Are resultant methods numerically unstable for multidimensional rootfinding? University of Oxford, 16th June 2015 NA internal seminar Alex Townsend MIT Joint work with Vanni Noferini Betteridge’s law of headlines “Any headline that ends in a question mark can be answered by the word no.” Are resultant methods numerically unstable for multidimensional rootfinding? I do not know What we do know: The two most popular resultant-based numerical rootfinders are numerically unstable. Including the one in chebfun2v/roots... Including Chebfun2’s competitors... [Sommese & Wampler, 05], [Sorber, Barel, Lathauwer, 14], [Nakasukasa, Noferini, T., 14], [Bini & Mario, 13], [Wallack, Emiris, Manocha, 98], [Bujnak, Kukelova, & Pajdla, 08], [Allgower, Georg, & Miranda, 92], [Strumfels, 98]. Alex Townsend @ MIT 1/12 Betteridge’s law of headlines “Any headline that ends in a question mark can be answered by the word no.” Are resultant methods numerically unstable for multidimensional rootfinding? I do not know What we do know: The two most popular resultant-based numerical rootfinders are numerically unstable. Including the one in chebfun2v/roots... Including Chebfun2’s competitors... [Sommese & Wampler, 05], [Sorber, Barel, Lathauwer, 14], [Nakasukasa, Noferini, T., 14], [Bini & Mario, 13], [Wallack, Emiris, Manocha, 98], [Bujnak, Kukelova, & Pajdla, 08], [Allgower, Georg, & Miranda, 92], [Strumfels, 98]. Alex Townsend @ MIT 1/12 Betteridge’s law of headlines “Any headline that ends in a question mark can be answered by the word no.” Are resultant methods numerically unstable for multidimensional rootfinding? I do not know What we do know: The two most popular resultant-based numerical rootfinders are numerically unstable. Including the one in chebfun2v/roots... Including Chebfun2’s competitors... [Sommese & Wampler, 05], [Sorber, Barel, Lathauwer, 14], [Nakasukasa, Noferini, T., 14], [Bini & Mario, 13], [Wallack, Emiris, Manocha, 98], [Bujnak, Kukelova, & Pajdla, 08], [Allgower, Georg, & Miranda, 92], [Strumfels, 98]. Alex Townsend @ MIT 1/12 Betteridge’s law of headlines “Any headline that ends in a question mark can be answered by the word no.” Are resultant methods numerically unstable for multidimensional rootfinding? I do not know What we do know: The two most popular resultant-based numerical rootfinders are numerically unstable. Including the one in chebfun2v/roots... Including Chebfun2’s competitors... [Sommese & Wampler, 05], [Sorber, Barel, Lathauwer, 14], [Nakasukasa, Noferini, T., 14], [Bini & Mario, 13], [Wallack, Emiris, Manocha, 98], [Bujnak, Kukelova, & Pajdla, 08], [Allgower, Georg, & Miranda, 92], [Strumfels, 98]. Alex Townsend @ MIT 1/12 Betteridge’s law of headlines “Any headline that ends in a question mark can be answered by the word no.” Are resultant methods numerically unstable for multidimensional rootfinding? I do not know