21
Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’1 5

Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Embed Size (px)

Citation preview

Page 1: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Salt Lake City, Oct. 14th, 2015

Re-parameterization reduces irreducible geometric constraint systems

H. Barki, L. Fang, D. Michelucci,S. Foufou

GDSPM’15

Page 2: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Summary

• Introduction and motivation

• Understanding Re-Parameterization (RP)

• How reduction speeds up Linear Algebra (LA)?

• How Re-Parameterization (RP) speeds up LA?

• How Re-Parameterization speeds-up P-adic methods?

• How Re-Parameterization speeds-up interval solvers?

• Experimental results

• Re-Parameterization at a higher level

• Conclusions, emerging questions and future work 2

Page 3: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Introduction and motivation

Background• Seminal work [8] about the Locus Intersection Method (LIM)

– Gao et al. generated all the possible irreducible and structurally well-constrained sys. involving up to 6 primitives (683 basic configurations)

– Basic configurations correspond to 3D sub-problems in GCSP and CAD but in general, they do not have a closed form solution

– Gao et al. proposed LIM1 and LIM2 with 1 and 2 key unknowns (parameters) for solving them -> Re-Parameterization (RP)

• Decomposition techniques of Ait Aoudia et al. [9]– Usage of bipartite graphs to decompose large systems into sub-systems– Efficient decompose or reduce well-constrained sys. into irreducible well-

constrained sys. -> efficient solving

3

Page 4: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Introduction and motivation

• Main problems– The methods of Ait Aoudia et al. [9] do not apply to (re-parameterized

irreducible) sys. of Gao et al. [8] and also those of [10-12]– For basic configurations with more than 6 primitives/complex primitives,

using LIM1 (k=1 key unknown) is not enough– LIMk (k >= 2) become much less convenient

• Contribution– We propose a technique for efficiently reducing/unlocking irreducible RP sys.

like those of [8,10–12] -> the decomposition methods proposed in [9] become applicable

– We also show that it is possible to benefit from the decomposition of [9] even when the values of the key unknowns are not known and also for k > 2 -> we propose to exploit RP at the lowest Linear Algebra (LA) level because a wide range of solvers uses these LA routines

4

Page 5: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Understanding Re-Parameterization

• A 2D example of a GCSP (9 edge lengths)– Under-constrained sys. (12 eqs./9 vars.) -> well-

constrained sys. (9 eqs./vars.) by fixing A’B’C’– Finite number of solutions for generic values– The sys. is structurally irreducible

• If the length u (key unknown) was known– The sys. would be easily solved by computing point

A, then B, then C (CC’ equation can be ignored)– The sys. becomes decomposable/reducible (1 eq.

in 1 unknown u)– Problem: the value of u is unknown!– Other complications due to multi-functions A(u),

B(u), and C(u)!

• A complex 3D hexahedron example (cf. paper) 5

Page 6: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Reduction speeds-up linear algebra

• Reduction/decomposition methods of [9]– Re-order unknowns/equations so that the Jacobian becomes block lower

triangular -> the decomposition becomes more visible– Consequence: the processes of solving a LS become more efficient, thanks to

the use of forward block substitution (complexity drop from O(n3) using Gauss pivoting to O(n2) for lower triangular sys. of size n [17])

– Consider the block lower triangular matrix

– Let us show how to exploit the decomposition of M in order to solve a system MX = B or to invert matrix M

– Note: floating point arithmetic is involved for Newton-Raphson or homotopy, interval arithmetic for interval solvers, or or rational/modular arithmetic for p-adic methods)

6

Page 7: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Reduction speeds-up linear algebra

• Inverting a block lower triangular matrix– K = M-1 is also block lower triangular

with and – It is it always possible to avoid the inversion of matrix M as it suffices to just

solve less than n linear systems (n is the size of M) as follows:

• Solving a linear sys. MX = B with X = (X1, X2, . . .)t, B = (B1, B2, . . .)t

– We have to successively solve the systems

– The smaller are the blocks Ml,l -> the greater is the number of null blocks under the diagonal -> the more important is the solving process speed-up

