betaFeedback

WorksMathematics

Observable functions of rational ODE models and how to find them

Abstract

Consider a parametric ODE control model. A function of the states and parameters is called observable if its value can in principle be reconstructed from input-output data. The observable functions form a field, called the observation field, represented naturally by a set of generators. Even when the model is not fully observable, this field captures the information still accessible from input-output data. We present an algorithm for computing a concise generating set for the observation field of a model with rational dynamics. The algorithm relies on two new results: one allows observable functions to be extracted from the coefficients of repeated Lie derivatives of the outputs, while the other reduces the required orders of differentiation by exploiting identifiable parameter combinations. We implement the resulting algorithm in StructuralIdentifiability..jl (https://github.com/SciML/StructuralIdentifiability.jl). For computational efficiency, we employ recent techniques for differential elimination and rational function field simplification. Using models from epidemiology, chemical kinetics, and cancer modeling, we show that the algorithm produces generators with domain-specific interpretations that can inform model analysis and development.

Introduction

Consider a dynamical model given by a system of ordinary differential equations:

Σ={x′(t)=f(μ,x(t),u(t)),y(t)=g(μ,x(t),u(t)),\Sigma= \begin{cases} \mathbf{x}'(t) = \mathbf{f}(\boldsymbol{\mu},\mathbf{x}(t),\mathbf{u}(t)),\\ \mathbf{y}(t) = \mathbf{g}(\boldsymbol{\mu},\mathbf{x}(t),\mathbf{u}(t)), \end{cases}

where x=(x1,…,xn)\mathbf{x}=(x_1,\ldots,x_n) are state variables, μ=(μ1,…,μk)\boldsymbol{\mu}=(\mu_1,\ldots,\mu_k) are time-independent parameters, u=(u1,…,uℓ)\mathbf{u}=(u_1,\ldots,u_\ell) are inputs, and y=(y1,…,ym)\mathbf{y}=(y_1,\ldots,y_m) are outputs. Roughly speaking, structural global observability (in what follows, simply “observability”) asks whether the state can be uniquely recovered from the known input and observed output under ideal conditions [19, 23]. Identifiability is the analogous property for the time-independent parameters of the model [4, 20].

In practice, models frequently lack observability and/or identifiability [7, 26, 28]. Yet this need not render such models completely opaque: even when individual states and parameters are not observable, their combinations might be. In particular, a function h(μ,x)h(\boldsymbol{\mu},\mathbf{x}) is said to be observable (or functionally observable [24, 31]) if its value can in principle be reconstructed from input-output data. Together, such functions form the observation field [2, 6] of the model,

FΣ:={h∈C(μ,x)∣h is observable},\mathcal{F}_{\Sigma}:=\{h\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x})\mid h\text{ is observable}\},

which is a nonlinear analogue of the classical notion of an observation space for linear systems [17]. Besides observable states and parameters, FΣ\mathcal{F}_{\Sigma} may contain some other functions of practical interest (e.g., the basic reproduction number [39] or the binding density [22], Section 3.3). Furthermore, knowledge of generators of FΣ\mathcal{F}_{\Sigma} can guide system identification [34], the selection of complementary data streams [39], and model reparametrization [32]. In this paper, we consider the problem of finding a concise generating set of the observation field FΣ\mathcal{F}_{\Sigma} for a given rational system (that is, when f\mathbf{f} and g\mathbf{g} are rational functions).

Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.

From a computational point of view, algorithms and software are available for testing observability of state variables [20, 44]. Furthermore, Němcová, Petreczky, and van Schuppen [32] proposed (building upon earlier works [2, 3, 33]) a complete and rigorous procedure for computing generators of the observation field of input-affine rational systems from sufficiently many Lie derivatives of the outputs. The procedure is constructive but relies on heavy computational tools from commutative algebra and, to the best of our knowledge, has not been implemented. A related problem of finding locally observable functions [5, 16, 29, 48] has typically been reduced to a search for analytic solutions of certain PDEs or ODEs, which itself is a hard problem with no complete algorithmic solution available. Consequently, these methods do not guarantee to find a generating set (see also [1 Corollary 3]).

The main result of the present paper is an algorithm for computing generators of the observation field for models with rational dynamics. In particular, we allow arbitrary rational dependence on inputs. Our algorithm is designed to be practical and we provide an implementation within the Julia package StructuralIdentifiability.jl. We apply it to models from the literature arising in epidemiology, chemistry, and cancer modeling, and show that it is efficient and produces generators that admit natural model-specific interpretations. As in previous works [20, 32], repeated Lie derivatives of the outputs serve as one of the building blocks in our algorithm, which we combine with the following two key technical contributions.

The first technical contribution allows us to extract observable functions from the Lie derivatives even when the dependence on the inputs is not affine. More precisely, we show that if an observable function is written as a differential rational function in uu with coefficients in C(μ,x)\mathbb{C}(\mu,x), then all coefficients of its numerator and denominator are themselves observable (Theorem 4.1). This can be viewed as an extension of the classical result for input-affine systems due to Sontag and Wang [50]. Alternatively, one could reduce the problem to the input-affine case via state augmentation. However, this would not only increase the dimension and hence the computational cost (see [27] for a related discussion), but also introduce the artificial state variables that would have to be filtered from the output of the algorithm.

Still, high-order Lie derivatives tend to be large expressions jeopardizing the efficiency and interpretability. Our second contribution reduces the required orders of Lie derivatives. We show that instead of generating the observation field from scratch using repeated Lie derivatives, one may start with some known observable functions and complete them with the Lie derivatives of lower order (Theorem 4.12). We propose to obtain such a starting set using identifiable functions, that is, observable functions depending only on the parameters, which can be read off from the input-output equations of the model [35, 37].

Our implementation builds on a recent efficient approach to computing input-output equations [13] and on our prior work on field generators simplification [11]. The latter simplification step turns out to be instrumental in making the result of the computation concise and interpretable.

The rest of the paper is organized as follows. § 2 contains the necessary background and introduces the problem formally. In § 3, we give an informal overview of the main ideas illustrating them with a specific example. § 4 contains the main theoretical results on input separation and completion of the observation field and the description of the algorithm. § 5 describes our implementation of the algorithm and demonstrates the impact of the aforementioned improvements on its efficiency. Finally, § 6 contains computational experiments and case studies involving models from the literature.

Acknowledgements. We would like to thank Oren Bassik, Hoon Hong, Alexey Ovchinnikov, Binod Pant, and Alejandro F. Villaverde for helpful discussions.

Preliminaries and problem statement

Consider an ODE system with control in the state-space form:

Σ={x′(t)=f(μ,x(t),u(t)),y(t)=g(μ,x(t),u(t)),(1)\Sigma= \left\{ \begin{aligned} x'(t) &= \mathbf{f}(\mu,x(t),u(t)),\\ y(t) &= \mathbf{g}(\mu,x(t),u(t)), \end{aligned} \right. \tag*{(1)}

where: x(t)x(t) are state variables, μ\mu are time-independent parameters, u(t)u(t) are input variables, y(t)y(t) are output variables, and f\mathbf{f} and g\mathbf{g} are vectors of elements of C(μ,x(t),u(t))\mathbb{C}(\mu,x(t),u(t)), that is, rational functions in u,x(t),u(t)u,x(t),u(t). In what follows, we may, for the sake of brevity, omit the explicit time dependence for x,y,ux,y,u.

The central notion of this paper is observability which is usually defined for the state variables or functions of them. On the other hand, for functions in parameters one typically uses term identifiability. In our context, this distinction is artificial, and we are interested in considering combinations potentially involving both states and parameter. We will, therefore, use the term observability for states and parameters.

In this section we will give a formal definition of observability using algebraic language which will be more convenient in our case. This definition is mathematically equivalent to the standard analytic one [20 Proposition 3.4] (see also [37 Definition 2.5]).

Definition 2.1 (Differential rings and fields). A differential ring (R, ′)(R,\,^\prime) is a commutative ring with a derivation ′:R→R^\prime: R \to R, that is, a map such that, for all a,b∈Ra,b \in R, (a+b)′=a′+b′(a+b)^\prime=a^\prime+b^\prime and (ab)′=a′b+ab′(ab)^\prime=a^\prime b+ab^\prime (Leibniz rule). A differential field is a differential ring that is a field. For i≥0i \ge0, a(i)a^{(i)} denotes the ii-th order derivative of a∈Ra \in R.

Example 2.2.

  • The sets of smooth or analytic functions on an interval are natural examples of differential rings with respect to the standard derivation. The set of functions meromorphic on a domain is a differential field.

  • Any polynomial differential equation in an unknown function xx can be represented as an element of an infinite-dimensional polynomial ring C[x,x′,x′′,…]\mathbb{C}[x,x^\prime,x^{\prime\prime},\ldots] in formal variables x,x′,x′′,…x,x^\prime,x^{\prime\prime},\ldots which has a natural structure of a differential ring with the derivation defined as (x(i))′=x(i+1)(x^{(i)})^\prime=x^{(i+1)} for all i≥0i \ge0. Analogously, one can define the field C(x,x′,x′′,…)\mathbb{C}(x,x^\prime,x^{\prime\prime},\ldots) with the derivation (x(i))′=x(i+1)(x^{(i)})^\prime=x^{(i+1)}.

Notation 2.3. Let xx be an element of a differential ring (e.g., a smooth function or a differential indeterminate) and h∈Z≥0h \in\mathbb{Z}_{\ge0}. We introduce

x(<h):=(x,x′,…,x(h−1)),x(∞):=(x,x′,x′′,…).x^{(<h)} := (x,x^\prime,\ldots,x^{(h-1)}), \qquad x^{(\infty)} := (x,x^\prime,x^{\prime\prime},\ldots).

x(≤h)x^{(\le h)} is defined analogously. If x=(x1,…,xn)x=(x_1,\ldots,x_n) is a tuple of elements of a differential ring, then

x(<h):=(x1(<h),…,xn(<h)),x(∞):=(x1(∞),…,xn(∞)).\mathbf{x}^{(<h)} := (x_1^{(<h)},\ldots,x_n^{(<h)}), \qquad\mathbf{x}^{(\infty)} := (x_1^{(\infty)},\ldots,x_n^{(\infty)}).

In particular, C[x(∞)]\mathbb{C}[x^{(\infty)}] and C(x(∞))\mathbb{C}(x^{(\infty)}) will denote the ring of polynomials and the field of rational functions in xx and its derivatives, respectively. Elements of this ring (resp., field) are polynomials (resp., rational functions) in finitely many of the x(∞)x^{(\infty)}.

Definition 2.4 (Observability). A function h∈C(μ,x,u(∞))h \in\mathbb{C}(\mu,\mathbf{x},\mathbf{u}^{(\infty)}) is called observable if there exist polynomials P1,P2∈C[y(∞),u(∞)]P_1,P_2 \in\mathbb{C}[\mathbf{y}^{(\infty)},\mathbf{u}^{(\infty)}] such that P2P_2 does not vanish identically on all the analytic solutions of (1) and, for every analytic solution μ∗,x∗,u∗,y∗\mu^*,\mathbf{x}^*,\mathbf{u}^*,\mathbf{y}^* of (1) on which P2P_2 does not vanish, we have

h(μ∗,x∗,u∗)=P1(y∗,u∗)P2(y∗,u∗).(2)h(\mu^*,\mathbf{x}^*,\mathbf{u}^*) = \frac{P_1(\mathbf{y}^*,\mathbf{u}^*)}{P_2(\mathbf{y}^*,\mathbf{u}^*)}. \tag*{(2)}

Remark 2.5. Intuitively, Definition 2.4 corresponds to the possibility (under the assumption of complete and noise-free data) to reconstruct the value of the function in question. On one hand, if the quality of the data allows to estimate high-order derivatives of y∗\mathbf{y}^* and u∗\mathbf{u}^*, the equality (2) already provides a theoretical approach to such estimation. But furthermore, as has been shown in [20 Proposition 3.4] (and extended in [37 Proposition 3.1]), the possibility to reconstruct the value by any means (as given by the analytic definition [20 Definition 2.3], [37 Definition 2.5]) always implies the existence of a formula (2).

The set of all observable functions is an important object describing what we in principle can know about the state of the system. Since the set of observable functions is closed under arithmetic operations (one can simply perform the corresponding arithmetic operations on the right-hand sides of (2)), it forms a field. One typically considers not this whole field but its restriction to states and parameters, thus excluding the inputs, which leads to the following definition.

Definition 2.6 (Observation field [2, 47]). For a model (1), the set

FΣ:={h∈C(μ,x)∣h is observable}(3)\mathcal{F}_{\Sigma}:=\{h \in\mathbb{C}(\mu,\mathbf{x}) \mid h \text{ is observable}\} \tag*{(3)}

is called the observation field of the model.

It may seem that, by restricting ourselves to C(μ,x)\mathbb{C}(\mu,\mathbf{x}) in (3) we may lose some interesting observable functions from C(μ,x,u(∞))\mathbb{C}(\mu,\mathbf{x},\mathbf{u}^{(\infty)}). As we shall see, this is not the case because all observable functions can be expressed in terms of FΣ\mathcal{F}_{\Sigma} and u(∞)\mathbf{u}^{(\infty)} (Theorem 4.1).

