## 1 Introduction

We consider a system of delay differential equations (DDEs) with constant delays and the initial condition of the following form:

\begin{aligned} {\left\{ \begin{array}{ll} x'(t) = f\left( x(t), x(t - \tau _1), x(t - \tau _2), \ldots , x(t - \tau _m)\right) , &{} t \ge 0\\ x(t) = \psi (t), &{} t \in [-\tau , 0], \end{array}\right. } \end{aligned}
(1)

where $$m \in \mathbb {N}$$ and $$\tau = \tau _1> \tau _2> \ldots > \tau _m \ge 0$$ are the delays, $$x'$$ is understood as a right derivative and $$\psi : [-\tau , 0] \rightarrow \mathbb {R}^d$$ is of class $$C^0$$ on $$[-\tau , 0]$$, $$x(t) \in \mathbb {R}^d$$, $$f: \mathbb {R}^{(m+1)d} \rightarrow \mathbb {R}^d$$.

In [34], we have presented a method of producing rigorous estimates on the function x(t) for $$t \ge 0$$ for the simplest scalar ($$d = 1$$) DDE with a single delay:

\begin{aligned} x'(t) = f(x(t), x(t-\tau )) \end{aligned}
(2)

The algorithm presented in [34] is an explicit Taylor method with piecewise Taylor representation of the solution over a fixed step size grid and with Lohner-type control of the wrapping effect encountered in interval arithmetic [22]. The method consists of two algorithms: one computing Taylor coefficients of the solutions at equally spaced grid points (together with the rigorous estimate on the error size), and the second one to compute enclosures of the solution segment after an arbitrary step of size $$\varepsilon$$, smaller than the grid step size h. The method is suited to construct Poincaré maps in the phase space of the DDEs and it was successfully applied to prove several (apparently stable) periodic solutions to scalar DDEs [32, 34] (among them to Mackey–Glass equation). However, the second method—$$\varepsilon$$ step part—is not optimal in the sense of the local error order. Essentially, the local error of some of the coefficients in the Taylor representation of solution is O(h). The reason is that some of the coefficients are computed using just an explicit Euler method with very rough estimates on the derivative. With this apparent loss of accuracy, the images of Poincaré maps in [34] are computed with less than optimal quality and are not well suited to handle more diverse spectrum of dynamical results.

In this work, we provide an effective way to increase the local order of the full-step algorithm after each full delay of the integration procedure to significantly reduce the error size later on, when applying the second $$\varepsilon$$ step procedure. Under some additional but reasonable assumptions about the integration time being long enough (see Definition 3 and Sect. 3.4), the modification allows to decrease the local error size of the $$\varepsilon$$ step method for all coefficients to $$O(h^{n+1})$$ where n is the order of representation of the initial function and h the step size of interpolation grid (compared to O(h) for the previous version from [34]).

All those enhancements are done without a significant increase in computational complexity of the most time consuming part of the algorithm: a Lohner-type method of controlling the wrapping effect. What is more, we present an elegant and more general Lohner-type method of wrapping effect control to handle both systems of equations and many delays such as Eq. (1). We also employ more elaborate Lohner sets to further reduce undesirable effects of interval arithmetic. With all those improvements, the method produces estimates on solutions of several orders of magnitude better than the previous one.

As a presentation of effectiveness of the new method, we give proofs of the existence of periodic solutions to Mackey–Glass equation for a wider spectrum of parameters than in [34]. The proofs are done for parameters in the chaotic regime and the orbits are apparently unstable. To this end, we need to expand on the theory, so we extend the concept of covering relations [5] to infinite Banach spaces, and we use Fixed Point Index in absolute neighbourhood retracts (ANRs) [8] to prove Theorem 25 about the existence of orbits for compact mappings in infinite-dimensional spaces following chains of covering relations. We use this technique to show existence of symbolic dynamics in a perturbed model $$x'(t) = f(x(t)) + \varepsilon \cdot g(x(t-\tau ))$$, where f is a chaotic ODE in three dimensions (Rössler system) and for a couple of g’s which are some explicitly bounded functions. We hope similar techniques will allow to prove chaos in Mackey–Glass equation [24].

The paper is organized as follows: in Sect. 2 we present some basic theory for DDEs with constant delays, and we recall shortly the basic structure of (pn)-functions sets to represent objects in the phase space of (1). We also generalize this structure and we discuss its properties. In Sect. 3, we recall algorithm from [34] within a new, more general notation and we introduce several modifications that will be crucial for the complexity and accuracy of the algorithm. This new algorithm will form a base to some improvements in the construction of Poincaré maps in the phase space, especially to enhance the quality of the estimates. We present some benchmarks to show how the new estimates are in comparison with the old algorithm. In Sect. 4, we present topological tools to prove existence of a special kind of solutions to DDEs (1). We go beyond the Schauder fixed-point theorem used in [34]: we use fixed-point index on ANRs [8] and we adapt the notion of covering relations [5] to the setting of (pn)-function sets describing the infinite-dimensional phase space of the DDEs. The compactness of the solution operator in the phase space for times bigger than delay allows to apply the Schauder fixed-point index in our case. We establish theorems to prove existence of symbolic dynamics conjugated to the sequences of covering relations on (pn)-functions sets. In Sect. 5, we apply presented methods to prove existence of (apparently unstable) periodic solutions to the Mackey–Glass equation, for the original values of parameters for which Mackey and Glass observed numerically chaotic attractor [24]. We also prove existence of symbolic dynamics in a delay-perturbed chaotic ODE (Rössler system).

### 1.1 Our Results in the Perspective of Current Research in the Field

There are many important works that establish the existence and the shape of the (global) attractor under some assumptions on (2), for example if it is of the form $$x'=-\mu x(t) + f (x(t - 1))$$ and under the assumption that f is strictly monotonic, either positive or negative, or if f has a simple explicit formula, usually piecewise constant or affine. We would like here to point out some results, but the list is for sure not exhaustive (we refer to the mentioned works and references therein). Mallet-Paret and Sell used discrete Lyapunov functionals to prove a Poincaré–Bendixson type of theorem for special kind of monotone systems [25]. Krisztin, Walther and Wu have conducted an intensive study on systems having a monotone positive feedback, including studies on the conditions needed to obtain the shape of a global attractor, see [17] and references therein. Krisztin and Vas proved that in the case of a monotonic positive feedback f, under some assumptions on the stationary solutions, there exists large amplitude slowly oscillatory periodic solutions (LSOPs) which revolve around more than one stationary solution [16]. Vas continued this work and showed a method to construct f such that the structure of the global attractor may be arbitrarily complicated (containing an arbitrary number of unstable LSOPs) [37]. On the other hand, Lani–Wayda and Walther were able to construct systems of the form $$x'=f(x(t - 1))$$ for which they proved the existence of a dynamic which is conjugate to a symbol shift (Smale’s horseshoe) [19]. Srzednicki and Lani-Wayda proved the existence of multiple periodic orbits and the existence of chaos for some periodic, tooth-shaped (piecewise linear) f by the use of the generalized Lefshetz fixed point theorem [18]. A nice review of works that deal with the question of existence of chaos in Mackey–Glass and similar systems are compiled in Walther review [38]. Recently, a new approach have been used to prove the existence of some periodic orbits to the Mackey–Glass equation in a limiting case when $$n \rightarrow \infty$$ [15].

While impressive, all mentioned analytical/theoretical results are usually hard to apply in the context of general functions f, so we might look for other means of obtaining rigorous results in such cases, for example, by employing computers for this task. In recent years, there were many computer-assisted proofs of various dynamical properties for maps, ODEs and (dissipative) partial differential equations ((d)PDEs) by application of the theory of dynamical systems with estimates obtained from rigorous numerical methods and interval arithmetic, see for example [12] and references therein. A big achievement of the rigorous computations are proofs of the existence of chaos and strange attractors, for example the paper by Tucker [35], and recently to prove chaos in Kuramoto–Shivasinski PDE [39]. The application of rigorous numerical methods to DDEs started to appear a few years ago and are steadily getting more attention. Probably the first method used to prove existence of periodic orbits by the expansion in Fourier modes was given in [41], and then in a more general framework and by a different theoretical approach in [14, 20]. Other methods, strongly using the form of r.h.s. f in (2), were used in [16] to prove the structure of the global attractor; then in [36] to close a gap in the proof of the Wright conjecture; and finally recently in [1] to show the existence of many stable periodic orbits for a DDE equation that is the limiting case of Mackey–Glass equation when $$n \rightarrow \infty$$. To the author’s knowledge, the results from our work [34] are the first application of rigorous integration (forward in time) of DDEs in the full phase space for a general class of problems to prove the existence of some dynamics, namely the existence of apparently stable periodic orbits in Mackey–Glass equation. A different approach to one presented in our work [34] was recently published which uses Chebyshev polynomials to describe solutions in the phase space and a rigorous fixed-point finding argument to produce estimates on the solutions to DDEs forward in time, together with estimates on the Frechét derivative of the time-shift operator $$\varphi (\tau , \cdot )$$ [21]; however, the presented approach has one disadvantage: it can find solutions only on full-delay intervals and therefore cannot be used directly to construct Poincaré maps. Recently, the extension of those methods was used to prove persistence of periodic solutions under small perturbations of ODEs [7], and a similar approach was used in a rigorous method of numerically solving initial value problems to state-dependent DDEs [3]. This last work uses similar technique as our work to subdivide the basic interval into smaller pieces and piecewise polynomial interpolation of the functions in the phase space, but instead of Taylor it uses Chebyshev polynomials and a fixed-point finding argument to prove existence of a true solution nearby. On the other hand, the parametrization method was used to prove the persistence of periodic orbits in delay-perturbed differential equations, including the state-dependent delays [40]; however, it assumes that $$\tau$$ is relatively small. Our method has an advantage over those methods, as it allows for a larger amplitude of the perturbation and to prove theorems beyond the existence of periodic orbits, as we are showing persistence of symbolic dynamics in a perturbed ODE. Finally, there are also some methods to obtain rigorous bounds on the solutions, e.g. [28]; however, as authors say, they do not produce estimates of quality good enough to prove theorems.

### 1.2 Notation

For reader’s convenience, we include here all the basic notions used in this paper. We will also remind them the first time they are used in the text, if necessary.

We will denote by $$C^k([-\tau , 0], \mathbb {R}^d)$$ the set of functions which are $$C^{k}$$ on $$(-\tau , 0)$$ and right and left derivatives up to k exist at $$t = -\tau$$ and $$t = 0$$, respectively. For short, we will usually write $$C^k$$ to denote $$C^k([-\tau , 0], \mathbb {R}^d)$$ when d and $$\tau$$ is known from the context.

We use standard convention in DDEs to denote the segment of $$x: (a-\tau , b) \rightarrow \mathbb {R}^d$$ at $$t \in (a, b)$$ by $$x_t$$, where $$x_t(s) = x(t+s)$$ for all $$s \in [-\tau , 0]$$. Then, we will denote by $$\varphi$$ the semiflow generated by DDE (1) on the space $$C^0$$, $$\varphi (t, x_0):= x_t$$ for a solution x of Eq. (1) with initial data $$x_0$$.

The algorithms presented in this paper produce estimates on various quantities; particularly, we often work with sets of values that only enclose some quantity. Therefore, for convenience, by $$\mathbb {I}$$ we will denote the set of all closed intervals $$[a, b]: a \le b, a, b \in \mathbb {R}$$ and we will denote sets by capital letters like XYZ,  etc., and values by lower case letters xyz, etc. Usually, the value $$x \in X$$ for easier reading, but it will be always stated explicitly in the text for clarity.

Sometimes, instead of using subscripts $$x_i$$, we will write projections to coordinates as $$\pi _i x$$ or $$\pi _{\mathcal {X}} x$$ (projection on some subspace $$\mathcal {X}$$ of some bigger space. This will be applied to increase readability of formulas.

Let $$Z \subset \mathbb {R}^M$$. By $$\texttt{hull}(Z)$$ we denote the interval hull of Z, that is, the smallest set $$[Z] \in \mathbb {I}^M$$ such that $$Z \subset [Z]$$. By $${\overline{Z}}$$ we denote the closure of set Z, by $${{\,\textrm{int}\,}}Z$$ we denote the interior of Z and by $$\delta Z$$ we denote boundary of Z. If Y is some normed vector space and $$Z \subset Y$$, then we will write $$\delta _Y Z, {{\,\textrm{int}\,}}_Y Z, {{\,\textrm{cl}\,}}_Y {Z}$$ to denote boundary, interior and closure of Z in space Y. By $${{\,\mathrm{\texttt{Dom}}\,}}f$$ we denote the domain of f.

For multi-index vectors $$\eta , \zeta \in \mathbb {N}^p$$ we will write $$\eta \ge \zeta$$ iff $$\eta _i \ge \zeta _i$$ for all $$i \in \{1,\dots ,p\}$$.

By $$\mathcal {M}(k, l)$$, we denote the set of matrices of dimensions $$k \times l$$ (rows $$\times$$ columns), while by $$Id_{d \times d}$$ the identity matrix and by $$0_{d \times d}$$ the zero matrix in $$\mathcal {M}(d, d)$$. When d is known from the context we will drop the subscript in Id.

By $${{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert }_{D}(p, r)$$ we denote the (open) ball in $$\mathbb {R}^D$$ in the given norm $$\Vert \cdot \Vert$$ at a point $$p\in R^D$$ with radius r. In the case when the norm is known from the context, we simply use $${{\,\mathrm{\textbf{B}}\,}}_{D}(p, r)$$, and eventually $${{\,\mathrm{\textbf{B}}\,}}_{D}(r)$$ for 0-centered balls.

## 2 Finite-Dimensional Description of the Phase Space

In the beginning, we will work with Eq. 2 (single delay) for simplicity of presentation, but all the facts can be applied to a more general Eq. (1).

As we are interested in computer-assisted proofs of dynamical phenomena for (2), we assume that f is a simple/elementary function, so that it and its derivatives can be given/obtained automatically as computer programs (subroutines). Many equations encountered in theory and applications are of this form, two well-known examples that fit into this category are Wright and Mackey–Glass equations. We will also assume that f is sufficiently smooth, usually $$C^\infty$$ in both variables. Under this assumptions, the solution x(t) of (2) with $$x_0 = \psi \in C^0$$ exists forward in time (for some maximal time $$T_{max}(\psi ) \in [0, +\infty ]$$) and is unique, see e.g. [4].

The crucial property of DDEs with f smooth (for simplicity we assume $$f \in C^\infty$$) is the smoothing of solutions [4]. If the solution exists for a long enough time, then it is of class at least $$C^{k}$$ on the interval $$(-\tau + \tau \cdot k, \tau \cdot k)$$ and it is of class at least $$C^{k}$$ at $$t = \tau \cdot k$$. If $$\psi$$ is of class $$C^{m}$$ then x is of class $$C^{m+k}$$ on any interval $$(-\tau + \tau \cdot k, \tau \cdot k)$$. Moreover, the solutions on the global attractor of (2) must be of class $$C^\infty$$ (for $$f \in C^\infty$$). From the topological methods point of view, the smoothing of solutions implies the semiflow $$\varphi (t, \cdot ): C^0 \rightarrow C^0$$ is a compact operator for $$t \ge \tau$$, essentially by the Arzela–Ascoli theorem, see e.g. [34] (in general, $$\varphi (t, \cdot ): C^k \rightarrow C^k$$ is well defined and compact in $$C^k$$ if $$t \ge (k+1) \cdot \tau$$).

On the other hand, the solution can still be of a lower class, in some cases—even only of class $$C^0$$ (at $$t = 0$$). It happens due to the very nature of the DDE (2), as the right derivative at $$t = 0$$ is given by (2) whereas the left derivative of the initial data $$\psi$$ at 0 can be arbitrary. This discontinuity propagates in time so the solution x, in general, is only of class $$C^k$$ at $$t = k \cdot \tau$$. In other words, a solution to DDE with an initial segment of higher regularity can sometimes “visit” the lower regularity subset of the phase space. This behaviour introduces some difficulties in the treatment of the solutions of DDEs and the phase space, especially when one is interested in finding $$\varphi (t, x)$$ for $$t \ne m \cdot \tau$$, $$m \in \mathbb {N}$$.

In the rest of this section, we will recall the notion of (pn)-functions sets from [34] used in our method to represent functions in the phase space of DDE (2). However, we use a slightly different notation and we introduce some generalizations that will be suitable for the new integration algorithm in Sect. 3.

### 2.1 Basic Definitions

The algorithm we are going to discuss in Sect. 3 is a modified version of the (explicit) Taylor rigorous method for ODEs, that is, we will be able to produce the Taylor coefficients of the solution at given times using only the well-known recurrent relation resulting from the successively differentiating formula (2) w.r.t. t. For this recurrent formula [presented later in the text, in Eq. (11)], it is convenient to use the language of jets.

Let $$m \in \mathbb {N}$$ and let $$g: \mathbb {R}^m \rightarrow \mathbb {R}$$ be of class $$C^{n}$$ and $$z \in \mathbb {R}^d$$. We denote by $$\alpha$$ the m-dimensional multi-index $$\alpha = \left( \alpha _1, \dots , \alpha _m \right) \in \mathbb {N}^m$$ and we denote $$z^{\alpha } = \Pi _{i=1}^{m} z_i^{\alpha _i}$$, $$|\alpha | = \sum _{i=1}^{m} \alpha _i$$, $$\alpha ! = \Pi _{i=1}^{n} \alpha _{i}!$$, and

\begin{aligned} g^{(\alpha )} = \frac{\partial ^{|\alpha |}g}{{{\partial z_1^{\alpha _1}}\dots {\partial z_m^{\alpha _m}}}}. \end{aligned}

By $$J^{[n]}_{z}{g}$$ we denote the d-dimensional jet of order n of g at z, i.e.:

\begin{aligned} \left( J^{[n]}_z{g}\right) (y) = \sum _{|\alpha | \le n} \frac{g^{(\alpha )}(z)}{\alpha !} \cdot (y - z)^\alpha . \end{aligned}
(3)

We will identify $$J^{[n]}_{z}{g}$$ with the collection of the Taylor coefficients $$J^{[n]}_{z}{g} = \left( g^{[\alpha ]}(z)\right) _{|\alpha | \le n}$$, where

\begin{aligned} g^{[\alpha ]}(z):= \frac{g^{(\alpha )}(z)}{\alpha !}. \end{aligned}

We will use $$J^{[n]}_{z}{g}$$ either as a function defined by (3) or a collection of numbers depending on the context. For a function $$g: \mathbb {R}^m \rightarrow \mathbb {R}^d$$ the jet $$J^{[n]}_z(g) = \left( J^{[n]}_z{g_1}, \ldots , J^{[n]}_z{g_d}\right)$$ is a collection of jets of components of g.

In the sequel we will use extensively the following properties of jets:

### Proposition 1

The following are true:

1. 1.

if $$g: \mathbb {R}\rightarrow \mathbb {R}$$ then $$J^{[k]}_z\left( {J^{[n]}_z{g}}\right) = J^{[k]}_z{g}$$ for $$k \le n$$;

2. 2.

if $$f = g \circ h: \mathbb {R}\rightarrow \mathbb {R}$$ for $$g: \mathbb {R}^d \rightarrow \mathbb {R}$$ and $$h: \mathbb {R}\rightarrow \mathbb {R}^d$$, then

\begin{aligned} J^{[n]}_{t_0} f = J^{[n]}_{t_0} \left( \left( J^{[n]}_{h(t_0)} g\right) \circ \left( J^{[n]}_{t_0}h_1, \dots , J^{[n]}_{t_0} h_d\right) \right) . \end{aligned}
(4)

In other words, Eq. (4) tells us that, in order to compute nth-order jet of the composition, we only need to compose jets (polynomials) of two functions and ignore terms of order higher than n. For a shorter formulas, we will denote by $$\circ _J$$ the composition of jets in (4), i.e. if $$a = J^{[n]}_{h(t_0)} g$$ and $$b_i = J^{[n]}_{t_0} h_i$$, for $$i \in \{1,\dots ,d\}$$ then:

\begin{aligned} a \circ _J b:= & {} J^{[n]}_{t_0} \left( a \circ b \right) \nonumber \\= & {} J^{[n]}_{t_0} \left( \left( J^{[n]}_{b_{[0]}} g\right) \circ \left( J^{[n]}_{t_0}h_1, \dots , J^{[n]}_{t_0} h_d\right) \right) . \end{aligned}
(5)

### Remark 2

Operation from Eq. (4) can be effectively implemented in an algorithmic and effective way by means of Automatic Differentiation [26, 27].