– For a matrix M of size n and blocks of small size t O(1)∈ , the complexity drops from n3 to n (full and detailed complexity study in the paper) -> the speed-up is considerable

7

Page 8: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

RP speeds-up linear algebra

• Benefits of RP to Linear Algerba: case of Newton-Raphson method– Let us assume a well-constrained system of equations, with a known set of

key unknowns and its Jacobian structure already computed by [9]– Each Newton-Raphson iteration solves a linear system having the structure:

where A is non-singular block lower triangular, U is the vector of key unknowns, and X is the vector of other unknowns

8

❑⇔ {𝐻𝑈+𝐴𝑋=𝑅

𝐶𝑈+𝐿𝑋=𝑆

Page 9: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

RP speeds-up linear algebra

• Concept of R-structure– Matrix is said to have the R-structure (R for Re-parameterization)

– Two matrices have the same R-structure iff the sizes of their four blocks H, A, C, and L are equal, and the block structures of their A parts are equal

– From what precedes, it follows that:

– By using re-parameterization, we could solve the last system for U and then deduce X (cf. paper for the detailed steps), without inverting A

• Conclusion: RP makes A highly reducible and the decomposition of A by methods of [9] allows to speed-up the solving process of a linear system Ax = B

9

Page 10: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

RP speeds-up linear algebra

• Results– The cost of solving a RP linear system is O(k(α+kn+k2)), which equals O(kn2) in

the worst case and is always less than O(n3) (Gauss pivoting, LU) as k << n

– It is even less than the cost O(n2.8...) of the inversion by the Strassen method and less than the cost O(n2.375...) of Coppersmith–Winograd method

• Remarks– We assumed than k << n and we didn’t discuss a max value of k in order for

the speed-up of RP to remain interesting

– The complexity study suggests a bound k = O(n0.375...) so that our RP has the same complexity as that of Coppersmith–Winograd -> it is probably easier to find sets of key unknowns of size O(n0.375...) rather than sets of size O(1) or O(log n)

10

Page 11: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

RP speeds-up p-adic methods

Used to solve diophantine problems [19–23], computer algebra, etc.

• RP benefits Hensel lifting used in P-adic methods– Assume X0 is a root of some RP alg. sys. F(X) = 0 modulo p and that we aim to

compute X such that X0 + p X1 is a root of F(X) = 0 modulo p2

– Then, F (X0) = 0 mod p implies that F (X0) (taken modulo p2) is a multiple of p

– After computing the vector λ = F (X0)/p, we find that F (X0 + p X1) = 0 which implies that λ + F (X′ 0)X1 = 0 [mod p] which is a linear sys. solvable modulo p

Conclusion: if F(x) = 0 is a RP system, then Hensel lifting can benefit from RP (as Hensel lifting is nothing else than the Newton–Raphson method but for the p-adics [26])

11

Page 12: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

RP speeds-up interval solvers

• Interval analysis– Let f(x) = 0 be a large well-constrained RP system for which we want to

compute all its real roots within a given box B. If we denote by x0 the center of box B, then:

– To compute Δx solution of , where :1. We compute f (B)′ by intervals analysis (sparsity of the Jacobian f )′2. We solve Δx := solve(f (B)′ Δ x = −f (x0)) by using the Jacobian R-structure

• A main difficulty– Any of the sub-matrices of the diagonal of the Jacobian f (B)′ is singular

12

Page 13: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

RP speeds-up interval solvers

• Solution: using the centered evaluation form– Let J = f (B)′ and J0 the center of J which is non-singular, if we denote by ΔJ = J

− J0, and substituting in what precedes, it follows after some computations:

– We have a non-singular linear sys. with unknown Δx with J0 having the R-structure of f -> solving this linear sys. can benefit from RP

• Conclusions– Interval solvers like ALIAS-C++ [27], RealPaver [28], IBEX [29], and QUIMPER

[30]) can benefit from RP– Further work needed to address the wrapping effect issue for RP systems

13

Page 14: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Experimental results

Benchmarks at the lowest linear Algebra level– Comparing to naïve Gaussian method for solving a Linear Sys. M=XB– Gaussian method implemented through Matlab command X=M\B

• 1st benchmark series: the hexahedron– No significant performance improvement -> sys. Too is small

• 2nd benchmark series: bigger random sys. Having R-structure– Different combinations of n (sys. size) and m (block size) parameters, 100

random sys./combination (k being the number of key unknowns)

14

Page 15: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Experimental results

• Results of 2nd benchmark series: bigger random sys.– RP is significantly faster (x26 average speed-up) over naïve Gaussian– The bigger is the sys., the more performant is our RP (x73.4 for n = 4002)

15

Page 16: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Experimental results

Benchmarks at the highest geometric level: the pentahedron [37, 38]– A GCSP of 6 vertices, 9 edges, and 5 faces -> 12 eqs. in 18 vars. -> an

irreducible sys. of 9 eqs. in 9 vars. (after placement rules)– Difficult to solve by interval solvers, Cayley-Manger determinants, and even

the re-parameterization technique of [9] with 1 key unknown

• Our RP of the pentahedron GCSP– By using the concurrency property at point I and law of cosines -> a new NLS

of 3 eqs. In 3 vars. Only!– A significant performance gain (x41) for our RP technique

16

Page 17: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

RP at a higher level

• Benefits of RP the lowest level of LA– Simplicity and factorization -> reuse of classical LA methods at a small cost of

minor implementation modification or just by recompiling code

• RP can also be used at a higher level of interval solvers by– Handling only boxes involving the parameters U– Boxes involving variables X are computed in function of U through interval

computations– Main difficulty: a big change in the solver implementation

17

Page 18: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Conclusion

• A new RP technique for reducing geometric irreducible systems

• No need for the values of the key unknowns and no limit on their number as

opposed to the existing methods [8]-[12].

• Enabling the usage of decomposition methods [9] on irreducible RP systems

• Usage at the lowest level and significant performance improvement

• Benefits for numerous solvers (Newton–Raphson, homotopy, p-adic methods,

etc.) with only small implementation modifications -> solving larger systems

18

Page 19: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

Emerging questions/future work

• Can interval Newton solvers benefit from both RP/preconditioning for limiting the wrapping effect? Is there any suitable contraction operator?

• Similar question for Bernstein-based solvers [40, 41]?

• Which redundant equations should be ignored (highest degree)?

• Is there a way to specify interesting values of k compared to n?

• How to compute a smaller set of key unknowns k? -> The generalization of the work done in [10, 12, 14] is an open problem!

• RP application for other LA computations (SVD and QR), for symbolic computations, and for (under-constrained) GCSPs?

• …19

Page 21: Salt Lake City, Oct. 14th, 2015 Re-parameterization reduces irreducible geometric constraint systems H. Barki, L. Fang, D. Michelucci, S. Foufou GDSPM’15

RP speeds-up linear algebra

• Steps for solving for U and X1. Zn×1 := solve(An×nZn×1 = Rn×1) costs O(α) and yields Z=A-1R, avoiding A-1

2. Kn×k := solve(An×nKn×k = Hn×k) costs O(kα) and yields K = A−1H, avoiding A-1

3. Compute C − LA−1H = Ck×k −Lk×nKn×k costs O(k2n)

4. Compute S − LA−1R = Sk×1 − Lk×nZn×1 costs O(kn)

5. Uk×1 := solve((C − LA−1H)k×k Uk×1 = (S − LA−1R)k×1), where only U is unknown, costs O(k3) and yields U

6. Compute Rn×1 − Hn×kUk×1 costs O(nk)

7. Xn×1 := solve(An×nXn×1 = (R − HU)n×1), where only X is unknown, costs O(α), and yields X, avoiding A−1 21