Example 2.7. Whenever the states of the model are observable and its parameters are identifiable, the observation field coincides with the whole field of rational functions, that is, FΣ=C(μ,x)\mathcal{F}_{\Sigma}=\mathbb{C}(\mu,\mathbf{x}).

Example 2.8. Consider a model with a single state x′(t)=μ1x(t)x^\prime(t)=\mu_1x(t) and single output y(t)=μ2x(t)y(t)=\mu_2x(t). The equation admits explicit solution x(t)=x0eμ1tx(t)=x_0e^{\mu_1t}, y(t)=μ2x0eμ1ty(t)=\mu_2x_0e^{\mu_1t}. Therefore, if μ2x0≠0\mu_2x_0\ne0, the knowledge of y(t)y(t) and its derivatives allows to reconstruct not only μ1\mu_1 and μ2x0\mu_2x_0 but not x0x_0. In terms of structural observability, it is possible to reconstruct μ1\mu_1 and μ2x(t)\mu_2x(t) but not x(t)x(t). Furthermore, every observable function is a combination of these, since the output is unchanged under the symmetry x↦λx,μ2↦μ2λx\mapsto\lambda x,\mu_2\mapsto\frac{\mu_2}{\lambda} for every λ≠0\lambda\ne0. Thus, the observation field is generated by μ1\mu_1 and μ2x(t)\mu_2x(t).

Remark 2.9 (On other definitions). A different definition of the observation field, as the field containing the outputs and invariant under the family of vector fields obtained from ff via evaluation of uu, also appears in the literature (e.g. [32 Definition III.3], [2 Definition 4]). Our Theorem 4.1 implies that it is equivalent to Definition 2.6.

Main problem

Given: an ODE model Σ\Sigma, as in (1), given by x′=f(μ,x,u)\mathbf{x}' = \mathbf{f}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}) and y=g(μ,x,u)\mathbf{y} = \mathbf{g}(\boldsymbol{\mu},\mathbf{x},\mathbf{u})

Find: a set of generators h1,…,hℓ∈C(μ,x)h_1,\ldots,h_\ell\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x}) of the observation field FΣ\mathcal{F}_{\Sigma}

Remark 2.10. Different solutions to the main problem may not be equally concise and useful. In the next section, for the example (4), we will construct three different sets of generators: (6), (8), and (9). The first one, (6), contains seven generators of degrees up to eleven, while the final one, (9), has only four generators of degree two, making it potentially more amenable to interpretation.

Informal description of solution

In this section we present a high-level informal description of our approach to computing generators of the observation field. To keep notation simple, we consider system as in (1) with a single output y(t)y(t) and a single input u(t)u(t). The description will be illustrated using the following example system Σ\Sigma:

{x′(t)=(μ1x(t)+μ2)2μ3+u(t),y(t)=μ4x(t)+μ5(4)\begin{cases} x'(t) = \dfrac{(\mu_1x(t)+\mu_2)^2}{\mu_3+u(t)},\\ y(t) = \mu_4x(t)+\mu_5 \tag*{(4)} \end{cases}

We recall that the Lie derivative of a function h(μ,x,u,u′,…)h(\boldsymbol{\mu},\mathbf{x},\mathbf{u},\mathbf{u}',\ldots) with respect to the vector field (1) is defined as

LΣ(h):=∑i=1nfi∂∂xi(h)+∑j=1ℓ∑k=0∞uj(k+1)∂∂uj(k)(h).\mathcal{L}_{\Sigma}(h) := \sum_{i=1}^{n} f_i\frac{\partial}{\partial x_i}(h) + \sum_{j=1}^{\ell}\sum_{k=0}^{\infty}u_j^{(k+1)}\frac{\partial}{\partial u_j^{(k)}}(h).

It represents the total derivative of hh with respect to time.

Generating all observable functions. By (1), yy is a rational function of μ,x,u\boldsymbol{\mu},\mathbf{x},\mathbf{u}. Furthermore, by computing the Lie derivative, one can show that, for any i≥0i \ge0, y(i)y^{(i)} is a rational function of μ,x,u,u′,…,u(i)\boldsymbol{\mu},\mathbf{x},\mathbf{u},u',\ldots,u^{(i)}. Then Definition 2.4 implies that all observable functions can be obtained from g,LΣ(g),LΣ2(g),…g,\mathcal{L}_{\Sigma}(g),\mathcal{L}_{\Sigma}^{2}(g),\ldots and u,u′,…u,u',\ldots by arithmetic operations. In other words, these derivatives can be viewed as the Taylor coefficients of the functions y(t)y(t) and u(t)u(t), respectively, which, in the case of analytic solutions, determine them uniquely [42].

In fact, general field-theoretic considerations show [2 Proposition 1] that it is sufficient to take only finitely many Lie derivatives. Furthermore, using [20 Theorem 3.16], [32 Corollary IV.4], or Theorem 4.12 one can show that Lie derivatives up to order ∣x∣+∣μ∣|\mathbf{x}|+|\boldsymbol{\mu}| (inclusive) suffice. For specific systems, this bound can be refined: Theorem 4.12 implies that every observable function in our running example is a rational function of u(∞)u^{(\infty)} and of the Lie derivatives of the order up to four (although ∣x∣+∣μ∣=6|\mathbf{x}|+|\boldsymbol{\mu}|=6):

μ4x+μ5⏟=g,μ4(μ1x+μ2)2μ3+u,μ4(μ1x+μ2)2(μ3+u)2(2μ1(μ1x+μ2)−u′)⏟=LΣ2(g),LΣ3(g),LΣ4(g).(5)\underbrace{\mu_4x+\mu_5}_{=g},\quad\mu_4\frac{(\mu_1x+\mu_2)^2}{\mu_3+u},\quad\underbrace{\mu_4\frac{(\mu_1x+\mu_2)^2}{(\mu_3+u)^2}\left(2\mu_1(\mu_1x+\mu_2)-u'\right)}_{=\mathcal{L}_{\Sigma}^{2}(g)},\quad\mathcal{L}_{\Sigma}^{3}(g),\quad\mathcal{L}_{\Sigma}^{4}(g). \tag*{(5)}

Since Theorem 4.12 is formulated in terms of the existence of an algebraic relation between the Lie derivatives, its condition can be verified using the observability rank condition [19].

Carving out the observation field. The outlined approach does not completely solve the main problem of computing the observation field since the produced expressions (5) may contain inputs and their derivatives. If the original system (1) is input-affine (that is, f(μ,x,u)=f0(μ,x)+f1(μ,x)u\mathbf{f}(\boldsymbol{\mu},\mathbf{x},\mathbf{u})=\mathbf{f}_0(\boldsymbol{\mu},\mathbf{x})+\mathbf{f}_1(\boldsymbol{\mu},\mathbf{x})\mathbf{u}), then the results of Sontag and Wang [50 Theorem 1] imply that the observation field is generated by the coefficients of the Lie derivatives viewed as polynomials in u(∞)u^{(\infty)}. This result is not applicable in (5), where dependency on the input is rational.

We extend the result of Sontag and Wang on ODE models with arbitrary rational dependence on inputs: if an observable rational function is written as a rational function in the input and its derivatives uu, u′u', u′′u'', …\ldots with coefficients in C(μ,x)\mathbb{C}(\mu,x), then all coefficients of its numerator and denominator are themselves observable (Theorem 4.1). This implies, in particular, that the coefficients (with respect to u(∞)u^{(\infty)}) of the Lie derivatives are observable. In the running example, for the first order Lie derivative we get

LΣ(g)=AB+u⟹A and B are observable,L_{\Sigma}(g)=\frac{A}{B+u}\quad\Longrightarrow\quad A\text{ and }B\text{ are observable,}

where A=μ4(μ1x+μ2)2A=\mu_4(\mu_1x+\mu_2)^2 and B=μ3B=\mu_3.

Furthermore, instead of successive Lie derivatives as in (5), one can alternate between taking Lie derivatives and applying Theorem 4.1. With this approach, fewer differentiation steps may be needed (that is, the condition of Theorem 4.12 would be satisfied earlier). In the context of our running example, it turns out to be sufficient to take the Lie derivative only up to order three. This results in the following generators of the observation field:

xμ4+μ5⏟=g,μ4(μ1x+μ2)2, μ3⏟coefficients of LΣ(g),2μ1μ4(μ1x+μ2)3, μ3⏟coefficients of LΣ(μ4(μ1x+μ2)2),6μ12μ4(μ1x+μ2)4, μ3⏟coefficients of LΣ(2μ1μ4(μ1x+μ2)3)(6)\underbrace{x\mu_4+\mu_5}_{=g},\quad \underbrace{\mu_4(\mu_1x+\mu_2)^2,\ \mu_3}_{\text{coefficients of }L_{\Sigma}(g)},\quad \underbrace{2\mu_1\mu_4(\mu_1x+\mu_2)^3,\ \mu_3}_{\text{coefficients of }L_{\Sigma}(\mu_4(\mu_1x+\mu_2)^2)},\quad \underbrace{6\mu_1^2\mu_4(\mu_1x+\mu_2)^4,\ \mu_3}_{\text{coefficients of }L_{\Sigma}(2\mu_1\mu_4(\mu_1x+\mu_2)^3)} \tag*{(6)}

Making the computation more efficient and the result more useful. The two preceding subsections already yield an algorithm for computing generators of the observation field (reminiscent of the one from [32]). However, in practice, high-order Lie derivatives are usually huge symbolic expressions, which makes this approach not very efficient computationally and its output too difficult to use.

Our second theoretical result, Theorem 4.12, provides a sort of scaffolding: it is sufficient to take fewer Lie derivatives given that we have already obtained some observable functions in another way. This is accomplished by determining the observable functions that depend only on the parameters, i.e., elements of the subfield FΣ∩C(μ)\mathcal{F}_{\Sigma}\cap\mathbb{C}(\mu). For such functions, observability is usually referred to as identifiability. One standard approach to this problem uses input-output equations [35], that is, differential equations satisfied by y(t)y(t) and u(t)u(t) with coefficients depending on μ\mu. For our running example, eliminating the states gives

(μ1μ5−μ2μ4)2μ4+2μ1μ2μ4−μ1μ5μ4y+μ12μ4y2−μ3y′−uy′=0(7)\frac{(\mu_1\mu_5-\mu_2\mu_4)^2}{\mu_4} +2\mu_1\frac{\mu_2\mu_4-\mu_1\mu_5}{\mu_4}y +\frac{\mu_1^2}{\mu_4}y^2 -\mu_3y' -uy'=0 \tag*{(7)}

For a sufficiently exciting input, the known functions y,y2,y′,uy′y,y^2,y',uy' are linearly independent over constants, so the coefficients in (7) can be uniquely determined. For example, this can be done by evaluating (7) at sufficiently many time points and solving the resulting linear system for the coefficients. Hence the coefficients are observable.

Theorem 4.12 implies that these coefficients can be complemented to a generating set of the observation field by considering only the output and its first order Lie derivative and eliminating the inputs as described above. The resulting generators are

μ4x+μ5, μ3, μ4(μ1x+μ2)2⏟from Lie derivatives,(μ1μ5−μ2μ4)2μ4, μ1(μ2μ4−μ1μ5)μ4, μ12μ4, μ3⏟from input-output equation xref‘(7)‘(8)\underbrace{\mu_4x+\mu_5,\ \mu_3,\ \mu_4(\mu_1x+\mu_2)^2}_{\text{from Lie derivatives}},\quad \underbrace{\frac{(\mu_1\mu_5-\mu_2\mu_4)^2}{\mu_4},\ \frac{\mu_1(\mu_2\mu_4-\mu_1\mu_5)}{\mu_4},\ \frac{\mu_1^2}{\mu_4},\ \mu_3}_{\text{from input-output equation {xref}`(7)`}} \tag*{(8)}

These expressions are simpler compared to the ones obtained by a more straightforward method (6): they have lower degree and the role of different parameter combinations is easier to see. However, there is still some room for simplification: for example, the fourth generator can be expressed in terms of the fifth and the sixth. Therefore, we finish the computation by applying a simplification algorithm developed in our previous work [11] and obtaining:

μ4x+μ5,μ3,μ12μ4,μ1μ5−μ2μ4μ1.(9)\mu_4x+\mu_5,\quad\mu_3,\quad\frac{\mu_1^2}{\mu_4},\quad\frac{\mu_1\mu_5-\mu_2\mu_4}{\mu_1}. \tag*{(9)}

Note that we could have applied this simplification procedure to our earlier result (6) and obtain the same generators. However, as we show in § 6 (see Table 2), the size of expressions typically makes this computation much harder and even out of reach.