What we do know: The two most popular resultant-based numerical rootfinders are numerically unstable. Including the one in chebfun2v/roots... Including Chebfun2’s competitors... [Sommese & Wampler, 05], [Sorber, Barel, Lathauwer, 14], [Nakasukasa, Noferini, T., 14], [Bini & Mario, 13], [Wallack, Emiris, Manocha, 98], [Bujnak, Kukelova, & Pajdla, 08], [Allgower, Georg, & Miranda, 92], [Strumfels, 98]. Alex Townsend @ MIT 1/12 Multidimensional rootfinding setting Find all the solutions to: ¨ ˛ p1 px1 , . . . , xd q ˚ ‹ .. ˝ ‚ “ 0, . px1 , . . . , xd q P Ωd Ă Cd . pd px1 , . . . , xd q Polynomials are of maximal degree n (at most degree n in each variable) Isolated solutions ùñ ď d!nd solutions (Bernstein’s Theorem) SIAM digit challenge problem [Trefethen, 02] 1 Easiest possible scenario: 1. Simple solutions, i.e., the Jacobian J px ˚ q´1 exists 2. No solutions at infinity 0.5 w −0.5 −1 −1 Alex Townsend @ MIT 257s 0 −0.5 0 0.5 1 2/12 Inherit robustness from eigenvalue solver Main idea: Inherit robustness from eigenvalue solver ˛ p1 px1 , . . . , xd q ˚ ‹ .. ˝ ‚“ 0 . ¨ pd px1 , . . . , xd q Av “ λv For a simple root x ˚ , absolute condition number is: For a semisimple eigenvalue λ˚ , absolute condition number is: κ2 px ˚ q “ }J px ˚ q´1 }2 κpλ˚ , A q Alex Townsend @ MIT 3/12 Overview d“1 dě2 p px q “ 0 p1 px q “ ¨ ¨ ¨ “ pd px q “ 0 Not much going on here. n n ď d!nd v “ λv companion, colleague, comrade v “ λv ď d!nd linearization of a (Cayley) resultant matrix Conditioning analysis: [Van Dooren & Dewilde, 83], [Edelman & Murakami, 95], [Terán, Dopico, & Mackey, 10], [Noferini & Pérez, 15] Alex Townsend @ MIT Conditioning analysis: [Noferini & T., 15] 4/12 Overview d“1 dě2 p px q “ 0 p1 px q “ ¨ ¨ ¨ “ pd px q “ 0 Lots going on inside here! Not much going on here. n n ď d!nd v “ λv companion, colleague, comrade v “ λv ď d!nd linearization of a (Cayley) resultant matrix Conditioning analysis: [Van Dooren & Dewilde, 83], [Edelman & Murakami, 95], [Terán, Dopico, & Mackey, 10], [Noferini & Pérez, 15] Alex Townsend @ MIT Conditioning analysis: [Noferini & T., 15] 4/12 Hidden-variable resultant methods p “ 0, q “ 0 ”Hide” one of the variables, say x1 : pj rx1 spx2 , . . . , xd q “ n ÿ i2 “¨¨¨“id “0 ci2 ,...,id px1 q d ź φis pxs q s “2 A resultant R is a polynomial such that (ignoring solutions at infinity): Rpx1˚ q “ 0 ðñ Dpx2˚ , . . . , xd˚ q P Cd ´1 s.t. p1 px ˚ q “ ¨ ¨ ¨ “ pd px ˚ q “ 0, Alex Townsend @ MIT R “ resultant ? ? ? 5/12 Hidden-variable resultant methods p “ 0, q “ 0 ”Hide” one of the variables, say x1 : pj rx1 spx2 , . . . , xd q “ n ÿ i2 “¨¨¨“id “0 ci2 ,...,id px1 q d ź φis pxs q s “2 A resultant R is a polynomial such that (ignoring solutions at infinity): Rpx1˚ q “ 0 ðñ Dpx2˚ , . . . , xd˚ q P Cd ´1 s.t. p1 px ˚ q “ ¨ ¨ ¨ “ pd px ˚ q “ 0, Alex Townsend @ MIT R “ resultant ? ? ? 