Abstract
We propose a probabilistic way for reducing the cost of classical projection-based model order reduction methods for parameter-dependent linear equations. A reduced order model is here approximated from its random sketch, which is a set of low-dimensional random projections of the reduced approximation space and the spaces of associated residuals. This approach exploits the fact that the residuals associated with approximations in low-dimensional spaces are also contained in low-dimensional spaces. We provide conditions on the dimension of the random sketch for the resulting reduced order model to be quasi-optimal with high probability. Our approach can be used for reducing both complexity and memory requirements. The provided algorithms are well suited for any modern computational environment. Major operations, except solving linear systems of equations, are embarrassingly parallel. Our version of proper orthogonal decomposition can be computed on multiple workstations with a communication cost independent of the dimension of the full order model. The reduced order model can even be constructed in a so-called streaming environment, i.e., under extreme memory constraints. In addition, we provide an efficient way for estimating the error of the reduced order model, which is not only more efficient than the classical approach but is also less sensitive to round-off errors. Finally, the methodology is validated on benchmark problems.
Similar content being viewed by others
Notes
We have \(\forall \mathbf {x} \in \mathcal {S}, \exists \mathbf {y} \in \mathcal {N}\) such that ∥x −y∥≤ γ.
Indeed, \(\exists \mathbf {n}_{0} \in \mathcal {N}\) such that ∥n −n0∥ := α1 ≤ γ. Let α0 = 1. Then assuming that \(\|\mathbf {n} - {\sum }^{m}_{i=0} \alpha _i \mathbf {n}_{i} \|:=\alpha _{m+1} \leq \gamma ^{m+1}\), \(\exists \mathbf {n}_{m+1} \in \mathcal {N}\) such that \(\| \frac {1}{\alpha _{m+1}} (\mathbf {n} - {\sum }^{m}_{i=0} \alpha _i \mathbf {n}_{i}) - \mathbf {n}_{m+1} \| \leq \gamma \) ⇒ \(\|\mathbf {n} - {\sum }^{m+1}_{i=0} \alpha _i \mathbf {n}_{i} \| \leq \alpha _{m+1} \gamma \leq \gamma ^{m+2} \).
References
Achlioptas, D.: Database-friendly random projections: Johnson-Lindenstrauss with binary coins. J. Comput. Syst. Sci. 66(4), 671–687 (2003)
Ailon, N., Liberty, E.: Fast dimension reduction using Rademacher series on dual bch codes. Discret. Comput. Geom. 42(4), 615 (2009)
Alla, A., Kutz, J.N.: Randomized model order reduction. tech. report, arXiv:http://arxiv.org/abs/1611.02316 (2016)
Baker, C.G., Gallivan, K.A., Dooren, P.V.: Low-rank incremental methods for computing dominant singular subspaces. Linear Algebra Appl. 436(8), 2866–2888 (2012)
Balabanov, O., Nouy, A.: Randomized linear algebra for model reduction. Part ii: minimal residual methods and dictionary-based approximation. arXiv:http://arxiv.org/abs/1910.14378 (2019)
Bebendorf, M.: Why finite element discretizations can be factored by triangular hierarchical matrices. SIAM J. Numer. Anal. 45(4), 1472–1494 (2007)
Bebendorf, M.: Hierarchical Matrices. Springer, Berlin (2008)
Benner, P., Cohen, A., Ohlberger, M., Willcox, K. (eds.): Model reduction and approximation: theory and algorithms. SIAM, Philadelphia, PA (2017)
Benner, P., Gugercin, S., Willcox, K.: A survey of projection-based model reduction methods for parametric dynamical systems. SIAM Rev. 57(4), 483–531 (2015)
Boman, E.G., Hendrickson, B., Vavasis, S.: Solving elliptic finite element systems in near-linear time with support preconditioners. SIAM J. Numer. Anal. 46 (6), 3264–3284 (2008)
Bourgain, J., Lindenstrauss, J., Milman, V.: Approximation of zonoids by zonotopes. Acta Math. 162(1), 73–141 (1989)
Boutsidis, C., Gittens, A.: Improved matrix algorithms via the subsampled randomized Hadamard transform. SIAM J. Matrix Anal. Appl. 34(3), 1301–1340 (2013)
Braconnier, T., Ferrier, M., Jouhaud, J.-C., Montagnac, M., Sagaut, P.: Towards an adaptive POD/SVD surrogate model for aeronautic design. Comput. Fluids 40(1), 195–209 (2011)
Buhr, A., Engwer, C., Ohlberger, M., Rave, S.: A numerically stable a posteriori error estimator for reduced basis approximations of elliptic equations. In: Onate, X.O.E., Huerta, A. (eds.) Proceedings of the 11th World Congress on Computational Mechanics. CIMNE, pp 4094–4102, Barcelona (2014)
Buhr, A., Smetana, K.: Randomized local model order reduction. SIAM J. Sci. Comput. 40(4), A2120–A2151 (2018)