ModelAlgorithm 1 without (Step 1)Full Algorithm 1Speedup
Example 6.10.47 s.0.2 s.2.35
Example 6.2 (y2=Qy_2 = Q)13.7 s.2.2 s.6.2
Example 6.2 (y2=vy_2 = v)161.2 s.234.6 s.0.69
Example 6.329.6 s.0.56 s.52.9
Example 6.43.7 s.2.5 s.1.48
Example 6.5>48>48 h.355.7 s.∞\infty

Table 2. Speedup achieved by (Step 1)

Formal description of solution

Splitting with respect to inputs

The goal of this section is to prove Theorem 4.1 with the following rationale behind it. While we have defined observability for functions in parameters, states, and inputs (see Definition 2.4), the theorem implies that an observable function involving inputs is a function of input-free observable functions and inputs themselves. This, in particular, shows that the observable functions inside C(μ,x)\mathbb{C}(\mu,x) as in Definition 2.6 together with u(∞)u^{(\infty)} generate all the observable functions in the sense of Definition 2.4.

Theorem 4.1. Let h∈C(μ,x,u(∞))h \in\mathbb{C}(\mu,x,u^{(\infty)}) be an observable function. We write h=ABh = \frac{A}{B}, where A,B∈C(μ,x)[u(∞)]A,B \in\mathbb{C}(\mu,x)[u^{(\infty)}] are coprime and at least one of them has a coefficient with respect to u(∞)u^{(\infty)} equal to one. Then all the coefficients of AA and BB viewed as polynomials in u(∞)u^{(\infty)} are observable.

Example 4.2. Consider the second Lie derivative of the output of the system (4) from § 3 (see (6)):

LΣ2(g)=μ4(μ1x+μ2)2(μ3+u)2(2μ1(μ1x+μ2)−u′)=2μ1μ4(μ1x+μ2)3−μ4(μ1x+μ2)2u′μ32+2μ3u+1⋅u2.\mathcal{L}_{\Sigma}^{2}(g)=\mu_{4}\frac{(\mu_{1}x+\mu_{2})^{2}}{(\mu_{3}+u)^{2}}\left(2\mu_{1}(\mu_{1}x+\mu_{2})-u'\right)=\frac{\boxed{2\mu_{1}\mu_{4}(\mu_{1}x+\mu_{2})^{3}}-\boxed{\mu_{4}(\mu_{1}x+\mu_{2})^{2}}u'}{\boxed{\mu_{3}^{2}}+\boxed{2\mu_{3}}u+\boxed{1}\cdot u^{2}}.

The coefficients with respect to u(∞)u^{(\infty)} are boxed in the equation above. One of them (in front of u2u^{2}) is equal to one, so Theorem 4.1 implies that all the other coefficients, 2μ1μ4(μ1x+μ2)32\mu_{1}\mu_{4}(\mu_{1}x+\mu_{2})^{3}, μ4(μ1x+μ2)2\mu_{4}(\mu_{1}x+\mu_{2})^{2}, μ32\mu_{3}^{2}, 2μ32\mu_{3} are observable.

In order to prove the theorem, we will first reformulate the definition of identifiability in terms of abstract differential fields.

Notation 4.3 (Differential field of an ODE system). Consider an ODE model as in (1),

x′=f(μ,x,u),y=g(μ,x,u).x' = f(\mu,x,u), \qquad y = g(\mu,x,u).

We will construct the differential field of the system as follows. We start with the differential field C(M)\mathbb{C}(M), where MM are constants (i.e., M′=0M'=0) independent over C\mathbb{C} with ∣M∣=∣μ∣|M|=|\mu|, where ∣M∣|M| denotes the number of elements in MM. Next, we adjoin infinitely many tuples of independent elements U,U′,U′′,…U,U',U'',\ldots, each of size ∣u∣|u|, and define the derivation on C(M,U(∞))\mathbb{C}(M,U^{(\infty)}) by (U(i))′=U(i+1)(U^{(i)})'=U^{(i+1)} for i≥0i\geq0. Finally, we adjoin ∣x∣|x| independent variables XX and define the derivation on C(M,X,U(∞))\mathbb{C}(M,X,U^{(\infty)}) by

X′=f(M,X,U).X'=f(M,X,U).

The resulting differential field C(M,X,U(∞))\mathbb{C}(M,X,U^{(\infty)}) contains elements YY defined by Y=g(M,X,U(∞))Y=g(M,X,U^{(\infty)}).

Lemma 4.4. Using Notation 4.3, a function h(μ,x,u)∈C(μ,x,u(∞))h(\mu,x,u)\in\mathbb{C}(\mu,x,u^{(\infty)}) is observable if and only if, the following holds inside the differential field of the system:

h(M,X,U)∈C(Y(∞),U(∞)).h(M,X,U)\in\mathbb{C}(Y^{(\infty)},U^{(\infty)}).

Remark 4.5. The elements M,X,U,YM,X,U,Y of the differential field are sometimes referred as a generic solution of the model (e.g., [36 Definition 6], [43 Section II.6]). The characterization of observability given by Lemma 4.4 is often used as a definition of observability [36 Definition 7]. Combining Lemma 4.4 with [20 Proposition 3.4] one can show that our Definition 2.4 is equivalent to the analytic definition of observability (e.g., [20 Definition 2.3]).

Lemma 4.6. Let P(μ,x,y,u)∈C[μ,x,y(∞),u(∞)]P(\mu,x,y,u)\in\mathbb{C}[\mu,x,y^{(\infty)},u^{(\infty)}] be a polynomial. Then P(M,X,Y,U)=0P(M,X,Y,U)=0 if and only if P(μ∗,x∗,y∗,u∗)=0P(\mu^{*},x^{*},y^{*},u^{*})=0 for any analytic solution μ∗,x∗,y∗,u∗\mu^{*},x^{*},y^{*},u^{*} of (1).

Proof. Assume that P(M,X,Y,U)≠0P(M,X,Y,U)\ne0. Let F\mathcal{F} be the differential subfield of C(M,X,U(∞))\mathbb{C}(M,X,U^{(\infty)}) generated by M,X,U(∞)M,X,U^{(\infty)} and the coefficients of P,f,gP,f,g over C\mathbb{C}. By Seidenberg embedding theorem [40, 46], this field can be embedded into the field of meromorphic function on some domain. Since X′=f(M,X,U)X'=f(M,X,U) and Y=g(M,X,U)Y=g(M,X,U), the images of M,X,Y,UM,X,Y,U under this embedding constitute an analytic solution of (1). Since F\mathcal{F} contains P(M,X,Y,U)≠0P(M,X,Y,U)\ne0, this solution does not vanish PP either.

Assume that P(M,X,Y,U)=0P(M,X,Y,U)=0, and let μ∗,x∗,y∗,u∗\mu^{*},x^{*},y^{*},u^{*} be an analytic solution of (1) at which PP does not vanish. By differentiating g(M,X,U)g(M,X,U) and eliminating X′X' using X′=f(M,X,U)X'=f(M,X,U), one can show by induction that, for every i≥0i\geq0, there exists gi∈C(μ,x,u(∞))g_i\in\mathbb{C}(\mu,x,u^{(\infty)}) such that Y(i)=gi(M,X,U)Y^{(i)}=g_i(M,X,U) (where g0=gg_0=g). Since every analytic solution satisfies (1), we also have (y∗)(i)=gi(μ∗,x∗,u∗)(y^{*})^{(i)}=g_i(\mu^{*},x^{*},u^{*}). Let Q∈C(μ,x,u(∞))Q\in\mathbb{C}(\mu,x,u^{(\infty)}) be the rational function obtained in the proof by replacing every gig_i by yiy_i. On the one hand, QQ is a nonzero polynomial since it does not vanish at μ∗,x∗,u∗\mu^{*},x^{*},u^{*}. On the other hand, Q(M,X,U)=0Q(M,X,U)=0 which contradicts the algebraic independence of M,X,U(∞)M,X,U^{(\infty)}. The obtained contradiction proves that P(M,X,Y,U)=0P(M,X,Y,U)=0 implies the same equality for any analytic solution of (1).

Proof of Lemma 4.4. Assume that function hh is identifiable, and let P1,P2∈C[y(∞),u(∞)]P_1,P_2\in\mathbb{C}[y^{(\infty)},u^{(\infty)}] be the corresponding differential polynomials from Definition 2.4. Let h=H1H2h=\frac{H_1}{H_2} for coprime H1,H2∈C[μ,x,u(∞)]H_1,H_2\in\mathbb{C}[\mu,x,u^{(\infty)}]. Lemma 4.6 implies that P2(Y,U)≠0P_2(Y,U)\ne0, so we can consider

h(M,X,U)−P1(Y,U)P2(Y,U)=H1(M,X,U)P2(Y,U)−H2(M,X,U)P1(Y,U)H2(M,X,U)P2(Y,U).h(M,X,U)-\frac{P_1(Y,U)}{P_2(Y,U)} =\frac{H_1(M,X,U)P_2(Y,U)-H_2(M,X,U)P_1(Y,U)}{H_2(M,X,U)P_2(Y,U)}.

Identifiability of hh implies that the product of the numerator and denominator of the latter expression vanishes for any analytic solution of (1). Then, by Lemma 4.6, it is equal to zero in the differential field. Since the denominator is nonzero in the field, the numerator vanishes implying that h(M,X,U)∈C(Y(∞),U(∞))h(M,X,U) \in\mathbb{C}(Y^{(\infty)}, U^{(\infty)}).

Assume h(M,X,U)∈C(Y,U)h(M,X,U) \in\mathbb{C}(Y,U). Then there exist coprime P1,P2∈C[y(∞),u(∞)]P_1, P_2 \in\mathbb{C}[y^{(\infty)},u^{(\infty)}] such that P2(Y,U)≠0P_2(Y,U) \ne0 and h(M,X,U)=P1(Y,U)P2(Y,U)h(M,X,U) = \frac{P_1(Y,U)}{P_2(Y,U)}. By Lemma 4.6, the numerator of h(μ,x,u)−P1(y,u)P2(y,u)h(\mu,x,u) - \frac{P_1(y,u)}{P_2(y,u)} vanishes on any analytic solution of (1). This implies identifiability of hh.

In this field-theoretic reformulation, Theorem 4.1 follows directly from the following proposition. The proposition is a general statement about differential fields of special form, so we do not use Notation 4.3.

Proposition 4.7. Let kk be a differential field of characteristic zero4 and consider the field F0\mathcal{F}_0 of differential rational functions k(u(∞))k(u^{(\infty)}). Let F2=k(x,u(∞))\mathcal{F}_2 = k(x,u^{(\infty)}) be a transcendental extension by xx with the derivation extended by

x′=f(x,u)x' = f(x,u)

for some ff, a tuple of elements from k(x,u)k(x,u). Let F1\mathcal{F}_1 be any differential field with F0⊆F1⊆F2\mathcal{F}_0 \subseteq\mathcal{F}_1 \subseteq\mathcal{F}_2. Let p∈F1p \in\mathcal{F}_1 be an element, and we write it as a differential rational function in uu, that is,

p=A(u)B(u)p = \frac{A(u)}{B(u)}

where A,B∈k(x)[u(∞)]A, B \in k(x)[u^{(\infty)}] are coprime and for at least one of AA and BB at least one of their coefficients is equal to 1. Then all the coefficients of AA and BB belong to F1\mathcal{F}_1.

The proof of Proposition 4.7 will be based on two lemmas.

Notation 4.8. For x=(x1,…,xn)x = (x_1,\ldots,x_n) and a differential polynomial PP from k[x(∞)]k[x^{(\infty)}], define the order of PP with respect to xix_i, denoted by ord⁡xiP\operatorname{ord}_{x_i} P, to be the largest jj such that xi(j)x_i^{(j)} appears in PP. If none of the derivatives of xix_i appears in PP, we set ord⁡xiP=−1\operatorname{ord}_{x_i} P = -1.

Lemma 4.9. In the notation of Proposition 4.7, assume that ord⁡u1p=h\operatorname{ord}_{u_1} p = h. Then ∂∂u1(h)(p)∈F1\frac{\partial}{\partial u_1^{(h)}}(p) \in\mathcal{F}_1.

Proof. Since F1\mathcal{F}_1 is a subfield of a finitely generated extension of F0\mathcal{F}_0, there exist q1,…,qℓq_1,\ldots,q_\ell such that F1=F0(q1,…,qℓ)\mathcal{F}_1 = \mathcal{F}_0(q_1,\ldots,q_\ell). Let HH be the maximum of the orders of p,q1,…,qℓp,q_1,\ldots,q_\ell with respect to u1u_1. Since F1\mathcal{F}_1 is a differential field, p(H−h+1)∈F1p^{(H-h+1)} \in\mathcal{F}_1. We have

p′=u1(h+1)∂∂u1(h)(p)+Q1⟹p(H−h+1)=u1(H+1)∂∂u1(h)(p)+Q2,p' = u_1^{(h+1)} \frac{\partial}{\partial u_1^{(h)}}(p) + Q_1 \qquad\Longrightarrow\qquad p^{(H-h+1)} = u_1^{(H+1)} \frac{\partial}{\partial u_1^{(h)}}(p) + Q_2,

where ord⁡u1∂∂u1(h)(p)≤h\operatorname{ord}_{u_1} \frac{\partial}{\partial u_1^{(h)}}(p) \le h, ord⁡u1Q1≤h\operatorname{ord}_{u_1} Q_1 \le h, ord⁡u1∂∂u1(h)(p)≤H\operatorname{ord}_{u_1} \frac{\partial}{\partial u_1^{(h)}}(p) \le H, and ord⁡u1Q2≤H\operatorname{ord}_{u_1} Q_2 \le H.