5/12 A linear example ¨ ˛ x1 ˚ ‹ A ˝ ... ‚` b “ 0 xd ¨ ˛ 1 ˚ “ ‰ ˚ x2 ‹ ‹ x A p :, 1 q ` b A p :, 2 : end q ˚ .. ‹ “ 0 1 looooooooooooooooooomooooooooooooooooooon ˝.‚ “A1 `x1 A p:,1qe1T xd ` ˘ det A1 ` x1 A p:, 1qe1T “ detpA1 q` x1 detpA q “ 0 Linear rootfinding problem ”Hide x1 ” Matrix determinant lemma Cramer’s rule: x1 “ ´ detpA1 q{ detpA q Alex Townsend @ MIT 6/12 A linear example ¨ ˛ x1 ˚ ‹ A ˝ ... ‚` b “ 0 xd ¨ ˛ 1 ˚ “ ‰ ˚ x2 ‹ ‹ x A p :, 1 q ` b A p :, 2 : end q ˚ .. ‹ “ 0 1 looooooooooooooooooomooooooooooooooooooon ˝.‚ “A1 `x1 A p:,1qe1T xd ` ˘ det A1 ` x1 A p:, 1qe1T “ detpA1 q` x1 detpA q “ 0 Linear rootfinding problem ”Hide x1 ” Matrix determinant lemma Cramer’s rule: x1 “ ´ detpA1 q{ detpA q Alex Townsend @ MIT 6/12 A linear example ¨ ˛ x1 ˚ ‹ A ˝ ... ‚` b “ 0 xd ¨ ˛ 1 ˚ “ ‰ ˚ x2 ‹ ‹ x A p :, 1 q ` b A p :, 2 : end q ˚ .. ‹ “ 0 1 looooooooooooooooooomooooooooooooooooooon ˝.‚ “A1 `x1 A p:,1qe1T xd ` ˘ det A1 ` x1 A p:, 1qe1T “ detpA1 q` x1 detpA q “ 0 Linear rootfinding problem ”Hide x1 ” Matrix determinant lemma Cramer’s rule: x1 “ ´ detpA1 q{ detpA q Alex Townsend @ MIT 6/12 A linear example ¨ ˛ x1 ˚ ‹ A ˝ ... ‚` b “ 0 xd ¨ ˛ 1 ˚ “ ‰ ˚ x2 ‹ ‹ x A p :, 1 q ` b A p :, 2 : end q ˚ .. ‹ “ 0 1 looooooooooooooooooomooooooooooooooooooon ˝.‚ “A1 `x1 A p:,1qe1T xd ` ˘ det A1 ` x1 A p:, 1qe1T “ detpA1 q` x1 detpA q “ 0 Linear rootfinding problem ”Hide x1 ” Matrix determinant lemma Cramer’s rule: x1 “ ´ detpA1 q{ detpA q Gabriel Cramer Alex Townsend @ MIT 6/12 A linear example ¨ ˛ x1 ˚ ‹ A ˝ ... ‚` b “ 0 xd ¨ ˛ 1 ˚ “ ‰ ˚ x2 ‹ ‹ x A p :, 1 q ` b A p :, 2 : end q ˚ .. ‹ “ 0 1 looooooooooooooooooomooooooooooooooooooon ˝.‚ “A1 `x1 A p:,1qe1T xd ` ˘ det A1 ` x1 A p:, 1qe1T “ detpA1 q` x1 detpA q “ 0 Linear rootfinding problem ”Hide x1 ” Matrix determinant lemma Cramer’s rule: x1 “ ´ detpA1 q{ detpA q Gabriel Cramer Alex Townsend @ MIT 6/12 Resultant method with matrix polynomials [Nakatsukasa, Noferini, & T., 14] Undo the determinant: Rpx1 q “ detpR px1 qq Matrix polynomial Polynomial eigenvalue problem: Consider ˜ ¸ T7 px qT7 py qp2 ` xy q “ 0, T5 px qT5 py qp1 ` xy q where Tk px q “ cospk cos´1 px qq. R px1 q has semisimple eigenvalues R px1 qv “ 0 Work by (and many others): Leiba Rodman Alex Townsend @ MIT Israel Gohberg Françoise Tisseur Absolute error “ 8.2 ˆ 10´16 7/12 Cayley resultant matrix Skipping the eye-watering tensor manipulations RCAYLEY px1 q is a matrix polynomial of size at most pd ´ 1q!nd ´1 of degree dn. It is the matricization of the tensor of expansion coefficients of ¨ ˛ p1 rx1 sps1 , s2 , . . . , sd ´1 q . . . pd rx1 sps1 , s2 . . . , sd ´1 q ˚ p rx spt , s . . . , s q . . . p rx spt , s . . . , s q ‹ O d ´1 ź d ´1 d 1 1 2 d ´1 ‹ ˚ 1 1 1 2 ‹ fCayley “ det ˚ psi ´ ti q. .. .. ... ˚ ‹ . . ˝ ‚ i “1 p1 rx1 spt1 , t2 , . . . , td ´1 q . . . pd rx1 spt1 , t2 , . . . , td ´1 q Devastating: In any degree-graded basis, there exists a polynomial system and a simple root