Casenave, F., Ern, A., Lelièvre, T.: Accurate and online-efficient evaluation of the a posteriori error bound in the reduced basis method. ESAIM: Mathematical Modelling and Numerical Analysis 48(1), 207–229 (2014)
Elman, H.C., Silvester, D.J., Wathen, A.J.: Finite elements and fast iterative solvers: with applications in incompressible fluid dynamics. Numer. Math. Sci. Comput. (2014)
Engquist, B., Ying, L.: Sweeping preconditioner for the Helmholtz equation: hierarchical matrix representation. Commun. Pure Appl. Math. 64(5), 697–735 (2011)
Gross, D., Nesme, V.: Note on sampling without replacing from a finite collection of matrices. arXiv:http://arxiv.org/abs/1001.2738 (2010)
Haasdonk, B.: Reduced basis methods for parametrized PDEs – a tutorial introduction for stationary and instationary problems. In: Benner, P., Cohen, A., Ohlberger, M., Willcox, K. (eds.) Model Reduction and Approximation, pp 65–136. SIAM, Philadelphia (2017)
Hackbusch, W.: Hierarchical matrices: algorithms and analysis, vol. 49. Springer, Berlin (2015)
Halko, N., Martinsson, P.-G., Tropp, J.A.: Finding structure with randomness: probabilistic algorithms for constructing approximate matrix decompositions. SIAM Rev. 53(2), 217–288 (2011)
Hesthaven, J.S., Rozza, G., Stamm, B.: Certified reduced basis methods for parametrized partial differential equations, 1st edn. Springer Briefs in Mathematics, Switzerland (2015)
Himpe, C., Leibner, T., Rave, S.: Hierarchical approximate proper orthogonal decomposition. SIAM J. Sci. Comput. 40(5), A3267–A3292 (2018)
Hochman, A., Villena, J.F., Polimeridis, A.G., Silveira, L.M., White, J.K., Daniel, L.: Reduced-order models for electromagnetic scattering problems. IEEE Trans. Antennas Propag. 62(6), 3150–3162 (2014)
Knezevic, D.J., Peterson, J.W.: A high-performance parallel implementation of the certified reduced basis method. Comput. Methods Appl. Mech. Eng. 200(13), 1455–1466 (2011)
Lee, Y.T., Sidford, A.: Efficient accelerated coordinate descent methods and faster algorithms for solving linear systems. In: 2013 IEEE 54th Annual Symposium on Foundations of Computer Science (FOCS), vol. 00, pp 147–156 (2014)
Maday, Y., Nguyen, N.C., Patera, A.T., Pau, S.H.: A general multipurpose interpolation procedure: the magic points. Communications on Pure & Applied Analysis 8(1), 383 (2009)
Martinsson, P.-G.: A fast direct solver for a class of elliptic partial differential equations. J. Sci. Comput. 38(3), 316–330 (2009)
Oxberry, G.M., Kostova-Vassilevska, T., Arrighi, W., Chand, K.: Limited-memory adaptive snapshot selection for proper orthogonal decomposition. Int. J. Numer. Methods Eng. 109(2), 198–217 (2017)
Quarteroni, A., Manzoni, A., Negri, F.: Reduced basis methods for partial differential equations: an introduction, vol. 92. Springer, Berlin (2015)
Rozza, G., Huynh, D.B.P., Patera, A.T.: Reduced basis approximation and a posteriori error estimation for affinely parametrized elliptic coercive partial differential equations. Arch. Comput. Meth. Eng. 15(3), 229 (2008)
Sarlos, T.: Improved approximation algorithms for large matrices via random projections. In: 2006 IEEE 47th annual symposium on foundations of computer science (FOCS), pp 143–152. IEEE (2006)
Sirovich, L.: Turbulence and the dynamics of coherent structures. I. coherent structures. Q. Appl. Math. 45(3), 561–571 (1987)
Tropp, J.A.: Improved analysis of the subsampled randomized Hadamard transform. Adv. Adapt. Data Anal. 3(01n02), 115–126 (2011)
Tropp, J.A.: User-friendly tail bounds for sums of random matrices. Found. Comput. Math. 12(4), 389–434 (2012)
Tropp, J.A., et al.: An introduction to matrix concentration inequalities. Foundations and Trends®;in Machine Learning 8(1-2), 1–230 (2015)
Woodruff, D.P., et al.: Sketching as a tool for numerical linear algebra. Foundations and Trends®;in Theoretical Computer Science 10(1–2), 1–157 (2014)
Xia, J., Chandrasekaran, S., Gu, M., Li, X.S.: Superfast multifrontal method for large structured linear systems of equations. SIAM J. Matrix Anal. Appl. 31(3), 1382–1411 (2009)
Zahm, O., Nouy, A.: Interpolation of inverse operators for preconditioning parameter-dependent equations. SIAM J. Sci. Comput. 38(2), A1044–A1074 (2016)
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by: Anthony Patera
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Appendix
Appendix
Here we list the proofs of propositions and theorems from the paper.