We write p(H−h+1)=F(q1,…,qℓ)p^{(H-h+1)} = F(q_1,\ldots,q_\ell), where F∈F0(q1,…,qℓ)F \in\mathcal{F}_0(q_1,\ldots,q_\ell). Then we take an equality

u1(H+1)∂∂u1(h)(p)+Q2=F(q1,…,qℓ)(10)u_1^{(H+1)} \frac{\partial}{\partial u_1^{(h)}}(p) + Q_2 = F(q_1,\ldots,q_\ell) \tag*{(10)}

and, considering it as an equality of rational functions in u(∞)u^{(\infty)} over k(x)k(x), perform a substitution u1(H+1)→u1(H+1)+1u_1^{(H+1)} \to u_1^{(H+1)} + 1. Since the orders of Q2,q1,…,qℓQ_2,q_1,\ldots,q_\ell in u1u_1 do not exceed HH, they will not be affected by the substitution, so we have

(u1(H+1)+1)∂∂u1(h)(p)+Q2=F~(q1,…,qℓ),(11)\left(u_1^{(H+1)} + 1\right) \frac{\partial}{\partial u_1^{(h)}}(p) + Q_2 = \widetilde{F}(q_1,\ldots,q_\ell), \tag*{(11)}

where F~\widetilde{F} is a result of applying the substitution to FF. Subtracting (10) from (11), we obtain

∂∂u1(h)(p)=F~(q1,…,qℓ)−F(q1,…,qℓ)∈F1.\frac{\partial}{\partial u_1^{(h)}}(p) = \widetilde{F}(q_1,\ldots,q_\ell) - F(q_1,\ldots,q_\ell) \in\mathcal{F}_1.

Lemma 4.10. Let K\mathcal{K} be a field of characteristic zero and K(t)\mathcal{K}(t) be the field of rational functions over K\mathcal{K} equipped with the standard derivation ∂∂t\frac{\partial}{\partial t}. Let p=a(t)b(t)∈K(t)p = \frac{a(t)}{b(t)} \in\mathcal{K}(t) be a rational function such that a(t)a(t) and b(t)b(t) are coprime and b(t)b(t) is monic. Then the field Q(t,p(∞))\mathbb{Q}(t,p^{(\infty)}) contains all the coefficients of a(t)a(t) and b(t)b(t).

Proof. We add an extra transcendental indeterminate TT. Now we will view p(t),p′(t),p′′(t),…p(t),p'(t),p''(t),\ldots as evaluations of the rational function p(T)p(T) and its derivatives at point T=tT=t. Let d:=deg⁡a+deg⁡bd:=\deg a+\deg b. Then the properties of the Padé approximation [49 Corollary 5.21] imply that the coefficients of p(T)p(T) can be computed using the extended Euclidean algorithm in the field Q(t,p(t),…,p(d)(t))\mathbb{Q}(t,p(t),\ldots,p^{(d)}(t)) and thus belong to this field.

Proof of Proposition 4.7. Let u=(u1,…,us)\mathbf{u}=(u_1,\ldots,u_s). For every 1≤i≤s1\le i\le s, we denote by hih_i the order of pp w.r.t. uiu_i. We will prove the proposition by induction in H:=h1+…+hsH:=h_1+\ldots+h_s. If H=−sH=-s, pp does not depend on u\mathbf{u} and the statement of the proposition is true. Assume that H>−sH>-s and consider any ii with hi≥0h_i\ge0. Without loss of generality, we will take i=1i=1.

Replacing AB\frac{A}{B} with BAB2\frac{BA}{B^2} if necessary, we will assume that one of the coefficients of BB with respect to u(∞)\mathbf{u}^{(\infty)} is equal to one. Let B0B_0 be a coefficient of BB as a polynomial in u1(h1)u_1^{(h_1)} such that one of the coefficients of B0B_0 with respect to u(∞)\mathbf{u}^{(\infty)} is equal to one. We denote the remaining coefficients of BB in u1(h1)u_1^{(h_1)} by B1,…,BℓBB_1,\ldots,B_{\ell_B}. The coefficients of AA in u1(h1)u_1^{(h_1)} will be denoted A0,…,AℓAA_0,\ldots,A_{\ell_A}. By applying Lemma 4.9 to pp repeatedly, we find that ∂i∂(u1(h1))i(p)∈F1\frac{\partial^i}{\partial(u_1^{(h_1)})^i}(p)\in\mathcal{F}_1 for every i≥0i\ge0. We apply Lemma 4.10 with p=pp=p, t=u1(h1)t=u_1^{(h_1)}, and K=k(x,u1(<h1),u2(∞),…,us(∞))\mathcal{K}=k(x,u_1^{(<h_1)},u_2^{(\infty)},\ldots,u_s^{(\infty)}) and obtain that B1B0,…,BℓBB0,A0B0,…,AℓAB0∈F1\frac{B_1}{B_0},\ldots,\frac{B_{\ell_B}}{B_0},\frac{A_0}{B_0},\ldots,\frac{A_{\ell_A}}{B_0}\in\mathcal{F}_1.

Coprimality of AA and BB as polynomials in u(∞)\mathbf{u}^{(\infty)} implies that B0,…,BℓB,A0,…,AℓAB_0,\ldots,B_{\ell_B},A_0,\ldots,A_{\ell_A} are coprime as well. This implies that there exists a Q\mathbb{Q}-linear combination CC of B1,…,BℓB,A0,…,AℓAB_1,\ldots,B_{\ell_B},A_0,\ldots,A_{\ell_A} coprime with B0B_0. Since ord⁡u1C,ord⁡u1B0<h\operatorname{ord}_{u_1}C,\operatorname{ord}_{u_1}B_0<h, we can apply the induction hypothesis to CB0\frac{C}{B_0} and conclude that the coefficients of B0B_0 and CC belong to F1\mathcal{F}_1, so B0∈F1B_0\in\mathcal{F}_1. Therefore, B0,…,BℓB,A0,…,AℓA∈F1B_0,\ldots,B_{\ell_B},A_0,\ldots,A_{\ell_A}\in\mathcal{F}_1 and, applying the induction hypothesis again, we conclude that their coefficients as polynomials in u(∞)\mathbf{u}^{(\infty)}, which coincide with the coefficients of AA and BB in u(∞)\mathbf{u}^{(\infty)}, belong to F1\mathcal{F}_1.

Proof of Theorem 4.1. The theorem follows from Proposition 4.7 applied to F0=C(M,U(∞))\mathcal{F}_0=\mathbb{C}(\mathcal{M},\mathbf{U}^{(\infty)}), F1=C(M,Y(∞),U(∞))\mathcal{F}_1=\mathbb{C}(\mathcal{M},\mathbf{Y}^{(\infty)},\mathbf{U}^{(\infty)}), and F2=C(M,X,U(∞))\mathcal{F}_2=\mathbb{C}(\mathcal{M},\mathbf{X},\mathbf{U}^{(\infty)}) and p=Hp=H.

Generating all observable functions

The goal of this section is to prove Theorem 4.12. For a system Σ\Sigma as defined in (1), we introduce the Lie derivative: for h∈k(μ,x,u(∞))h\in k(\boldsymbol{\mu},\mathbf{x},\mathbf{u}^{(\infty)}), we define it as

LΣ(h):=∑i=1nfi∂∂xi(h)+∑j=1ℓ∑k=0∞uj(k+1)∂∂uj(k)(h).(12)L_{\Sigma}(h):=\sum_{i=1}^{n}f_i\frac{\partial}{\partial x_i}(h)+\sum_{j=1}^{\ell}\sum_{k=0}^{\infty}u_j^{(k+1)}\frac{\partial}{\partial u_j^{(k)}}(h). \tag*{(12)}

Notation 4.11. Let p=(p1,…,pm)\mathbf{p}=(p_1,\ldots,p_m) with p1,…,pm∈C(μ,x,u(∞))p_1,\ldots,p_m\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}^{(\infty)}). Let h=(h1,…,hm)\mathbf{h}=(h_1,\ldots,h_m) with h1,…,hm∈N∪{0,−1}h_1,\ldots,h_m\in\mathbb{N}\cup\{0,-1\}.

  • We denote LΣ≤h(p)=(p1,LΣ(p1),…,LΣh1−1(p1),…,pm,LΣ(pm),…,LΣhm−1(pm))\mathcal{L}_{\Sigma}^{\leq\mathbf{h}}(\mathbf{p})=(p_1,L_{\Sigma}(p_1),\ldots,L_{\Sigma}^{h_1-1}(p_1),\ldots,p_m,L_{\Sigma}(p_m),\ldots,L_{\Sigma}^{h_m-1}(p_m)).

  • For every i∈Ni\in\mathbb{N}, we denote h+i=(h1+i,…,hm+i)\mathbf{h}+i=(h_1+i,\ldots,h_m+i). We define LΣ≤h(p)=LΣ≤h+1(p)\mathcal{L}_{\Sigma}^{\leq\mathbf{h}}(\mathbf{p})=\mathcal{L}_{\Sigma}^{\leq\mathbf{h}+1}(\mathbf{p}).

Theorem 4.12. Consider the system Σ\Sigma as in (1) with the outputs y=(y1,…,ym)\mathbf{y}=(y_1,\ldots,y_m) with yi=gi(μ,x,u)y_i=g_i(\boldsymbol{\mu},\mathbf{x},\mathbf{u}) for gi∈C(μ,x,u)g_i\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}) for i=1,…,mi=1,\ldots,m. Let E⊆C(μ,x)\mathcal{E}\subseteq\mathbb{C}(\boldsymbol{\mu},\mathbf{x}) be a subfield of the observation field closed under taking Lie derivatives. We write E=C(c)\mathcal{E}=\mathbb{C}(\mathbf{c}) with c=(c1,…,cp)∈C(μ,x)\mathbf{c}=(c_1,\ldots,c_p)\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x}).

Let h1,…,hm∈N∪{0,−1}h_1,\ldots,h_m\in\mathbb{N}\cup\{0,-1\} be such that for every i=1,…,mthevariablesi=1,\ldots,m the variables\mathbf{y},\mathbf{u}$ satisfy a differential equation over E\mathcal{E} of order hi+1h_i+1 in yiy_i and of order at most hjh_j in yjy_j for j≠ij\ne i for j=1,…,mj=1,\ldots,m. We denote h:=(h1,…,hm)\mathbf{h}:=(h_1,\ldots,h_m). Then h∈C(μ,x,u(∞))h\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}^{(\infty)}) is observable if and only if

h∈C(c,LΣ≤h+1(g),u(∞)).h\in\mathbb{C}(\mathbf{c},\mathcal{L}_{\Sigma}^{\leq\mathbf{h}+1}(\mathbf{g}),\mathbf{u}^{(\infty)}).

The following definition can be seen as a generalization of observability indices of the system [9 §2.1 and §4.6].

Definition 4.13 (Parametric profile, cf. [13 Definition 2.8]). For a system Σ\Sigma as in (1) and a subfield E⊆C(μ,x)\mathcal{E}\subseteq\mathbb{C}(\boldsymbol{\mu},\mathbf{x}) such that LΣ(E)⊆E\mathcal{L}_{\Sigma}(\mathcal{E})\subseteq\mathcal{E}, a tuple h=(h1,…,hm)∈(N∪{0,−1})m\mathbf{h}=(h_1,\ldots,h_m)\in(\mathbb{N}\cup\{0,-1\})^m is called a parametric profile for Σ\Sigma and E\mathcal{E} if

  • y,u\mathbf{y},\mathbf{u} do not satisfy any nontrivial differential equation over E\mathcal{E} of order at most hih_i in yiy_i for every i=1,…,mi=1,\ldots,m, and

  • for every i=1,…,mi=1,\ldots,m, there exists a differential equation QiQ_i over E\mathcal{E} satisfied by y,u\mathbf{y},\mathbf{u} such that ord⁡yiQi=hi+1\operatorname{ord}_{y_i}Q_i=h_i+1 and ord⁡yjQi≤hj\operatorname{ord}_{y_j}Q_i\leq h_j for every j≠ij\ne i, j=1,…,mj=1,\ldots,m.

Example 4.14 (Dependence on the ground field E\mathcal{E}). Consider the model x′(t)=μ1x(t)x'(t)=\mu_1x(t) and y(t)=x(t)y(t)=x(t). Naturally, y(t)y(t) satisfies the differential equation y′−μ1y=0y'-\mu_1y=0 over E=C(μ1)\mathcal{E}=\mathbb{C}(\mu_1), hence h=0h=0 is a parametric profile for E=C(μ1)\mathcal{E}=\mathbb{C}(\mu_1). However, if we take E=C\mathcal{E}=\mathbb{C} (that is, prohibit μ1\mu_1 to appear in the coefficients), then yy′′−(y′)2=0yy''-(y')^2=0 is the equation of the smallest order satisfied by y(t)y(t), and the parametric profile is h=1h=1.

