Subspace-based dimension reduction for forward and inverse uncertainty quantification

Slides for talk at Isaac Newton Institute workshop on "UQ for inverse problems in complex systems"

Transcript

1. Subspace-based dimension reduction for forward and inverse uncertainty quantification PAUL CONSTANTINE Assistant Professor Department of Computer Science University of Colorado, Boulder activesubspaces.org! @DrPaulynomial! SLIDES AVAILABLE UPON REQUEST DISCLAIMER: These slides are meant to complement the oral presentation. Use out of context at your own risk.
2. What I’m not working on and why
3. What I’m not working on and why surrogate / reduced-order models adaptivity / partitioning the parameter space
4. Troubles in high dimensions
5. Troubles in high dimensions the information-based complexity (IBC) notion of tractable
6. Troubles in high dimensions the information-based complexity (IBC) notion of tractable volume of a unit ball in m dimensions: 0 10 20 30 40 50 Dimension 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 m-ball volume 0 10 20 30 40 50 Dimension 10-12 10-10 10-8 10-6 10-4 10-2 100 m-ball volume ⇡ m 2 (m 2 + 1)
7. Database Theory --- ICDT'99, Springer (1999)
8. When Is “Nearest Neighbor” Meaningful? 0 10 20 30 40 50 Dimension 100 101 102 103 104 105 106 107 E[max dist / min dist] 1e1 1e2 1e3 1e4 1e5 0 10 20 30 40 50 Dimension 10-2 100 102 104 106 108 1010 Std[max dist / min dist] 1e1 1e2 1e3 1e4 1e5
9. The best way to fight the curse is to reduce the dimension. But what is dimension reduction?
10. ( u( s , t; x ) ) spatial / temporal classical physics, mechanics, applied mathematics parameter principal component analysis, Karhunen-Loève, dimensional analysis state model reduction POD, reduced basis, empirical interpolation, … see, e.g., Benner, Cohen, Ohlberger, and Willcox (SIAM, 2017) PDE solution space time parameters functional of interest What is dimension reduction? x 2 X ✓ Rm t 2 [0, T] ⇢ R s 2 ⌦ ⇢ R3 u 2 Wk,2 ⇢ L2 2 R
11. ( u( s , t; x ) ) x 2 X ✓ Rm t 2 [0, T] ⇢ R s 2 ⌦ ⇢ R3 u 2 Wk,2 ⇢ L2 2 R functional of interest What is dimension reduction? x 7! How can we exploit the map from parameters to quantity-of-interest to reduce the parameter dimension? f( x ) parameters ASSUME PARAMETERS ARE INDEPENDENT
12. Hypersonic scramjet models Constantine, Emory, Larsson, and Iaccarino (2015) Aerospace design Lukaczyk, Palacios, Alonso, and Constantine (2014) Integrated hydrologic models Jefferson, Gilbert, Constantine, and Maxwell (2015) Solar cell models Constantine, Zaharatos, and Campanelli (2015) Magnetohydrodynamics models Glaws, Constantine, Shadid, and Wildey (2017) Ebola transmission models Diaz, Constantine, Kalmbach, Jones, and Pankavich (arXiv, 2016) Lithium ion battery model Constantine and Doostan (2017) Automobile design Othmer, Lukaczyk, Constantine, and Alonso (2016) f( x )
13. Z f( x ) d x APPROXIMATION OPTIMIZATION INTEGRATION ˜ f( x ) ⇡ f( x ) minimize x f( x ) INVERSION given y, find x such that y ⇡ f( x )
14. f ( x ) ⇡ r X k=1 fk,1( x1) · · · fk,m( xm) f( x ) ⇡ p X k=1 ak k( x ), k a k0 ⌧ p f ( x ) ⇡ f1( x1) + · · · + fm( xm) Structure-exploiting methods STRUCTURE METHODS Separation of variables [Beylkin & Mohlenkamp (2005)], Tensor-train [Oseledets (2011)], Adaptive cross approximation [Bebendorff (2011)], Proper generalized decomposition [Chinesta et al. (2011)], … Compressed sensing [Donoho (2006), Candès & Wakin (2008)], … Sparse grids [Bungartz & Griebel (2004)], HDMR [Sobol (2003)], ANOVA [Hoeffding (1948)], …
15. “Even more understanding is lost if we consider each thing we can do to data only in terms of some set of very restrictive assumptions under which that thing is best possible—assumptions we know we CANNOT check in practice.” “Many algorithms … aim to diminish the ‘curse of dimensionality.’ Such algorithms take advantage of special properties of the functions being treated, such as alignment with the axes, but their authors do not always emphasize this aspect of their methods.”