Proof of Proposition 2.2 (modified Cea’s lemma)
For all x ∈ Ur, it holds
where the first and last inequalities directly follow from the definitions of αr(μ) and βr(μ), respectively. Now,
which completes the proof. □
Proof of Proposition 2.3
For all \(\mathbf {a} \in \mathbb {K}^{r}\) and x := Ura, it holds
Then the proposition follows directly from definitions of αr(μ) and βr(μ). □
Proof of Proposition 2.4
We have
and the result follows from definition (2.10). □
Proof of Proposition 2.5
To prove the first inequality, we notice that \(\mathbf {Q}\mathbf {P}_{U_{r}}\mathbf {U}_{m}\) has rank at most r. Consequently,
For the second inequality, let us denote the i-th column vector of Br by bi. Since \(\mathbf {Q}\mathbf {R}_{U}^{-1}\mathbf {Q}^{\mathrm {H}}= \mathbf {Q}\mathbf {Q}^{\dagger }\), with Q‡ the pseudo-inverse of Q, is the orthogonal projection onto range(Q), we have
□
Proof of Proposition 3.3
It is clear that \(\langle \cdot , \cdot \rangle ^{\boldsymbol {\Theta }}_{X}\) and \(\langle \cdot , \cdot \rangle ^{\boldsymbol {\Theta }}_{X^{\prime }}\) satisfy (conjugate) symmetry, linearity, and positive semi-definiteness properties. The definitenesses of \(\langle \cdot , \cdot \rangle ^{\boldsymbol {\Theta }}_{X}\) and \(\langle \cdot , \cdot \rangle ^{\boldsymbol {\Theta }}_{X^{\prime }}\) on Y and \(Y^{\prime }\), respectively, follow directly from Definition 3.1 and Corollary 3.2. □
Proof of Proposition 3.4
Using Definition 3.1, we have
which yields the right inequality. To prove the left inequality, we assume that \( \| \mathbf {y}^{\prime } \|_{Z^{\prime }}- \varepsilon \| \mathbf {y}^{\prime } \|_{X^{\prime }}\geq 0\). Otherwise, the relation is obvious because \(\| \mathbf {y}^{\prime } \|^{\boldsymbol {\Theta }}_{Z^{\prime }}\geq 0\). By Definition 3.1, we have
which completes the proof. □
Proof of Proposition 3.7
Let us start with the case \(\mathbb {K} = \mathbb {R}\). For the proof, we shall follow standard steps (see, e.g., [38, Section 2.1]). Given a d-dimensional subspace \(V \subseteq \mathbb {R}^{n}\), let \(\mathcal {S} = \{ \mathbf {x} \in V : \ \| \mathbf {x} \| = 1 \}\) be the unit sphere of V. According to [11, Lemma 2.4], for any γ > 0 there exists a γ-net \(\mathcal {N}\) of \(\mathcal {S}\)Footnote 1 satisfying \(\# \mathcal {N} \leq (1+2/\gamma )^{d}\). For η such that 0 < η < 1/2, let \(\boldsymbol {\Theta } \in \mathbb {R}^{k \times n}\) be a rescaled Gaussian or Rademacher matrix with \(k\geq 6\eta ^{-2} ({2} d \log (1+2/\gamma )+ \log (1/\delta ))\). By [1, Lemmas 4.1 and 5.1] and an union bound argument, we obtain for a fixed x ∈ V
Consequently, using a union bound for the probability of success, we have
holds with probability at least 1 − δ. Then, we deduce that
