On the Adjoint of the Eulerian Idempotent in an Analytic Context

We generalize Gehrig-Kawski theorem connecting the adjoint of the Eulerian idempotent with the logarithm of identity operator in the convolution product algebra . This has application in dynamical systems, control theory, coordinates of the first kind, generalized BCH-formula, Magnus expansion, etc., and is connected with iterated integrals and the signature of a path. We also show certain algebraic identities, which are meaningful in context of control and path-signature theory.


Introduction
The Eulerian idempotent can be explicitly defined as a linear endomorphism satisfying for all w ∈ A * . Its importance comes, however, from a more implicate (but natural) definition π 1 = log Id, where the logarithm is taken in the algebra with the convolution product * . The name "Eulerian idempotent" emanates from the Eulerian numbers, which are related with coefficients (in symmetric group algebras) of higher order Eulerian idempotents π n = 1 n! π * n 1 [1]. π 1 is interesting from many points of view. Purely mathematical aspects concern free Lie algebras, symmetric algebras, Solomon algebras, preLie algebras, etc. In this article, we focus our attention on algebraic applications in dynamical systems, control theory, coordinates of the first kind, generalized BCH-formula, Magnus expansion, etc., which are connected with iterated integrals, the signature of a path and so on. In detail, Sussmann [22] showed a product expansion for the solution of a non-linear control-affine system in terms of Lyndon basis of the free Lie algebra on words assigned to the controls. Then Melançon and Reutenauer [16] discovered the same expansion in purely algebraic context, and Reutenauer [20] generalized it to a Hall basis setting. This was rewritten in control-theoretic setting by Kawski and Sussmann [11]. The generalized algebraic version of these results is as follows: In this formula, it is crucial that exp is taken with respect to a product sh ⊗ conc (see Section 2.2 for the definitions), and H is a Hall set (see Section 2.4 for the definitions of H, P h , S h ). After a decade, Gehrig [8] and Kawski with Gehrig [7] proved that the adjoint homomorphism gives rise to another formula as follows: with the same data as previously. They work in a context of control theory. Namely, for a control system with the following: the above formula gives solution as follows: where φ t is a linear mapping defined for a 1 · · · a k ∈ A * k by and X h are appropriate vector fields. In particular, for fixed controls u i and a fixed t > 0, the solution x(t) is an image of x(0) under the t-time flow of a certain vector field. For varying t, the vector field is also t-varying. This is also connected with the theory of rough paths [12][13][14]. Namely, for a basis (e i ) of , take an alphabet A = {e i }; define a linear isomorphism by ι ⊗ (e i 1 · · · e i k ) = e i 1 ⊗ · · · ⊗ e i k . The signature of a path is as follows: and Gehrig-Kawski formula gives the logarithm of the signature as follows: where is a homomorphism defined by φ γ (P ) = (X(γ )|ι ⊗ (P )). In this article, we firstly generalize Gehrig-Kawski's result to a non-Hall basis (which means to all basis) of free Lie algebra Lie(A) (this is a more general answer to the Problem 1 stated in [10]). We prove that for a basis { P h | h ∈ B } in Lie(A) and a projection , there exist the canonical subspace and its basis { S h | h ∈ B } for which is as follows: We state this result in Theorem 1. After this, we give examples for this theorem choosing different basis in Lie(A) and a projection ρ. In particular, we obtain the Gehrig-Kawski formula for a Hall basis.
In the second part of the article, we focus on the aforementioned mappings φ t • π 1 and φ γ • π 1 as particular cases of a mapping φ • π 1 , with an algebra homomorphism (R is a ). We show in Theorem 3 that for w ∈ A * m , it follows: where φ n (w) = w 1 ···w k =w φ(w 1 ) · · · φ(w k ). From this theorem, we deduce that for a path γ , the homogeneous components of the logarithmic signature of γ can be expressed through the following: defined for w ∈ A * m . This formula explicitly connects the m-homogeneous component of log X(γ ) with the signatures X(γ n ) (n = 1, . . . , m) of the nth powers concatenation of γ .

Preliminary
In the article, we assume to be the set of natural numbers beginning from 1, to be a field of characteristic 0. For a unitary associative with unit 1, we use standard notations as follows: for any Q ∈ Alg, for which it makes sense. Here, Q •k is defined recursively by Q •0 = 1, and Q •k = Q • Q •k−1 for .