From the Taylor’s Theorem with integral form of the remainder it follows:

\begin{aligned} x(t) = \left( J^{[n]}_{a}x\right) (t) + (n+1) \cdot \int _{a}^{t} x^{[n+1]}(s) \cdot (t-s)^n ds. \end{aligned}
(6)

Equation (6) motivates the following:

### Definition 1

We say that a function $$x: \mathbb {R}\rightarrow \mathbb {R}$$ has a forward Taylor representation of order n on interval $$I = [a, a+\delta )$$, $$\delta > 0$$ iff formula (6) is valid for $$x|_I$$.

We say that $$x: \mathbb {R}\rightarrow \mathbb {R}^d$$ has a forward Taylor representation on I, iff each component $$x_j: \mathbb {R}\rightarrow \mathbb {R}$$ has the representation on I.

Mostly, we will be using jets to describe (parts of) functions $$g: I \rightarrow \mathbb {R}^d$$ with forward Taylor representations; therefore, in such cases we understand that in

\begin{aligned} \left( J^{[n]}_z g\right) (y) = \sum _{k=0}^{n} \frac{g^{(k)}(z)}{k!} \cdot (y - z)^k \end{aligned}

the $$g^{(k)}$$ is computed as a right-side derivative.

It is easy to see and it will be often used in the algorithms:

### Proposition 3

Assume $$x: \mathbb {R}\rightarrow \mathbb {R}$$ has a forward Taylor representation over $$[t, t+\delta )$$ of order n. Then, for $$k \in \{0, \dots , n\}$$ the function $$x^{[k]} = \frac{x^{(k)}}{k!}$$ has a forward Taylor representation over $$[t, t+\delta )$$ of order $$m = n-k$$ and

\begin{aligned} J^{[m]}_{t}(x^{[k]})&= \left( c^0,\dots ,c^{n-k}\right)&\\ (x^{[k]})^{[m+1]}(s)&= \left( {\begin{array}{c}n+1\\ k\end{array}}\right) \cdot x^{[n+1]}(s)&s \in [t,t +\delta ) \end{aligned}

where

\begin{aligned} c^l&= \left( {\begin{array}{c}l+k\\ k\end{array}}\right) \cdot x^{[l+k]}(t),&l \in 0,\dots ,n-k. \end{aligned}

### Proposition 4

Assume $$x: \mathbb {R}\rightarrow \mathbb {R}$$ has a forward Taylor representation over $$I = [t, t+\delta )$$ of order n. Then, for $$k=0,\dots ,n$$

\begin{aligned} x^{[k]}(t + \varepsilon )&= \sum _{l=0}^{n-k} \left( {\begin{array}{c}l+k\\ k\end{array}}\right) \cdot \left( J^{[n]}_{t} x\right) _{[l+k]} \cdot \varepsilon ^l \ \nonumber \\&\quad + (n+1-k) \cdot \int _{0}^{\varepsilon } \left( {\begin{array}{c}n+1\\ k\end{array}}\right) \cdot x^{[n+1]}(t+s) \cdot (\varepsilon -s)^{n-k} ds \end{aligned}
(7)

for $$\varepsilon \in [0, \delta )$$.

### Remark 5

(On treating jets as vectors and vice versa) As mentioned earlier, for $$g: \mathbb {R}\rightarrow \mathbb {R}$$ the Taylor series $$J^{[n]}_{t_0} g$$ (which is formally also a function $$\mathbb {R}\rightarrow \mathbb {R}$$) can be uniquely identified with the collection of the Taylor coefficients $$\left( g^{[k]}(t_0)\right) _{0 \le k \le n}$$, and this collection might be identified with a vector in $$\mathbb {R}^{n+1}$$. One have a freedom how to organize the sequence into the vector (up to a permutation of coefficients), but in computer programs we will use the standard ordering from $$k = 0$$ at the first coordinate of the vector and $$k = n$$ at the last coordinate. Conversely, for any vector $$j \in \mathbb {R}^n$$:

\begin{aligned} j = \left( j_{[0]}, j_{[1]}, \ldots , j_{[n]}\right) , \end{aligned}
(8)

we can build a jet (at some point $$t_0$$) given by

\begin{aligned} \left( J_{t_0}^{[n]}g\right) (t) = \sum _{k=0}^{n} j_{[k]} (t - t_0)^k. \end{aligned}
(9)

This notion will be convenient when we would have some estimates on the jet, in particular, we can write that a jet $$J^{[n]}_{t_0}{g} \in X \subset \mathbb {R}^{n+1}$$, meaning, that there exists a vector $$j \in X$$ such that (9) is true for j interpreted as a jet at a given $$t_0$$. Also, we can use the convention to do algebraic operations on jets, such as vector–matrix multiplication to describe jets in suitable coordinates, etc.

We will use convention with square brackets $$j_{[k]}$$ to denote the relevant coefficient from the sequence $$j = J^{[n]}_{z}{g}$$, and to underline the fact that we are using the vector j as its jet interpretation.

For $$g: \mathbb {R}\rightarrow \mathbb {R}^d$$, the jet $$J^{[n]}_{t_0} g$$ can be represented as a vector in a high-dimensional space $$\mathbb {R}^M$$, where $$M = d \cdot (n+1)$$. We organize such jets into vectors in the same manner as in Eq. (8), but each $$j_{[k]}$$ represents d consecutive values.

### 2.2 Outline of the Method and the Motivation for Phase Space Description

In a numerical Taylor method for ODEs, one produces the jet of solution at the current time $$t_0$$ by differentiating the equation $$x'(t) = f(x(t))$$ w.r.t. t on both sides at $$t_0$$, as long as the differentiation makes sense. For $$f \in C^\infty$$ we can get any order of the jet at $$t_0$$ and the situation is similar in the case of DDE (2). If f has a jet at $$z = \left( x(t_0), x(t_0 - \tau )\right)$$ and x has a jet at $$(t_0 - \tau )$$, both of order n, then we can proceed as in the case of ODEs to obtain jet at $$t_0$$. In the following Lemma we underline the fact that this jet can be computed from $$x(t_0)$$ and $$J^{[n]}_{t_0 - \tau } x$$:

### Lemma 6

Let $$t_0$$ be fixed and z be a solution to (2) with f of class at least $$C^{n}$$. Assume z exists on $$[t_0 - \tau , t_0 + \delta ]$$, and z is of class $$C^{n}$$ on some past interval $$I = [t_0 - \tau , t_0 - \tau + \delta )$$ for some $$\delta > 0$$. Then z is of class $$C^{n+1}$$ on $$I = [t_0, t_0 + \delta )$$, $$J^{[n+1]}_{t_0} z$$ exists and it is given explicitly in terms of $$z(t_0)$$, $$J^{[n]}_{t_0 - \tau } z$$ and r.h.s. f of Eq. (2).

### Proof

The continuity $$C^{n+1}$$ on $$[t_0, t_0 + \delta )$$ follows directly from (2), since $$x'$$ is of class $$C^n$$ on $$[t_0, t_0 + \delta )$$. Let $$F(t):= f(z(t-\tau ), z(t))$$ and denote the coefficients of jets $$J^{[n]}_{t_0}F$$, $$J^{[n+1]}_{t_0}z$$ and $$J^{[n]}_{t_0 - \tau }z$$ by $$F_{[0]},\ldots ,F_{[n]}$$, $$x_{[0]},\ldots ,x_{[n]},x_{[n+1]}$$ and $$y_{[0]},\ldots ,y_{[n]}$$, respectively, that is

\begin{aligned} \left( J^{[n]}_{t_0} F\right) (t)= & {} F_{[0]} + F_{[1]} \cdot (t-t_0) + \dots + F_{[n]} \cdot (t-t_0)^n, \\ \left( J^{[n+1]}_{t_0} z\right) (t)= & {} x_{[0]} + x_{[1]} \cdot (t-t_0) + \dots + x_{[n]} \cdot (t-t_0)^n + x_{[n+1]} \cdot (t-t_0)^{n+1} \\ \left( J^{[n]}_{t_0 - \tau } z\right) (t)= & {} y_{[0]} + y_{[1]} \cdot (t-t_0) + \dots + y_{[n]} \cdot (t-t_0)^n \end{aligned}

Now Eq. (2) implies that

\begin{aligned} \left( J^{[n+1]}_{t_0} z\right) ' = J^{[n]}_{t_0}{F}, \end{aligned}

or more explicitly:

\begin{aligned}{} & {} \left( x_{[0]}+ x_{[1]}(t-t_0)+ \dots + x_{[n+1]}(t-t_0)^{n+1}\right) '\nonumber \\ {}{} & {} \quad = F_{[0]} + F_{[1]}(t-t_0) + \dots + F_{[n]}(t-t_0)^n. \end{aligned}

Using the obvious fact that $$(z^{(k)})' = z^{(k+1)}$$, we have $$(z^{[k]})' = (k+1) z^{[k+1]}$$ and matching coefficients of the same powers we end up with:

\begin{aligned} x_{[k]}=\frac{1}{k} F_{[k-1]}. \end{aligned}
(10)

Finally, using Proposition 1 on $$J^{[n]}_{t_0}{F}$$ we get:

\begin{aligned} J^{[n]}_{t_0}{F}= & {} \left( J^{[n]}_{(z(t_0), z(t_0-\tau ))}{f}\right) \circ _J \left( J^{[n]}_{t_0}{z}, J^{[n]}_{t_0-\tau }{z} \right) \\= & {} \left( J^{[n]}_{(x_{[0]}, y_{[0]})}{f}\right) \circ _J \left( x, y \right) . \end{aligned}

Now, we get the following recurrent formula:

\begin{aligned} F^{[0]}(x_{[0]}, y):= & {} f(x_{[0]}, y_{[0]}), \nonumber \\ F^{[k]}(x_{[0]}, y):= & {} \left( \left( J^{[k]}_{(x_{[0]}, y_{[0]})}{f}\right) \circ _J \left( \left( x_{[0]}, w_k * F^{[k-1]}(x_{[0]}, y) \right) , \left( y_{[0]},\ldots ,y_{[k]}\right) \right) \right) , \nonumber \\ \end{aligned}
(11)

for $$1 \le k \le n$$ with operation $$w_n * j$$ defined for a jet j as:

\begin{aligned} w_n * j:= \left( \frac{1}{1} j_{[0]}, \frac{1}{2} j_{[1]}, \ldots , \frac{1}{n} j_{[n-1]}\right) . \end{aligned}

Obviously $$F^{[k]}(x_{[0]}, y) = (F_{[0]},\ldots , F_{[k]}) = J^{[k]}_{t_0} F$$, and together with (10) we get:

\begin{aligned} \left( x_{[0]}, \ldots , x_{[n]}, x_{[n+1]}\right) = \left( x_{[0]}, w_{n+1} \cdot F^{[n]}(x_{[0]}, y)\right) , \end{aligned}
(12)

that depends only on the formula for f, $$x_{[0]} = z(t_0)$$ and the jet $$y = J^{[n]}_{t_0 - \tau } z$$. $$\square$$

We note two important facts. Firstly, the a priori existence of the solution z over $$[t_0, t_0 + \delta )$$ is assumed in Lemma 6 and, when doing the integration step, it needs to be achieved by some other means—we will later show one way to do that. Secondly, Eq. (12) gives recipe to produce $$J^{[n+1]}_{t_0}$$—a jet of order one higher than the order of the input jet $$y = J^{[n]}_{t_0 - \tau } x$$. This simple observation will lead to a significant improvement in the rigorous integration algorithm in comparison with the first version presented in [34]. To have a complete rigorous method, we will need also formulas to estimate Taylor remainder in (6)—we will do this later in Sect. 3.

As the jet at $$t_0 - \tau$$ and the value at $$t_0$$ allows to compute the jet of the solution x at $$t_0$$, the reasonable choice for the description of functions in the phase space is to use piecewise Taylor representation of the solutions at grid points that match the step size of the method. Uniform step size over the integration time will assure that the required jets of the solution in the formula (12) are always present in the description of the solution. This approach have been proposed in [34] with the uniform order of the jets at each grid point. Now, we are going to elaborate how to implement and use the extra derivative we get in Eq. (12) to improve the method. For this, we will need a representation of solutions with non-uniform order of jets.

### 2.3 Representation of the Phase Space

Previously, in [34], we have proposed to describe sets in the phase space by piecewise Taylor forward representation of a fixed order n on a uniform grid of points over basic interval $$[-\tau , 0]$$. Our definition was stated for $$d = 1$$ (scalar equations), but the notion can be extended to any number of dimensions—just by assuming each of the Taylor coefficients in equations are in fact d-dimensional vectors. No formula will be different in that case. In the rest of the paper we will assume that d is known from the general context, so we will omit it from the definitions.

We start with a key definition from [34] and then we will propose some generalization that will be relevant to many important improvements proposed later in this paper.

### Definition 2

Let $$p \ge 1$$, $$n \ge 0$$ be given natural numbers. Let $$h = \frac{\tau }{p}$$ be a grid step, $$t_i = -i \cdot h$$ be grid points for $$i \in \{0, \dots , p\}$$ and let intervals $$I_i = [t_i, t_{i-1})$$ for $$i \in \{1, \dots , p\}$$.

We define $$C^n_p([-\tau , 0], \mathbb {R}^d)$$ to be a set of functions $$x: [-\tau , 0] \rightarrow \mathbb {R}^d$$ such that x has a forward Taylor representation of order n on all $$I_i$$ and such that $$x^{(n+1)}$$ (understood as a right derivative) is bounded over whole $$[-\tau , 0]$$.

From now on, we will assume that $$\tau$$ is fixed and we will write $$C^n_p$$ and $$C^{k}$$ to denote $$C^n_p([-\tau , 0], \mathbb {R}^d)$$ and $$C^k([-\tau , 0], \mathbb {R}^d)$$, respectively. Moreover, whenever we use p and h without an additional assumption, we assume that h is given by $$h = h(p, \tau ) = \frac{\tau }{p}$$ as in Definition 2.

Note that $$x \in C^{n}_{p}$$ might be discontinuous at $$t = t_i$$, $$i \in \{-p, \dots , 0\}$$. However, $$C^n_p \cap C^k$$ is a linear subspace of $$C^k$$ for any $$k \in \mathbb {N}$$ and if $$k > n$$ then obviously $$C^k_p \subset C^n_p$$ (see [34]). Therefore, $$\mathcal {X} = C^n_p \cap C^0$$ can be used as a suitable subspace of the phase space $$C^0$$ for solutions of Eq. (2). In fact, following two lemmas, proved in [34], state that $$\varphi (h, \cdot )$$ and $$\varphi (t, \cdot )$$ for t large enough are well defined maps $$\mathcal {X} \rightarrow \mathcal {X}$$:

### Lemma 7

Assume f in (2) is $$C^\infty$$ (or smooth enough). Let $$\psi \in C^{n}_p$$ be an initial function to (2). If $$\varphi (h, \psi )$$ exists then $$\varphi (h, \psi ) \in C^{n}_{p}$$. Moreover, if $$\psi \in C^n_p \cap C^0$$ and $$i = k \cdot p$$ for some $$k \in \mathbb {N}$$ then $$\varphi (i \cdot h, \psi ) \in C^{n+k}_{p} \cap C^k$$.

### Lemma 8

Assume f in (2) is $$C^\infty$$ (or smooth enough). Let $$\psi \in C^{n}_p \cap C^0$$ be initial function so that the solution to (2) exists up to some $$t \ge T$$, where $$T = T(n, \tau ) = (n+1) \cdot \tau$$. Then $$\varphi (t, \psi ) \in C^{n}_{p} \cap C^0$$.

Time $$T(n, \tau )$$ will be important when constructing Poincaré maps later in the paper, so to underline its importance, we state the following:

### Definition 3

We call T(n) in Lemma 8 a long enough integration time.

In the current work, we generalize the notion of the space $$C^n_p$$ to allow different order of the jets at different points of the grid. This will be beneficial to the final estimates later, as the representation of functions will take advantage of the smoothing of solutions:

### Definition 4

Let p be fixed, $$\eta = (n_1, \dots , n_p) \in \mathbb {N}^p$$ and let $$t_i, I_i, h$$ be as in Definition 2. We define space of functions $$C^\eta _p$$ so that $$x \in C^\eta _p$$ iff x has a forward Taylor representation of order $$n_i$$ on $$I_i$$ and $$x^{(n_i + 1)}(I_i)$$ is bounded for $$i \in \{1,\dots ,p\}$$.

The discussion from Sect. 2.2 about the smoothing of solutions of DDEs shows that if we have nth-order Taylor representation at $$t = -\tau$$ then we can obtain $$(n+1)$$th-order representation of x at $$t = 0$$. Therefore, the order of the representation of solution will not decrease during the integration, and it can increase, in general, only by one at a time (after integration over a full delay). Therefore, we introduce the following special class of $$C^\eta _p$$ spaces. Let $$q \in \{0, \dots , p\}$$ by $$C^n_{p, q}$$ we will denote the space $$C^{\eta }_{p}$$ with

\begin{aligned} \eta _i = {\left\{ \begin{array}{ll} n+1 &{} i \le q \\ n &{} i > q \end{array}\right. }, \end{aligned}

that is, the Taylor representation would be of order n on grid points $$-\tau = t_p, t_{p-1}, \dots , t_{q-1}$$ and of order $$n+1$$ on $$t_{q}, t_{q+1}, \dots , t_1 = h$$. Among all $$C^\eta _p$$ spaces, spaces $$C^n_{p,q}$$ will be used most extensively in the context of rigorous integration of DDEs, but we keep the general notation of Definition 4 for simplicity of formulas later.

Now, it is easy to see that $$C^{n}_{p, p} = C^{n+1}_{p, 0}$$ and so that $$C^n_p = C^n_{p,0}$$. Analogously we can write for $$q > p$$ that $$C^{n}_{p, q} = C^{n+{\overline{q}}}_{p, {\underline{q}}}$$ with $${\overline{q}} = \left\lfloor \frac{q}{p} \right\rfloor$$ and $${\underline{q}} = q \bmod p$$. With that in mind the analogue of Lemma 7 can be stated as:

### Lemma 9

Let $$\psi \in C^{n}_{p, q}$$ be an initial function to (2) and let h be as in Definition 2. If $$\varphi (h, \psi )$$ exists then $$\varphi (h, \psi ) \in C^{n}_{p,q+1}$$. Moreover, if $$\psi \in C^n_{p,q} \cap C^0$$ and $$m = k \cdot p$$ for some $$k \in \mathbb {N}$$ then $$\varphi (m \cdot h, \psi ) \in C^{n+k}_{p,q} \cap C^k$$.