holds with probability at least 1 − δ. Now, let n be some vector in \(\mathcal {S}\). Assuming γ < 1, it can be proven by induction that \( \mathbf {n} = {\sum }_{i\ge 0} \alpha _{i} \mathbf {n}_{i},\) where \(\mathbf {n}_{i} \in \mathcal {N}\) and 0 ≤ αi ≤ γi.Footnote 2 If (7.1) is satisfied, then
and similarly \( \| \boldsymbol {\Theta } \mathbf {n} \|^{2} \geq 1-\frac {\eta }{(1-\gamma )^{2}}\). Therefore, if (7.1) is satisfied, we have
For a given ε ≤ 0.5/(1 − γ)2, let η = (1 − γ)2ε. Since (7.2) holds for an arbitrary vector \(\mathbf {n} \in \mathcal {S}\), using the parallelogram identity, we easily obtain that
holds for all x,y ∈ V if (7.1) is satisfied. We conclude that if \(k\geq 6 \varepsilon ^{-2} (1-\gamma )^{-4} ({2} d \log (1+2/\gamma )+ \log (1/\delta ))\) then Θ is a ℓ2 → ℓ2ε-subspace embedding for V with probability at least 1 − δ. The lower bound for the number of rows of Θ is obtained by taking \(\gamma = \arg \min \limits _{x \in (0,1)}(\log (1+2/x)/(1-x)^{4})\approx 0.0656\).
The statement of the proposition for the case \(\mathbb {K} = \mathbb {C}\) can be deduced from the fact that if Θ is (ε,δ, 2d) oblivious ℓ2 → ℓ2 subspace embedding for \(\mathbb {K} = \mathbb {R}\), then it is (ε,δ,d) oblivious ℓ2 → ℓ2 subspace embedding for \(\mathbb {K} = \mathbb {C}\). A detailed proof of this fact is provided in the supplementary material. To show this, we first note that the real part and the imaginary part of any vector from a d-dimensional subspace \(V^{*} \subseteq \mathbb {C}^{n}\) belong to a certain subspace \(W \subseteq \mathbb {R}^{n}\) with \(\dim (W) \leq 2d\). Further, one can show that if Θ is ℓ2 → ℓ2ε-subspace embedding for W, then it is ℓ2 → ℓ2ε-subspace embedding for V∗. □
Proof of Proposition 3.9
Let \(\boldsymbol {\Theta } \in \mathbb {R}^{k\times n}\) be a P-SRHT matrix, let V be an arbitrary d-dimensional subspace of \(\mathbb {K}^{n}\), and let \(\mathbf {V} \in \mathbb {K}^{n\times d}\) be a matrix whose columns form an orthonormal basis of V. Recall, Θ is equal to the first n columns of matrix \(\boldsymbol {\Theta }^{*} = k^{-1/2} (\mathbf {R}\mathbf {H}_{s}\mathbf {D}) \in \mathbb {R}^{k \times s}\). Next we shall use the fact that for any orthonormal matrix \(\mathbf{V} ^{*} \in \mathbb {K}^{s\times d}\), all singular values of a matrix Θ∗V∗ belong to interval \([\sqrt {1-\varepsilon }, \sqrt {1+\varepsilon }]\) with probability at least 1 − δ. This result is basically a restatement of [12, Lemma 4.1] and [35, Theorem 3.1] including the complex case and with improved constants. It can be shown to hold by mimicking the proof in [35] with a few additional algebraic operations. For a detailed proof of the statement, see the supplementary material.