Remark 4.15. Theorem 4.12 gives an upper bound on the required orders of Lie derivatives. There may exist a tuple h∗≤h+1\mathbf{h}^*\leq\mathbf{h}+1, with hi∗<hi+1h_i^*<h_i+1 for some i=1,…,mi=1,\ldots,m, such that the statement still holds with h+1\mathbf{h}+1 replaced by h∗\mathbf{h}^*. However, if additionally h\mathbf{h} is a parametric profile for E\mathcal{E}, for any such h∗\mathbf{h}^*, definition of parametric profile implies (see also [13 Definition 2.8]) that h≤h∗≤h+1\mathbf{h}\leq\mathbf{h}^*\leq\mathbf{h}+1.

Lemma 4.16. Let kk be a field of characteristic zero. Let k⊆E⊆F⊆k(μ,x,u(∞))k\subseteq\mathcal{E}\subseteq\mathcal{F}\subseteq k(\boldsymbol{\mu},\mathbf{x},\mathbf{u}^{(\infty)}) be field extensions, such that LΣ(E)⊆E\mathcal{L}_{\Sigma}(\mathcal{E})\subseteq\mathcal{E} and F=k(LΣ(∞)(g),u(∞))\mathcal{F}=k(\mathcal{L}_{\Sigma}^{(\infty)}(\mathbf{g}),\mathbf{u}^{(\infty)}) for g=(g1,…,gm)\mathbf{g}=(g_1,\ldots,g_m) with g1,…,gm∈k(μ,x,u)g_1,\ldots,g_m\in k(\boldsymbol{\mu},\mathbf{x},\mathbf{u}). Let h=(h1,…,hm)∈(N∪{0,−1})m\mathbf{h}=(h_1,\ldots,h_m)\in(\mathbb{N}\cup\{0,-1\})^m be such that for every i=1,…,mi=1,\ldots,m, LΣhi+1(gi)\mathcal{L}_{\Sigma}^{h_i+1}(g_i) is algebraic over E(LΣ≤h(g),u(∞))\mathcal{E}(\mathcal{L}_{\Sigma}^{\leq\mathbf{h}}(\mathbf{g}),\mathbf{u}^{(\infty)}). Then

F=E(LΣ≤h+1(g),u(∞)).\mathcal{F}=\mathcal{E}(\mathcal{L}_{\Sigma}^{\leq\mathbf{h}+1}(\mathbf{g}),\mathbf{u}^{(\infty)}).

Proof. By decrementing some hih_i’s if necessary, it follows from [25 Ch. 8, §1, Theorem 1.1] that there exists a parametric profile h∗=(h1∗,…,hm∗)\mathbf{h}^*=(h_1^*,\ldots,h_m^*) for Σ\Sigma and E\mathcal{E} such that hi∗≤hih_i^*\leq h_i for i=1,…,mi=1,\ldots,m. If the statement of the lemma holds in the case h=h∗\mathbf{h}=\mathbf{h}^*, then it holds in all cases. Thus it suffices to prove the lemma in the case h=h∗\mathbf{h}=\mathbf{h}^*. Assume h=h∗\mathbf{h}=\mathbf{h}^*.

Let Kh=E(LΣ≤h(g),u(∞))\mathcal{K}_{\mathbf{h}}=\mathcal{E}(\mathcal{L}_{\Sigma}^{\leq\mathbf{h}}(\mathbf{g}),\mathbf{u}^{(\infty)}). It suffices to show that

Kh+1=Kh+2=Kh+3=⋯ .\mathcal{K}_{\mathbf{h}+1}=\mathcal{K}_{\mathbf{h}+2}=\mathcal{K}_{\mathbf{h}+3}=\cdots.

By induction, it suffices to show that Kh+1=Kh+2\mathcal{K}_{\mathbf{h}+1}=\mathcal{K}_{\mathbf{h}+2}.

Let i∈{1,…,m}i\in\{1,\ldots,m\} be arbitrary. Per hypothesis, LΣhi+1(gi)\mathcal{L}_{\Sigma}^{h_i+1}(g_i) is algebraic over Kh\mathcal{K}_{\mathbf{h}}. Choose Q∈Kh[T]Q\in\mathcal{K}_{\mathbf{h}}[T] of minimal degree in TT (and this degree is not zero due to h=h∗\mathbf{h}=\mathbf{h}^*) such that

Q(LΣhi+1(gi))=0.Q\left(\mathcal{L}_{\Sigma}^{h_i+1}(g_i)\right)=0.

Let H>0H>0 be arbitrary. By HH applications of LΣ\mathcal{L}_{\Sigma} to this equality and using the chain rule, we obtain

LΣhi+1+H(gi)(∂Q∂T)(LΣhi+1(gi))+R(LΣhi+1(gi),…,LΣhi+H(gi))=0.\mathcal{L}_{\Sigma}^{h_i+1+H}(g_i)\left(\frac{\partial Q}{\partial T}\right)\left(\mathcal{L}_{\Sigma}^{h_i+1}(g_i)\right)+R\left(\mathcal{L}_{\Sigma}^{h_i+1}(g_i),\ldots,\mathcal{L}_{\Sigma}^{h_i+H}(g_i)\right)=0.

where R∈Kh+H[T0,…,TH−1]R\in\mathcal{K}_{\mathbf{h}+H}[T_0,\ldots,T_{H-1}]. Therefore,

LΣhi+1+H(gi)=−R(LΣhi+1(gi),…,LΣhi+H(gi))(∂Q∂T)(LΣhi+1(gi))(13)\mathcal{L}_{\Sigma}^{h_i+1+H}(g_i)=-\frac{R\left(\mathcal{L}_{\Sigma}^{h_i+1}(g_i),\ldots,\mathcal{L}_{\Sigma}^{h_i+H}(g_i)\right)}{\left(\frac{\partial Q}{\partial T}\right)\left(\mathcal{L}_{\Sigma}^{h_i+1}(g_i)\right)} \tag*{(13)}

where the denominator is not zero: since char⁡k=0\operatorname{char} k=0, ∂Q∂T\frac{\partial Q}{\partial T} is not identically zero, so if it vanished at the point, this would contradict minimality of the degree of QQ in TT. Both the numerator and the denominator on the right-hand side in (13) are polynomials in LΣhi+1(gi),…,LΣhi+H(gi)\mathcal{L}_{\Sigma}^{h_i+1}(g_i),\ldots,\mathcal{L}_{\Sigma}^{h_i+H}(g_i) with coefficients in Kh+H\mathcal{K}_{\mathbf{h}+H}, and thus both belong to Kh+H\mathcal{K}_{\mathbf{h}+H}. Therefore, LΣhi+1+H(gi)∈Kh+H\mathcal{L}_{\Sigma}^{h_i+1+H}(g_i)\in\mathcal{K}_{\mathbf{h}+H}.

We have shown that for every H>0H>0 we have LΣhi+1+H(g)∈E(LΣ≤h+H(g),u(∞))\mathcal{L}_{\Sigma}^{h_i+1+H}(\mathbf{g})\in\mathcal{E}(\mathcal{L}_{\Sigma}^{\leq\mathbf{h}+H}(\mathbf{g}),\mathbf{u}^{(\infty)}). In particular, Kh+1=Kh+2\mathcal{K}_{\mathbf{h}+1}=\mathcal{K}_{\mathbf{h}+2}, and the claim follows. □

Proof of Theorem 4.12. Subfield E\mathcal{E} and tuple h\mathbf{h} satisfy the assumptions of Lemma 4.16. We apply Lemma 4.16 to F=C(LΣ(∞)(g),u(∞))\mathcal{F}=\mathbb{C}(\mathcal{L}_{\Sigma}^{(\infty)}(\mathbf{g}),\mathbf{u}^{(\infty)}), where y=g(μ,x,u)\mathbf{y}=\mathbf{g}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}) are as in (1), which yields

F=E(LΣ≤h+1(g),u(∞)).\mathcal{F}=\mathcal{E}(\mathcal{L}_{\Sigma}^{\leq\mathbf{h}+1}(\mathbf{g}),\mathbf{u}^{(\infty)}).

Consider the differential field C(M,X,U(∞))\mathbb{C}(\mathbf{M},\mathbf{X},\mathbf{U}^{(\infty)}) of the system Σ\Sigma defined using Notation 4.3. By construction, for every i=1,…,mi=1,\ldots,m and h≥0h\geq0 repeated differentiation of gi(M,X,U)g_i(\mathbf{M},\mathbf{X},\mathbf{U}) yields the equality in this field:

Yi(h)=LΣh(gi)(M,X,U).Y_i^{(h)}=\mathcal{L}_{\Sigma}^{h}(g_i)(\mathbf{M},\mathbf{X},\mathbf{U}).

By Lemma 4.4, h∈C(μ,x,u(∞))h\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}^{(\infty)}) is observable if and only if

h(M,X,U)∈C(Y(∞),U(∞)).h(\mathbf{M},\mathbf{X},\mathbf{U})\in\mathbb{C}(\mathbf{Y}^{(\infty)},\mathbf{U}^{(\infty)}).

Therefore, hh is observable if and only if h∈Fh\in\mathcal{F}, and the claim follows. □

Where identifiability and observability meet

We combine the results of the previous sections to describe how to obtain generators of the observation field of a given model. As before, we are considering a system Σ\Sigma as in (1), given by

x′=f(μ,x,u),y=g(μ,x,u),\mathbf{x}'=\mathbf{f}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}),\quad\mathbf{y}=\mathbf{g}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}),

with parameters μ\boldsymbol{\mu}, state variables x\mathbf{x}, input variables u\mathbf{u}, and output variables y\mathbf{y}. While the Lie derivatives of the outputs generate the observation field (e.g., by Theorem 4.12), the required orders may be quite high (see § 5). Roughly speaking, the following corollary shows that, if we start with a set of observable functions of parameters, the required orders become lower (controlled by a parametric profile, see Example 4.14).

Corollary 4.17. Let E⊆C(μ)\mathcal{E}\subseteq\mathbb{C}(\boldsymbol{\mu}) be a field whose elements are observable. We write E=C(c1,…,cp)\mathcal{E}=\mathbb{C}(c_1,\ldots,c_p) with c1,…,cp∈C(μ)c_1,\ldots,c_p\in\mathbb{C}(\boldsymbol{\mu}). Let h∈(N∪{0,−1})m\mathbf{h}\in(\mathbb{N}\cup\{0,-1\})^m be a parametric profile for Σ\Sigma and E\mathcal{E}. Then, the observation field of Σ\Sigma is generated by c1,…,cpc_1,\ldots,c_p and by the coefficients of g,…,LΣh+1(g)\mathbf{g},\ldots,\mathcal{L}_{\Sigma}^{\mathbf{h}+1}(\mathbf{g}) in the sense of Theorem 4.1.

Proof. Let q1,…,qη∈C(μ,x)q_1,\ldots,q_\eta\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x}) denote the coefficients of g,…,LΣh+1(g)\mathbf{g},\ldots,\mathcal{L}_{\Sigma}^{\mathbf{h}+1}(\mathbf{g}) in the sense of Theorem 4.1. It suffices to show that

h∈C(μ,x) is observable⟺h∈C(c1,…,cp,q1,…,qη).h\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x})\text{ is observable}\quad\Longleftrightarrow\quad h\in\mathbb{C}(c_1,\ldots,c_p,q_1,\ldots,q_\eta).

Assume h∈C(μ,x)h\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x}) is observable. Let c=(c1,…,cp)\mathbf{c}=(c_1,\ldots,c_p) and q=(q1,…,qη)\mathbf{q}=(q_1,\ldots,q_\eta). It suffices to show that h∈C(c,q)h\in\mathbb{C}(\mathbf{c},\mathbf{q}). Voilà:

h∈C(c,LΣh+1(g),u(∞))(by xref‘Theorem 4.12‘)h\in\mathbb{C}\left(\mathbf{c},\mathcal{L}_{\Sigma}^{\mathbf{h}+1}(\mathbf{g}),\mathbf{u}^{(\infty)}\right)\qquad\text{(by {xref}`Theorem 4.12`)}
⟺h∈C(c,q,u(∞))(since LΣh+1(g)⊆C(q,u(∞)))\Longleftrightarrow\quad h\in\mathbb{C}\left(\mathbf{c},\mathbf{q},\mathbf{u}^{(\infty)}\right)\qquad\text{(since }\mathcal{L}_{\Sigma}^{\mathbf{h}+1}(\mathbf{g})\subseteq\mathbb{C}(\mathbf{q},\mathbf{u}^{(\infty)})\text{)}
⟺h∈C(c,q)□\Longleftrightarrow\quad h\in\mathbb{C}(\mathbf{c},\mathbf{q})\qquad\Box

Remark 4.18 (Prolongations only approach). The observation field can be generated by the coefficients of g,…,LΣ∣x∣+∣μ∣(g)\mathbf{g},\ldots,\mathcal{L}_{\Sigma}^{|\mathbf{x}|+|\boldsymbol{\mu}|}(\mathbf{g}) in the sense of Theorem 4.1 alone, which follows from Corollary 4.17 with E:=C\mathcal{E}:=\mathbb{C}.