### Proof

It follows from the smoothing of solutions, the definition of $$C^{n}_{p,q}$$, equality of spaces $$C^{n}_{p,p} = C^{n+1}_{p,0}$$ and by applying method of steps (see e.g. [4]) to solve (2). $$\square$$

In the rigorous method we will use Lemma 9 as follows: we will start with some set $$X_0 \subset C^n_p = C^n_{p,0}$$ defined with a finite number of constraints. Then we will in sequence produce representations of sets $$X_i = \varphi (h, X_{i-1}) \in C^{n}_{p,i} = C^{n + {\overline{i}}}_{p,{\underline{i}}}$$. Finally, to compare sets defined in different $$C^\eta _p$$ spaces we would need the following simple fact:

### Proposition 10

$$C^{\eta }_{p} \subset C^{\zeta }_{p}$$ iff $$\eta _i \ge \zeta _i$$ for all $$i \in \{1,\dots ,p\}$$.

Now we show how to describe sets in $$C^{\eta }_p$$. Obviously, by the Taylor’s theorem, we have that $$x \in C^\eta _{p}$$ is uniquely described by a tuple $${\bar{x}} = \left( z(x), j(x), \xi (x)\right)$$, where

• $$z(x):= x(0) \in \mathbb {R}^d$$,

• $$j(x):= (j_1(x),\dots ,j_p(x))$$ with $$j_i(x):= J^{[n_i]}_{t_i}(x) \in \mathbb {R}^{d \cdot (n_i+1)}$$,

• $$\xi (x):= (\xi _1(x),\dots ,\xi _p(x))$$ and $$\xi _i(x):= x^{[n_i+1]}|_{I_i} \in C^0(I_i, \mathbb {R}^d)$$ are bounded.

Please note that the subscript i denotes the grid point here, not the component of the x in $$\mathbb {R}^d$$. We will usually use subscript j for this purpose and we will write $$z(x)_j$$, $$j_i(x)_j$$, etc., but for now, all formulas can be interpreted simply for $$d = 1$$, generalization to many dimensions being straightforward. We will use notation of z(x), j(x), $$\xi (x)$$ etc. for a shorthand notation in formulas, sometimes dropping the argument x if it is known from the context. For example, we will say that we have a solution described by a tuple $$(z, j, \xi ) \in \mathbb {R}^M \times (C^{0})^{p\cdot d}$$, then we will know how to interpret them to get the function x. Here $$M = M(p, \eta , d) = d \cdot (1 + \sum _{i=1}^{p} (\eta _i + 1))$$. A direct consequence is that:

### Proposition 11

The space $$C^\eta _p$$ is a Banach space isomorphic to $$\mathbb {R}^M \times (C^{0})^{p \cdot d}$$ by $$x \mapsto (z(x), j(x), \xi (x))$$, and with a natural norm on x given by

\begin{aligned} \Vert x\Vert _{C^\eta _p}:= \left\| \left( z(x), j(x)\right) \right\| + \sum _{i = 1}^{p}\sum _{j = 1}^{d} \sup _{t \in I_i} |\xi (x)_j(t)|, \end{aligned}

where $$\Vert \cdot \Vert$$ denotes any norm in $$\mathbb {R}^M$$ (all equivalent). We will use $$\max$$ norm in $$\mathbb {R}^M$$.

Let now $$\mathbb {I}$$ be a set of all closed intervals over $$\mathbb {R}$$. We define:

\begin{aligned}{}[\xi ]_i(x)_j:= & {} \left[ \min _{\varepsilon \in [0, h]} \xi _i(x)_j(\varepsilon ), \max _{\varepsilon \in [0, h]} \xi _i(x)_j(\varepsilon ) \right] \in \mathbb {I}, \nonumber \\ [\xi ]_i(x):= & {} [\xi ]_i(x)_1 \times \cdots \times [\xi ]_i(x)_d \in \mathbb {I}^{d} \end{aligned}
(13)

and $$[\xi ](x) = \left( [\xi ]_1(x),\dots ,[\xi ]_p(x)\right) \in \mathbb {I}^{d \cdot p}$$. That is a very complicated way to say $$[\xi ](x)$$ is the collection of bounds on the remainder terms in the Taylor representation of x. The interval $$[\xi ]_i(x)_j$$ is well defined, since we assumed each $$x^{(n_i+1)}$$ bounded in Definition 4. Now, we can describe $$x \in C^{\eta }_{p}$$ by the following finite set of numbers:

### Definition 5

Let $$M = M(p, \eta , d) = d \cdot \left( 1 + \sum _{i=1}^{p} (\eta _i+1) \right)$$.

We say that $${\bar{x}} = (z(x), j(x), [\xi ](x)) \in \mathbb {R}^M \times \mathbb {I}^{d \cdot p}$$ is a (p,$$\eta$$)-representation of $$x \in C^\eta _p$$.

Given $${\bar{x}} \in \mathbb {R}^M \times \mathbb {I}^{d \cdot p}$$ by $$X({\bar{x}}) \subset C^\eta _p$$ we denote the set of all functions whose $${\bar{x}}$$ is their (p,$$\eta$$)-representation.

The number M is called the size of the representation and we will omit parameters if they are known from the context. We will use shorthand notation of $$\mathbb {R}^{n}_{p}$$, $$\mathbb {R}^{\eta }_{p}$$ or $$\mathbb {R}^{n}_{p,q}$$ to denote appropriate $$\mathbb {R}^{M}$$ in context of spaces $$C^{n}_{p}$$, $$C^{\eta }_{p}$$ and $$C^{n}_{p,q}$$, respectively. We will write $$\mathbb {I}_p$$ to denote $$\mathbb {I}^{p \cdot d}$$. Note that we are dropping d because it is always well known from the context.

Observe that, in general, $$X({\bar{x}})$$ contains infinitely many functions. We will identify $${\bar{x}}$$ and $$X({\bar{x}})$$, so that we could use notion of $$z({\bar{x}})$$, $$j({\bar{x}})$$, etc. Moreover, we will further generalize the notion of $$X({\bar{x}})$$:

### Definition 6

Let $$A \subset \mathbb {R}^\eta _p$$, $$R \in \mathbb {I}_p$$ be a product of closed intervals. We define set X(AR) as

\begin{aligned} X(A, R) = \left\{ x \in C^{\eta }_{p}: (z(x), j(x)) \in A, [\xi ](x) \subset R \right\} \end{aligned}

We call X(AR) a (p,$$\eta$$)-functions set (or (p,$$\eta$$)-fset for short) and (AR) its (p,$$\eta$$)-representation.

If A is convex then X(AR) is also a convex subset of $$C^{\eta }_p$$, so $$X(A, R) \cap C^k$$ is also convex for any $$k \in \mathbb {N}$$, see [34]. For a space $$C^{n}_{p,q}$$ we will use the term (pqn)-representation and (pqn)-fsets when needed, but usually we will use just names like “fset” and “representation”.

Finally, we introduce the following shorthand symbols used for evaluation of terms:

\begin{aligned} \texttt{T}^n(j; \varepsilon ):= & {} \sum _{k=0}^{n} j_{[k]} \cdot \varepsilon ^k\, \end{aligned}
(14)
\begin{aligned} \texttt{S}^n(\xi ; \varepsilon ):= & {} (n+1) \cdot \int _0^{\varepsilon } \xi _i(s) \cdot (\varepsilon -s)^n \textrm{d}s, \end{aligned}
(15)
\begin{aligned} \texttt{E}^n(j, \xi ; \varepsilon ):= & {} \texttt{T}^n(j; \varepsilon ) + \texttt{S}^n(\xi ; \varepsilon ), \end{aligned}
(16)

for any function $$\xi \in C^{0}([0, h), \mathbb {R}^d)$$ and any jet $$j \in \mathbb {R}^{N \cdot d}$$ of order $$N \ge n$$. The letters should be coined to the terms $$\texttt{T}$$—(T)aylor sum, $$\texttt{S}$$—(S)umma, formal name for the integral symbol, $$\texttt{E}$$—(E)valuation of the function. We use superscript n to underline order to which the operation applies, but in general, it can be simply inferred from the arguments (for example—maximal order of the jet j in $$\texttt{T}$$). Also, the superscript argument might be used to truncate computation for higher-order jets, e.g. let $$j = J^{[2n]}_t x$$ and consider applying $$\texttt{T}^{n}(j)$$ to Taylor-sum only part of the jet. This will be used in algorithms later. If we omit the parameter n then it is assumed that we use the biggest possible n (for that argument, inferred from the representation itself).

Then we will write formally for any $$x \in C^\eta _p$$:

\begin{aligned} \texttt{T}^n(x; t):= & {} \texttt{T}^n(j_i(x); \varepsilon ), \\ \texttt{S}^n(x; t):= & {} \texttt{S}^n(\xi _i(x); \varepsilon ), \\ \texttt{E}^n(x; t):= & {} \texttt{T}^n(j_i(x); \varepsilon ) + \texttt{S}^n(\xi _i(x)(\cdot - t_i); \varepsilon ), \end{aligned}

where $$t = t_i + \varepsilon$$, $$\varepsilon \in [0, h)$$. For $$X = X(A,R)$$ we will write $$a(X) = A$$ and $$[\xi ](X) = R$$ and for $$x \in X(A, R)$$ we will write $$a(x) = (z(x), j(x)) \in A$$. We will also extend the notion of operators $$\texttt{T}$$, $$\texttt{S}$$ and $$\texttt{E}$$ to ($$p,\eta$$)-fsets:

\begin{aligned} \texttt{T}(X; t):= & {} \texttt{T}^{\eta _i}(j_i(X); \varepsilon ), \\ \texttt{S}(X; t):= & {} [\xi ]_i(X) \cdot \varepsilon ^{\eta _i+1}, \\ \texttt{E}(X; t):= & {} \texttt{T}^{\eta _i}(j_i(X); \varepsilon ) + \texttt{S}^{\eta _i}([\xi ]_i(X); \varepsilon ). \end{aligned}

where $$t = t_i + \varepsilon$$, $$\varepsilon \in [0, h)$$. Note that $$\texttt{T}(x; t) = \texttt{T}({\bar{x}}; t)$$, $$\texttt{S}(x; t) \in \texttt{S}({\bar{x}}; t)$$ and of course $$\texttt{E}(x; t) \in \texttt{E}({\bar{x}}; t)$$. In the rigorous computation we as well might use intervals or whole sets in the computation (e.g. $$t = [t] = t_i + [0, \varepsilon ]$$)—in such circumstances we will get sets representing all possible results and in that way an estimate for the true value. From now on, we will also drop bar in $${\bar{x}}$$ wherever we treat x as an element of $$C^{\eta }_p$$ with a known bounds in form of some X(AR).

Finally, we make an observation that for x—a solution to DDE (2) such that $$x_{t_0} \in C^\eta _p$$—the kth derivative $$x_{t_0}^{[k]}$$ must also by representable by piecewise Taylor representation. In fact, since we know x(0) and all jets of the representation of $$x_{t_0}$$ we can obtain $$x_{t_0}^{[k]}(0)$$ by applying Lemma 6, namely Eq. (12). Then, the value of all other jets and remainders follows from Proposition 3:

### Proposition 12

Let $$x \in C^{\eta }_{p}$$ be a segment of a solution to DDE (2) and for $$k \in \mathbb {N}$$ define $$\eta - k:= (\eta _1-k, \dots , \eta _p -k)$$. Then for $$1 \le k \le \min _i{\eta _i}$$ the derivative $$x^{[k]}(t)$$ (interpreted as a right derivative) exists for $$t \in [-\tau , 0]$$ and $$x^{[k]} \in C^{\eta -k}_{p}$$, with a (p,$$\eta -k$$)-representation given in terms of the (p,$$\eta$$)-representation of x:

\begin{aligned} j_i(x^{[k]})&= \left( c^0_i,\dots ,c^{\eta _i-k}_i\right) \nonumber \\ \xi _i(x^{[k]})&= \left( {\begin{array}{c}\eta _i+1\\ k\end{array}}\right) \cdot \texttt{S}^{\eta _i-k}(\xi _i(x); \ \cdot \ ) \nonumber \\ [\xi ]_i(x^{[k]})&\subset \left( {\begin{array}{c}\eta _i+1\\ k\end{array}}\right) \cdot [\xi ]_i(x), \nonumber \\ z(x^{[k]})&= \frac{1}{k} \cdot \left( F_{k-1} \left( j_p(x), z(x) \right) \right) _{[k-1]}, \end{aligned}
(17)

for $$i \in \{1,\dots ,p\}$$, where

\begin{aligned} c^l_i&= \left( {\begin{array}{c}l+k\\ k\end{array}}\right) \cdot j_i(x)_{[l+k]},&l \in 0,\dots ,\eta _i-k. \end{aligned}

## 3 Rigorous Integrator: Basic Algorithms and Some Improvements

Now we are ready to show how to obtain estimates on the representation Y of $$\varphi (h, X)$$ for a given set of initial functions $$X \in C^{\eta }_{p}$$. Due to the finite nature of the description of the set Y we will only have the relation $$\varphi (h, X) \subset Y$$, in general.

First, we want to recall in short the details of the integrator from [34] as those are crucial in the improvements presented later. Then, we will show how to incorporate new elements: the extension of the representation from (12) and the spaces $$C^\eta _p$$, the generalization to systems of equations (i.e. $$d > 1$$), and to multiple delays (under the assumption that they match the grid points). Then, we will discuss the Lohner-type method for the generalized algorithm.

### 3.1 ODE Tools

We start with describing some ODE tools to be used in rigorously solving (18) using the computer. For this we will need a method to find rigorous enclosures of the solution x (and its derivatives w.r.t. t) over compact intervals [0, h]. A straightforward method here is to consider Eq. (2) on $$[t_0, t_0 + h]$$, $$h \le \tau$$ as a non-autonomous ODE, just as in the case of method of steps [4]. If we plug-in a known initial function $$x_{t_0}$$ into (2) and we denote $${\hat{f}}(z, t):= f(z, x_{t_0}(t-\tau ))$$ for $$t\in [0, h]$$ we end up with non-autonomous ODE:

\begin{aligned} {\left\{ \begin{array}{ll} z'(t) = {\hat{f}}(z, t), &{} t \in [0, h], \\ z(0) = x_{t_0}(0). \end{array}\right. } \end{aligned}
(18)

Please note that $$t - \tau \in {{\,\mathrm{\texttt{Dom}}\,}}\left( x_{t_0}\right) = [-\tau , 0]$$ so $${\hat{f}}$$ is well defined, and $${\hat{f}}$$ is of class $$C^{k}$$ as long as the solution segment $$x_{t_0}$$ is of class $$C^{k}$$ (for f sufficiently smooth). Therefore, in view of (10) and (11), to find estimates on the Taylor coefficients of x over $$I_{t_0} = [t_0, t_0 + h)$$ it suffices only to ascertain the existence of z over $$I_{t_0}$$ and to have some finite a priori bounds Z on it, as the estimates on the higher-order coefficients will follow from recurrent formulas (10) and (11). Luckily, the existence of the solution to Eq. (18) and a good a priori bounds over $$I_{t_0}$$ can be obtained using existing tools for ODEs [22, 43] as was shown in [34] and efficient implementations are already available [11, 12]. We have the following:

### Lemma 13

(see Theorem 1 in [22]) We consider $${\hat{f}}$$ as in non-autonomous ODE (18).

Let $$B \subset \mathbb {R}^d$$ be a compact set. If a set $$W \subset \mathbb {R}^d$$ is such that

\begin{aligned} B + [0,\varepsilon ] \cdot {\hat{f}}(W, [t_0,t_0+\varepsilon ]) =: Z \subset W, \end{aligned}

then, any solution z of (18) such that $$z(t_0) \in B$$ has $$z(t_0 + \delta ) \in Z$$ for all $$\delta \in [0,\varepsilon ]$$.

By $$\texttt{roughEncl}$$ we denote a procedure (heuristic) to find the set Z:

\begin{aligned} \texttt{roughEncl}(f, B, t_0, \varepsilon ):= Z, \text { as in Lemma}~13. \end{aligned}

We do not go into the details of this algorithm nor the proof of Lemma 13, but we refer to [11, 22, 43] and references therein.

### Remark 14

Please note that finding a rough enclosure is a heuristic procedure, and therefore it is the point where the algorithm can fail (in fact the only one). If that happens, we must abort computations or apply some strategy to overcome the problem. In the ODE context, it is possible to shorten the step or to subdivide the set of initial conditions. Those strategies can be difficult to adopt in the DDE context: we cannot shorten step because of the definition of $$C^\eta _p$$ spaces and the loss of continuity problems discussed earlier; and we could not afford extensive subdivision as we work with very high-dimensional representations (projections) of functions. This makes obtaining the higher-order methods even more useful.

Consider now $$x_{t_0} \in C^n_p$$, so that $${\hat{f}} \in C^{n+1}$$. Applying Eqs. (10) and (11) allows to obtain $$J^{[n+1]}_{t_0} x$$, where rough enclosure procedure gives Z such that $$x(I_{t_0}) \subset Z$$. In what follows, we will sum up all the formulas needed to obtain (guaranteed enclosures on) the forward Taylor representation of x on the interval $$I_{t_0}$$ of order $$n+1$$.

### 3.2 The Rigorous Integrator in $$C^{n}_{p,q}$$

Assume now that we are given some $$x_0 \in C^n_{p,q}$$. We will show how to compute rigorous estimates on a set $$X(A_h, R_h) \subset C^n_{p,q+1}$$, with an explicitly given $$A_h \subset \mathbb {R}^n_{p,q+1}$$ and $$R_h \in \mathbb {I}^{p\cdot d}$$, representing $$\varphi (h, x_0)$$, i.e. $$\varphi (h, x_0) \in X(A_h, R_h)$$. The sets $$A_h$$ and $$R_h$$ will be computed using only data available in $$(z(x), j(x), \xi (x))$$. The subscript h in $$A_h$$, $$R_h$$ is used to underline that we are making a full step $$h = \frac{1}{p}$$. In what follows, we will use the convention that $$X_h = X(A_h, R_h)$$.

This is an analogue to the algorithm described in Sect. 2.2 in [34], but we account for the effect of smoothing of the solutions in DDEs (Lemma 9), so that $$\varphi (h, x_0) \in C^{n}_{p,q+1}$$ (and we remind that $$C^n_{p,p} = C^{n+1}_{p,0} = C^{n+1}_p$$):

### Theorem 15

Let $$x \in C^n_{p,q}$$, with $$0 \le q < p$$ and the representation $$(z(x), j(x), [\xi ](x)) \in \mathbb {R}^n_{p,q} \times \mathbb {I}^{d \cdot p}$$.

We define the following quantities:

\begin{aligned} {\hat{f}}&:= \text { as in Eq.~18 } \nonumber \\ [c]_{[k]}&:= \texttt{E}\left( x^{[k]}; [-\tau , -\tau +h]\right) = \texttt{E}\left( x^{[k]}, [t_p, t_p + h]\right) \in \mathbb {I}^d,&0 \le k \le n \nonumber \\ [c]_{[n+1]}&:= [\xi ]_p(x) \in \mathbb {I}^d \nonumber \\ [Z]&:= \texttt{roughEncl}({\hat{f}}, z(x), t_0, h) \in \mathbb {I}^d \end{aligned}
(19)
\begin{aligned}{}[F]&:= F^{[n+1]}\left( [Z], [c]\right) \end{aligned}
(20)

Then, we have for $$y = \varphi (x, h)$$ the following:

\begin{aligned} j_i(y)&= j_{i-1}(x) \ =: \ j_i(X_h)&i \in \{2,\dots ,p\} \end{aligned}
(21)
\begin{aligned}{}[\xi ]_i(y)&= [\xi ]_{i-1}(x) \ =: \ [\xi ]_i(X_h)&i \in \{2,\dots ,p\} \end{aligned}
(22)
\begin{aligned} j_1(y)&= \left( z(x), w_{n+1} * F^{[n]}\left( z(x), j_p(x)\right) \right) \ =: \ j_1(X_h)&\end{aligned}
(23)
\begin{aligned}{}[\xi ]_1(y)&\subset \frac{1}{n+2} \cdot [F]_{[n+1]} \ =: \ [\xi ]_1(X_h),&\end{aligned}
(24)
\begin{aligned} z(y)&\in \texttt{T}(j_1(y); h) + \left( [F]_{[n+1]} \cdot [0, h]\right) \cdot h^{n+1} \ =: z(X_h) \ {}&\end{aligned}
(25)

or, in other words, $$y \in X_h \subset C^{n}_{p,q+1}$$.