x ˚ such that Absolute conditioning can increase by an exponential factor Alex Townsend @ MIT κpRCAYLEY , x1˚ q ě ||J px ˚ q´1 ||d . [Noferini & T., 15] 8/12 Cayley resultant matrix Skipping the eye-watering tensor manipulations RCAYLEY px1 q is a matrix polynomial of size at most pd ´ 1q!nd ´1 of degree dn. It is the matricization of the tensor of expansion coefficients of ¨ ˛ p1 rx1 sps1 , s2 , . . . , sd ´1 q . . . pd rx1 sps1 , s2 . . . , sd ´1 q ˚ p rx spt , s . . . , s q . . . p rx spt , s . . . , s q ‹ O d ´1 ź d ´1 d 1 1 2 d ´1 ‹ ˚ 1 1 1 2 ‹ fCayley “ det ˚ psi ´ ti q. .. .. ... ˚ ‹ . . ˝ ‚ i “1 p1 rx1 spt1 , t2 , . . . , td ´1 q . . . pd rx1 spt1 , t2 , . . . , td ´1 q Devastating: In any degree-graded basis, there exists a polynomial system and a simple root x ˚ such that Absolute conditioning can increase by an exponential factor Alex Townsend @ MIT κpRCAYLEY , x1˚ q ě ||J px ˚ q´1 ||d . [Noferini & T., 15] 8/12 Cayley resultant matrix Skipping the eye-watering tensor manipulations RCAYLEY px1 q is a matrix polynomial of size at most pd ´ 1q!nd ´1 of degree dn. It is the matricization of the tensor of expansion coefficients of ¨ ˛ p1 rx1 sps1 , s2 , . . . , sd ´1 q . . . pd rx1 sps1 , s2 . . . , sd ´1 q ˚ p rx spt , s . . . , s q . . . p rx spt , s . . . , s q ‹ O d ´1 ź d ´1 d 1 1 2 d ´1 ‹ ˚ 1 1 1 2 ‹ fCayley “ det ˚ psi ´ ti q. .. .. ... ˚ ‹ . . ˝ ‚ i “1 p1 rx1 spt1 , t2 , . . . , td ´1 q . . . pd rx1 spt1 , t2 , . . . , td ´1 q Devastating: In any degree-graded basis, there exists a polynomial system and a simple root x ˚ such that Absolute conditioning can increase by an exponential factor Alex Townsend @ MIT κpRCAYLEY , x1˚ q ě ||J px ˚ q´1 ||d . [Noferini & T., 15] 8/12 Sylvester resultant matrix A popular alternative in 2D is the Sylvester resultant ř řn k k RSYLV px1 q associated with p1px1, x2 q “ m k “0 akpx1 qx2 and p2px1, x2 q “ k “0 bkpx1 qx2 : ¨ a0 px1 q ˚ ˚ ˚ ˚ ˚ ˚ RSylv px1 q “ ˚ ˚b0 px1 q ˚ ˚ ˚ ˚ ˝ a 1 px 1 q ... a m px 1 q .. .. .. . . . a 0 px 1 q a 1 px 1 q b1px1 q ... bn px1 q .. .. .. . . b 0 px 1 q . b 1 px 1 q .. . ... .. . ... ˛ , / / . ‹ / ‹ ‹ n rows ‹ / / ‹ / a m px 1 q‹ ‹, ‹/ ‹/ ‹/ . ‹ ‹ m rows ‚/ / / bn px1 q - Devastating: For any degree-graded basis, there exists a polynomial system and a simple root x ˚ such that Absolute conditioning can be squared Alex Townsend @ MIT κpRSYLV , x1˚ q ě ||J px ˚ q´1 ||2 . [Noferini & T., 15] 9/12 Sylvester resultant matrix A popular alternative in 2D is the Sylvester resultant ř řn k k RSYLV px1 q associated with p1px1, x2 q “ m k “0 akpx1 qx2 and p2px1, x2 q “ k “0 bkpx1 qx2 : ¨ a0 px1 q ˚ ˚ ˚ ˚ ˚ ˚ RSylv px1 q “ ˚ ˚b0 px1 q ˚ ˚ ˚ ˚ ˝ a 1 px 1 q ... a m px 1 q .. .. .. . . . a 0 px 1 q a 1 px 1 q b1px1 q ... bn px1 q .. .. .. . . b 0 px 1 q . b 1 px 1 q .. . ... .. . ... ˛ , / / . ‹ / ‹ ‹ n rows ‹ / / ‹ / a m px 1 q‹ ‹, ‹/ ‹/ ‹/ . ‹ ‹ m rows ‚/ / / bn px1 q - Devastating: For any degree-graded basis, there exists a polynomial system and a simple root x ˚ such that Absolute conditioning can be squared Alex Townsend @ MIT κpRSYLV , x1˚ q ě ||J px ˚ q´1 ||2 . [Noferini & T., 15] 9/12 Absolute versus relative conditioning Quote: ”Usually, it is the relative condition number that is of interest, but it is more convenient to state results for the absolute condition number” [N. Higham] Devastating example in absolute and relative conditioning: Let u be a small real positive parameter and d ě 2. Take ¯ ´? ? 2 2 2 p2i ´1 px q “ x2i ´1 ` u 2 x2i ´1 ` 2 x2i , Ωd “ r´1, 1sd , ´? ¯ ? p2i px q “ x2i2 ` u 22 x2i ´1 ´ 22 x2i , 1 ď i ď td {2u, where if d is odd take pd px q “ xd2 ` uxd . Alex Townsend @ MIT d 2 3 5 10 u “ ||J px ˚ q´1 ||2 6.7 ˆ 107 1.6 ˆ 105 1.4 ˆ 103 36.7 10/12 Absolute versus relative conditioning Quote: ”Usually, it is the relative condition number that is of interest, but it is more convenient to state results for the absolute condition number” [N. Higham] Devastating example in absolute and relative conditioning: Let u be a small real positive parameter and d ě 2. Take ¯ ´? ? 2 2 2 p2i ´1 px q “ x2i ´1 ` u 2 x2i ´1 ` 2 x2i , Ωd “ r´1, 1sd , ´? ¯ ? p2i px q “ x2i2 ` u 22 x2i ´1 ´ 22 x2i , 1 ď i ď td {2u, where if d is odd take pd px q “ xd2 ` uxd . Alex Townsend @ MIT d 2 3 5 10 u “ ||J px ˚ q´1 ||2 6.7 ˆ 107 1.6 ˆ 105 1.4 ˆ 103 36.7 10/12 Absolute versus relative conditioning Quote: ”Usually, it is the relative condition number that is of interest, but it is more convenient to state results for the absolute condition number” [N. Higham] Devastating example in absolute and relative conditioning: Let u be a small real positive parameter and d ě 2. Take ¯ ´? ? 2 2 2 p2i ´1 px q “ x2i ´1 ` u 2 x2i ´1 ` 2 x2i , Ωd “ r´1, 1sd , ´? ¯ ? p2i px q “ x2i2 ` u 22 x2i ´1 ´ 22 x2i , 1 ď i ď td {2u, where if d is odd take pd px q “ xd2 ` uxd . Alex Townsend @ MIT d 2 3 5 10 u “ ||J px ˚ q´1 ||2 6.7 ˆ 107 1.6 ˆ 105 1.4 ˆ 103 36.7 10/12 Negative result with (hopefully) a positive impact dě2 A BIG bag of numerical tricks Numerical trick Will it help? Newton polishing No Homogenization No Change of basis No Domain subdivision For low d Regularization For large solutions Local refinement Only sometimes Rotate coordinates Not always Variable transform It may or may not Randomization Not analyzed Reformulation as min Not analyzed p1 px q “ ¨ ¨ ¨ “ pd px q “ 0 It’s got to happen before here! ď d!nd v “ λv ď d!nd linearization of a resultant matrix Conditioning analysis: [Noferini & T., 15] Alex Townsend @ MIT 11/12 Thank you If you know of a practical and robust numerical multidimensional rootfinding. Call me! +1 (857) 204-3609 Thanks also to: Anthony Austin, Daniel Bates, John Boyd, Martin Lotz, Nick Higham, Gregorio Malajovich, Yuji Nakatsukasa, Bor Plestenjak, Andrew Sommese. Alex Townsend @ MIT 12/12