algebraic aspects of signatures fg6
play

Algebraic aspects of signatures FG6 Nikolas Tapia based on joint - PowerPoint PPT Presentation

Algebraic aspects of signatures FG6 Nikolas Tapia based on joint work with J. Diehl & K. Ebrahimi-Fard Weierstra-Institut fr angewandte Analysis und Stochastik July 23, 2019 Introduction Classification of time series is a very active


  1. Algebraic aspects of signatures FG6 Nikolas Tapia based on joint work with J. Diehl & K. Ebrahimi-Fard Weierstraß-Institut für angewandte Analysis und Stochastik July 23, 2019

  2. Introduction Classification of time series is a very active field of research. Most methods rely on extraction of features . Signatures a , b provide features that are interesting for a number of applications. Also useful for other tasks such as analysing control systems, pathwise solutions to Stochastic Differential Equations, among others. a K.-T. Chen. „Integration of paths, geometric invariants and a generalized Baker-Hausdorff formula“. In: Ann. of Math. (2) 65 (1957), pp. 163–178. b T. Lyons. „Differential equations driven by rough signals“. In: Revista Matemática Iberoamericana 14 (1998), pp. 215–310. 2/16 Algebraic aspects of signatures

  3. Continuous-time signatures Let X : [ 0 , 1 ] → � continuous path. Definition Given p ≥ 1 , define the p -variation of X over the interval [ s , t ] ⊆ [ 0 , 1 ] by � 1 / p � X � p ; [ s , t ] ≔ � | X v − X u | p � � � sup . � � π ∈P[ s , t ] [ u , v ]∈ π The space of all paths such that � X � p ; [ s , t ] < ∞ is denoted by V p ([ s , t ]) . Can be generalized to functions Ξ : [ 0 , 1 ] 2 → � by replacing the increment X v − X u by Ξ u , v . This generalization is an essential piece in T. Lyon’s theory of rough paths . a a T. Lyons. „Differential equations driven by rough signals“. In: Revista Matemática Iberoamericana 14 (1998), pp. 215–310. 3/16 Algebraic aspects of signatures

  4. Young integration Theorem (Young a ) Let X ∈ V p , Y ∈ V q with 1 p + 1 q > 1 . The integral ∫ t � n ( π ) Y u d X u ≔ Y t i ( X t i +1 − X t i ) lim | π |→ 0 s i =0 � � ∫ t � � is well defined and � � � � ≤ C p , q � X � p ; [ s , t ] � Y � q ; [ s , t ] . Y u d X u − Y s ( X t − X u ) s ∫ X d X is well defined as long as X ∈ V p for 1 ≤ p < 2 . In particular, the iterated integral ∫ More generally, if X = ( X 1 , . . . , X d ) takes values in � d then the integrals X i d X j are defined. a L. C. Young. „An inequality of the Hölder type, connected with Stieltjes integration“. In: Acta Mathematica 67.1 (1936), p. 251. 4/16 Algebraic aspects of signatures

  5. Signatures Definition The signature of the path X : [ 0 , 1 ] → � d is the collection of iterated integrals ∫ t ∫ t ∫ u ∫ ∫ S ( X ) s , t ≔ 1 + d X u + d X v ⊗ d X u + · · · + · · · d X u 1 ⊗ · · · ⊗ d X u n + · · · s s s s < u 1 < ··· < u n < t Theorem The signature satisfies 1. Chen’s identity: S ( X ) s , u ⊗ S ( X ) u , t = S ( X ) s , t . 2. Reparametrization invariance: S ( X ◦ ϕ ) s , t = S ( X ) s , t . 3. It is the unique solution to the fixed-point equation ∫ t S ( X ) s , t = 1 + S ( X ) s , u ⊗ d X u . s 5/16 Algebraic aspects of signatures

  6. Signatures � Additionally, the shuffle relations are satisfied: S ( X ) I s , t S ( X ) J S ( X ) K s , t = s , t . K ∈ Sh ( I , J ) This introduces some redundancy, e.g. S ( X ) ji s , t S ( X ) j s , t − S ( X ) ij s , t = S ( X ) i s , t . A way to compress the available information is to work with the so-called log-signaure Ω ( X ) s , t ≔ log ⊗ S ( X ) s , t ∈ L( � d ) . Ω ( X ) corresponds to a pre-Lie Magnus expansion w.r.t. the pre-Lie product ∫ t ∫ u [ d X v , d Y u ] X ⊲ Y ≔ s s 6/16 Algebraic aspects of signatures

  7. Signatures The map I �→ S ( X ) I s , t defines a linear map from the tensor algebra to the reals. The shuffle relations then mean that this map is a character , i.e. S ( X ) I s , t S ( X ) J s , t = S ( X ) I ✁ J s , t where the shuffle product is recursively defined, for I = ( i 1 , . . . , i n ) , J = ( j 1 , . . . , j m ) , by I ✁ J = ( I ′ ✁ J ) i n + ( I ✁ J ′ ) j m where I ′ = ( i 1 , . . . , i n − 1 ) , J ′ = ( j 1 , . . . , j m − 1 ) and ∫ ∫ d X i 1 u 1 · · · d X i n S ( X ) I · · · s , t = u n . s < u 1 < ··· < u n < t Remark We can think of S ( X ) as a formal series � S ( X ) I S ( X ) s , t = s , t I . I 7/16 Algebraic aspects of signatures