16. www.youtube.com/watch?v=mJvKzjT6lmY
17. Design a jet nozzle under uncertainty (DARPA SEQUOIA project) 10-parameter engine performance model (See animation at https://youtu.be/Fek2HstkFVc)
18. Do these structures arise in real models? (Yes.)
19. Hypersonic scramjet models Constantine, Emory, Larsson, and Iaccarino (2015) Evidence of 1d ridge structures across science and engineering models
20. Integrated jet nozzle models Alonso, Eldred, Constantine, Duraisamy, Farhat, Iaccarino, and Jakeman (2017) Evidence of 1d ridge structures across science and engineering models
21. Integrated hydrologic models Jefferson, Gilbert, Constantine, and Maxwell (2015) Evidence of 1d ridge structures across science and engineering models
22. −2 −1 0 1 2 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Active Variable 1 Lift Lukaczyk, Constantine, Palacios, and Alonso (2014) −2 −1 0 1 2 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 Active Variable 1 Drag Aerospace vehicle geometries Evidence of 1d ridge structures across science and engineering models
23. In-host HIV dynamical models T-cell count Loudon and Pankavich (2016) Evidence of 1d ridge structures across science and engineering models
24. Solar cell circuit models −2 −1 0 1 2 0 0.05 0.1 0.15 0.2 0.25 Active Variable 1 P max (watts) Constantine, Zaharatos, and Campanelli (2015) Evidence of 1d ridge structures across science and engineering models
25. Atmospheric reentry vehicle model Cortesi, Constantine, Magin, and Congedo (hal, 2017) −1 0 1 ˆ wT q x 0.4 0.6 0.8 1.0 1.2 Stagnation heat flux qst ×107 −1 0 1 ˆ wT p x 20000 40000 60000 80000 100000 Stagnation pressure pst Evidence of 1d ridge structures across science and engineering models
26. Magnetohydrodynamics generator model -1 0 1 wT 1 x 0 5 10 15 f(x) Average velocity Glaws, Constantine, Shadid, and Wildey (2017) -1 0 1 wT 1 x 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 f(x) Induce magnetic field Evidence of 1d ridge structures across science and engineering models
27. Lithium ion battery model 2 0 2 wT x 3.65 3.70 Voltage [V] Constantine and Doostan (2017) 2 0 2 wT x 2.0 2.2 Capacity [mAh·cm 2] Evidence of 1d ridge structures across science and engineering models
28. Automobile geometries Othmer, Lukaczyk, Constantine, and Alonso (2016) Evidence of 1d ridge structures across science and engineering models
29. -4 -2 0 2 4 Quantity of interest #10-3 0 1 2 3 4 5 -4 -2 0 2 4 Quantity of interest #10-4 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 Long length scale Short length scale Constantine, Dow, and Wang (2014) r · (aru) = 1, s 2 D u = 0, s 2 1 n · aru = 0, s 2 2 Input field Solution Short corr Long corr Evidence of 1d ridge structures across science and engineering models