### Proof

Equations (21) and (22) are representing the shift in time by h (one full grid point): from segment $$x_0$$ to segment $$x_h$$ (of the solution x), therefore, we simply reassign appropriate jets $$j_i$$ and remainders $$[\xi ]_i$$, as the appropriate grid points in both representations overlap. The rest of formulas are an easy consequence of Lemmas 9 and 13, the recurrence relation (10) for $$F^{[n]}$$ and Proposition 12 to obtain estimates on $$x^{[k]}$$ over intervals $$[-\tau , -\tau + h)$$ in (19). Note that the second term in (25) is formally given by the integral remainder in Taylor formula (6), namely for $$s \in [0, h)$$ we have $$\xi _1(y)(s) \in [\xi ]_1(y) = \frac{1}{n+2} \cdot [F]_{[n+2]}$$ (by the recurrence formula 10) and

\begin{aligned} \texttt{S}\left( \xi _1(y), s\right) \in \texttt{S}\left( [\xi ]_1(y), [0, h]\right)= & {} (n+2) \cdot \left( [\xi ]_1(y) \cdot [0, h]\right) \cdot h^{n+1} \nonumber \\ {}= & {} \left( [F]_{[n+2]} \cdot [0, h]\right) \cdot h^{n+1}. \end{aligned}

$$\square$$

We denote the procedure of computing $$X(A_h, R_h)$$ for a given initial data $$x \in C^n_{p,q}$$ by $$\mathcal {I}$$, i.e. $$\mathcal {I}(x) = X(A_h, R_h)$$. Clearly, it is a multivalued function $$\mathcal {I}: C^n_{p,q} \rightrightarrows C^n_{p,q+1}$$. We are abusing the notation here, as $$\mathcal {I}$$ is a family of maps (one for each domain space $$C^n_{p,q}$$), but it is always known from the context (inferred from the input parameters).

We would like to stress again that the increase in the order of representation at $$t = h$$ in the solution x will be very important for obtaining better estimates later. It happens in Eq. (23), as the resulting jet is of order $$n+1$$ instead of order n as it was in [34]. Please remember that $$F^{[n]}$$ is a recurrent formula for computing whole jet of order n of function $$F = f \circ (x(\cdot ), x(\cdot - \tau ))$$ at the current time t, so it produces a sequence of coefficients, when evaluating (20) and (23). Obviously, each of those coefficients belongs to $$\mathbb {R}^d$$.

The nice property of the method is that the Taylor coefficients at $$t = 0$$, i.e. $$j_1(y)$$ are computed exactly, just like in the corresponding Taylor method for ODEs (or in other words, if x is a true solution to (2) and $$X_h = \mathcal {I}(x)$$ then $$J^{(n+1)}_h(x) = j_1(X_h)$$). It is easy to see, as formulas (21) and (23) does not involve a priori any interval sets (bracketed notation, e.g. $$[\xi ]$$,[Z], etc.). Therefore, to assess local error made by the method we only need to investigate Eq. (25), which is essentially the same as in the Taylor method for ODEs. As the interval bounds are only involved in the remainder part $$\left( [F]_{[n+1]} \cdot [0, h]\right) \cdot h^{n+1}$$, the local error of the method is $$O(h^{n+2})$$. Since $$J^{(n+1)}_h(x) = j_1(X_h)$$ for a true solution x, this error estimation also applies to all the coefficients in the $$j_1(y)$$ computed in Eq. (23) in the next integration step, when computing $$X_{2h} = \mathcal {I}(X_h)$$, as they depend on $$z(X_h)$$ that already contains the error. It will be also easily shown in numerical experiments (benchmarks) presented at the end of this section.

### 3.3 Extension to Many Delays

Now, we are in a position to show how our algorithm can be generalized to include the dependence on any number of delays $$\tau _i$$ as in Eq. (1), as long as they match with the grid points: $$\tau _i = i \cdot h$$. Therefore, we consider the following:

\begin{aligned} x'(t) = f\left( x(t), x(t - p_1 h), x(t - p_2 h), \ldots , x(t - p_m h)\right) , \end{aligned}
(26)

where $$1 \le m \le p$$ and $$p = p_1> p_2> \ldots > p_m \ge 1$$. We will denote by $$u(x_t) = u_f(x_t):= (x(t), x(t - p_1 h), x(t - p_2 h), \ldots , x(t - p_mh))$$ the set of variables that are actually used in the evaluation of the r.h.s. f in Eq. (26) [as opposed to “unused” variables, those at grid points not corresponding to any delays $$\tau _i = p_i \cdot h$$ in (26)]. This distinction will be important to obtain good computational complexity later on. In case of Eq. (2), we have $$u(x_t) = (x(t), x(t-\tau ))$$. Please note that since $$u(x_t)$$ contains variables at grid points, it is easy to obtain $$J_{t}^{(n)} u$$ of appropriate order n. If $$u = u(x_t)$$, we will use subscripts $$u_{0}$$, $$u_{p_1}$$, etc. to denote respective projections onto given delayed arguments, and we will use $$u_{p_i,[k]}$$ to denote their appropriate coefficients of the jet $$J^{(n)}_{-\tau _i} x_t$$.

In order to present the method for many delays we need to redefine $$F(t) = (f \circ u \circ x) (t)$$ and investigate Eqs. (19)–(25). It is easy to see, that the only thing which is different is F and computation of its jets. Thus, we rewrite the algorithm $$F^{[n]}$$ from Eq. (11) in terms of $$u = u(x)$$:

\begin{aligned} F^{[0]}(u):= & {} f(u), \nonumber \\ F^{[k]}(u):= & {} \left( J^{[k]}_{(u)}{f}\right) \circ _J \nonumber \\{} & {} \left( \left( u_0, w_k * F^{[k-1]}(u) \right) , \left( u_{{p_1},[l]}\right) _{0 \le l \le k}, \ldots , \left( u_{{p_m},[l]}\right) _{0 \le l \le k} \right) . \end{aligned}
(27)

Now, the algorithm from (19)–(25) for an $$x \in C^{\eta }_p$$ consists of two parts. First, the enclosure of the solution and all used variables over the basic interval [0, h]:

\begin{aligned} {\hat{f}} (t, z)&:= f(z, x(t-p_1 h), x(t-p_2 h),\ldots , x(t-p_m h))&\nonumber \\ n&:= \min _{1 \le i \le m } \eta _{p_i} =: n(\eta ,f)&\nonumber \\ [U]_{{p_i},[k]}&:= \texttt{E}\left( x^{[k]}, [t_{p_i}, t_{p_i} + h]\right) \in \mathbb {I}^d,&1 \le i \le m, \quad 0 \le k \le n \nonumber \\ [U]_{{p_i},[n+1]}&:= [\xi ]_{p_i}(x) \in \mathbb {I}^d&1 \le i \le m \end{aligned}
(28)
\begin{aligned}{}[U]_0&:= \texttt{roughEncl}({\hat{f}}, z(x), t_0, h) \in \mathbb {I}^d&\end{aligned}
(29)
\begin{aligned}{}[F]&:= F^{[n+1]}\left( [U]\right) ,&\end{aligned}
(30)

then, building the representation after the step h:

\begin{aligned} j_i(y)&= j_{i-1}(x) \ =: \ j_i(X_h)&i \in \{2,\dots ,p\} \\ [\xi ]_i(y)&= [\xi ]_{i-1}(x) \ =: \ [\xi ]_i(X_h)&i \in \{2,\dots ,p\} \\ j_1(y)&= \left( z(x), w_{n+1} * F^{[n]}\left( u(x)\right) \right) \ =: \ j_1(X_h)&\\ [\xi ]_1(y)&\subset \frac{1}{n+2} \cdot [F]_{[n+1]} \ =: \ [\xi ]_1(X_h),&\\ z(y)&\in \texttt{T}(j_1(y); h) + \left( [F]_{[n+1]} \cdot [0, h]\right) \cdot h^{n+1} \ =: z(X_h) \&\end{aligned}

Please note that we used in (29) symbol $$[U]_0$$ to denote enclosure of x over [0, h] (computed by the $$\texttt{roughEncl}$$ procedure). All other components of [U] are computed estimates on jets $$j_{p_i}(x)$$ over the same interval [0, h] using Proposition 3. That way, we can think of [U] as the enclosure of u over interval [0, h]. We have also generalized the algorithm to be valid for any $$C^\eta _p$$ by introducing the notion of $$n(\eta ,f)$$ in Eq. (28). The $$n(\eta ,f)$$ depends on f in the sense, the minimum is computed only for $$n_i$$ that are actually used in computations.

### 3.4 Steps Smaller than h

In this section, we consider computation of the (pn)-representations of $$\varphi (t, x_0)$$ where t is not necessary the multiple of the basic step size $$h = \frac{\tau }{p}$$, and for the initial $$x_0 \in C^{\eta }_{p}$$, where the apparent connection between $$\eta$$, n and t will be discussed soon. This problem arises naturally in the construction of Poincaré maps. Roughly speaking, the Poincaré map P for a (semi)flow in the phase space $$\mathcal {X}$$ is defined as $$P(x) = \varphi (t_P(x), x)$$, where $$x \in S \subset \mathcal {X}$$ and $$t_P: S \rightarrow (0, \infty )$$—the return time to the section S - is a continuous function such that $$\varphi (t_P(x), x) \in S$$ (we skip the detailed definition and refer to [34]). We see that the algorithm presented so far is insufficient for this task, as it can produce estimates only for discrete times $$t = i \cdot h$$, $$i \in \mathbb {N}$$, not for a possible continuum of values of $$t_P(S)$$. It is obvious that we can express $$t = m \cdot h + \varepsilon$$ with $$m \in \mathbb {N}$$ and $$0< \varepsilon < h$$ and the computation of $$\varphi (t, x_0)$$ can be realized as a composition $$\varphi (\varepsilon , \varphi (m \cdot h, x_0))$$. Therefore, we assume that the initial function is given as $$x_m = \varphi (m \cdot h, x)$$ and we focus on the algorithm to compute (estimates on) $$x_\varepsilon = \varphi (\varepsilon , x_m)$$.

First, we observe that, for a general $$x_m$$ in some $$(p, \eta )$$-fset, we cannot expect that $$x_\varepsilon \in C^{\zeta }_p$$ for any $$\zeta$$. The reason is that the solution x of DDE (2) with initial data in $$C^\eta _p$$ can be of class as low as $$C^0$$ at $$t=0$$, even when the r.h.s. and the initial data is smooth (as we have discussed in the beginning of Sect. 2). The discontinuity appears at $$t = 0$$ due to the very nature of Eq. (2). This discontinuity is located at $$s = -\varepsilon$$ in the segment $$x_\varepsilon$$ of the solution and, of course, we have $$-\varepsilon \in [-h, 0]$$. Therefore, the function $$x_\varepsilon$$ does not have any Taylor representation (in the sense of Definition 1) on the interval $$I_1 = [-h, 0]$$, as the first derivative of x is discontinuous there.

On the other hand, we are not working with a general initial function, but with $$x_m = \varphi (m \cdot h, x_0)$$, with $$x_0 \in C^\eta _p$$. From Lemma 9 we get that $$x_m \in C^{\eta + n + 1}_{p} \cap \mathcal {C}^{n + 1}$$, where $$n \in \mathbb {N}$$ be the largest value such that $$m \ge (n + 1) \cdot p$$. Moreover, the same is true for $$x_{m+1} = \varphi (h, x_m)$$. Therefore, $$x_\varepsilon = \varphi (\varepsilon , x_m) \in \mathcal {C}^{n+1}$$, so that it has a $$C^{n}_{p}$$ representation.

Now, the question is: can we estimate this (pn)-representation in terms of the coefficients of representations of $$x_m$$ (and maybe $$x_{m+1}$$)? The answer is positive, and we have:

### Lemma 16

Assume x is a solution to (2) with a segment $$x_0 \in C^{\eta }_{p} \cap \mathcal {C}^{0}$$. Let $$t \in \mathbb {R}$$ be given with $$t = m \cdot h + \varepsilon$$, $$m \in \mathbb {N}$$, $$0< \varepsilon < h$$. Let $$n = \lfloor \frac{m}{p}\rfloor - 1$$ and assume $$n \ge 0$$, i.e. $$m \ge p$$ and $$t \ge \tau$$.

Let denote $$x_m = \varphi (m \cdot h, x_0)$$ and $$x_{m+1} = \varphi (m \cdot h + h, x_0)$$ and for $$i \in \{1,\ldots , p\}$$ let

\begin{aligned}{}[L]_{i}&= \texttt{E}\left( j_i(x_m^{[n+1]}), [\xi ]_i(x_m^{[n+1]}), [0, h]\right) , \end{aligned}
(31)
\begin{aligned}{}[R]_{i}&= \texttt{E}\left( j_i(x_{m+1}^{[n+1]}), [\xi ]_i(x_{m+1}^{[n+1]}), [0, \varepsilon ]\right) . \end{aligned}
(32)

Then we have $$x_t \in X_\varepsilon \subset C^{n}_{p} \cap C^{n+1}$$ for $$X_\varepsilon$$ given by:

\begin{aligned} z\left( X_\varepsilon \right)&:= \texttt{T}\left( j_1(x_{m+1}); \varepsilon \right) + \texttt{S}\left( [\xi ]_1(x_{m+1}); \varepsilon \right) ,&\end{aligned}
(33)
\begin{aligned} j_{i,[k]}\left( X_\varepsilon \right)&:= \texttt{T}\left( j_i(x_m^{[k]}); \varepsilon \right) + \texttt{S}\left( [\xi ]_{i}(x_m^{[k]}); \varepsilon \right) ,&i \in \{1,\ldots , p\}, k \in \{0, \ldots , n\}, \end{aligned}
(34)
\begin{aligned}{}[\xi ]_i\left( X_\varepsilon \right)&:= \texttt{hull}\left( [L]_{i}, [R]_{i}\right) ,&i \in \{1,\ldots , p\}. \end{aligned}
(35)

Before the proof, we would like to make a small comment. The representation of $$x_{m+1}$$ is used for optimization and simplification purposes, as usually we have it computed nevertheless (when finding the crossing time of the Poincaré map). It contains the representation of x over $$[mh, mh + h)$$ in $$j_1$$. Otherwise we would need to expand the jet of solution x at $$t = 0$$ to compute $$[R]_1$$ and z in (33). Also, the formula (32) would be less compact.

### Proof of Lemma 16

It is a matter of simple calculation. To focus the attention on the $$\varepsilon$$ step, let us abuse notation and denote $$x_t = x_\varepsilon = \varphi (\varepsilon , x_m)$$. We have

\begin{aligned} x^{[k]}_\varepsilon (-i \cdot h) = x^{[k]}_m(-i \cdot h + \varepsilon ), \quad i \in \{1, \ldots , p \} \end{aligned}

so we get a straightforward formula:

\begin{aligned} j_i(x_\varepsilon )_{[k]}&= \texttt{E}\left( x_m^{[k]}; -i \cdot h + \varepsilon \right) \nonumber \\ {}&= \ \texttt{T}^{\eta _i + n + 1 - k}\left( j_i(x^{[k]}_m); \varepsilon \right) + \texttt{S}^{\eta _i + n + 1 - k}\left( \xi _i(x_m^{[k]}); \varepsilon \right) \end{aligned}
(36)

where representations of $$x^{[k]}_{m}$$ are obtained by applying Proposition 12. Similarly, one can find that

\begin{aligned} z(x_\varepsilon ) = x_\varepsilon (0) = x_{m+1}(-h + \varepsilon ) = \ \texttt{T}(j_1(x_{m+1}); \varepsilon ) + \texttt{S}([\xi ]_1(x_{m+1}); \varepsilon ), \end{aligned}
(37)

and for $$s \in [0, h)$$, $$i \in \{1, \ldots , p \}$$:

\begin{aligned}{} & {} \xi _i(x_\varepsilon )(s) = x^{[n+1]}_\varepsilon (-i \cdot h + s) \nonumber \\{} & {} = {\left\{ \begin{array}{ll} x^{[n+1]}_m(-i \cdot h + \varepsilon + s) = \texttt{E}\left( x^{[n+1]}_m; \varepsilon + s\right) &{} \varepsilon + s < h \\ x^{[n+1]}_{m+1}(-i \cdot h + (\varepsilon + s - h)) = \texttt{E}\left( x^{[n+1]}_{m+1}; (\varepsilon + s - h)\right) &{} \varepsilon + s \ge h \end{array}\right. }.\nonumber \\ \end{aligned}
(38)

Note, in the second case of Eq. (38), we have $$0 \le (\varepsilon + s - h) < h$$. Now, we exchange each $$\xi$$ with $$[\xi ]$$ in Eqs. (36)–(38) to get the corresponding estimates in Eqs. (33)–(35). $$\square$$

This algorithm is valid for any number of dimensions and for any number of delays (i.e. for any definition of used variables u(nf))—in fact, there is no explicit dependence on the r.h.s of (2) in the formulas—the dynamics is “hidden” implicitly in the already computed jets $$j_{i}(x_m)$$ and $$j_1(x_{m+1})$$. This form of the algorithm will allow in the future to make general improvements to the method, without depending on the actual formula for the projection of used variables u(nf) in the r.h.s. of DDE (1), or even when constructing methods for other forms of functional differential equations. We will denote the $$\varepsilon$$ step algorithm given by (33)–(35) by $$\mathcal {I}_\varepsilon$$.

As a last remark, similarly to the discussion in the last paragraph of Sect. 3.2, let us consider the order of the local error in the method $$\mathcal {I}_\varepsilon$$. This local error will have a tremendous impact on the computation of Poincaré maps, and thus on the quality of estimates in computer-assisted proofs. To see why, set the order n and let us consider two maps: $$T = \varphi (m h, \cdot )$$ and $$T_\varepsilon = \varphi (m h + \varepsilon , \cdot )$$, where, without loss of generality, we choose $$m = p \cdot (n+1)$$ (in applications, return time in Poincaré maps will be required to be greater than this) and we fix some $$0< \varepsilon < h$$. It is of course sufficient to use full-step method $$\mathcal {I}$$ to rigorously compute map T, while $$T_\varepsilon$$ is a good model of computing estimates on a real Poincaré Map and will require usage of $$\mathcal {I}_\varepsilon$$ in the last step. Let us denote $$x_m = T(x_0)$$, $$x_{m+1} = \varphi (h, T(x_0))$$ and $$x_\varepsilon = T_{\varepsilon }(x_0)$$. Obviously, we have $$x_{m+1} = \varphi (h, x_m) \in \mathcal {I}(x_m)$$ and $$x_\varepsilon = \varphi (\varepsilon , x_m) \in \mathcal {I}_\varepsilon (x_m)$$ Assume $$x_0 \in C^\eta _p$$ with uniform order on all grid points, $$\eta = n$$. From Lemma 9 for both maps, we end up with $$x_m \in C^{2n + 1}_{p} \cap \mathcal {C}^{n+1}$$, $$x_{m+1} \in C^{2n + 2}_{p,1} \cap \mathcal {C}^{n+1}$$ and $$x_\varepsilon \in C^{n+1}_{p} \cap \mathcal {C}^{n+1}$$. From discussion in the last paragraph of Sect. 3.2,we can infer that the local error introduced in $$\mathcal {I}(x_m)$$ is of order $$O(h^{2n + 2})$$, as the only term with nonzero Taylor remainder is $$z(\mathcal {I}_\varepsilon (x_m))$$. Therefore, we can expect that the accumulated error of estimating map T with $$\mathcal {I}^m$$ (m steps of the full-step integrator $$\mathcal {I}$$) is of order $$O(h^n)$$ [9], as this is the accumulated error of covering the first delay interval $$[0, \tau )$$ in the beginning of the integration process. Later, thanks to smoothing of solutions and expanded space, the subsequent errors would be of higher order. This in general should apply even if we do not expand the representation, as in such case the local error in each step (even after $$[0, \tau ]$$) in $$\mathcal {I}$$ is still just $$O(h^{n+1})$$.