  8. Signatures Why should we care? 1. Useful for the description of the solutions of controlled systems : if � Y t = V ( Y t ) � � X t then V I ( Y s ) S ( X ) I Y t − Y s = s , t + R s , t . I 2. Captures features of X , useful for applications to Machine Learning, pattern recognition, time series analysis, etc. In principle hard to compute. However, if X is piecewise linear then S ( X ) s , t = exp ⊗ ( v 1 ) ⊗ · · · ⊗ exp ⊗ ( v k ) and we can use the Baker–Campbell–Hausdorff formula. a a J. Reizenstein and B. Graham. „The iisignature library: efficient calculation of iterated-integral signatures and log signatures“. In: (2018). arXiv: 1802 . 08252 [cs.DS] . 8/16 Algebraic aspects of signatures

  9. Signatures However: ∫ ∫ 1. For a one-dimensional signal: d X u 1 · · · d X u n = ( X t − X s ) n · · · . n ! s < u 1 < ··· < u n < t This can be cured to some extent by introducing more dimensions a and other tricks b . 2. In practice we are confronted with discrete data. This can also be avoided by interpolation . 3. A more severe problem is tree-like equivalence . c We propose instead a new framework operating directly at the discrete level. a T. Lyons and H. Oberhauser. „Sketching the order of events“. In: (2017). arXiv: 1708 . 09708 [stat.ML] . b F . J. Király and H. Oberhauser. „Kernels for sequentially ordered data“. In: Journal of Machine Learning Research 20.31 (2019), pp. 1–45. c B. Hambly and T. Lyons. „Uniqueness for the signature of a path of bounded variation and the reduced path group“. In: Ann. Math. 171.1 (Mar. 2010), pp. 109–167. 9/16 Algebraic aspects of signatures

  10. Discrete signatures A composition of an integer n is a sequence ( i 1 , . . . , i k ) with i 1 + · · · + i k = n . Definition (Gessel a ) Given a composition I = ( i 1 , . . . , i k ) define � z i 1 j 1 · · · z i k M I ( z ) ≔ j k . j 1 < j 2 <...< j k � � � For example z 2 M ( 1 ) ( z ) = M ( 2 ) ( z ) = z j , M ( 1 , 1 ) = z j 1 z j 2 , j . j j 1 < j 2 j Note that M ( 1 ) ( z ) 2 = M ( 2 ) ( z ) + 2 M ( 1 , 1 ) . a I. M. Gessel. „Multipartite P -partitions and inner products of skew Schur functions“. In: Combinatorics and algebra (Boulder, Colo., 1983) . Vol. 34. Contemp. Math. Amer. Math. Soc., Providence, RI, 1984, pp. 289–317. 10/16 Algebraic aspects of signatures

  11. Discrete signatures The map I �→ M I ( z ) defines a linear map from compositions to the reals. The product rule above can be expressed as M I ( z ) M J ( z ) = M I ⋆ J ( z ) where the quasi-shuffle product a ⋆ is recursively defined, for I = ( i 1 , . . . , i n ) , J = ( j 1 , . . . , j m ) , by I ⋆ J = ( I ′ ⋆ J ) i n + ( I ⋆ J ′ ) j m + c ( I , J ) where I ′ and J ′ are defined as before, and c ( I , J ) ≔ ( i 1 , . . . , i n − 1 , j 1 , . . . , j m − 1 , i n + j m ) . Definition Given a discrete time series x = ( x 0 , x 1 , . . . , x N ) , its discrete signature is � M I ( ∆ m DS ( x ) n , m = n x ) I I where ∆ m n x = ( x n +1 − x n , . . . , x m − x m − 1 ) . a M. E. Hoffman. „Quasi-shuffle products“. In: J. Algebraic Combin. 11.1 (2000), pp. 49–68. 11/16 Algebraic aspects of signatures

  12. Discrete signatures Why should we care? 1. Can be used to analyse solutions of controlled recurrence equations of the form y k +1 = y k + V ( y k )( x k +1 − x k ) relevant e.g. for applications to Residual Neural Networks. a , b , c 2. Invariant under time warping , useful for applications to time series classification. d 3. No need to transform the data in any way. Even if x is one-dimensional we get more information, e.g. � N DS ( x ) ( 2 ) ( x j − x j − 1 ) 2 � ( x N − x 0 ) 2 . 0 , N = j =1 4. No need for BCH formula. a Current project with P . Friz (TU) and C. Bayer (WIAS) b K. He et al. „Deep Residual Learning for Image Recognition“. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . 2016. c E. Haber and L. Ruthotto. „Stable architectures for deep neural networks“. In: Inverse Problems 34.1 (2018), pp. 014004, 22. d J. Diehl, K. Ebrahimi-Fard, and N. Tapia. „Time warping invariants of multidimensional time series“. In: (2019). arXiv: 1906 . 05823 [math.RA] . 12/16 Algebraic aspects of signatures

  13. Discrete signatures We can actually count the number of invariants. Theorem (Diehl, Ebrahimi-Fard, T.; Novelli, Thibon a ) The number of time-warping invariants of a d -dimensional time series has generating function � ∞ t 2 + d ( 13 d 2 + 9 d + 2 ) ( 1 − t ) d 2 ( 1 − t ) d − 1 = 1 + dt + d ( 3 d + 1 ) c n ( d ) t n = t 3 + · · · . G ( t ) ≔ 2 6 n =0 Compare with the corresponding generating function for the shuffle algebra: 1 1 − dt = 1 + dt + d 2 t 2 + d 3 t 3 + · · · . H ( t ) = a J.-C. Novelli and J.-Y. Thibon. „Free quasi-symmetric functions and descent algebras for wreath products, and noncommutative multi-symmetric functions“. In: Discrete Math. 310.24 (2010), pp. 3584–3606. 13/16 Algebraic aspects of signatures

Recommend


More recommend