sequence; evaluates training, scattered, boundary, hole, and out-of-domain errors at every $n$; repeats critical decompositions at higher precision; compares overlaps with an independent quadrature; and emits the basis, singular values, conditioning, costs, worst-case parameters, thresholds, and actual discrepancies. Failure of a boundary or hole test returns a failed in-domain compression claim even if average training error is tiny.

The benchmark is limited to the versioned waveform modes, parameter box, alignment, and norm. It does not establish low rank for precessing dynamics, eccentricity, matter effects, other modes, longer waveforms, a new detector curve, or another data representation without a new coverage study.

Check your understanding

Training snapshots all lie within $10^{-8}$ of the vector $e_1$, so a one-vector POD basis has negligible training error. A physically admissible but unsampled boundary waveform is $e_2$, orthogonal to $e_1$. What does the small singular-value tail establish?

Solution

It establishes that the sampled and weighted catalog is nearly rank one. It says nothing about the omitted $e_2$ direction: that waveform has relative projection error one. Add boundary and leave-region-out tests, state the parameter domain, and densify the catalog adaptively around the detected failure. Only held-out coverage evidence can connect empirical singular-value decay to the physical solution family.

Continue with Empirical Interpolation and Reduced Operators.

References

• Field, Scott E., Chad R. Galley, Frank Herrmann, Jan S. Hesthaven, Evan Ochsner, and Manuel Tiglio. “Reduced Basis Catalogs for Gravitational Wave Templates.” Physical Review Letters 106 (2011): 221102. arXiv:1101.3765. DOI.
• Sirovich, Lawrence. “Turbulence and the Dynamics of Coherent Structures. Part I: Coherent Structures.” Quarterly of Applied Mathematics 45, no. 3 (1987): 561–571. DOI.