In comparison, algorithm $$I_\varepsilon$$ evaluates Taylor expansion with nonzero remainder not only at $$z(\cdot )$$ in (33), but also at every grid point and every coefficient order of the representation in (34). What is more, the impact of the remainder term $$[\xi ]$$ is of different order at different Taylor coefficients. Here, we use Proposition 12 to get that kth Taylor coefficient $$x_m^{[k]}$$ has a (p,l)-representation with $$l = 2n + 1 - k$$, so the local error of $$j_{i,{[k]}}(X_\varepsilon )$$ is of order $$O(h^{2n + 1 - k})$$. Since $$k \in \{0, \ldots , n\}$$, then in the worst case of $$k = n$$, the local error size is $$O(h^{n+1})$$. This is of course worse than $$O(h^{2n + 2})$$ of the full-step method, but it is a significant improvement over the first version of the algorithm presented in [34], where the local error of the last $$\varepsilon$$ step was O(h) (basically, because $$x^{[n]}_\varepsilon$$ was computed by explicit Euler method in the non-expanded representation of $$x_m \in C^n_p$$). Current error is of the order comparable to the accumulated error over the course of a long-time integration $$\mathcal {I}^m$$ and therefore has a lot less impact on the resulting estimates.

Exemplary computations, supporting the above discussion, are presented in Sect. 3.7.

### 3.5 Computation of Poincare Maps

In this section, we would like to discuss shortly some minor changes to the algorithm of computing images of Poincaré map using algorithms $$\mathcal {I}$$ (full step h) and $$\mathcal {I}_\varepsilon$$ ($$\varepsilon < h$$), particularly, we discuss the case when the estimate on $$t_P(S)$$ has diameter bigger than h—this will be important in one of the application discussed in this paper.

In the context of using rigorously computed images of Poincaré maps in computer-assisted proofs in DDEs, we will usually do the following (for details, see [34]):

1. 1.

We choose subspace of the phase space of the semiflow $$\varphi$$ as $$C^n_p \cap \mathcal {C}^0$$ with p, n fixed.

2. 2.

We choose sections $$S_1, S_2 \subset C^{n}_{p}$$, usually as some hyperplanes $$S_i = \{ x \in C^{n}_{p}: S_i(x):= (s_i. a(x)) - c_i = 0 \}$$, with $$s_i \in \mathbb {R}^{M(d, p, n)}$$, $$c \in \mathbb {R}$$ and (.) denoting the standard scalar product in $$\mathbb {R}^{M(d, p, n)}$$ (we remind $$a(x) = (z(x), j(x)) \in \mathbb {R}^{M(d,p,n)}$$, $$M(d,p,n) = d \cdot (1 + (n+1) \cdot p)$$). Of course, in the simplest case, we can work only with a single section, $$S_1 = S_2$$.

3. 3.

We choose some initial, closed and convex set $$X_0 \subset S_1 \subset C^n_p$$ on the section $$S_1$$.

4. 4.

We construct $$[t] \in \mathbb {I}$$ such that $$t_P(X_0) \subset [t]$$, where $$t_P: X_0 \rightarrow \mathbb {R}_+$$ is the return time function from $$X_0$$ to $$S_2$$, so that $$\varphi (t_P(x_0), x_0) \in S_2 \subset C^n_p$$ for all $$x_0 \in X_0$$. This is done usually alongside the computation of the image $$P(X_0)$$, by successive iterating $$X_{j+1} = \mathcal {I}(X_{j})$$ until $$X_m$$ is before and $$X_{m+1}$$ is after the section $$S_2$$ (i.e. $$S_2(X_m) < 0$$ and $$S_2(X_{m+1}) > 0$$ or $$S_2(X_m) > 0$$ and $$S_2(X_{m+1}) < 0$$). In such a case, $$[t] = m \cdot h + [\varepsilon ]$$, where $$[\varepsilon ] \subset [0, h)$$.

In view of Lemma 16, we require $$t_P(X_0) \ge (n+1) \cdot \tau$$—the return time to the section is long enough. Moreover, $$X_m$$ and $$X_{m+1}$$ are already computed to be used in the formulas (33)–(35). The tight estimates on [t] can be obtained for example with the binary search algorithm, in the same manner as it was done in [34].

Finally, using formulas from Lemma 16 we get $$X_\varepsilon \subset C^n_p$$ such that $$\varphi ([\varepsilon ], X_m) \subset X_\varepsilon$$.

5. 5.

We use sets $$X_0$$ and $$X_\varepsilon$$ together with the estimates on $$P(X_0) \subset \varphi ([t], X_0)$$ to draw conclusion on existence of some interesting dynamics. For example, if $$S_1 = S_2$$ and $$P(X_0) \subset X_0$$ we can use Schauder fixed-point theorem to show existence of a periodic point of P (the compactness of the operator P plays here a crucial role).

Now, we have already mentioned that the computation of the Poincaré map $$P(x_0) = \varphi (t_P(x_0), x_0)$$ can be done by splitting the return time $$t_P(x_0) = m(x_0) \cdot h + \varepsilon (x_0)$$ with $$m(x_0) \in \mathbb {N}$$ and $$\varepsilon (x_0) \in (0, h)$$. This leads to a rough idea of rigorous algorithm to compute estimates on $$P(x_0)$$ in the following form:

\begin{aligned} P(x_0) \in \mathcal {I}_{\varepsilon (x_0)} \circ \mathcal {I}^{m(x_0)} \left( x_0\right) . \end{aligned}
(39)

However, in the case of computing (estimates on) $$P(X_0)$$ for a whole set $$X_0 \subset C^n_p$$, we can face the following problem: for $$x, y \in X_0$$ we can have $$m(x) \ne m(y)$$, especially when $$X_0$$ is large. In [34], we have simply chosen $$X_0$$ so small, such that m(x) is constant in $$X_0$$. Then, we have $$[t] = m \cdot h + [\varepsilon ]$$, with $$[\varepsilon ] = [\varepsilon _1, \varepsilon _2]$$, $$0< \varepsilon _1 \le \varepsilon _2 < h$$. In such a situation, formula (39) could be applied with $$m(x) = m$$ and $$\varepsilon = [\varepsilon ]$$. In the current work, we propose to take the advantage of all the data already stored in the $$(p,\eta )$$-fsets and to extend the algorithm in Lemma 16 to produce rigorous estimates on $$\varphi ([\varepsilon ], x_0)$$ for $$[\varepsilon ] = [\varepsilon ] = [\varepsilon _1, {\bar{m} \cdot h} + \varepsilon _2]$$, $$0< \varepsilon _i < h$$ $${\bar{m}} \in \mathbb {N}$$. It is not difficult to see that we have the following:

### Proposition 17

Let $$[t] = m \cdot h + [\varepsilon _1, {\bar{m}} \cdot h + \varepsilon _2]$$ with $$0< \varepsilon _1, \varepsilon _2 < h$$, $$m, {\bar{m}} \in \mathbb {N}$$ with $${\bar{m}} > 0$$. Let assume $$X_j$$ are such that $$\varphi (j \cdot h, X_0) \subset X_M$$ for $$j = m, m+1, \ldots , m+{\bar{m}}+1$$. Finally, let n be as in Lemma 16.

We define ($$k \in \{ 0, \ldots , n+1 \}$$, $$j \in \{0, \ldots , {\bar{m}} + 1\}$$, $$i \in \{ 1, \ldots , p\}$$):

\begin{aligned}{}[L]^{[k]}_{i,j}&= \texttt{E}\left( j_i(x_{m+j}^{[k]}), [\xi ]_i(x_{m+j}^{[k]}), [\varepsilon _1, h]\right) , \\ [C]^{[k]}_{j,i}&= \texttt{E}\left( j_i(x_{m+j}^{[k]}), [\xi ]_i(x_{m+j}^{[k]}), [0, h]\right) , \\ [R]^{[k]}_{i,j}&= \texttt{E}\left( j_i(x_{m+j}^{[k]}), [\xi ]_i(x_{m+j}^{[k]}), [0, \varepsilon _2]\right) , \end{aligned}

and a set $$X_\varepsilon$$ given by:

\begin{aligned} z(X_\varepsilon )&:= \texttt{hull}\left( [L]^{[0]}_{1,1}, [C]^{[0]}_{1,2}, \ldots , [C]^{[0]}_{1,{\bar{m}}}, [R]^{[0]}_{1,{\bar{m}} + 1}\right) ,&\end{aligned}
(40)
\begin{aligned} j_{i,[k]}(X_\varepsilon )&:= \texttt{hull}\left( [L]^{[k]}_{i,0}, [C]^{[k]}_{i,1}, \ldots , [C]^{[k]}_{i,{\bar{m}}-1}, [R]^{[k]}_{i,{\bar{m}}} \right) , \nonumber \\ {}&i \in \{1,\ldots , p\}, k \in \{0, \ldots , n\}, \end{aligned}
(41)
\begin{aligned} _i(X_\varepsilon )&:= \texttt{hull}\left( [L]^{[n+1]}_{i,0}, [C]^{[n+1]}_{i,1}, \ldots , [C]^{[n+1]}_{i,{\bar{m}}}, [R]^{[n+1]}_{i,{\bar{m}}+1}\right) , \nonumber \\ {}&i \in \{1,\ldots , p\}. \end{aligned}
(42)

Then for all $$t \in [t]$$ we have $$x_t \in X_\varepsilon \subset C^{n}_{p} \cap C^{n+1}$$.

Of course, in the case $$m(X_0) = const$$ we use algorithm from Lemma 16.

### 3.6 The Lohner-Type Control of the Wrapping Effect

An important aspect of the rigorous methods using interval arithmetic is an effective control of the wrapping effect. The wrapping effect occur in interval numerics, when the result of some nonlinear operation or map needs to be enclosed in an interval box. When this box is chosen naively, then a huge overestimates may occur, see Fig. 6 in “Appendix A”.

To control wrapping effect in our computations, we employ the Lohner algorithm [22] by representing sets in a good local coordinate frame: $$X = x_0 + C \cdot r + E$$, where $$x_0$$ is a vector in $$\mathbb {R}^M$$, $$C \in \mathcal {M}(M, N)$$, $$r_0 \in \mathbb {I}^N$$—an interval box centred at 0, and E some representation of local error terms. As it was shown in [34], taking $$E \in \mathbb {I}^M$$ (an interval form of the error term) was enough to prove existence of periodic orbits. Moreover, taking into account the form of the algorithm given by (21)–(25) (especially the shift part (21)–(22)) to properly reorganize computations was shown to be crucial to obtain an algorithm of optimal computational complexity.

In this work, we not only adopt this optimized Lohner algorithm to the systems of equations and to many delays, but we also propose another form of the error term E to get better estimates on the solutions in case of systems of equations, $$d > 1$$, much in the same way it is done for systems of ODEs [12, 22]. The proposed algorithm does not sacrifice the computational complexity to obtain better estimates. We use this modified algorithm in our proof of the symbolic dynamics in a delay-perturbed Rössler system.

The details of the algorithm are highly technical, so we decided to put them in “Appendix A”, to not overshadow the presentation of the theoretical aspects, but on the other hand to be accessible for people interested in actual implementation details and/or in re-implementing presented methods on their own.

### 3.7 Benchmarks

As the last remark in this section, we present the numerical experiment showing the effect of using the new algorithm with expanding representation in comparison with the old algorithm in [34]. As a test, we use a constant initial function $$x_0(t) = 1.1$$ for $$t \in [-\tau , 0]$$ and the Mackey–Glass equation with parameter values $$\beta = 2$$, $$\gamma = 1$$, $$n = 8$$ and $$\tau = 2$$. The configuration of (dpn)-fset $$X_0$$ has $$n = 4$$ (order 4 method), $$p = 128$$, $$d = 1$$ (scalar equation). The initial diameter of the set $$X_0$$ is 0. The test does integration over the 3n full delays (so that the final solution is smoothed enough). Then, an $$\varepsilon$$-step is made, with the step $$\varepsilon = \frac{h}{2}$$, where $$h = \frac{\tau }{p}$$ is the grid size (full step). In Table 1, we present the maxima over all diameters of the coefficients of the sets: $$X_{3n} = \mathcal {I}^{3n}(X_0)$$ that contains the segment $$x_{3n}$$ of the solution, and $$X_{3n + \varepsilon } = \mathcal {I}_{\varepsilon } \left( \mathcal {I}^{3n}(X_0)\right)$$. We remind that $$\mathcal {I}$$ denotes the full-step integrator method that does one step of size h, while $$\mathcal {I}_{\varepsilon }$$ is the $$\varepsilon$$-step method. Each maximum diameter is computed over all Taylor coefficients of a given order $$0 \le k \le 4$$. We also show the maximum diameter of the $$\Xi$$ part (order $$k = 5$$).

We test several maximal orders of the expanded representations: 2n, $$2n+1$$ and 3n. The last one is the maximal order obtainable with the 3n full-delay integration steps, while the first one is the minimal reasonable one—taking into account the long enough integration time, see Definition 3 and Lemma 8.

### Remark 18

Using the diameter 0 of the set $$X_0$$ in the test will show how the local errors of the method at each step affect the final outcome.

From Table 1, we see that the diameters of the sets integrated with the new algorithm are far superior to the old one. One can observe in (a) that for the fixed number of full steps both methods produce results with coefficients of all orders of a comparable diameter. This indicates that both methods are of order $$h^4$$. However, new algorithm produces estimates of three orders of magnitude better. This is because internally, the algorithm becomes of higher order after each full delay. After k full delays, the actual order of the method is $$n + k$$. The second big advantage is shown in the (b) part, where we have diameters of coefficients after a small $$\varepsilon$$ step. This simulates for example computation of a Poincaré map. The old algorithm produces estimates that depend on the order of coefficient: the coefficient 0 has a diameter proportional to $$h^n$$; however, other coefficients are computed with worse accuracy. The fourth-order coefficient is computed with the lowest accuracy of order $$h^1$$. On the contrary, the new algorithm still retains the accuracy of the full-step size algorithm and produce far superior estimates (several orders of magnitude better).

The data and programs used in those computations are described more in detail in “Appendix B”.

## 4 Topological Tools

In [34], we have proven the existence of periodic orbits (apparently stable) using the Schauder fixed-point theorem. Here, we are interested in a more general way to prove existence of particular solutions to DDEs with the use of Poincaré maps generated with semiflow $$\varphi$$ of (2). For this, we will recall the concept of covering relations from [5], but we will adopt it to the setting of infinite-dimensional spaces and compact mappings, similarly to a recent work [39]. The main theoretical tool to prove the existence of solutions, in particular the fixed points of continuous and compact maps in $$C^n_p$$, will be the Leray–Schauder degree, which is an extension of fixed-point index (i.e. the local Brouwer degree of $$Id - F$$) to infinite-dimensional Banach spaces. We only recall the properties of the degree that are relevant to our applications. For a broader description of the topic together with the proofs of presented theorems, we point out to [2, 8] and references therein. In particular, in what follows, we will use the notion of absolute neighbourhood retract (ANR) [8]. We do not introduce the formal definition but we only note that (1) any Banach space is ANR and (2) any convex, closed subset of a Banach space (or a finite sum of such) is an ANR (Corollaries 4.4 and 5.4 in Sect. 11 of [8], respectively).

### 4.1 Fixed-Point Index for Compact Maps in ANRs

Let $$\mathcal {X}$$ be a Banach space. We recall that a continuous function $$f: \mathcal {X}\supset V \rightarrow \mathcal {X}$$ is a compact map iff $$\overline{f(V)}$$ is compact in $$\mathcal {X}$$. With $$Fix(f, U) = \{ x \in U: f(x) = x \}$$ we denote the set of fixed points of f in U. Let now X be an ANR [8], in particular $$\mathcal {X}$$ can be X, and let U be open subset of X, $$f: {\overline{U}} \rightarrow X$$. Following [8], by $$\mathcal {K}\left( {\overline{U}}, X\right)$$ we denote the set of all compact maps $${\overline{U}} \rightarrow X$$, and by $$\mathcal {K}_{\delta U}\left( {\overline{U}}, X\right)$$ the set of all maps $$f \in \mathcal {K}({\overline{U}}, X)$$ that have no fixed points on $$\delta U$$, $$Fix(f, \delta U) = \emptyset$$. We will denote $$Fix(f) = Fix(f, U) = Fix(f, {\overline{U}})$$. Let $$V \subset \mathcal {X}$$ be any set in the Banach space $$\mathcal {X}$$. We say that a map is admissible in V iff Fix(fV) is a compact set. The following stronger assumption that implies admissibility is often used in applications:

### Lemma 19

Let $$\mathcal {X}$$ be a Banach space (can be infinite dimensional) and $$U \subset \mathcal {X}$$ be an open set. Assume $$f: {\overline{U}} \rightarrow \mathcal {X}$$ is a continuous, compact map. If $$f(x) \ne x$$ for all $$x \in \delta U$$ then f is admissible.

### Proof

Let $$F = (Id - f)^{-1}(\{0\})$$ be the set of fixed points of f. By assumption $$f(x) \ne x$$ on $$\delta U$$, we have $$F \cap \delta U = \emptyset$$ so $$F \cap {\bar{U}} = F \cap U$$. The set F is closed as a preimage of the closed set $$\{0\}$$ under continuous function $$Id - f$$, and so is $$F \cap {\overline{U}}$$. Therefore, $$F \cap U$$ is closed and thus compact as a subset of a compact set $$\overline{f({\overline{U}})}$$: $$F \cap U = F \cap {\overline{U}} = f(F \cap {\overline{U}}) \subset \overline{f({\overline{U}})}$$. $$\square$$

By Lemma 19, we see that all functions $$f \in \mathcal {K}_{\delta U}({\overline{U}}, X)$$ are admissible, so that the fixed-point index is well defined on them [8]:

### Theorem 20

(Theorem 6.2 in [8]) Let X be an ANR. Then, there exists an integer-valued fixed-point index function $$\iota (f, U) \in \mathbb {Z}$$ (Leray–Schauder degree of $$Id - f$$) which is defined for all $$U \subset X$$ open and all $$f \in \mathcal {K}_{\delta U}\left( {\overline{U}}, X\right)$$ with the following properties:

1. (I)

(Normalization) If f is constant $$f(x) = x_0$$ then, $$\iota (f, U) = 1$$ iff $$x_0 \in U$$ and $$\iota (f, U) = 0$$ iff $$x_0 \notin U$$.

2. (II)

(Additivity) If $$Fix(f) \subset U_1 \cup U_2 \subset U$$ with $$U_1, U_2$$ open and $$U_1 \cap U_2 = \emptyset$$, then $$\iota (f,U) = \iota (f,U_1) + \iota (f,U_2)$$.

3. (III)

(Homotopy) If $$H: [0, 1] \times {\overline{U}} \rightarrow X$$ is an admissible compact homotopy, i.e. H is continuous, $$H_t = H(t, \cdot )$$ is compact and admissible for all t, then $$\iota (H_t) = \iota (H_0)$$ for all $$t \in [0, 1]$$.