30. Do these structures arise in real models? (Yes, but not every model.)
31. Gilbert, Jefferson, Constantine, and Maxwell (2016) No evidence of 1d structure: A subsurface hydrology problem 0 100 200 300 0 100 200 300 0 20 40 x (m) y (m) z (m) Student Version of MATLAB Domain Hydraulic conductivities Unsaturated case Saturated case
32. Constantine, Hokanson, and Kouri (2018) D DC R 5 0 −5 0 5 −5 Domain Desired state (Re, Im) u k2u = s No evidence of 1d structure: An acoustic scattering model f(⇠)
33. f( x ) Jupyter notebooks: github.com/paulcon/as-data-sets
34. f( x ) ⇡ g(UT x ) Ridge approximations UT : Rm ! Rn g : Rn ! R where Constantine, Eftekhari, Hokanson, and Ward (2017)
35. Ridge approximations A subset of relevant literature Approximation theory: Mayer et al. (2015), Pinkus (2015), Diaconis and Shahshahani (1984), Donoho and Johnstone (1989) Compressed sensing: Fornasier et al. (2012), Cohen et al. (2012), Tyagi and Cevher (2014) Statistical regression: Friedman and Stuetzle (1981), Ichimura (1993), Hristache et al. (2001), Xia et al. (2002) Uncertainty quantification: Tipireddy and Ghanem (2014); Lei et al. (2015); Stoyanov and Webster (2015); Tripathy, Bilionis, and Gonzalez (2016); Li, Lin, and Li (2016); … f( x ) ⇡ g(UT x )
36. f( x ) ⇡ g(UT x ) What is U? What is the approximation error? What is g? Constantine, Eftekhari, Hokanson, and Ward (2017) Ridge approximations
37. C = Z rf( x ) rf( x )T ⇢( x ) d x = W ⇤W T Define the active subspace The average outer product of the gradient and its eigendecomposition, Partition the eigendecomposition, Rotate and separate the coordinates, ⇤ =  ⇤1 ⇤2 , W = ⇥ W 1 W 2 ⇤ , W 1 2 Rm⇥n x = W W T x = W 1W T 1 x + W 2W T 2 x = W 1y + W 2z active variables inactive variables f = f( x ), x 2 Rm, rf( x ) 2 Rm, ⇢ : Rm ! R + Constantine, Dow, and Wang (2014) Some relevant literature Statistical regression: Samarov (1993), Hristache et al. (2001) Machine learning: Mukerjee, Wu, and Xiao (2010); Fukumizu and Leng (2014) Signal processing: van Trees (2001) The function, its gradient vector, and a given weight function:
38. C = Z rf( x ) rf( x )T ⇢( x ) d x = W ⇤W T Define the active subspace The function, its gradient vector, and a given weight function: The average outer product of the gradient and its eigendecomposition: f = f( x ), x 2 Rm, rf( x ) 2 Rm, ⇢ : Rm ! R + Constantine, Dow, and Wang (2014) i = Z w T i rf( x ) 2 ⇢( x ) d x , i = 1, . . . , m average, squared, directional derivative along eigenvector eigenvalue Eigenvalues measure ridge structure with eigenvectors:
39. Poincaré constant eigenvalues associated with inactive subspace f( x ) µ(W T 1 x ) L2(⇢)  C ( n+1 + · · · + m)1 2 Constantine, Dow, and Wang (2014) The eigenvalues measure the approximation error conditional expectation first n eigenvectors (i.e., the active subspace)
40. (1) Draw samples: (2) Compute: and fj = f( xj) (3) Approximate with Monte Carlo, and compute eigendecomposition Equivalent to SVD of samples of the gradient Called an active subspace method in T. Russi’s 2010 Ph.D. thesis, Uncertainty Quantification with Experimental Data in Complex System Models C ⇡ 1 N N X j=1 rfj rfT j = ˆ W ˆ ⇤ ˆ W T 1 p N ⇥ rf1 · · · rfN ⇤ = ˆ W p ˆ ⇤ ˆ V T rfj = rf( xj) Constantine, Dow, and Wang (2014), Constantine and Gleich (2015, arXiv) xj ⇠ ⇢( x ) Estimate the active subspace with Monte Carlo
41. N = ⌦ ✓ L2 1 2 k "2 log( m ) ◆ = ) | k ˆk |  k " How many gradient samples? number of samples eigenvalue error (w.h.p.) subspace error (w.h.p.) Constantine and Gleich (2015) via Gittens and Tropp (2011), Stewart (1973) N = ⌦ ✓ L2 1"2 log( m ) ◆ = ) dist( W 1, ˆ W 1)  4 1" n n+1 bound on gradient dimension number of samples bound on gradient dimension