Tensor Algebra
Let V be a finite dimensional linear space over . For , denote by V ⊗k the kth tensor product of . Let T ⊗ V = ∞ k=0 V ⊗k and T⊗V be its algebraic closure. Consider ⊗ : T⊗V × T⊗V → T⊗V as a product, i.e., V ⊗k × V ⊗l (x, y) → x ⊗ y ∈ V ⊗k+l . Then (T ⊗ V , ⊗) and (T⊗V , ⊗) are associative K-algebras with as the unit.

Shuffle Algebra
Let A be a certain finite set of cardinality ≥ 2, called the alphabet. Denote by , the set of words of length n, the set of all words, the K-algebras of non-commutative polynomials, and series in the letters A, respectively; by 1 ∈ A * 0 , we denote the empty word; by , the subset of non-trivial (non-empty) words, the submodule spanned on non-trivial words, and its algebraic closure, respectively. The module is dual to , and we identify with the functional by writing P = w∈A * (P |w)w. The product of two series is therefore defined by (PQ|w) := uv=w (P |u)(Q|v). Since is a finite-free generated algebra, the module is dual to , and we also identify with by Q = w∈A * (Q|w)w. Clearly, (P |Q) = (Q|P ), and it also plays a role of a scalar product in . For a , we denote by the module of series representing functionals vanishing on Y . If are -submodules then for an endomorphism ρ : X → Y , we denote ker ⊥ ρ := (ker ρ) ⊥ .
Since , and is the algebraic closure of , we define all objects in the larger algebra. The concatenation product in has its tensorial version which we denote by , i.e., conc(P ⊗ Q) = PQ. Introduce another-shuffle-product defined recursively for words by for any w ∈ A * , and for all a 1 , a 2 ∈ A and w 1 , w 2 ∈ A * . In what follows, we will use another standard notation for the shuffle product , i.e., , and its generalized version given by the following: for w 1 , . . . , w k ∈ A * . We denote by (resp. ) the commutative of non-commutative polynomials (resp. series) in the letters A with as the product. The adjoint coproduct of the shuffle product and its generalized version are defined by the following: Similarly, we define the adjoint coproduct of the concatenation and its generalized version by the following; where the sum is taken over all w i ∈ A * . Note that and are mutually adjoint bialgebras [20, Prop. 1.9].

Let
be the standard Lie bracket, i.e., a bilinear mapping given by [P , Q] := PQ − QP. We denote by Lie(A) the smallest of , which contains A, and is closed under the Lie bracket, i.e., the free Lie algebra generated by A, and by Lie((A)) its algebraic closure. By Lie * (A), we denote the module of functionals . Denote by (S X) the action of a on an element X ∈ Lie(A). Let B be an totally ordered set, and { P h | h ∈ B } be a basis in Lie(A). Let { S h | h ∈ B } be the dual basis of Lie * (A), i.e., the one defined by the following: for any X ∈ Lie(A).

Hall Sets & Basis
Let M(A) be the set of binary, complete, planar, rooted trees with leaves labeled by A. Each such tree can be naturally identified with the unique expression in the set E(A) defined by the following two conditions: In the sequel, we will not distinguish between these sets, i.e., we assume M(A) = E(A). Define the mapping F(·) : M(A) → A * , which assigns to a tree t ∈ M(A), the word given by dropping all brackets in it, i.e., F(a) = a for all a ∈ A, and F(( Define also the mapping P · : M(A) → Lie(A), which changes the rounded brackets into the Lie brackets, i.e., P a = a for all a ∈ A, and P (t,t ) := [P t , P t ] for all t, t ∈ M(A). We will generalize this definition in the sequel. A Hall set H on the letters A [9] 1 , is a subset of M(A) totally ordered by ≤ and satisfying: Fix a Hall set H on the letters A totally ordered by ≤. Each Hall tree h ∈ H corresponds to a word F(h) ∈ A * called a Hall word. Denote by W, the set of Hall words with ordering ≤ inherited from the ordering on H in the natural way. Each word w ∈ A * , is the unique concatenation of a unique non-increasing series of Hall words, that is, w = h 1 · · · h k for some unique , and h i ∈ W such that h 1 ≥ · · · ≥ h k . Let be the mapping defined by the following is the Hall basis of Lie(A) corresponding to the Hall set H. By the Poincaré-Birkhoff-Witt theorem, the set of ordered products P h 1 · · · P h k , where h 1 ≥ · · · ≥ h k are Hall words, creates a basis for the enveloping algebra of Lie(A), which in the free case is isomorphic to . Therefore, P w | w ∈ A * is a basis in . Consider the dual basis S w | w ∈ A * in .