Remark 4.19 (Encoding parameters). Given a model Σ\Sigma with parameters, one can convert it to an equivalent model without parameters by making every parameter a state variable satisfying the equation x′(t)=0x'(t)=0. Then, Corollary 4.17 reduces to the approach from Remark 4.18.

Example 4.20. Consider a model of the DC motor [12 Chapter VI, Exercise 6.28]:

{Li′(t)=−Ri(t)−Keω(t)+u(t),Jω′(t)=Kti(t)−Bω(t),y(t)=ω(t).\begin{cases} L i'(t)=-R i(t)-K_e\omega(t)+u(t),\\ J\omega'(t)=K_t i(t)-B\omega(t), \end{cases} \qquad y(t)=\omega(t).

Here, x=(i(t),ω(t))\mathbf{x}=(i(t),\omega(t)) and μ=(R,L,Ke,Kt,J,B)\boldsymbol{\mu}=(R,L,K_e,K_t,J,B). The output is y(t)y(t) and u(t)u(t) is the input. If Ω(s)\Omega(s), U(s)U(s) are the Laplace transforms of y(t)y(t) and u(t)u(t), respectively, then the transfer function [12 Chapter VI] of the system is

Ω(s)U(s)=1c1s2+c2s+c3,\frac{\Omega(s)}{U(s)}=\frac{1}{c_1s^2+c_2s+c_3},

where

(c1,c2,c3)=(JLKt,JR+BLKt,BR+KeKtKt).(c_1,c_2,c_3)=\left(\frac{JL}{K_t},\frac{JR+BL}{K_t},\frac{BR+K_eK_t}{K_t}\right).

Then c1,c2,c3c_1,c_2,c_3 are identifiable since for linear models the coefficients of the transfer function coincide with the coefficients of the input-output equation and also due to [38 Theorem 1]. Equivalently, we could have obtained these coefficients from the input-output equation itself by eliminating the states. We take E:=C(c1,c2,c3)\mathcal{E}:=\mathbb{C}(c_1,c_2,c_3). To apply Corollary 4.17 we also need to know a parametric profile h\mathbf{h} for E\mathcal{E} (as in Definition 4.13) or an upper bound for it. Note that h≤1\mathbf{h}\leq1 since y(t)y(t) and u(t)u(t) satisfy a differential equation over E\mathcal{E} of order two. It follows from Corollary 4.17 that the observation field is generated by c1,c2,c3c_1,c_2,c_3 together with the coefficients of ω,LΣ2(ω)\omega,\mathcal{L}_{\Sigma}^{2}(\omega) in terms of uu. On the other hand, with the straightforward method (that applies Corollary 4.17 with E=C\mathcal{E}=\mathbb{C}, as in Remark 4.18) one needs to take derivatives up to order ∣x∣+∣μ∣=8|\mathbf{x}|+|\boldsymbol{\mu}|=8.

Main algorithm

The corollary above yields the following approach to computing the observation field:

Algorithm 1 Computing the observation field

Input: a system Σ\Sigma, given by x′=f(μ,x,u)\mathbf{x}' = \mathbf{f}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}), y=g(μ,x,u)\mathbf{y} = \mathbf{g}(\boldsymbol{\mu},\mathbf{x},\mathbf{u}), as in (1).

Output: elements of C(μ,x)\mathbb{C}(\boldsymbol{\mu},\mathbf{x}) generating the observation field of Σ\Sigma.

(Step 1) Compute some observable functions of parameters (c1,…,cp∈C(μ)(c_1,\ldots,c_p \in\mathbb{C}(\boldsymbol{\mu}) in Corollary 4.17):

(a) Compute a parametric profile h=(h1,…,hm)\mathbf{h}=(h_1,\ldots,h_m) for Σ\Sigma and C(μ)\mathbb{C}(\boldsymbol{\mu}) and the corresponding differential equations Q1,…,QmQ_1,\ldots,Q_m as in Definition 4.13, for instance, using differential elimination [13 Algorithm 4.5].

(b) Normalize each of Q1,…,QmQ_1,\ldots,Q_m so that each has at least one coefficient with respect to y,u\mathbf{y},\mathbf{u} equal to one. Let c1,…,cp∈C(μ)c_1,\ldots,c_p \in\mathbb{C}(\boldsymbol{\mu}) be the coefficients of Q1,…,QmQ_1,\ldots,Q_m with respect to y,u\mathbf{y},\mathbf{u}.

(c) Check if c1,…,cpc_1,\ldots,c_p are observable, for instance, using [13 Algorithm 5.2] (see also [37 Lemma 1]). If they are not, then convert parameters into state variables as in Remark 4.19, and return the output of Algorithm 1 applied to the modified model.

(Step 2) Compute q1,…,qη∈C(μ,x)q_1,\ldots,q_\eta\in\mathbb{C}(\boldsymbol{\mu},\mathbf{x}), the coefficients of the Lie derivatives g,…,LΣh+1(g)\mathbf{g},\ldots,L_{\Sigma}^{h+1}(\mathbf{g}) in the sense of Theorem 4.1, by repeatedly differentiating the outputs g\mathbf{g} symbolically.

(Step 3) Return a simple generating set of the field generated by c1,…,cp,q1,…,qηc_1,\ldots,c_p,q_1,\ldots,q_\eta by applying the simplification algorithm [11 Algorithm 8] (alternatively, use [30 Steps 1–3 of the Algorithm in Section 5]).

Remark 4.21 (Early termination). In (Step 2), one may alternate between taking Lie derivatives, extracting their coefficients using Theorem 4.1, and taking Lie derivatives of these coefficients. Roughly speaking, this amounts to augmenting the model on the fly with outputs corresponding to these coefficients. As a result, differentiation orders lower than h+1h+1 might suffice. By Theorem 4.12, one may terminate once higher order derivatives are algebraic over the field generated so far; this can be tested, for example, using [14 Theorem 2.2].

Remark 4.22 (Observability of individual parameters and states). A particular state xix_i or parameter μj\mu_j is observable if and only if it is among the generators returned by Algorithm 1. This is due to the following property of the simplification procedure from [11]: if the field generated by g1(z),…,gℓ(z)∈Q(z)g_1(\mathbf{z}),\ldots,g_\ell(\mathbf{z}) \in\mathbb{Q}(\mathbf{z}) with z=(z1,…,zm)\mathbf{z}=(z_1,\ldots,z_m) contains a variable, say z1z_1, then the simplified set of generators will contain z1z_1.

Proposition 4.23. Algorithm 1 is correct.

Proof. First consider the case when c1,…,cpc_1,\ldots,c_p are observable in (Step 1)(c). Note h\mathbf{h} is a parametric profile for Σ\Sigma and C:=C(c1,…,cp)\mathbb{C}:=\mathbb{C}(c_1,\ldots,c_p). Then the observation field is generated by c1,…,cpc_1,\ldots,c_p and q1,…,qηq_1,\ldots,q_\eta due to Corollary 4.17.

Now consider the case when c1,…,cpc_1,\ldots,c_p are not observable in (Step 1)(c), so Algorithm 1 is applied again to the modified model. Consider this second application. Since the modified model has no parameters, c1,…,cpc_1,\ldots,c_p are elements of C\mathbb{C}, and are thus observable. Then the observation field of the modified model is generated by c1,…,cpc_1,\ldots,c_p and q1,…,qηq_1,\ldots,q_\eta due to Corollary 4.17. The observation fields of the original and modified models coincide, and the claim follows.

Example 4.24. Consider a variant of the Lotka-Volterra model,

x1′(t)=x1(t)⋅(α+βx2(t)−κu(t)),x2′(t)=x2(t)⋅(−δ+γx1(t)),y(t)=x1(t),\begin{aligned} x_1'(t) &= x_1(t)\cdot\left(\alpha+\beta x_2(t)-\kappa u(t)\right),\\ x_2'(t) &= x_2(t)\cdot\left(-\delta+\gamma x_1(t)\right), \end{aligned} \qquad y(t)=x_1(t),

where u(t)u(t) is the input and y(t)y(t) is the output. We have x=(x1(t),x2(t))\mathbf{x}=(x_1(t),x_2(t)) and μ=(α,β,δ,γ,κ)\boldsymbol{\mu}=(\alpha,\beta,\delta,\gamma,\kappa).

In the first step, Algorithm 1 finds some observable functions of parameters. By eliminating x1x_1 and x2x_2, it calculates that y(t)y(t) and u(t)u(t) satisfy a differential equation of order two:

yy′′−(y′)2+c1yy′−c2y2y′+c3y2u′+c4y3−c5y2−c6y3u+c7y2u=0yy''-(y')^2+c_1yy'-c_2y^2y'+c_3y^2u'+c_4y^3-c_5y^2-c_6y^3u+c_7y^2u=0

where (c1,…,c7)=(δ,γ,κ,αγ,αδ,γκ,δκ)(c_1,\ldots,c_7)=(\delta,\gamma,\kappa,\alpha\gamma,\alpha\delta,\gamma\kappa,\delta\kappa). It follows that the parametric profile for C(μ)\mathbb{C}(\boldsymbol{\mu}) is h=1\mathbf{h}=1. It can be checked by [13 Algorithm 5.2] that the coefficients c1,…,c7c_1,\ldots,c_7 of this input-output equation are identifiable, which concludes (Step 1) of the algorithm.

At (Step 2), the coefficients of x1x_1, LΣ(x1)\mathcal{L}_{\Sigma}(x_1), LΣ2(x1)\mathcal{L}_{\Sigma}^{2}(x_1) in terms of uu, u′u' are computed. This yields that q1,…,q7q_1,\ldots,q_7 correspond to the underlined coefficients in:

x1=x1LΣ(x1)=−(κx1)u+(αx1+βx1x2),x_1=x_1 \qquad\mathcal{L}_{\Sigma}(x_1)=-(\kappa x_1)u+(\alpha x_1+\beta x_1x_2),
LΣ2(x1)=(κ2x1)u2−(2ακx1+2βκx1x2)u−(κx1)u′+(α2x1+(2αβ−βδ)x1x2+β2x1x22+βγx12x2)\mathcal{L}_{\Sigma}^{2}(x_1)=(\kappa^2x_1)u^2-(2\alpha\kappa x_1+2\beta\kappa x_1x_2)u-(\kappa x_1)u'+(\alpha^2x_1+(2\alpha\beta-\beta\delta)x_1x_2+\beta^2x_1x_2^2+\beta\gamma x_1^2x_2)

Using Corollary 4.17 with E:=C(c1,…,c7)\mathcal{E}:=\mathbb{C}(c_1,\ldots,c_7), we may conclude that the observation field of the model is generated by c1,…,c7c_1,\ldots,c_7 and q1,…,q7q_1,\ldots,q_7, that is, FΣ=C(c1,…,c7,q1,…,q7)\mathcal{F}_{\Sigma}=\mathbb{C}(c_1,\ldots,c_7,q_1,\ldots,q_7).

The final step is to apply algebraic simplification to FΣ\mathcal{F}_{\Sigma}. Using [11], Algorithm 8, we obtain that FΣ\mathcal{F}_{\Sigma} can be generated simply by

α, γ, δ, κ, x1, βx2,\alpha,\ \gamma,\ \delta,\ \kappa,\ x_1,\ \beta x_2,

which is the output of Algorithm 1.

Implementation

Algorithm 1 has been implemented in the StructuralIdentifiability.jl package in Julia. Since we use the algorithms from [11], our implementation is Monte-Carlo probabilistic in the sense that it may return an incorrect result but the user can specify any positive upper bound on the probability of this event. In this section, we will apply our implementation to several models (presented in detail in the next section) arising in applications with varying complexity and type of control and present some statistics illustrating the design choices made in Algorithm 1. The computations were done using the version 0.5.21 of the software on MacBook Air with Intel i5 core (1,6 GHz) and 16 Gb RAM5

First, we will examine the impact of the identifiable function computation at (Step 1) and of the final result simplification (which are optional as far as only the correctness of the output is concerned) on the complexity of the obtained generators. To this end, we collect in Table 1 the highest degrees and term counts among the computed generators for various versions of the approach:

ModelProlongations onlyProlongations onlyProlongations only(Step 1) and prolongations(Step 1) and prolongations(Step 1) and prolongationsAlgorithm 1Algorithm 1
ordersdegree# termsordersdegree# termsdegree# terms
Example 6.1714313361533
Example 6.2 (y2=Qy_2=Q)(5, 1)25958(3, 1)139422
Example 6.2 (y2=vy_2=v)(5, 1)25958(3, 1)139432
Example 6.3555672191023
Example 6.46211624144134
Example 6.5122249381361074857

Table 1. Degrees and sizes of generators of the observation field

  • Prolongations only (the combination of Remarks 4.18 and 4.21) stands for using the approach described in the first two subsections of § 3: the generators are obtained only by taking Lie derivatives and applying Theorem 4.1 (similarly to [32]). The orders columns display the order of differentiation needed for every output.