42. 1 p N ⇥ rf1 · · · rfN ⇤ ⇡ ˆ W 1 q ˆ ⇤1 ˆ V T 1 Low-rank approximation of the collection of gradients: Low-dimensional linear approximation of the gradient: f( x ) ⇡ g ⇣ ˆ W T 1 x ⌘ Approximate a function of many variables by a function of a few linear combinations of the variables: ✔ ✖ ✖ Remember the problem to solve span ( ˆ W 1) ⇡ { rf( x ) : x 2 supp ⇢( x ) }
43. f( x ) ⇡ g(UT x ) What is U? Define the error function: R(U) = 1 2 Z (f( x ) µ(UT x ))2 ⇢( x ) d x Minimize the error: minimize U R ( U ) subject to U 2 G ( n, m ) Grassmann manifold of n-dimensional subspaces Constantine, Eftekhari, Hokanson, and Ward (2017) Ridge approximations best approximation
44. (1) Draw samples: (2) Compute: fj = f( xj) (3) Minimize the misfit Minimize over polynomials and subspaces Constantine, Eftekhari, Hokanson, and Ward (2017), Hokanson and Constantine (2018) xj ⇠ ⇢( x ) Estimate the optimal subspace with discrete least squares minimize g2P p(Rn) U2G(n,m) N X j=1 ⇣ fj g(UT xj) ⌘2
45. Assessing ridge or near-ridge structure Z rf( x ) rf( x )T ⇢( x ) d x Derivative-based ideas: eigenvalues suggest structure, eigenvectors give directions Active subspaces [Constantine et al. (2014), Russi (2010)], Gradient outer product [Mukherjee et al. (2010)], Outer product of gradient [Hristache et al. (2001)] Z r2f( x ) ⇢( x ) d x Principal Hessian directions [Li (1992)], Likelihood-informed subspaces [Cui et al. (2014)] Ideas for approximating these without gradients: finite differences [Constantine & Gleich (2015), Lewis et al. (2016)], polynomial approximations [Yang et al (2016), Tippireddy & Ghanem (2014)], kernel approximations [Fukumizu & Leng (2014)] See Samarov’s average derivative functionals [Samarov (1993)]
46. Assessing ridge or near-ridge structure Sufficient dimension reduction ideas: eigenvalues suggest structure, eigenvectors give directions Sliced inverse regression [Li (1991), Glaws et al. (2018)] Sliced average variance estimation [Cook & Weisberg (1991), Glaws et al. (2018)] E ⇥ E[ x |f] E[ x |f]T ⇤ E h ( I Cov[x |f ]) 2 i E ⇥ ( x1 x2) ( x1 x2)T | |f( x1) f( x2)|  ⇤ Contour regression [Li et al. (2005)] These are population metrics; data produces sample estimates.
47. minimize g, U f( x ) g(UT x ) Assessing ridge or near-ridge structure Optimization ideas: optimum residual suggests structure, optimizer gives directions Ridge approximation [Constantine et al. (2017, 2018)], Minimum average variance estimation [Xia et al. (2002)], Gaussian processes [Vivarelli & Wiliams (1999), Tripathy et al. (2016)] Projection pursuit regression [Friedman & Stuetzle (1981), Huber (1985)] Likelihood-based sufficient dimension reduction [Cook & Forzani (2009)] minimize gi, ui f( x ) X i gi( u T i x ) ! maximize U E [ k PU Cov[x |f ] PU k ⇤ ] All nonconvex optimizations. Some on Grassmann manifold of subspaces.
48. (1)  Exploitable + for dimension reduction, not just cheap surrogate (2)  Insights + which variables are important (3)  Discoverable / checkable + eigenvalues + non-residual metrics: + plots in 1 and 2d E[ Var[ f | UT x ] ] Why I like ridge structure
49. The best way to fight the curse of dimensionality is to reduce the dimension! There are many notions of important subspaces; active subspaces are one of them. Ridge structures are discoverable and exploitable for forward and inverse UQ. TAKE HOMES
50. Low-dimensional subspaces in inverse UQ How do these compare? See Zahm and Marzouk (in prep) And lots of applications! Related sloppy models:
51. How well can you estimate the subspaces? What if my model doesn’t fit your setup? (no gradients, multiple outputs, correlated inputs, …) How does this help with emulation? PAUL CONSTANTINE Assistant Professor University of Colorado Boulder activesubspaces.org! @DrPaulynomial! QUESTIONS? Active Subspaces SIAM (2015)