Hall words and , then
One of the consequences of this proposition is that is the algebraic closure of the free commutative algebra over , where h 1 ≥ · · · ≥ h k are Hall words, creates a basis for .

Algebra of Endomorphisms
Consider -the of linear endomorphisms from to . For define their convolution product as follows: Let be a projection (Q) = (Q|1)1. Then, is an associative with unit . Introduce the complete tensor product as follows: and a product * = sh ⊗ conc : A ⊗ A → A, i.e., for Then, (A, * ) is an associative with unit 1 = 1 ⊗ 1. The canonical isomorphism of modules given by the following: is a homomorphism of algebras and (A, * ), i.e., Note that in the definition of Im, we choose the most natural basis { u | u ∈ A * } in and its dual basis { u | u ∈ A * } in , but in general one can take a different basis and its dual.

Eulerian Idempotent and its Adjoint
Let be a projection given by I (Q) = Q − (Q). We are particularly interested in an endomorphism called the Eulerian idempotent. A straightforward calculation shows that Im(π 1 ) equals It is known that π 1 is a projection, , and for all . The second statement means . From Proposition 1, we conclude that is a basis of this space. The kernel of π 1 have another characterization: , where [20,Thm. 3.7]. The main results in this article concern -the adjoint endomorphism to π 1 , i.e., the one defined by the following; (π 1 (P )|Q) = (P |π 1 (Q)) for all . A straightforward calculation using adjointness of conc k and δ k , adjointness of sh k and δ k , and self-adjointness of I , brings to the following formula Since π 1 is a projection, the same is true for π 1 . In particular, and , which we denote by Lie ⊥ (A). Since π 1 is onto Lie(A), we also consider its adjoint as an epimorphism. Namely, for , we introduce its adjoint endomorphism . It is a monomorphism onto ker ⊥ π 1 satisfying the following: for all .

bases in Lie(A) and its dual in
Proof We prove that This ends the proof.

Theorem 1 Let { P h | h ∈ B } a basis in Lie(A), a projection on Lie(A), and { S h | h ∈ B } the basis in Lie * ρ (A) given by (4). It follows that
A Im(Id) = exp h∈B π 1 (S h ) ⊗ P h .
A few remarks are in order. First of all, we recall that in the algebra A the product is sh ⊗ conc, so the shuffle product is used to compute the left side of the tensor product. It means that good properties of π 1 (S h ) with respect to sh are welcome. Secondly, the proved formula is similar to the quite clear formula (6.2.1) in Reutenauer's book [20] (in which the sum is taken over all words). In our case, however, the sum is taken over the basis of the Lie algebra, which contains essential information about the logarithm of a series [4,18]. Therefore, this theorem generalizes the Gehrig-Kawski theorem (see Theorem 2 beneath) to its most extent. The advantage of the theorem is that it can be used to compute BCH-formula, Magnus expansion, logarithm of the signature, coordinates of the second kind, etc., both in general case, as well as for particular situations. In each case, one can try to choose a basis in Lie(A) and ρ to utilize specific features of a given problem.
There are several natural choices for ρ. The first one is ρ = π 1  The second one is to take ρ = ρ ⊥ as the orthogonal projection with respect to the scalar product (·|·) in (for more details about this projection see [5]). In this case, Lie * ρ ⊥ (A) = ker ⊥ ρ ⊥ = Lie((A)). In particular, ρ ⊥ = π 1 . Take a basis { P h | h ∈ B } in Lie(A) on a well-ordered set B. Using Gram-Schmidt process obtain the orthonormal basis P h | h ∈ B . Then, this is also the dual basis in Lie(A). Therefore, from the above theorem, we get Im(Id) = exp h∈B π 1 (P h ) ⊗P h . The same is clearly true for any orthonormal basis in Lie(A).
The third one is to take ρ = ρ PBW as the projection with kernel derived from Poincare-Birkhoff-Witt theorem. More precisely, universal enveloping algebra of the free Lie algebra } is a basis in Lie(A) then, using Poincaré-Birkhoff-Witt theorem, the set of ordered products P h 1 · · · P h k , where h 1 ≥ · · · ≥ h k ∈ B, is a basis for the enveloping algebra of Lie(A), i.e., in . Denote by S h 1 ,...,h k , h 1 ≥ · · · ≥ h k ∈ B, the elements of the dual basis in . We define ρ PBW as a projection on Lie(A) with the kernel . From PBW theorem, the elements written in this formula are linearly independent, hence creates a basis of ker ρ PBW . On the dual side, ker ⊥ ρ PBW = span { S h | h ∈ B }, but in general case, there is not known explicit formulas for S h . It follows that ker ρ PBW = ker π 1 , because for h > h ∈ B, we have ker π 1 (P h +P h ) 2 Therefore, ρ PBW = π 1 . It also follows that ker ⊥ ρ PBW = ker ⊥ ρ ⊥ , because for h > h ∈ B, we have (P h P h |[P h , P h ]) = (P h P h |P h P h ) − (P h P h |P h P h ) > 0, since P h P h , P h P h are linearly independent and have the same norm. Therefore, ρ PBW = ρ ⊥ . Note that each choice of the basis in Lie(A) gives a different projection ρ PBW , so in this case, we actually define a class of examples.
The fourth one, a subclass of the previous one, is to take ρ = ρ Hall , the projection derived as above taking a Hall basis in Lie(A). Recall that we described these basis in Section 2.4. In this case, Theorem 1 is equivalent to the following Gehrig-Kawski theorem.