4. (IV)

(Existence) If $$\iota (f, U) \ne 0$$ then $$Fix(f) \ne \emptyset$$.

5. (V)

(Excision) If $$V \subset U$$ is open, and f has no fixed points in $$U {\setminus } V$$ then $$\iota (f, U) = \iota (f, U {\setminus } V)$$.

6. (VI)

(Multiplicativity) Assume $$f_i: \mathcal {X}_i \supset X_i \supset {\overline{U}}_i \rightarrow X_i$$, $$i = 1, 2$$ are admissible compact maps, and define $$f(x_1, x_2) = (f_1(x_1), f_2(x_2)) \in X_1 \times X_2$$ for $$(x_1, x_2) \in U:= {\overline{U}}_1 \times {\overline{U}}_2$$. Then f is a continuous, compact and admissible map with $$\iota (f, U) = \iota (f_1, U_1) \cdot \iota (f_2, U_2)$$.

7. (VII)

(Commutativity) Let $$U_i \subset X_i \subset \mathcal {X}_i$$, for $$i =1, 2$$ be open and assume $$f_i: U_1 \rightarrow X_2$$, $$g: U_2 \rightarrow \mathcal {X}_1$$ and at least one of the maps fg is compact. Define $$V_1 = U_1 \cap f^{-1}(U_2)$$ and $$V_2 = U_2 \cap g^{-1}(U_1)$$, so that we have maps $$g \circ f: \overline{V_1} \rightarrow X_1$$ and $$f \circ g: \overline{V_2} \rightarrow X_2$$.

Then $$f \circ g$$ and $$g \circ f$$ are compact and if $$Fix(g \circ f) \subset V_1$$ and $$Fix(f \circ g) \subset V_2$$ then

\begin{aligned} \iota (g \circ f, V_1) = \iota (f \circ g, V_2). \end{aligned}

For us, the key and the mostly used properties are the Existence, Homotopy and Multiplicativity properties. First one states that, if the fixed-point index is nonzero, then there must be a solution to the fixed-point problem $$f(x) = x$$ in the given set. The homotopy allows to relate the fixed-point index $$\iota (f, U)$$ to some other, usually easier and better understood map, for example $$\iota (A, U)$$, where A is some linear function in finite-dimensional space. Normalization and Multiplicativity are used to compute the fixed point index in the infinite-dimensional “tail part”.

The following is a well-known fact:

### Lemma 21

Let $$A: \mathbb {R}^n \rightarrow \mathbb {R}^n$$ be a linear map. Then for any $$U \subset \mathbb {R}^n$$:

\begin{aligned} \iota (A, U) = \textrm{sgn}\left( \det \left( Id - A\right) \right) . \end{aligned}
(43)

Applying Commutativity property to $$f = F \circ h^{-1}$$ and $$g = h$$ gives:

### Lemma 22

Let $$F: U \rightarrow \mathcal {X}$$ be admissible, continuous, compact map and let $$h: \mathcal {X}\rightarrow \mathcal {X}'$$ be a homeomorphism. Then $$h \circ F \circ h^{-1}: \mathcal {X}' \supset h(U) = V \rightarrow \mathcal {X}'$$ is admissible, and

\begin{aligned} \iota (F, U) = \iota (h \circ F \circ h^{-1}, V). \end{aligned}

### 4.2 Covering Relations in $$\mathbb {R}^d$$

In our application, we will apply the fixed-point index to detect periodic orbits of some Poincaré maps $$P: C^{n}_p \supset U \rightarrow C^{n}_p$$. We will introduce a concept of covering relations. A covering relation is a way to describe that a given map f stretches in a proper way one set over another. This notion was formalized in [5] for finite-dimensional spaces and recently extended to infinite spaces in [39] in the case of mappings between compact sets. In the sequel we will modify this slightly for compact mappings between (not compact) sets in the $$C^n_p$$ spaces.

To set the context and show possible applications, we start with the basic definitions from [5] in finite-dimensional space $$\mathbb {R}^d$$, and then we will move to extend the theory in case of $$C^n_p$$ spaces later in this section.

### Definition 7

(Definition 1 in [5]) An h-set N in $$\mathbb {R}^{d_N}$$ is an object consisting of the following data:

• |N|—a compact subset of $$\mathbb {R}^{d_N}$$;

• $$u_N, s_N \in \mathbb {N}$$ such that $$u_N + s_N = d_N$$;

• a homeomorphism $$c_N: \mathbb {R}^{d_N} \rightarrow \mathbb {R}^{d_N} = \mathbb {R}^{u_N} \times \mathbb {R}^{s_N}$$ such that

\begin{aligned} c_N(|N|) = \overline{{{\,\mathrm{\textbf{B}}\,}}_{u_N}}(0,1) \times \overline{{{\,\mathrm{\textbf{B}}\,}}_{s_N}}(0,1). \end{aligned}

We set:

\begin{aligned} N_c= & {} \overline{{{\,\mathrm{\textbf{B}}\,}}_{u_N}}(0,1) \times \overline{{{\,\mathrm{\textbf{B}}\,}}_{s_N}}(0,1) \\ N^-_c= & {} \delta \overline{{{\,\mathrm{\textbf{B}}\,}}_{u_N}}(0,1) \times \overline{{{\,\mathrm{\textbf{B}}\,}}_{s_N}}(0,1) \\ N^+_c= & {} \overline{{{\,\mathrm{\textbf{B}}\,}}_{u_N}}(0,1) \times \delta \overline{{{\,\mathrm{\textbf{B}}\,}}_{s_N}}(0,1) \\ N^-= & {} c_N^{-1}(N^-_c) \\ N^+= & {} c_N^{-1}(N^+_c). \end{aligned}

In another words, h-set N is a product of two closed balls in an appropriate coordinate system. The numbers $$u_N$$ and $$s_N$$ stands for the dimensions of exit (nominally unstable) and entry (nominally stable) directions. We will usually drop the bars from the support |N| of the h-set, and use just N (e.g. we will write f(N) instead of f(|N|).

The h-sets are just a way to organize the structure of a support into nominally stable and unstable directions and to give a way to express the exit set $$N^-$$ and the entry set $$N^+$$. There is no dynamics here yet—until we introduce some maps that stretch the h-sets across each other in a proper way.

### Definition 8

(Definition 2 in [5]) Assume N, M are h-sets, such that $$u_N = u_M = u$$. Let $$P: |N| \rightarrow \mathbb {R}^{d_M}$$ a continuous map. We say that N P-covers M, denoted by:

\begin{aligned} N {\mathop {\Longrightarrow }\limits ^{P}} M \end{aligned}

iff there exists continuous homotopy $$H: [0,1] \times |N| \rightarrow R^{d_M}$$ satisfying the following conditions:

• $$H(0, \cdot ) = P$$;

• $$h([0,1], N^-) \cap M = \emptyset$$;

• $$h([0,1], N) \cap M^+ = \emptyset$$;

• there exists a linear map $$A: \mathbb {R}^{u} \rightarrow \mathbb {R}^{u}$$ such that

\begin{aligned} H_c(1, (p, q))= & {} (A p, 0) \\ A(\delta {{\,\mathrm{\textbf{B}}\,}}_u(0,1))\subset & {} \mathbb {R}^u \setminus \overline{{{\,\mathrm{\textbf{B}}\,}}_u}(0,1) \end{aligned}

where $$H_c(t, \cdot ) = c_M \circ H(t, \cdot ) \circ c_N^{-1}$$ is the homotopy expressed in good coordinates.

A basic theorem about covering relations is as follows:

### Theorem 23

(Simplified version of Theorem 4 in [5]) Let $$X_i \subset \mathbb {R}^d$$ be h-sets and let

\begin{aligned} X_1 {\mathop {\Longrightarrow }\limits ^{P_1}} X_2 {\mathop {\Longrightarrow }\limits ^{P_2}} \ldots {\mathop {\Longrightarrow }\limits ^{P_k}} X_{k+1} = X_1 \end{aligned}

be a covering relations chain. Then there exists $$x \in X_1$$ such that

\begin{aligned} x\in & {} X_1 \\ (P_{r-1} \circ \ldots \circ P_{1}) (x)\in & {} X_r \quad \text { for } 2 \le r \le k, \\ (P_{k} \circ \ldots \circ P_{1}) (x)= & {} x. \end{aligned}

Before we move on, we would like to point out what results can be obtained using Theorem 23:

• Example 1. Let $$X {\mathop {\Longrightarrow }\limits ^{P}} X$$, where X is some h-set on a section $$S \subset \mathbb {R}^d$$ and P is a Poincare map $$S \rightarrow S$$ induced by the local flow $$\varphi$$ of some ODE $$x' = f(x)$$. Then, there exists a periodic solution x to this ODE, with initial value $$x_0 \in X$$. The parameter $$u_X$$ gives the number of apparently unstable directions for P at x.

• Example 2. Let $$X_1$$ and $$X_2$$ be h-sets on a common section $$S \subset \mathbb {R}^d$$, $$X_1 \cap X_2 = \emptyset$$, and assume $$X_i {\mathop {\Longrightarrow }\limits ^{P}} X_j$$ for all $$i, j \in \{1, 2\}$$ where again P is a Poincaré map $$S \rightarrow S$$ induced by the semiflow $$\varphi$$ of some ODE. Then, this ODE is chaotic in the sense that there exists a countable many periodic solutions of arbitrary basic period that visits $$X_1$$ and $$X_2$$ in any prescribed order. Also, there exist non-periodic trajectories with the same property, see for example [5, 42].

In what follows, we will show the same construction can be done under some additional assumptions in the infinite-dimensional spaces.

### 4.3 Covering Relations in Infinite-Dimensional Spaces

The crucial tool in proving Theorem 23 is the fixed-point index in finite-dimensional spaces. Therefore, similar results are expected to be valid for maps and sets for which the infinite-dimensional analogue, namely Leray–Schauder degree of $$Id - f$$, exists. This was used in [39] for maps on compact sets in infinite-dimensional spaces. In this work, we do not assume sets are compact, but we use the assumption that the maps are compact—the reasoning is almost the same. We will work on spaces $$\mathcal {X}= \mathcal {X}_1 \oplus \mathcal {X}_2$$, where $$\mathcal {X}_1$$ is finite dimensional (i.e. $$\mathcal {X}_1 \equiv \mathbb {R}^M$$) and $$\mathcal {X}_2$$ will be infinite dimensional (sometimes refereed to as the tail). In our applications, we will set $$\mathcal {X}= C^n_p = \mathbb {R}^{M(d,p,n)} \times (C^0([0,h], \mathbb {R}^d))^{d \cdot p}$$, with $$\mathcal {X}_1 = \mathbb {R}^{M(d,p,n)}$$. We will use the following definitions that are slight modifications of similar concepts from [39], where the tail was assumed to be a compact set.

### Definition 9

Let $$\mathcal {X}$$ be a real Banach space. An h-set with tail is a pair $$N = (N_1, |N_2|)$$ where

• $$N_1$$ is an h-set in $$\mathcal {X}_1$$,

• $$|N_2| \subset \mathcal {X}_2$$ is a closed, convex and bounded set.

Additionally, we set $$u_N = u_{N_1}$$, $$|N| = |N_1| \times |N_2|$$, $$c_N = (c_{N_1}, Id)$$ and

\begin{aligned} N_c= & {} c^{-1}_N\left( |N|\right) \ = \ N_{1,c} \times |N_2| \ = \\= & {} \overline{{{\,\mathrm{\textbf{B}}\,}}_{u_{N_1}}}(0,1) \times \overline{{{\,\mathrm{\textbf{B}}\,}}_{s_{N_1}}}(0,1) \times |N_2|. \end{aligned}

The tail in the definition refers to the part $$|N_2|$$. We will just say that N is an h-set when context is clear. Please note that each h-set N in $$\mathbb {R}^d$$ can be viewed as an h-set with tail, where the tail is set as the trivial space $$\mathbb {R}^{0} = \{ 0 \}$$.

### Definition 10

Let $$\mathcal {X}$$ be as in Definition 9. Let N, M be h-sets with tails in $$\mathcal {X}$$ such that $$u_N = u_M = u$$. Let $$P: N \rightarrow \mathcal {X}$$ be a continuous and compact mapping in $$\mathcal {X}$$.

We say that N P-covers M (denoted as before in Definition 8 by $$N {\mathop {\Longrightarrow }\limits ^{P}} M$$), iff there exists continuous and compact homotopy $$H: [0,1] \times |N| \rightarrow \mathcal {X}$$ satisfying the conditions:

• (C0) $$H\left( t, |N|\right) \subset \mathbb {R}^{d_{M_1}} \times |M_2|$$;

• (C1) $$H\left( 0, \cdot \right) = P$$;

• (C2) $$H\left( \left[ 0,1\right] , N_1^- \times |N_2|\right) \cap M = \emptyset$$;

• (C3) $$H\left( \left[ 0,1\right] , |N|\right) \cap \left( M_1^+ \times |M_2|\right) = \emptyset$$;

• (C4) there exists a linear map $$A: \mathbb {R}^{u} \rightarrow \mathbb {R}^{u}$$ and a point $${\bar{r}} \in M_2$$ such that for all $$(p, q, r) \in N_c = \overline{{{\,\mathrm{\textbf{B}}\,}}_{u_{N_1}}}(0,1) \times \overline{{{\,\mathrm{\textbf{B}}\,}}_{s_{N_1}}}(0,1) \times |N_2|$$ we have:

\begin{aligned} H_c(1, (p, q, r))= & {} (A p, 0, {\bar{r}}) \\ A(\delta {{\,\mathrm{\textbf{B}}\,}}_u(0,1))\subset & {} \mathbb {R}^u \setminus \overline{{{\,\mathrm{\textbf{B}}\,}}_u}(0,1) \end{aligned}

where again $$H_c(t, \cdot ) = c_M \circ H(t, \cdot ) \circ c_N^{-1}$$ is the homotopy expressed in good coordinates.

Let us make some remarks on Definition 10. In contrary to [39], we do not assume that the h-sets with tails N and M are compact in $$\mathcal {X}$$, but we assume that the map P is compact instead. However, the definition in [39] is a special case of Definition 10, if we have $$u_{N_1} = d_{N_1}$$ and $$|M_2|$$ is a compact set. The additional structure of the finite-dimensional part $$N_1$$ we assume in Definition 10 allows for a more general form of covering occurring in the finite-dimensional part, see Fig. 1.

Now we will state theorems, similar to Theorem 23, that joins the sequences of covering relations to the real dynamics happening in the underlying compact maps. We start with definitions:

### Definition 11

Let $$k > 0$$ be a fixed integer and let B be a transition matrix: $$B \in \mathcal {M}(k, k)$$ such that $${{\,\mathrm{\textbf{B}}\,}}_{ij} \in \{0, 1\}$$. Then define:

\begin{aligned} \Sigma ^+_B = \left\{ s \in \{1,\ldots ,k\}^\mathbb {N}: {{\,\mathrm{\textbf{B}}\,}}_{s_i,s_{i+1}} = 1, \quad \forall i \in \mathbb {N}\right\} \end{aligned}

and a shift function $$\sigma : \Sigma ^+_B \rightarrow \Sigma ^+_B$$ by

\begin{aligned} \sigma (s)_i = s_{i+1}. \end{aligned}

The pair $$(\Sigma ^+_B, \sigma )$$ is called a subshift of finite type with transition matrix B.

### Definition 12

Let $$\mathcal {F}$$ be a family of compact maps in a real Banach space $$\mathcal {X}$$.

We say that $$\Gamma = ( \mathcal {N}, \mathcal {F}, Cov )$$ is a set of covering relations on $$\mathcal {X}$$ iff

• $$\mathcal {F}$$ is a collection of continuous and compact maps on $$\mathcal {X}$$,

• $$\mathcal {N}$$ is a collection of h-sets with tails $$N_i \subset \mathcal {X}$$, $$i \in \{1,..,k\}$$,

• $$Cov \subset \mathcal {N} \times \mathcal {F} \times \mathcal {N}$$ is a collection of covering relations, that is, if $$(N_i, P_l, N_j) \in Cov$$ then $$N_j {\mathop {\Longrightarrow }\limits ^{P_l}} N_j$$.

A transition matrix $$B \in \mathcal {M}(k, k)$$ associated to $$\Gamma$$ is defined as:

\begin{aligned} B_{ij} = {\left\{ \begin{array}{ll} 1 &{} \text {if there exists covering relation } N_i {\mathop {\Longrightarrow }\limits ^{P_l}} N_j \in Cov \\ 0 &{} \text {otherwise}. \end{array}\right. } \end{aligned}
(44)

### Definition 13

A sequence $$(x_i)_{i \in \mathbb {N}}$$ is called a full trajectory with respect to family of maps $$\mathcal {F} = \{ f_i: 1 \le i \le m\}$$ if for all $$i \in \mathbb {N}$$ there is $$j(i) \in \{ 1,\ldots ,m \}$$ such that $$f_{j(i)}(x_i) = x_{i+1}$$.

Now we state two main theorems:

### Theorem 24

The claim of Theorem 23 is true for a covering relation chain where sets $$X_i$$ are h-sets with tail in a real Banach space $$\mathcal {X}$$.

### Theorem 25

Let $$\Gamma = ( \mathcal {N}, \mathcal {F}, Cov )$$ be a set of covering relations and let B be its transition matrix.

Then, for every sequence of symbols $$(\alpha _i )_{i \in \mathbb {N}} \in \Sigma ^+_B$$ there exist $$(x_i)_{i \in \mathbb {N}}$$—a full trajectory with respect to $$\mathcal {F}$$, such that $$x_i \in X_{\alpha _i}$$. Moreover, if $$(\alpha _i )_{i \in \mathbb {N}}$$ is T-periodic, then the corresponding trajectory may be chosen to be a T-periodic sequence too.

Before we do the proofs of Theorems 24 and 25, we note that the examples of results that can be obtained with covering relations on h-sets with tails are the same as given before in Sect. 4.2 in the case of a finite-dimensional space $$\mathbb {R}^d$$. In the context of DDEs we will use those theorems for h-sets with tails in the form of a (pn)-fset: $$N = (N_1, |N_2|) = X(A, R) \subset C^n_p$$. The natural decomposition is such that $$\{ \xi \in \left( C([0, h], \mathbb {R}^d)\right) ^p: [\xi ] \subset R \} = |N_2|$$ (the tail) and $$N_1 = A \subset \mathbb {R}^{M(d,p,n)}$$ (the finite-dimensional part). In each application presented later in the paper, we will decide on $$u_{N_1}$$ and on the coordinates $$c_{N_1}$$ on the finite-dimensional part A.

### Proof of Theorem 24

We proceed in a way similar to the proof of Theorem 2 in [39]. To focus the attention and get rid of too many subscripts at once, we assume without loss of generality that $$c_{X_i} = Id$$ for all i and $$X_i = N_i \times R_i$$, where $$N_i \in \mathbb {R}^M$$ is the finite-dimensional part.

Let now denote $$X = X_1 \times \cdots \times X_k$$, $$N = N_1 \times \cdots \ N_k$$ and $$R = R_1 \times \cdots \ R_k$$. Let also denote by $$\mathcal {Y} = \mathbb {R}^{M \cdot k} \times R$$. With a slight abuse of notation we can write $$X \subset \mathcal {Y}$$ and that $$\mathcal {Y} \subset \mathcal {X}^k$$. Since $$\mathcal {X}^k$$ is a Banach space (with the product maximum norm) so is $$\mathcal {Y}$$ with topology inherited from the space $$\mathcal {X}^k$$. Moreover, we have $$X \subset \mathcal {Y}$$ with $${{\,\textrm{int}\,}}_\mathcal {Y}X = {{\,\textrm{int}\,}}N_1 \times R_1 \times \cdots \times {{\,\textrm{int}\,}}N_k \times R_k$$. This will be important for proving that a fixed-point problem we are going to construct is solution-free on the boundary of X in $$\mathcal {Y}$$.

We construct zero finding problem:

\begin{aligned} \begin{array}{rcl} P_k(x_k) &{} = &{} x_1 \\ P_1(x_1) &{} = &{} x_2 \\ &{} \cdots &{} \\ P_{k-1}(x_{k-1}) &{} = &{} x_{k}, \\ \end{array} \end{aligned}
(45)

we denote the left side of (45) by F(x) and we are looking for a solution $$x = F(x)$$ with $$x = (x_1, x_2, \ldots , x_k) \in X$$. With the already mentioned abuse of notation, we can write $$F(a, \xi ) = (b, \zeta )$$ for $$a \in \mathbb {R}^{M \cdot k}$$, $$\xi \in R$$. In a similar way we construct a homotopy H, by pasting together homotopies from the definition of h-sets with tails $$X_i$$:

\begin{aligned} H(t, x) = \left( H_k\left( t, x_k\right) , H_1\left( t, x_1\right) , \ldots , H_{k-1}\left( t, x_{k-1}\right) \right) \end{aligned}

It is obvious that $$H(0,\cdot ) = F$$ and we will show that $$H(t, \cdot )$$ is fixed-point free (admissible) on the boundary $$\delta _{\mathcal {Y}}X$$. Indeed, since $${{\,\textrm{int}\,}}_\mathcal {Y}X = {{\,\textrm{int}\,}}N_1 \times R_1 \times \ldots \times {{\,\textrm{int}\,}}N_k \times R_k$$ then for $$(b, \zeta ) \in \delta _\mathcal {Y}X$$ there must be $$i \in \{1,\ldots ,k\}$$ such that $$b_i \in \delta N_i = N_i^+ \cup N_i^-$$. If $$b_i \in N_i^-$$ then (C2) gives $$H_i\left( t, (b_i,\zeta _i)\right) \notin X_{i+1}$$ and consequently $$({{\,\mathrm{\textbf{B}}\,}}_{i+1}, \zeta _{i+1}) \ne H\left( t, (b, \zeta )\right) _{i+1}$$ (note, if $$i = k$$, then we set $$i+1 = 1$$). If $$b_i \in N_i^+$$, then from (C3) it follows that $$H_{i-1}\left( t, ({{\,\mathrm{\textbf{B}}\,}}_{i-1}, \zeta _{i-1})\right) \notin \left( N_i^+ \times |R_i|\right)$$ and so $$H\left( t, (b, \zeta )\right) _i \ne (b_i, \zeta _i)$$ (note, if $$i = 1$$, then we set $$i-1 = k$$). Therefore, H is admissible, $$H(t, x) \ne x$$ for all $$x \in \delta _\mathcal {Y}X$$. Of course H is also continuous and compact.

Now, $$\mathcal {Y}$$ is an ANR (Corollary 4.4 in Sect. 11. of [8]) so fixed-point index $$\iota (H(t, \cdot ), X)$$ is well defined and constant for all $$t \in [0, 1]$$. Applying Multiplicativity, Normalization (on the tail part) and 21 on $$H(1,\cdot )$$ we get $$\iota (H(1, \cdot ), X) = \Pi \iota \left( A_i, B^{u}(0,1)\right) = \pm 1$$ (since $$det(Id - A_i) \ne 0$$ as $$\Vert A_i\Vert > 1$$ due to (C4)).

Finally, Existence property yields a fixed point $${\bar{x}}$$ to $$H(0, x) = F(x) = x$$. $$\square$$

### Proof of Theorem 25

is almost the same as of Theorem 3 in [39], with the exception that the sets $$X_i$$ are not compact. This is overcome by considering the convergence of sequences of points in the images $$P_i(X_i)$$, which are pre-compact by the assumption on $$P_i$$’s. $$\square$$

We conclude with a lemma that allows to easily check whether $$N {\mathop {\Longrightarrow }\limits ^{P}} M$$ in case $$u_N = u_M = 1$$. We will check the assumptions of this lemma later in Sect. 5, with the help of a computer.

### Lemma 26

For an h-set with tail N let define:

• $$N_c^{l} = \{-1\} \times {{\,\mathrm{\textbf{B}}\,}}_{s} \times |N|$$, $$N^l = c_N^{-1}(N_c^{l})$$—the left edge of N, and

• $$N_c^{r} = \{1\} \times {{\,\mathrm{\textbf{B}}\,}}_{s} \times |N|$$, $$N^r = c_N^{-1}(N_c^{r})$$—the right edge of N.

Let $$\mathcal {X}$$ be a Banach space, $$X \subset \mathcal {X}$$ be an ANR, $$N = (N_1, |N_2|)$$, $$M(M_1, |M_2|)$$ be h-sets with tails in X with $$u_N = u_M = 1$$ and $$P: |N| \rightarrow X$$ be a continuous and compact map such that the following conditions apply (with $$P_c = c_M \circ P \circ c_N^{-1}: N_c \rightarrow M_c$$):

1. 1.

(CC1) $$\pi _{\mathcal {X}_2} P\left( |N|\right) \subset |M_2|$$;

2. 2.

Either (CC2A)

\begin{aligned} P_c\left( N_c^l\right) \subset (-\infty , -1) \times \mathbb {R}_{s} \times |M_2|&and&P_c\left( N_c^r\right) \subset (1, \infty ) \times \mathbb {R}_{s} \times |M_2| \end{aligned}

or (CC2B)

\begin{aligned} P_c\left( N_c^l\right) \subset (1, \infty ) \times \mathbb {R}_{s} \times |M_2|&and&P_c\left( N_c^r\right) \subset (-\infty , -1) \times \mathbb {R}_{s} \times |M_2| \end{aligned}
3. 3.

(CC3) $$P_c\left( N_c\right) \cap \left( {{\,\mathrm{\textbf{B}}\,}}_s \times |M_2|\right) = \emptyset$$

Then, $$N {\mathop {\Longrightarrow }\limits ^{P}} M$$ with the homotopy given as $$H(t, \cdot ) = (1-t) \cdot P + t \cdot (A, 0, {\bar{r}})$$, where $$A: \mathbb {R}\rightarrow \mathbb {R}$$ such that $$Ax = 2x$$ (CC2A) or $$Ax = -2x$$ (CC2B) and $${\bar{r}}$$ is any selected point in $$|M_2|$$.

### Proof

(C0) and (C1) from Definition 10 are obviously satisfied. We also have (CC2) implies (C2) and (CC3) is the same as (C3). Therefore, we only need to show (C4), that is, the image of the homotopy computed on the set $$\delta {{\,\mathrm{\textbf{B}}\,}}_u \times \overline{{{\,\mathrm{\textbf{B}}\,}}_s} \times |N_2|$$ does not touch the set $$M_c$$. This is obvious from the definition of A in both cases (CC2A) and (CC2B). $$\square$$

Figure 1 presents such a covering in case $$u = s = 1$$ and $$N = M$$. The easiest way to assure (CC1) and (CC3) is to assume $$P_c(N_c) \subset \mathbb {R}\times {{\,\mathrm{\textbf{B}}\,}}_s \times |M_2|$$—in fact we check this in our computer-assisted proofs presented in the next section.

## 5 Applications

In this section, we present applications of the discussed algorithm to two exemplary problems. First one is a computer-assisted proof of symbolic dynamics in a delay-perturbed Rössler system [29]. The proof is done for two different choices of perturbations. The second application consists of proofs of (apparently) unstable periodic orbits in the Mackey–Glass equation for parameter values for which Mackey and Glass observed chaos in their seminal paper [24].

Before we state the theorems, we would like to discuss presentation of floating-point numbers in the article. Due to the very nature of the implementation of real numbers in current computers, numbers like 0.1 are not representable [10], i.e. cannot be stored in memory exactly. On the other hand, many numbers representable on the computer could not be presented in the text of the manuscript in a reasonable way, unless we adopt not so convenient digital base-2 number representation. However, the implementation IEEE-754 of the floating-point numbers on computers [10] guarantees that, for any real number x and its representation $${\tilde{x}}$$ in a computer format, there is always a number $$|\varepsilon | \le \varepsilon _{machine}$$ such that $${\tilde{x}} = x(1+\varepsilon )$$. The number $$\varepsilon _{machine}$$ defines the machine precision, and, for the double precision C++ floating-point numbers that we use in the applications, it is of the order $$10^{-16}$$. Finally, in our computations we use the interval arithmetic to produce rigorous estimates on the results of all basic operations such as $$+$$, −, $$\times$$, $$\div$$, etc. In principle, we operate on intervals [ab], where a and b are representable numbers, and the result of an operation contains all possible results, adjusting end points so that they are again representable numbers (for a broader discussion on this topic, see the work [34] and references therein). For a number $$x \in \mathbb {R}$$ we will write [x] to denote the interval containing x. If $$x \in \mathbb {Z}$$ then we have $$[z] = [z,z]$$, as integer numbers (of reasonably big value) are representable in floating point arithmetic.

Taking all that into account we use the following convention:

• Whenever there is an explicit decimal fraction defined in the text of the manuscript of the form $$d_1 d_2 \cdots d_k. q_1 q_2 \cdots q_m$$, that number appears in the computer implementation as

\begin{aligned}{}[d_1 d_2 \cdots d_k q_1 q_2 \cdots q_k] \div [10^m], \end{aligned}

where $$\div$$ is computed rigorously with the interval arithmetic. For example, number $$10^{-3} = 0.001$$ appears in source codes as Interval(1.)/Interval(1000.).

• Whenever we present a result from the output of the computer program x as a decimal number with nonzero fraction part, then we have in mind the fact that this represents some other number y—the true value, such that $$y = x(1+\varepsilon )$$ with $$|\varepsilon | \le \varepsilon _{machine}$$. This convention applies also to intervals: if we write interval $$[a_1, a_2]$$, then there are some representable computer numbers $$b_1$$, $$b_2$$ which are true output of the program, so that $$b_i = a_i (1 +\varepsilon _i)$$.

• If we write a number in the following manner: $$d_1. d_2 \cdots d_k {}^{u_1 u_2 \cdots u_m}_{l_1 l_2 \cdots l_m}$$ with digits $$l_i, u_i, d_i \in \{0,..,9\}$$ then it represents the following interval

\begin{aligned} \left[ d_1. d_2 \cdots d_k l_1 l_2 \cdots l_m, d_1. d_2 \cdots d_k u_1 u_2 \cdots u_m \right] . \end{aligned}

For example $$12.3_{456}^{789}$$ represents the interval [12.3456, 12.3789] (here we also understand the numbers taking into account the first two conventions).

The last comment concerns the choice of various parameters for the proof, namely the parameters of the space $$C^n_p$$ and the initial sets around the numerically found approximations of the dynamical phenomena under consideration. The latter strongly depend on the investigated phenomena, so we will discuss general strategy in each of the following sections, whereas the technical details are presented in Appendices A and B.

The choice of parameters n and p corresponds basically to the choice of the order of the numerical method and a fixed step size $$h = \frac{\tau }{p}$$, respectively.

Usually, in computer-assisted proofs, we want n to be high, so that the local errors are very small. In the usual case of ODEs with $$f \in C^\infty$$ we can use almost any order, and it is easy for example to set $$n = 40$$. However, in the context of $$C^n_p$$ spaces and constructing Poincaré maps for DDEs, we are constrained with the long enough time $$T = (n+1) \cdot \tau$$ (Definition 3) to obtain well defined maps. Therefore, the choice of n corresponds usually to the return time to section $$t_P$$ for a given Poincaré map, satisfying $$t_P(X_0) > (n+1) \cdot \tau$$, for some set of initial data $$X_0 \subset C^n_p$$.

The choice of the step size h is more involved. It should not be too small, to reduce the computational time and cumulative impact of all local errors after many iterations, and not so big, as to effectively reduce the size of the local error. Also, the dynamics of the system (e.g. stiff systems) can impact the size of the step size h. In the standard ODE setting, there are strategies to set the step size dynamically, from step to step, e.g. [9], but in the setting of our algorithm for DDEs, due to the continuity issues described in Sect. 3, we must stick to the fixed step size $$h = \frac{\tau }{p}$$. The step size must be also smaller than the (apparent) radius of convergence of the forward Taylor representation of the solution at each subinterval, but this is rarely an issue in comparison with other factors, e.g. the local error estimates. In our applications, we chose $$p = 2^{m}$$ for a fixed $$m \in \mathbb {N}$$, so that the grid points are representable floating point numbers (but the implementation can work for any p).

We also need to account for the memory and computing power resources. For d-dimensional systems (2), and with n, p fixed, we have that the representation of a Lohner-type set $$A = x_0 + C \cdot r_0 + E$$ in phase space of $$\varphi$$, where $$C \in \mathcal {M}{(M, M)}$$, requires at least $$O(M^2)$$, with $$M = O(d \cdot n \cdot p)$$. Then, doing one step of the full-step algorithm is of $$O(d^2 \cdot n^2 \cdot M)$$ computational complexity. Due to the long enough time integration, computation of a single orbit takes usually $$O(n \cdot p)$$ steps, and we get the computational complexity of computing image P(X) for a single set X of $$O(d \cdot n^2 \cdot d \cdot p \cdot n \cdot M) = O(d \cdot n^2 \cdot M \cdot M) = O(M^2)$$ (if we assume $$n, d \ll M$$). Therefore, we want to keep $$M^2 = (d \cdot n \cdot p)^2$$ of reasonable size, both because of time and memory constraints. Our choice here is $$M \le 10^3$$.

### 5.1 Symbolic Dynamics in a Delay-Perturbed Rössler System

In the first application, we use Rössler ODE of the form [29]:

\begin{aligned} \begin{aligned} x'=&{} -(y + z) \\ y'=\,&{} x + ay \\ z'=\,&{} b + z(x-c). \end{aligned} \end{aligned}
(46)

In what follows, we will denote r.h.s. of (46) by f and by $$v \in \mathbb {R}^3$$ we denote vector $$v = (x, y, z)$$. By $$\pi _x$$ we denote projection onto x coordinate, similarly for $$\pi _y, \pi _z$$.

We set the classical value of parameters $$a = b = 0.2$$, $$c = 5.7$$ [29]. For those parameter values, an evidence of a strange attractor was first observed numerically in [29], see Fig. 2. In [42], it was proved by computer-assisted argument that there is a subset of the attractor which exhibit symbolic dynamics. A more recent results for Rössler system can also be found in [6] (Sharkovskii’s theorem) and the methodologies there should be easily adaptable in the context of delay-perturbed systems presented in this paper.

We are going to study a delayed perturbation of the Rössler system (46) of the following form:

\begin{aligned} v'(t) = f(v(t)) + \epsilon \cdot g(v(t - 1)), \end{aligned}
(47)

where parameter $$\epsilon$$ is small. We consider two toy examples: first, where $$g = f$$ and the second one where g is given explicitly as

\begin{aligned} g(x, y, z) = \left( \sin (x \cdot y), \sin (y \cdot z), \sin (x \cdot z)\right) . \end{aligned}
(48)

We expect that for any bounded g there should be a sufficiently small $$\epsilon$$ [33] so that the dynamics of the perturbed system is preserved. However, in this work, we study explicitly given value for $$\epsilon$$.

### Remark 27

The source codes of the proof are generic. The interested reader can experiment with other forms of the perturbation by just changing the definition of the function g in the source codes of the example.

We will be studying the properties of a Poincaré map defined on the section $$S_0 \subset C^n_p$$ given by:

\begin{aligned} S_0 = \{ v \in C^n_p: \pi _x (v(0)) = 0 \}. \end{aligned}

The section $$S_0$$ is an extension to $$C^n_p$$ of the section $$S = \{ v \in \mathbb {R}^3: \pi _x v = 0 \} \subset \mathbb {R}^3$$ used in the proofs in [42]. The section S is drawn in green in Fig. 3, whereas the projection of the attractor onto section $$S_0$$ is drawn as a blue-red gradient (the solution segments v with $$\pi _x v(0) = 0$$).

In what follows, we set the parameters for the space $$C^n_p$$ to $$p = 32$$ and $$n = 3$$. We prove, with the computer assistance, the following theorems:

### Theorem 28

For parameter values $$a = b = 0.2$$, $$c = 5.7$$ in (46), there exist sets $$X_A = X(A, \Xi ), X_1 = X(N_1, \Xi ), X_2 = X(N_2, \Xi ) \subset S_0$$ with explicitly given $$A, N_1, N_2$$ and $$\Xi$$, such that for the system (47) with $$\epsilon = 10^{-3}$$ and perturbations: (a) $$g \equiv 0$$—original system treated as a DDE, (b) $$g = f$$, and (c) g given as in Eq. (48) we have the following:

1. 1.

$$P(X(A, \Xi )) \subset X(A, \Xi )$$ and, in consequence, there exists a non-empty invariant set in $$X(A, \Xi )$$ for the map $$P: S_0 \rightarrow S_0$$.

2. 2.

The invariant set $$I = Inv(P^2, X_1 \cup X_2)$$ of $$X_1 \cup X_2$$ under the map $$P^2$$ on I is non-empty and the dynamics of $$P^2$$ is conjugated to the shift on two symbols ($$\sigma : \Sigma _2 \rightarrow \Sigma _2$$, $$\sigma (e_k) = e_{k+1}$$), i.e. if we denote by $$g: I \rightarrow \Sigma _2$$ the function $$g(x)_k = i \iff P^{2k}(x) \in X_i$$, then we have $$g \circ P^2|_I = \sigma \circ g$$.

Before we present the proof(s), we would like to make a remark on the presentation of the data from the computer-assisted part:

### Remark 29

(Convention used in the proofs) The proofs of those theorems are computer-assisted and the parameters of the phase space $$C^n_p$$ of representations are $$d = 3$$, $$p = 32$$, $$n=3$$, giving in total the dimension of the finite-dimensional part of $$M(d, p, n) = d \cdot (1 + p \cdot (n+1)) = 387$$. Therefore, it is not convenient to present complete data of the proofs in the manuscript. Instead, we assume the sets are explicitly given in the following forms (and the interested reader is refereed to “Appendix B” for the details on how they are constructed):

\begin{aligned} X_A = X(A, R): \quad&A = v_\textrm{ref} + C \cdot \{ 0\} \times W_u \times {{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert _\infty }_{M-2}(0, 1) \\ X_i = X(N_i, R): \quad&N_i = v_\textrm{ref} + C \cdot \{ 0\} \times W_i \times {{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert _\infty }_{M-2}(0, 1) \\ \Xi = {{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert _\infty }_{d \cdot p}(0, 1) \quad{} & {} \end{aligned}

with $$v_\textrm{ref} \in S_0$$, $$W_u, W_1, W_2$$ closed intervals such that $$W_1 \cap W_2 = \emptyset$$ and $$W_i \subset W_u \subset \mathbb {R}$$, and we remind $${{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert _\infty }_{D}(0, 1)$$ denotes the unit radius ball in the $$\max$$ norm in $$\mathbb {R}^{D}$$ centred at 0. Note that this description of sets makes it clear they are h-sets with tails on $$S_0$$ (up to the scaling of nominally unstable direction W), where $$u = 1$$ and $$s_A = s_{N_i} = s = M(d, p,n)-2$$, the support set $$|A| = \{ 0\} \times W_u \times {{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert _\infty }_{M-2}(0, 1)$$ and the affine coordinate change $$c_A(\cdot ) = v_\textrm{ref} + C(\cdot )$$ with inverse change $$c_A^{-1}(\cdot ) = C^{-1}(\cdot - v_{ref})$$. Now, the computation of any Poincaré map $$P: X_A \rightarrow S_0$$ for the initial data $$X(A, \Xi )$$ produces a set $$X(B, \Omega ) = P(X(A, \Xi ))$$ and there exist sets

\begin{aligned} c_A^{-1}(B)&=\ B_c&\subset \ \{ 0 \} \times (B_c)_2 \times {{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert _\infty }_{M-2}(0, r_{B}) \\&\ \Omega&\subset \ {{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert _\infty }_{d \cdot p}(0, r_{\Omega }) \end{aligned}

for some $$r_{B}, r_{\Omega } \in \mathbb {R}_{+}$$. This allows to describe the geometry of $$X(A, \Xi )$$ and (estimates on) $$P(X(A, \Xi ))$$ by just a couple of numbers: $$W_u$$, $$\pi _2 B_c$$ (the size of set B in the nominally unstable direction), $$r_{B}$$ (upper bound on all coefficients in the finite nominally stable part) and $$r_{\Omega }$$ (upper bound on all $$\xi$$ in the tail part), which are suitable for a concise presentation in the manuscript.

The sets used in the computations are obtained by computing the appropriately enlarged enclosure on the set of segments of solutions to the unperturbed ODE (46). We choose a set $${\tilde{A}} \subset \mathbb {R}^3$$ such that $${\tilde{A}} \in \{ v \in \mathbb {R}^3 \pi _x v = 0\}$$ is a trapping region for the Poincaré map of the unperturbed ODE: $$P({\tilde{A}}) \subset {\tilde{A}}$$. Then we choose a set $$X(A, \Xi )$$ to contain the segments of $${\tilde{A}}$$ propagated back in time for a full delay with the unperturbed ODE:

\begin{aligned} \left\{ v:[-1, 0] \rightarrow \mathbb {R}: v(0) \in A, v(s) = \varphi _0(s, v(0)) \right\} \subset A, \end{aligned}

where $$\varphi _0$$ is the flow in $$\mathbb {R}^3$$ for (46). Detailed procedure how the set A was generated is described in “Appendix B”. The set $${\tilde{A}}$$ was chosen to be $$\{0 \} \times [-10.7, -2.2] \times [0.021, 0.041]$$, whereas the sets $${\tilde{N}}_1 = [-8.4, -7.6]$$ and $${\tilde{N}}_2 = [-5.7, -4.6]$$. Finally, the orbit $$v_0$$ with $$\pi _2 v_0(0) = -6.8$$ is selected among the orbits in the attractor as the reference point of the sets $$X_A, X_1, X_2$$. The set $$W_u$$ is chosen as $$W_u = \pi _2 A_c = \pi _y {\tilde{A}} - \pi _2 v_0(0) = [-3.9,4.6]$$. The same is true for sets $$N_1, N_2$$, with $$W_1 = [-1.6, -0.8]$$, $$W_2 = [1.1, 2.2]$$.

Now, we can proceed to the proofs.

### Proof o Theorem 28

The proofs for parts (a), (b), and (c) follow the same methodology; therefore, we present the details only for case (a) and then only the estimates from the other two cases. In principle, we will show that $$P(X_A) \subset X_A$$ and $$X_i {\mathop {\Longrightarrow }\limits ^{P^2}} X_j$$ for all $$i, j \in \{ 1, 2 \}$$ and then apply Theorem 25.

The set X(AR) and two other sets are given as described in Remark 29. The computer programs for the proof are stored in ./examples/rossler_delay_zero. The data for which presented values were computed is stored in ./data/rossler_chaos/epsi_0.001. See “Appendix B” for more information. Additionally to the estimates presented below, the computer programs verify that $$t_P(x) > (n+1)$$ (i.e. long enough for Poincaré maps to be well defined) and that the function $$t_P(\cdot )$$ is well defined. For details, see the previous work [34].

First, we prove that $$P_c(X(A, \Xi )) \subset (A_c, \Xi )$$. Let $$(B_c, \Omega )$$ will be output of the rigorous program rig_prove_trapping_region_exists run for the system in case (a) such that $$P_c(X(A, \Xi )) \subset (B_c, \Omega )$$. It suffices to show the following:

• $$\pi _2 P_c(X(A, \Xi )) = \pi _2 B_c \subset W_u = \pi _2 A_c$$;

• $$\pi _i P_c(X(A, \Xi )) = \pi _i B_c < 1$$ for all $$i > 2$$;

• $$\pi _{\Xi _i} P_c(X(A, \Xi )) = \pi _i \Omega < 1$$ for all $$i \in \{1, \ldots , p \cdot d \}$$.

Indeed, we have:

• $$\pi _2 P_c(X(A, \Xi )) = [-3.786230021035,3.92103823500285] \subset [-3.9,4.6] = W_u$$;

• $$\pi _i P_c(X(A, \Xi )) \le 0.910355124006778 < 1$$, for $$i > 2$$;

• $$\pi _{\Xi _i} P_c(X(A, \Xi )) \le 0.395102819146026 < 1$$ for all i.

Which finishes the proof of the first assertion.

For the second assertion we prove that we have a set of full covering relations:

\begin{aligned} X_i {\mathop {\Longrightarrow }\limits ^{P^2}} X_j, \quad i, j \in \{ 1, 2\}. \end{aligned}

We remind that the sets $$N_{i,c} = \{ 0 \} \times [W_{i}^{l}, W_{i}^r] \times {{\,\mathrm{\textbf{B}}\,}}^{\Vert \cdot \Vert _\infty }_{M-2}(0, 1)$$ with $$W_1 = [-1.6, -0.8]$$, $$W_2 = [1.1, 2.2]$$. The program ./rig_prove_covering_relations produces the following inequalities:

• (L1-L1) $$\pi _2 P^2_c(X(N_{1}^{l}, \Xi )) = -1._{696238902429803}^{708946819732338}< -1.6 = \pi _2 N_{1,c}^{l} < \pi _2 N_{2,c}^{l}$$

• (R1-R2) $$\pi _2 P^2_c(X(N_{1}^{r}, \Xi )) = 2.4_{09511664184434}^{17718805618395}> 2.2 = \pi _2 N_{2,c}^{r} > \pi _2 N_{1,c}^{r}$$

• (R2-L1) $$\pi _2 P^2_c(X(N_{2}^{r}, \Xi )) = -1.83_{8887194518363}^{9215629292839}< -1.6 = \pi _2 N_{1,c}^{l} < \pi _2 N_{2,c}^{l}$$

• (L2-R2) $$\pi _2 P^2_c(X(N_{2}^{l}, \Xi )) = 2.2_{69015891346912}^{70120359885664}> 2.2 = \pi _2 N_{2,c}^{r} > \pi _2 N_{1,c}^{l}$$,

where sets $$N^l$$, $$N^r$$ etc. are defined as in Lemma 26. It is ease to see that those inequalities, together with the existence of trapping region $$X_A$$, imply that for each $$i, j \in \{1, 2\}$$ conditions (CC1)-(CC3) in Lemma 26 are satisfied, that is $$X_i {\mathop {\Longrightarrow }\limits ^{P^2}} X_j$$, which finishes the proof for the case (a) after applying Theorem 25.

For the cases (b) and (c) we only present estimates:

• Case (b), $$g = f$$. Output from rig_prove_trapping_region_exists is:

• $$\pi _2 P_c(X(A, \Xi )) = [-3.82791635121864,3.90123013871349] \subset [-3.9,4.6] = W_u$$;

• $$\pi _i P_c(X(A, \Xi )) \le 0.960537051554584 < 1$$, for $$i > 2$$;

• $$\pi _{\Xi _i} P_c(X(A, \Xi )) \le 0.397264977921163 < 1 = r_\Xi$$, for all i.

Output from program ./rig_prove_covering_relations is:

• (L1-L1) $$\pi _2 P^2_c(X(N_{1}^{l}, \Xi )) = -1.6_{68486957556001}^{84410417326001}< -1.6 = \pi _2 N_{1,c}^{l} < \pi _2 N_{2,c}^{l}$$

• (R1-R2) $$\pi _2 P^2_c(X(N_{1}^{r}, \Xi )) = 2.4_{64065036803807}^{74268236696726}> 2.2 = \pi _2 N_{2,c}^{r} > \pi _2 N_{1,c}^{r}$$

• (R2-L1) $$\pi _2 P^2_c(X(N_{2}^{r}, \Xi )) = -1.76_{7206286440370}^{9151140189891}< -1.6 = \pi _2 N_{1,c}^{l} < \pi _2 N_{2,c}^{l}$$

• (L2-R2) $$\pi _2 P^2_c(X(N_{2}^{l}, \Xi )) = 2.36_{0282881761384}^{2685243092644}> 2.2 = \pi _2 N_{2,c}^{r} > \pi _2 N_{1,c}^{r}$$

• Case (c), g as in (48). Output from rig_prove_trapping_region_exists is:

• $$\pi _2 P_c(X(A, \Xi )) = [-3.78710970137727,3.92188126709857] \subset [-3.9,4.6] = W_u$$;

• $$\pi _i P_c(X(A, \Xi )) \le 0.951680057117636 < 1$$, for $$i > 2$$;

• $$\pi _{\Xi _i} P_c(X(A, \Xi )) \le 0.459753301095895 < 1$$, for all i.

Output from program ./rig_prove_covering_relations is:

• (L1-L1) $$\pi _2 P^2_c(X(N_{1}^{l}, \Xi )) = -1._{695427259804897}^{714200213156898}< -1.6 = \pi _2 N_{1,c}^{l} < \pi _2 N_{2,c}^{l}$$

• (R1-R2) $$\pi _2 P^2_c(X(N_{1}^{r}, \Xi )) = 2.4_{08774107762390}^{20396855111791}> 2.2 = \pi _2 N_{2,c}^{r} > \pi _2 N_{1,c}^{r}$$

• (R2-L1) $$\pi _2 P^2_c(X(N_{2}^{r}, \Xi )) = -1.8_{38300180457653}^{41157932729915}< -1.6 = \pi _2 N_{1,c}^{l} < \pi _2 N_{2,c}^{l}$$

• (L2-R2) $$\pi _2 P^2_c(X(N_{2}^{l}, \Xi )) = 2.2_{67377344403297}^{70144525622461}> 2.2 = \pi _2 N_{2,c}^{r} > \pi _2 N_{1,c}^{r}$$

$$\square$$

Figure 3 shows the numerical representations of the apparent strange attractor in the respective systems, while Fig. 4 depicts the computed estimates of the proof in a human-friendly manner. The total running time of the proof in (a) is around 16 min, and the cases (b) and (c) of around 23 min. Computations were done on a laptop with Intel® Core$${}^{\texttt{TM}}$$ i7-10750 H 2.60GHz CPU. The majority of the computations is done in the proof of trapping region $$X_A$$, which must be divided into 200 pieces along the vector $$W_u$$. Those computations are easily parallelized (each piece computed in a separate thread). The data and programs used in the proofs are described in more detail in “Appendix B”, together with the links to source codes.

### 5.2 Unstable Periodic Orbits in Mackey–Glass Equation

In this application, we study the following scalar equation:

\begin{aligned} x'(t) = -\gamma \cdot x(t) + \beta \cdot \frac{x(t-\tau )}{1 + (x(t-\tau ))^n}. \end{aligned}
(49)

In [24], the authors shown numerical evidence of chaotic attractor in that system, see Fig. 5a. In their work, Mackey and Glass used the following values of parameters: $$\tau = 2$$, $$n = 9.65$$, $$\beta = 2$$, $$\gamma = 1$$. In our previous work [34], we have shown existence of several (apparently) stable periodic orbit for $$n \le 8$$. In this work, we show that the new algorithm, together with the fixed-point index, can be used to prove more diverse spectrum of results. We prove existence of several (apparently) unstable periodic orbits for the classical values of parameters, for which the chaotic attractor is observed, $$\tau = 2$$, $$n = 9.65$$, $$\beta = 2$$, $$\gamma = 1$$.

### Remark 30

In what follows, we get rid of the variable delay $$\tau$$ and we rescale the system to have unit delay by the change of variables: $$y(t) = x(\tau \cdot t)$$. It is easy to see that Eq. (49) in the new variables becomes:

\begin{aligned} y'(t) = \tau \cdot f(y(t), y(t-1)), \end{aligned}

that is, we can remove parameter $$\tau$$ by rescaling $$\beta$$ and $$\gamma$$ to $${\bar{\beta }} = \tau \cdot \beta$$ and $${\bar{\gamma }} = \tau \cdot \gamma$$.

We state the following:

### Theorem 31

Each of the three approximate solutions $${\bar{T}}^i$$ shown in Fig. 5c, d has a small, explicitly given vicinity $$V_i \subset C^n_p$$ with $$n=4$$ and $$p=128$$ of the initial segment $${\bar{T}}^i_0$$ such that there exists a true periodic solution $$T^i$$ with the initial segment $$T^i_0 \in V_i$$ of the Mackey–Glass equation (49) for the classical parameter values $$\tau = 2$$, $$n = 9.65$$, $$\beta = 2$$, $$\gamma = 1$$ [24].

### Proof of Theorem 31

we use the parameters $$\beta = 4$$ and $$\gamma = 2$$, $$n = 9.65$$ and $$\tau = 1$$ in (49) and we use Remark 30. The proof is similar to that of Theorem 28 and boils down to checking appropriate covering relations. The initial segments $${\bar{T}}^i$$ lie on the section $$S = \left\{ x \in C([0,1],\mathbb {R}): x(0) = x(-1), x(t) < 0.96 \right\}$$. The index i corresponds to the basic period of the solution $$T^i$$ as a periodic point of a map $$P: S \rightarrow S$$. In the coverings, we use map $$P^2$$ to guarantee that the return time $$t_P$$ to the section is long enough.

Each of the $$V_i = X(N_i, \Xi _i)$$ is given with $$N_i = {\bar{T}}^i_0 + C_i \cdot r_i$$ with $$r_i = \{ 0 \} \times W_i^u \times [-1,1]^{M-2}$$. Additionally, in case of $$T^4$$ we have another set $$V'_4 = X(N'_4, \Xi '_4)$$ with $$N'_4$$ of the similar form: $$N'_4 = P^2({\bar{T}}^4) + C'_4 \cdot r'_4$$. In other words, the origin point of the set $$N'_4$$ is the second iteration of the Poincaré map $$P^2$$ of the initial segment of $$T^4$$. The sets are obtained as described in “Appendix B”. Each of these sets defines a section $$S_i = \{ x \in C^n_p: c_i \cdot (a(x) - {\bar{T}}^i_0) = 0 \}$$ (different from S), where $$c_i = (C_i)_{\cdot ,1}$$—the first column of the matrix $$C_i$$. The reason for that is described in “Appendix B” and boils down to assure that $${{\,\mathrm{\textbf{diam}}\,}}(t_P(X_i))$$ is as small as possible.

We will show that:

\begin{aligned} V_1 {\mathop {\Longrightarrow }\limits ^{P_{S_1 \rightarrow S_1}}} V_1, \quad \quad \quad V_2 {\mathop {\Longrightarrow }\limits ^{P_{S_2 \rightarrow S_2}}} V_2, \quad \quad \quad V_4 {\mathop {\Longrightarrow }\limits ^{P_{S_4 \rightarrow S'_4}}} V'_4 {\mathop {\Longrightarrow }\limits ^{P_{S'_4 \rightarrow S_4}}} V_4, \end{aligned}
(50)

where the Poincaré maps $$P_{S_i \rightarrow S_j}$$ are derived from the flow of Eq. (49) and maps indicated sections: $$P_{S_i \rightarrow S_j}: S_i \rightarrow S_j$$, with additional assumption that the return time $$t_P$$ is long enough. We will drop the subscripts if they are easily known from the context.

For $$T^1$$ we have:

• for all $$i > 2$$, $$\left| \pi _i P_c(X(N_{1}, \Xi _{i}))\right| = 0.614451801967851 < 1$$

• for all i, $$\left| \pi _{\Xi _i} P_c(X(N_{1}, \Xi _{i}))\right| = 0.999998174289212 < 1$$

• $$\pi _2 P_c(X(N_{1}^{r}, \Xi )) = -{}_{3.845940820239275}^{4.514877050431105} < -1 = \pi _2 N_{c}^{l}$$

• $$\pi _2 P_c(X(N_{1}^{r}, \Xi )) = {}_{3.827847664967472}^{4.496773405715568} > 1 = \pi _2 N_{c}^{r}$$

For $$T^2$$ we have:

• for all $$i > 2$$, $$\left| \pi _i P_c(X(N_{2}, \Xi _{2}))\right| \le 0.731193331043839 < 1$$

• for all i, $$\left| \pi _{\Xi _i} P_c(X(N_{2}, \Xi _{2}))\right| \le 0.999996951451891 < 1$$

• $$\pi _2 P_c(X(N_{2}^{r}, \Xi _2)) = {}_{3.995778903452447}^{5.033339010859840} > 1 = \pi _2 N_{2,c}^{r}$$

• $$\pi _2 P_c(X(N_{2}^{r}, \Xi _2)) = -{}_{3.978765934264806}^{5.016322912452834} < -1 = \pi _2 N_{2,c}^{l}$$

For $$T_4$$ we have:

• for all $$i > 2$$, $$\left| \pi _i P_c(X(N_{4}, \Xi _{4}))\right| \le 0.999948121260377 < 1$$

• for all i, $$\left| \pi _{\Xi _i} P_c(X(N_{4}, \Xi _{4}))\right| \le 0.956276660970399 < 1$$

• $$\pi _2 P_c(X(N_{4}^{l}, \Xi _4)) = -{}_{1.1221122505976317}^{3.2368193002087367} < -1 = {N'_{4}}^{l}_{c}$$

• $$\pi _2 P_c(X(N_{4}^{r}, \Xi _4)) = {}_{1.1239689716822263}^{3.2385726261859316} > 1 = {N'_{4}}^{r}_{c}$$

and

• for all $$i > 2$$, $$\left| \pi _i P_c(X(N'_{4}, \Xi '_{4}))\right| \le 0.898580326387734 < 1$$

• for all i, $$\left| \pi _{\Xi _i} P_c(X(N'_{4}, \Xi '_{4}))\right| \le 0.952378028038733 < 1$$

• $$\pi _2 P_c(X({N'_{4}}^{l}, \Xi '_4)) = {}_{1.6331410212899785}^{3.0485204456349866} > 1 = {N_{4}}^{r}_{c}$$

• $$\pi _2 P_c(X({N'_{4}}^{r}, \Xi '_4)) = -{}_{1.6341550945779965}^{3.0495636957165507} < -1 = {N_{4}}^{l}_{c}$$

All those inequalities satisfy appropriate assumptions of Lemma 26. Therefore, all the coverings from (50) exist and, from Theorem 25, we infer existence of appropriate periodic points $$T^i_0 \in V_i$$. $$\square$$

The diameters of the sets expressed in commonly used functional norms are presented in Table 2. The data and programs used in the proofs are described in more detail in “Appendix B”, together with the links to source codes.

### 5.3 A Comment About the Exemplary Systems

Both Rössler and Mackey–Glass systems studied as an exemplary application in this work are chaotic for the parameters used. However, Mackey–Glass system is a scalar equation, so the chaos present in the system must be a result of the infinite nature of the phase space and the delay plays a crucial role here. It is not clear if the dynamics can be approximated with a finite number of modes, and how to choose good coordinate frame to embed the attractor. The Rssler system on the other hand is a 3D chaotic ODE (for parameters specified), and the chaotic behaviour is the result of the dynamic in this explicitly finite dimension space. The systems of the form (47) are small perturbations of the ODE and thus one can expect the dynamics of the ODE persist in some sense, at least for $$\epsilon$$ small enough [33]. It is much easier to propose sets for the covering relations inherited directly from the coverings in finite dimension for unperturbed system, see “Appendix B”, where we use the flow of unperturbed ODE to generate the apparently unstable direction for the trapping region containing the attractor.