  • (Step 1) and prolongations refers to performing the steps of Algorithm 1 (and orders column contains the parametric profile) but without the final simplification.

  • Algorithm 1 corresponds to the full algorithm. We do not list the orders of the Lie derivatives since they coincide with the ones obtained in the previous item.

We observe that starting with identifiable functions of parameters allows to considerably reduce the orders of Lie derivatives and, consequently, the complexity of the expressions. Furthermore, the final simplification also allows to reduce the complexity even further and quite dramatically. When considering examples one by one in § 6, we will show that the resulting generators are not only compact but also often admit a domain-specific interpretation.

While the final simplification could have been applied to the generators computed using Prolongations only approach, Table 2 shows that computing identifiable function of parameters first in (Step 1) typically speeds up the computations (sometimes turning intractable problems into tractable ones!).

Examples

In this section, we apply our algorithm to several models from the literature and discuss interpretations of the obtained generators of the observation field.

Example 6.1 (Cholera model). We describe how our algorithm was applied in [39 Section 3.4] to study identifiability of the following model describing the spread of cholera proposed in [15]:

{S′(t)=−βWS(t)W(t)−βIS(t)I(t),I′(t)=βWS(t)W(t)+βIS(t)I(t)−γI(t),W′(t)=αI(t)−ζW(t),y(t)=W(t).(14)\begin{cases} S'(t) = -\beta_W S(t)W(t) - \beta_I S(t)I(t),\\ I'(t) = \beta_W S(t)W(t) + \beta_I S(t)I(t) - \gamma I(t),\\ W'(t) = \alpha I(t) - \zeta W(t),\\ y(t) = W(t). \tag*{(14)} \end{cases}

In this model S(t)S(t), I(t)I(t), W(t)W(t) are the state variables corresponding to susceptible and infected population and to the pathogen concentration in water, respectively. Parameters are α\alpha, βW\beta_W, βI\beta_I, γ\gamma, ζ\zeta, and the observed quantity is the pathogen concentration WW. Our Algorithm 1 finds the following generators for the observation field:

αβI, γ+ζ, γζ, αβW+βIζβI, W(t), βIS(t), βII(t)+βWW(t).(15)\frac{\alpha}{\beta_I},\ \gamma+\zeta,\ \gamma\zeta,\ \frac{\alpha\beta_W+\beta_I\zeta}{\beta_I},\ W(t),\ \beta_I S(t),\ \beta_I I(t)+\beta_W W(t). \tag*{(15)}

In this list, the first four quantities depending on parameters only are computed at (Step 1) and the time-dependent identifiable functions are obtained by performing the rest of the computation. By Remark 4.22, we see that none of S(t)S(t), I(t)I(t), α\alpha, βI\beta_I, βW\beta_W, γ\gamma, ζ\zeta is observable.

We remark that the quantities (15) produced by the algorithm allow for domain-specific interpretation. For example, βII(t)+βWW(t)\beta_I I(t)+\beta_W W(t) is the combined per-capita infection risk from human and contaminated water exposure (also called the force of infection) and βIS(t)\beta_I S(t) reflects the fact that measuring only the pathogen concentration one cannot distinguish between large population with low transmission rate and small population with high transmission rate. Furthermore, one can then use (15) to see that the effective reproduction number is identifiable. More precisely, for this model this time-dependent quantity is equal to αβW+βIζγζS(t)\frac{\alpha\beta_W+\beta_I\zeta}{\gamma\zeta}S(t) and it naturally decomposes as a product (see [39 Section 3.4]):

αβW+βIζγζS(t)=αβW+βIζβI⋅1γζ⋅βIS(t).\frac{\alpha\beta_W+\beta_I\zeta}{\gamma\zeta}S(t) = \frac{\alpha\beta_W+\beta_I\zeta}{\beta_I}\cdot\frac{1}{\gamma\zeta}\cdot\beta_I S(t).

Other applications of (15) in [39] include determining possible complementary data sufficient to restore identifiability.

Example 6.2 (Cancer modeling). We consider the following model of prostate cancer reproduced from [41 Eq. (12)-(15)], which incorporates the impact of androgen on the tumor growth (for the history and variations of the model, see the references in [41]).

{x′(t)=μm(1−qQ(t))x(t)−(v(t)RR+Q(t)+δx(t))x(t),v′(t)=−dv(t),Q′(t)=(γ1v(t)+γ2)(Qm−Q(t))−μm(Q(t)−q),P′(t)=bQ(t)+σx(t)Q(t)−ϵP(t).\begin{cases} x'(t) = \mu_m\left(1-\frac{q}{Q(t)}\right)x(t)-\left(v(t)\frac{R}{R+Q(t)}+\delta x(t)\right)x(t),\\ v'(t) = -dv(t),\\ Q'(t) = (\gamma_1 v(t)+\gamma_2)(Q_m-Q(t))-\mu_m(Q(t)-q),\\ P'(t) = bQ(t)+\sigma x(t)Q(t)-\epsilon P(t). \end{cases}

In this model, x(t)x(t) is the cancer population, Q(t)Q(t) is the intracellular androgen level, v(t)v(t) is the androgen-dependent death rate for the cancer cells, and P(t)P(t) is the serum PSA (prostate-specific antigen) level. The model is controlled through the drug injection u(t)u(t). The main measurement protocol considered in [41] involved two outputs y1(t)=P(t)y_1(t)=P(t), y2(t)=Q(t)y_2(t)=Q(t). In this case, the observation field is generated by σx(t)\sigma x(t), δσ\frac{\delta}{\sigma}, and the remaining states and parameters. If one takes instead, for example, y2(t)=v(t)y_2(t)=v(t), a generating set of observation field will be

bq, bR, bQm, bQ(t), δx(t), bδσbq,\ bR,\ bQ_m,\ bQ(t),\ \delta x(t),\ \frac{b\delta}{\sigma}

and the remaining states and parameters (γ1,γ2,ε,d,μm,v(t),P(t))(\gamma_1,\gamma_2,\varepsilon,d,\mu_m,v(t),P(t)). In both cases, the observable functions are products of integer powers of parameters and states meaning that all the symmetries of the model are scaling transformations.

Example 6.3 (Controlling enzyme kinetics). We consider a reaction with two competing substrates, S1(t)S_1(t) and S2(t)S_2(t), following the Michaelis-Menten kinetics (see, e.g. [45 Section 3]) non-affinely controlled by I(t)I(t), a common uncompetitive inhibitor [21 Eq. (4)]. By taking the total amount of the substrates as the output, we obtain:

{S1′(t)=V1S1(t)1+K1S1(t)(1+I(t)L1)+K2S2(t)(1+I(t)L2),S2′(t)=V2S2(t)1+K1S1(t)(1+I(t)L1)+K2S2(t)(1+I(t)L2),y(t)=S1(t)+S2(t).\begin{cases} S_1'(t)=\frac{V_1S_1(t)}{1+K_1S_1(t)\left(1+\frac{I(t)}{L_1}\right)+K_2S_2(t)\left(1+\frac{I(t)}{L_2}\right)},\\ S_2'(t)=\frac{V_2S_2(t)}{1+K_1S_1(t)\left(1+\frac{I(t)}{L_1}\right)+K_2S_2(t)\left(1+\frac{I(t)}{L_2}\right)},\\ y(t)=S_1(t)+S_2(t). \end{cases}

Applying Algorithm 1, we obtain the following generators of the observation field:

V1+V2,L1+L2,K1+K2,S1(t)+S2(t),L1V2+L2V1,V_1+V_2,\qquad L_1+L_2,\qquad K_1+K_2,\qquad S_1(t)+S_2(t),\qquad L_1V_2+L_2V_1,
V1V2,L1L2,K1K2,S1(t)S2(t),V1S1(t)+V2S2(t).V_1V_2,\qquad L_1L_2,\qquad K_1K_2,\qquad S_1(t)S_2(t),\qquad V_1S_1(t)+V_2S_2(t).

The presence of sums and product in the list can be naturally explained by the symmetry of the model exchanging the substrates. Interestingly, in addition to the observed sum S1(t)+S2(t)S_1(t)+S_2(t), we also see that the product S1(t)S2(t)S_1(t)S_2(t) and the sum weighted by the limiting rates V1S1(t)+V2S2(t)V_1S_1(t)+V_2S_2(t) is among the generators.

Example 6.4 (Modeling a disease with a relapse after remission). The following compartmental epidemiological model was considered in [10 Eq (4)] in order to take into account the possibility of a relapse after a period of remission:

{S′(t)=−βS(t)I(t)N−u(t)S(t)N,L′(t)=βS(t)I(t)N−αL(t),I′(t)=αL(t)−γI(t)+σQ(t),Q′(t)=(1−ν)γI(t)−σQ(t),y(t)=I(t)N.\begin{cases} S'(t)=-\beta\frac{S(t)I(t)}{N}-u(t)\frac{S(t)}{N},\\ L'(t)=\beta\frac{S(t)I(t)}{N}-\alpha L(t),\\ I'(t)=\alpha L(t)-\gamma I(t)+\sigma Q(t),\\ Q'(t)=(1-\nu)\gamma I(t)-\sigma Q(t),\\ y(t)=\frac{I(t)}{N}. \end{cases}

The states S(t)S(t), L(t)L(t), I(t)I(t), Q(t)Q(t) correspond to the susceptible, exposed, infectious individuals and the individuals at the remission stage, respectively. The model is controlled by the vaccination rate u(t)u(t) which is assumed to be sufficiently generic. Computation with Algorithm 1 returns the following generating observable functions:

σ, β, N, α+γ, ανγ, αγ+(ν−1)γσ, I(t), αS(t), (α−σ)Q(t), αL(t)−γI(t)+σQ(t).\sigma,\ \beta,\ N,\ \alpha+\gamma,\ \alpha\nu\gamma,\ \alpha\gamma+(\nu-1)\gamma\sigma,\ I(t),\ \alpha S(t),\ (\alpha-\sigma)Q(t),\ \alpha L(t)-\gamma I(t)+\sigma Q(t).

The observable functions involving state variables include the total number of infectious individuals (I(t)I(t)) as well as the rate of its change (αL(t)−γI(t)+σQ(t)\alpha L(t)-\gamma I(t)+\sigma Q(t); can be thought as “new daily cases”). The element αS(t)\alpha S(t) reflects the confounding of larger population with longer average latent period (which is equal to 1α\frac{1}{\alpha}) and smaller population with shorter latent period. Similarly, (α−σ)Q(t)(\alpha-\sigma)Q(t) reflects a tradeoff between the number of individuals at the remission stage and the difference between the remission and latent periods.

Example 6.5 (EAIHRD epidemiology model). Consider an epidemiological model from [18] (see also [8]) which takes into account multiple aspects such as hospitalization, mortality, and latent period. The model is given by the following ODE system ([18 Eq. (1)–(6)] and [8, eq:6a–(6f)]):

{A′(t)=aE(t)−r1A(t),I′(t)=sE(t)−(h+r2)I(t),H′(t)=hI(t)−(r3+d)H(t),R′(t)=r1A(t)+r2I(t)+r3H(t),D′(t)=dH(t),E′(t)=(N−A(t)−I(t)−H(t)−R(t)−D(t)−E(t))(c1A(t)+c2I(t))−(a+s)E(t),y(t)=D(t).(16)\begin{cases} A'(t)=aE(t)-r_1A(t),\\ I'(t)=sE(t)-(h+r_2)I(t),\\ H'(t)=hI(t)-(r_3+d)H(t),\\ R'(t)=r_1A(t)+r_2I(t)+r_3H(t),\\ D'(t)=dH(t),\\ E'(t)=(N-A(t)-I(t)-H(t)-R(t)-D(t)-E(t))(c_1A(t)+c_2I(t))-(a+s)E(t),\\ y(t)=D(t). \tag*{(16)} \end{cases}

The states A(t)A(t), I(t)I(t), H(t)H(t), R(t)R(t), D(t)D(t), E(t)E(t) correspond to asymptomatic, infected, hospitalized, recovered, dead, and exposed individuals. The observation function is taken to be the cumulative mortality. Algorithm 1 produces the following list of generators of the observation field:

r1,d+r3,a+h+s+r2,ah+ar2+hs+sr2,dhsac1+sc2,r_1,\quad d+r_3,\quad a+h+s+r_2,\quad ah+ar_2+hs+sr_2,\quad\frac{dhs}{ac_1+sc_2},
D(t),dH(t),dhI(t),c1A(t)+c2I(t),a+s+sE(t)I(t),D(t),\quad dH(t),\quad dhI(t),\quad c_1A(t)+c_2I(t),\quad a+s+\frac{sE(t)}{I(t)},
c2(h−r1+r2)I(t)−(ac1+sc2)E(t),s(A(t)+H(t)+R(t)+D(t)−N)I(t)−a.c_2(h-r_1+r_2)I(t)-(ac_1+sc_2)E(t),\quad\frac{s(A(t)+H(t)+R(t)+D(t)-N)}{I(t)}-a.