Theorem 2 ([8] Theorem 28) Let W be a set of Hall words on the letters A. It follows that
In Tables 1 and 2, we gather known informations about ker ρ, ker ⊥ ρ = Lie * ρ (A) and basis in Lie(A) and its dual in Lie * ρ (A) associated with projections ρ just described. Table 1 Kernel and dual space associated with a projection ρ on Lie(A) Let (R, μ) be a with a multiplication μ : R⊗R → R. If there is no confusion, we will use standard notation ab = μ(a ⊗ b). We also recursively introduce μ k : R ⊗k → R by μ 1 (a) = a, μ 2 = μ, μ k+1 = μ•(μ k ⊗μ 1 ). Let (R , ·) be an K-algebra with a multiplication · : R × R → R (we use simpler definition of this product, because we will not use its tensorial properties).
Let and be algebra homomorphisms. Then, it is easy to see that From Theorem 1, we conclude that Our aim is to give expression for φ • π 1 . For , let us define a linear mapping , for v ∈ A * given by the following: where the sum is taken over all v i ∈ A * . We emphasize the existence of the empty word 1 ∈ A * . In particular φ 1 (v) = φ(v). Recall , and  ρ ⊥ Any orthonormal one The same one where sh k and δ k are defined by (2) and (3), respectively. Therefore, For , let us define linear mappings , for v ∈ A * given by the following:φ where the sum is taken over all u i ∈ A * + -the set of non-trivial words. Since I annihilates empty word, it is easy to see that We conclude that In the following lemma, we deriveφ k 's in terms of φ k 's.

Lemma 4.1 It follows thatφ
Proof Since in the definition of φ k 's, we sum over all words and in the definition ofφ k 's, we sum over all non-empty words, we see that Now, we use induction on k to prove the hypothesis. For k = 1, this is clear. Then, using (6) and then induction hypothesis, we have the following: Since , and changing the order of summation, we have the following:φ The expression in the square brackets equals −1, and we are done.
Using this lemma, we see that If v m ∈ A * m , then clearlyφ n (v m ) = 0 for all n > m (since we can not divide m-letter word on n non-trivial words). It means that After changing the order of summation, we get A simple induction on m shows that The above reasoning brings us to the following theorem.

Signature of a Path
Let γ be a continuous path with finite variation in a finite dimensional linear space V over equipped with the metric d V . More precisely, take T > 0 and let t = (t 1 , . . . , t r ), such that 0 = t 0 < t 1 < · · · < t r ≤ T and . Denote the set off all such tuples by P. For a continuous mapping γ · : [0, T ] → V its length is defined by the following: If |γ | < +∞, then it is of finite variation. In particular, this implies that γ is differentiable almost everywhere and Denote X k (γ ) = X T k (γ ) for k > 0, and X 0 (γ ) = 1 the neutral element in T ⊗ V . The signature of the path [13] γ is in the tensor algebra T⊗V .
If for all h ∈ B, S h is homogeneous of order #h, i.e., S h ∈ span w ∈ A * #h (this is the case for Hall bases), then This means that having got the signature of a path, the logarithm of this signature can be computed in terms of the signatures of concatenations of the path. For more information about log signature (see [6,15,19]).
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.