By taking V∗ with the first n × d block equal to V and zeros elsewhere, and using the fact that ΘV and Θ∗V∗ have the same singular values, we obtain that
holds with probability at least 1 − δ. Using the parallelogram identity, it can be easily proven that relation (4) implies
We conclude that Θ is a (ε,δ,d) oblivious ℓ2 → ℓ2 subspace embedding. □
Proof of Proposition 3.11
Let V be any d-dimensional subspace of X and let V∗ := {Qx : x ∈ V }. Since the following relations hold 〈⋅,⋅〉U = 〈Q⋅,Q⋅〉 and \(\langle \cdot , \cdot \rangle ^{\boldsymbol {\Theta }}_{U} = \langle \mathbf {Q} \cdot , \mathbf {Q} \cdot \rangle _{2}^{\boldsymbol {\Omega }}\), we have that sketching matrix Θ is an ε-embedding for V if and only if Ω is an ε-embedding for V∗. It follows from the definition of Ω that this matrix is an ε-embedding for V∗ with probability at least 1 − δ, which completes the proof. □
Proof of Proposition 4.1 (sketched Cea’s lemma)
The proof exactly follows the one of Proposition 2.2 with \(\| \cdot \|_{U_{r}^{\prime }}\) replaced by \(\| \cdot \|^{\boldsymbol {\Theta }}_{U_{r}^{\prime }}\). □
Proof of Proposition 4.2
According to Proposition 3.4, and by definition of ar(μ), we have
Similarly,
□
Proof of Proposition 4.3
Let \(\mathbf {a} \in \mathbb {K}^{r}\) and x := Ura. Then
By definition,
Combining (7.4) and (7.5) we conclude
The statement of the proposition follows immediately from definitions of \(\alpha ^{\boldsymbol {\Theta }}_{r}(\mu )\) and \(\beta ^{\boldsymbol {\Theta }}_{r}(\mu )\). □
Proof of Proposition 4.4
The proposition directly follows from relations (2.10), (3.1), (3.2), and (4.7). □
Proof of Proposition 4.6
We have
and (4.10) follows by combining (7.6) with (2.15). □
Proof of Proposition 5.1
In total, there are at most \(\binom {m}{r}\)r-dimensional subspaces that could be spanned from m snapshots. Therefore, by using the definition of Θ, the fact that \(\dim (Y_{r}(\mu ))\leq 2 r+1\) and a union bound for the probability of success, we deduce that Θ is a U → ℓ2ε-subspace embedding for Yr(μ), for fixed \(\mu \in \mathcal {P}_{\text {train}}\), with probability at least 1 − m− 1δ. The proposition then follows from another union bound. □
Proof of Proposition 5.4
We have
Moreover, matrix \(\boldsymbol {\Theta } \mathbf {P}^{\boldsymbol {\Theta }}_{U_{r}} \mathbf {U}_{m}\) is the rank-r truncated SVD approximation of \(\mathbf {U}_{m}^{\boldsymbol {\Theta }}\). The statements of the proposition can be then derived from the standard properties of the SVD. □
Proof of Theorem 5.5
Clearly, if Θ is a U → ℓ2ε-subspace embedding for Y, then \(\text {rank}(\mathbf {U}^{\boldsymbol {\Theta }}_{m})\geq r\). Therefore Ur is well-defined. Let \(\{( \lambda _{i}, \mathbf {t}_{i}) \}_{i=1}^{l}\) and Tr be given by Definition 5.3. In general, \(\mathbf {P}^{\boldsymbol {\Theta }}_{U_{r}}\) defined by (5.3) may not be unique. Let us further assume that \(\mathbf {P}^{\boldsymbol {\Theta }}_{U_{r}}\) is provided for x ∈ Um by \( \mathbf {P}^{\boldsymbol {\Theta }}_{U_{r}}\mathbf {x}:= \mathbf {U}_{r}\mathbf {U}_{r}^{\mathrm {H}}\boldsymbol {\Theta }^{\mathrm {H}}\boldsymbol {\Theta }\mathbf {x}, \) where \(\mathbf {U}_{r}= \mathbf {U}_{m}[\frac {1}{\sqrt {\lambda _{1}}}\mathbf {t}_{1}, ..., \frac {1}{\sqrt {\lambda _{r}}}\mathbf {t}_{r}]\). Observe that \( \mathbf {P}^{\boldsymbol {\Theta }}_{U_{r}} \mathbf {U}_{m} = \mathbf {U}_{m} \mathbf {T}_{r} \mathbf {T}_{r}^{\mathrm {H}}. \) For the first part of the theorem, we establish the following inequalities. Let \(\mathbf {Q} \in \mathbb {K}^{n\times n}\) denote the adjoint of a Cholesky factor of RU, then
and
Now, we have
which is equivalent to (5.7).
The second part of the theorem can be proved as follows. Assume that Θ is U → ℓ2ε-subspace embedding for Um, then
which completes the proof. □
Rights and permissions
About this article
Cite this article
Balabanov, O., Nouy, A. Randomized linear algebra for model reduction. Part I: Galerkin methods and error estimation. Adv Comput Math 45, 2969–3019 (2019). https://doi.org/10.1007/s10444-019-09725-6
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10444-019-09725-6