Several of the elements of the list admit natural interpretation. For example, since only the cumulative mortality is observed, hospitalized and infected individuals appear in dH(t)dH(t) and dhI(t)dhI(t), respectively, scaled by the mortality and hospitalization rates. The expression c1A(t)+c2I(t)c_1A(t)+c_2I(t) is the force of infection (i.e. combined per-capita infection risk, see Example 6.1). Furthermore, the functions involving parameters only obtained by our algorithm are in an agreement with new parameters FF, R2R_2, R3R_3, C1C_1, C2C_2 introduced by hand in [18 Eq. (9)] (see also [8 Section 2]). For example, d+r3=R3d+r_3=R_3, a+h+r2+s=F+R2a+h+r_2+s=F+R_2, ah+ar2+hs+sr2=FR2ah+ar_2+hs+sr_2=FR_2, and dhsac1+sc2\frac{dhs}{ac_1+sc_2} is the harmonic mean of C1C_1 and C2C_2.

References

  1. [1]S. Balakrishnan, A. Hasnain, R. Egbert, and E. Yeung. Data-driven observability decomposition with koopman operators for optimization of output functions of nonlinear systems, 2022. URL https: //arxiv.org/abs/2210.098343.
  2. [2]Z. Bartosiewicz. Rational systems and observation fields. Systems & Control Letters, 9(5):379–386, 1987. URL https://doi.org/10.1016/0167-6911(87)90066-1.
  3. [3]Z. Bartosiewicz. Minimal polynomial realizations. Control Signal Systems, 1:227–237, 1988. URL https://doi.org/10.1007/BF02551285.
  4. [4]R. Bellman and K. Åström. On structural identifiability. Mathematical Biosciences, 7(3):329–339, 1970. URL https://doi.org/10.1016/0025-5564(70)90132-X.
  5. [5]J. G. Borgqvist, A. P. Browning, F. Ohlsson, and R. E. Baker. Framing local structural identifiability and observability in terms of parameter-state symmetries, 2026. URL https://arxiv.org/ abs/2603.11387.arxiv.org/abs/2603.11387
  6. [6]R. W. Brockett. Nonlinear systems and nonlinear estimation theory. In M. Hazewinkel and J. C. Willems, editors, Stochastic Systems: The Mathematics of Filtering and Identification and Applica- tions, pages 441–477, Dordrecht, 1981. Springer Netherlands.DOI
  7. [7]M. N. Chatzis, E. N. Chatzi, and A. W. Smyth. On the observability and identifiability of nonlinear structural and mechanical systems. Structural Control and Health Monitoring, 22(3):574–593, 2014. URL http://dx.doi.org/10.1002/stc.1690.
  8. [8]Q. Y. Chen, Z. Rapti, Y. Drossinos, J. Cuevas-Maraver, G. A. Kevrekidis, and P. G. Kevrekidis. Prac- tical identifiability and parameter estimation of compartmental epidemiological models, 2024. URL https://arxiv.org/abs/2406.17827.
  9. [9]G. Conte, C. H. Moog, and A. Maria Perdon. Algebraic Methods for Nonlinear Control Systems. Springer, 2007. URL https://doi.org/10.1007/978-1-84628-595-0.
  10. [10]E. A. Dankwa, A. F. Brouwer, and C. A. Donnelly. Structural identifiability of compartmental models for infectious disease transmission is influenced by data type. Epidemics, 41:100643, Dec. 2022. URL http://dx.doi.org/10.1016/j.epidem.2022.100643.
  11. [11]A. Demin and G. Pogudin. Simple generators of rational function fields, 2026. URL https:// arxiv.org/abs/2602.10878.arxiv.org/abs/2602.10878
  12. [12]J. J. DiStefano, I. J. Williams, and A. R. Stubberud. Feedback and Control Systems. McGraw-Hill, New York, 2 edition, 2012.
  13. [13]R. Dong, C. Goodbrake, H. A. Harrington, and G. Pogudin. Differential elimination for dynamical models via projections with applications to structural identifiability. SIAM Journal on Applied Algebra and Geometry, 7(1):194–235, 2023. URL https://doi.org/10.1137/22M1469067.
  14. [14]R. Ehrenborg and G.-C. Rota. Apolarity and canonical forms for homogeneous polynomials. European Journal of Combinatorics, 14(3):157–181, 1993. URL https://doi.org/10.1006/eujc.1993.1022.
  15. [15]M. C. Eisenberg, S. L. Robertson, and J. H. Tien. Identifiability and estimation of multiple transmission pathways in cholera and waterborne disease. Journal of Theoretical Biology, 324:84–102, 2013. URL https://doi.org/10.1016/j.jtbi.2012.12.021.
  16. [16]N. D. Evans and M. J. Chappell. Extensions to a procedure for generating locally identifiable reparameterisations of unidentifiable systems. Mathematical Biosciences, 168(2):137–159, 2000. URL http://dx.doi.org/10.1016/S0025-5564(00)00047-X.
  17. [17]M. Fliess. A note on the invertibility of nonlinear input-output differential systems. Systems & Control Letters, 8(2):147–151, 1986. URL https://doi.org/10.1016/0167-6911(86)90073-3.
  18. [18]A. Fokas, J. Cuevas-Maraver, and P. Kevrekidis. A quantitative framework for exploring exit strategies from the COVID-19 lockdown. Chaos, Solitons & Fractals, 140:110244, 2020. URL http://dx.doi.org/10.1016/j.chaos.2020.110244.
  19. [19]R. Hermann and A. J. Krener. Nonlinear controllability and observability. IEEE Transactions on Automatic Control, pages 728–740, 1977.DOI
  20. [20]H. Hong, A. Ovchinnikov, G. Pogudin, and C. Yap. Global identifiability of differential models. Communications on Pure and Applied Mathematics, 73(9):1831–1879, 2020. URL https://doi.org/10.1002/cpa.21921.
  21. [21]T. Hsu and G. T. Tsao. Convenient method for studying enzyme kinetics. Biotechnology and Bioengineering, 21(12):2235–2246, 1979. URL http://dx.doi.org/10.1002/bit.260211206.
  22. [22]I. Ilmer, A. Ovchinnikov, and G. Pogudin. Web-based structural identifiability analyzer. In E. Cinquemani and L. Paulevé, editors, Computational Methods in Systems Biology, pages 254–265. Springer International Publishing, 2021. URL https://doi.org/10.1007/978-3-030-85633-5_17.
  23. [23]S. R. Kou, D. L. Elliott, and T. J. Tarn. Observability of nonlinear systems. Information and Control, 22(1):89–99, 1973. URL https://doi.org/10.1016/S0019-9958(73)90508-1.
  24. [24]C. Kravaris. On functional observability of nonlinear systems and the design of functional observers with assignable error dynamics. arXiv preprint, 2024. URL https://arxiv.org/abs/2501.00167.
  25. [25]S. Lang. Algebra. Springer, 3rd edition, 2002. URL https://doi.org/10.1007/978-1-4613-0041-0.
  26. [26]A. Mahdi, N. Meshkat, and S. Sullivan. Structural identifiability of viscoelastic mechanical systems. PLoS ONE, 9(2):e86411, Feb. 2014. URL http://dx.doi.org/10.1371/journal.pone.0086411.
  27. [27]A. Martinelli. Local identifiability of nonlinear ODE models with time-varying parameters: The general algebraic condition. IEEE Transactions on Automatic Control, 71(4):2418–2433, 2026. ISSN 2334-3303. URL http://dx.doi.org/10.1109/TAC.2025.3627899.
  28. [28]G. Massonis, J. R. Banga, and A. F. Villaverde. Structural identifiability and observability of compartmental models of the COVID-19 pandemic. Annual Reviews in Control, 51:441–459, 2021. URL http://dx.doi.org/10.1016/j.arcontrol.2020.12.001.
  29. [29]G. Massonis, J. R. Banga, and A. F. Villaverde. AutoRepar: A method to obtain identifiable and observable reparameterizations of dynamic models with mechanistic insights. International Journal of Robust and Nonlinear Control, 33(9):5039–5057, 2023. URL https://doi.org/10.1002/rnc.5887.
  30. [30]N. Meshkat, C. Anderson, and J. J. DiStefano. Finding identifiable parameter combinations in nonlinear ODE models and the rational reparametrization of their input-output equations. Mathematical Biosciences, 233(1):19–31, 2011. URL https://doi.org/10.1016/j.mbs.2011.06.001.
  31. [31]A. N. Montanari, L. Freitas, D. Proverbio, and J. Goncalves. Functional observability and subspace reconstruction in nonlinear systems. Physical Review Research, 4:043195, 2022. URL https://doi.org/10.1103/PhysRevResearch.4.043195.
  32. [32]J. Němcová, M. Petreczky, and J. H. van Schuppen. Observability reduction algorithm for rational systems. In 2016 IEEE 55th Conference on Decision and Control (CDC). IEEE, 2016. URL http://dx.doi.org/10.1109/CDC.2016.7799151.
  33. [33]J. Němcová, M. Petreczky, and J. H. van Schuppen. System reduction and system identification of rational systems. IFAC Proceedings Volumes, 45(16):953–958, 2012. URL https://doi.org/10.3182/20120711-3-BE-2027.00388.
  34. [34]J. Norden, E. Oostwal, M. Chappell, P. Tiño, and K. Bunte. Structure is information: Structural identifiability mappings for machine learning with partially observed dynamical systems. IEEE Transactions on Cybernetics, page 1–13, 2026. URL http://dx.doi.org/10.1109/TCYB.2026.3694145.
  35. [35]F. Ollivier. Le problème de l’identifiabilité structurelle globale: approche théorique, méthodes effectives et bornes de complexité. Ph.D. thesis, 1990. URL https://www.lix.polytechnique.fr/~ollivier/PRODUCTION_SCIENT/THESE/these.pdf.
  36. [36]A. Ovchinnikov, A. Pillay, G. Pogudin, and T. Scanlon. Computing all identifiable functions of parameters for ode models. Systems & Control Letters, 157:105030, 2021. URL https://doi.org/10.1016/j.sysconle.2021.105030.
  37. [37]A. Ovchinnikov, G. Pogudin, and P. Thompson. Parameter identifiability and input-output equations. Applicable Algebra in Engineering, Communication and Computing, 34(2):165–182, 2021. URL https://doi.org/10.1007/s00200-021-00486-8.
  38. [38]A. Ovchinnikov, G. Pogudin, and P. Thompson. Input-output equations and identifiability of linear ode models. IEEE Transactions on Automatic Control, 68(2):812–824, 2023. URL https://doi.org/10.1109/TAC.2022.3145571.
  39. [39]B. Pant, O. Saucedo, and G. Pogudin. Uncovering identifiability of epidemiological models: basic reproduction number and complementary data streams. Jan. 2026. URL http://dx.doi.org/10.64898/2026.01.16.26344284.
  40. [40]D. Pavlov, G. Pogudin, and Y. Razmyslov. From algebra to analysis: New proofs of theorems by Ritt and Seidenberg. Proceedings of the American Mathematical Society, 150(12):5085–5095, 2022. URL http://dx.doi.org/10.1090/proc/16065.
  41. [41]T. Phan, J. Bennett, and T. Patten. Practical understanding of cancer model identifiability in clinical applications. Life, 13(2):410, Feb. 2023. URL http://dx.doi.org/10.3390/life13020410.
  42. [42]P. Pohjanpalo. System identifiability based on the power series expansion of the solution. Mathematical Biosciences, 41(1–2):21–33, 1978. URL http://dx.doi.org/10.1016/0025-5564(78)90063-9.
  43. [43]J. F. Ritt. Differential Algebra. AMS, 1950.DOI
  44. [44]M. P. Saccomani, G. Bellu, S. Audoly, and L. d’Angió. A new version of DAISY to test structural identifiability of biological models. page 329–334, 2019. URL https://doi.org/10.1007/978-3-030-31304-3_21.
  45. [45]S. Schnell and C. Mendoza. Enzyme kinetics of multiple alternative substrates. Journal of Mathematical Chemistry, 27(1-2):155–170, 2000. URL http://dx.doi.org/10.1023/A:1019139423811.
  46. [46]A. Seidenberg. Abstract differential algebra and the analytic case. Proceedings of the American Mathematical Society, 9(1):159–164, Feb. 1958. URL http://dx.doi.org/10.1090/S0002-9939-1958-0093655-0.
  47. [47]E. Sontag and Y. Wang. I/O equations for nonlinear systems and observation spaces. In [1991] Proceedings of the 30th IEEE Conference on Decision and Control, page 720–725. IEEE. URL http://dx.doi.org/10.1109/CDC.1991.261406.
  48. [48]J. D. Stigter and J. Molenaar. A fast algorithm to assess local structural identifiability. Automatica, 58:118–124, Aug. 2015. URL http://dx.doi.org/10.1016/j.automatica.2015.05.004.
  49. [49]J. von zur Gathen and J. Gerhard. Modern Computer Algebra. Cambridge University Press, 2013.
  50. [50]Y. Wang and E. D. Sontag. On two definitions of observation spaces. Systems & Control Letters, 13 (4):279–289, 1989. URL https://doi.org/10.1016/0167-6911(89)90116-3.

Paper details

vlnc:W_x8th7h95
Version
v2 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents