Identifiability of Differential-Algebraic Systems

Arthur N. Montanari∗,1 , François Lamoline2 , Robert Bereza3 , and Jorge Gonçalves2,4

1
Department of Physics and Astronomy, Northwestern University, Evanston, IL 60208, USA
2
Luxembourg Centre for Systems Biomedicine, University of Luxembourg, Belvaux L-4367, Luxembourg
3
Division of Decision and Control Systems, KTH Royal Institute of Technology, Stockholm SE-100 44, Sweden
4
Department of Plant Sciences, Cambridge University, Cambridge CB2 3EA, United Kingdom
arXiv:2405.13818v1 [eess.SY] 22 May 2024

May 24, 2024

Data-driven modeling of dynamical systems often faces numerous data-related challenges. A fundamental requirement
is the existence of a unique set of parameters for a chosen model structure, an issue commonly referred to as
identifiability. Although this problem is well studied for ordinary differential equations (ODEs), few studies have
focused on the more general class of systems described by differential-algebraic equations (DAEs). Examples of
DAEs include dynamical systems with algebraic equations representing conservation laws or approximating fast
dynamics. This work introduces a novel identifiability test for models characterized by nonlinear DAEs. Unlike
previous approaches, our test only requires prior knowledge of the system equations and does not need nonlinear
transformation, index reduction, or numerical integration of the DAEs. We employed our identifiability analysis
across a diverse range of DAE models, illustrating how system identifiability depends on the choices of sensors,
experimental conditions, and model structures. Given the added challenges involved in identifying DAEs when
compared to ODEs, we anticipate that our findings will have broad applicability and contribute significantly to the
development and validation of data-driven methods for DAEs and other structure-preserving models.
Keywords: identifiability; differential-algebraic equations; descriptor systems; system identification; observability.

1 Introduction perturbed systems like biological regulatory networks [7]).

DAE models directly preserve the structure of the un-
Differential-algebraic equations (DAEs) are a powerful derlying physical system—a feature that is often lost af-
tool for the modeling, analysis, and control of constrained ter reductionist methods are applied to obtain equivalent
dynamical systems [1, 2]. In the most general case, DAEs ODE models [2] (e.g., Kron reduction in power grids [8]).
are represented in implicit form F (ẋ, x) = 0, where x(t) To leverage the structure and constraint preservation in
is the n-dimensional state of the system and the Jaco- DAEs, many methods in control [9–12] and state esti-
bian matrix ∂F (ẋ, x)/∂x can be singular; in particular, mation [13–15] have been proposed for a wide variety of
when the Jacobian is nonsingular, DAEs can be reduced applications. Not surprisingly, controllability and observ-
to ordinary differential equations (ODEs). Due to its ability conditions were also established for linear [2,16,17]
singular form, DAE models require the development of and nonlinear [18, 19] DAE models.
novel mathematical and computational tools to handle
Closely related to the problems of state estimation and
the constraints imposed in the state-space exploration of
observability are the problems of parameter estimation
the system [3]. This is evident for DAEs represented in
and identifiability in system identification. Given the
semi-explict form, where the model is decomposed as a set
structure-preserving nature of most DAE models, data-
of ODEs ẋ1 = f1 (x1 , x2 ) and a set of algebraic equations
driven methods for system identification of DAEs could
f2 (x1 , x2 ) = 0. The latter equations directly capture con-
be proven very useful for a variety of applications that
straints imposed in a wide range of systems by conserva-
have been restricted to ODE models thus far; for instance,
tion laws (e.g., chemical processes [4] and power grids [5]),
to uncover structural relations between molecules in reg-
mechanical constraints (e.g., multibody mechanics and
ulatory networks [20–22] or interacting components in
robotics [6]), and timescale separations (e.g., singularly
cyber-physical systems [23]. Notwithstanding, the identi-
∗ Corresponding author: [email protected] fication of DAEs is a severely underdeveloped field, es-

1
pecially when compared to the identification of ODEs systems is used to refer to dynamical systems described
[24–26]. The majority of methods developed for parame- by ODEs [model (8)] and the term descriptor systems
ter estimation in DAEs are application-specific (cf. appli- to singular systems described by DAEs [implicit (1) and
cations in metabolic models [4] and power grids [5]), with semi-explicit (2) models]. Let In denote the identity ma-
only two methods available in the literature for general trix of size n and 0m×n denote an m × n null matrix
classes of linear [27] and nonlinear [28, 29] DAE models. (subscripts are omitted when self-evident). col(X) de-
However, even for the simple low-dimensional DAE model notes the column space of X. The vectorization of an
of a pendulum (which involves only three parameters), m × n matrix X is performed column-wise, as defined by
the estimation of certain combinations of parameters us- vec(X) = [X11 X21 . . . Xm1 . . . X1n X2n . . . Xmn ]T .
ing the above tools can be unfeasible (as shown in this The vectorization of multiple matrices is denoted by
paper). vec(X, . . . , Y ) = [vec(X)T . . . vec(Y )T ]T . The Kro-
The practical limitations in DAE identification lead to necker product is denoted by X ⊗ Y .
the following identifiability questions: How to determine Column vectors are represented by bold letters x. The
which parameters can be estimated from data in DAE vector y (ν) := dν y(t)/dtν denotes the νth derivative of
models? What data (with respect to measured variable, y with respect to time t. Composition maps are denoted
time-series length, or excitability conditions) should be as f ◦ h(x) := f (h(x)). Let Lνf hj (x) be the ν-th Lie
available for system identification? While there are many derivative of the j-th component of h(x) along the vector
analytical [30–34], simulation-based [35–37], and data- field f (x). By definition, L0f hj (x) := hj (x) and
driven [38] identifiability tests for ODE systems, just a
few studies have focused on DAE systems [39–43]. More- ∂Lν−1
f hj (x)
Lνf hj (x) := · f (x).
over, identifiability tests for DAEs have either involved ∂x
several nonlinear system transformations and/or index
For compactness, let Lνf h(x) := [Lνf h1 (x) . . . Lνf hq (x)]T .
reductions [39, 42, 44], been restricted to linearized mod-
els [40], or required extensive numerical integration of the
system for many initial conditions [41, 43]. As a result, 2 Observability of Descriptor Systems
such tests are often time-consuming, cumbersome, or in-
feasible to computationally implement and test for many Consider the continuous-time autonomous implicit de-
real-world applications. scriptor system
This paper presents an identifiability test for nonlin- 0 = F (x, ẋ), (1a)
ear DAE systems that only requires prior knowledge of
y = h(x), (1b)
the model structure, measurement functions, and their
successive derivatives (Section 3). The theoretical results where x ∈ X ⊆ Rn is the state vector, ẋ ∈ X̄ ⊆ Rn is
are based on a generalization of rank-based conditions for the time derivative of the state vector, and y ∈ Rq is the
the observability analysis of DAE models [18] (reviewed output vector (measurement signals). Functions F : X ×
in Section 2.1). As a side contribution, we rigorously X̄ 7→ F ⊆ Rn and h : X 7→ H ⊆ Rq are considered to be
show how the observability test for nonlinear DAE mod- sufficiently smooth. Here, we assume that the descriptor
els reduces to well-known observability tests for nonlin- system (1) can be represented in semi-explicit DAE form
ear ODEs as well as linear DAEs and ODEs—formalizing (see Refs. [1, 45] for sufficient conditions in which this
the relation between previous work in the literature (Sec- assumption holds):
tion 2.2). Analogously, our identifiability test can also
be employed for systems described by linear DAEs and ẋ1 = f1 (x1 , x2 ), (2a)
(non)linear ODEs. This allows us to establish some spe- 0 = f2 (x1 , x2 ), (2b)
cial conditions for the identifiability of linear DAEs based y = h(x1 , x2 ), (2c)
on the structure of the system matrices (Section 3.1).
where x1 ∈ X1 ⊆ Rn1 and x2 ∈ X2 ⊆ Rn2 are the differen-
The proposed identifiability analysis is demonstrated in
tial and algebraic variables, respectively (x := [xT
several applications (Section 4). These examples discuss T T
1 x2 ] ).
Uniqueness of solutions is considered in the context
the dependence of the system identifiability on the place-
of initial value problems, meaning that solutions are re-
ment of sensors, excitation regions, and system structure.
quired to satisfy a consistent initial condition x(t0 ). An
In particular, we show that, even for linear DAEs with
initial condition x(t0 ) is said to be consistent if the asso-
full-state measurement, the identification process can be
challenging and reliant on the system structure. ciated initial value problem has at least one solution [46].
Let us introduce the set of consistent states
Notation. The following terminology and notation are n o
adopted throughout the paper. The term nonsingular L := (t, x, ẋ) ẋ1 = f1 (x1 , x2 ), 0 = f2 (x1 , x2 ) ⊆ X .

2
From [1, Theorem 4.13], we deduce the existence and set of equations:
uniqueness of a local solution to (2a)–(2b):     
M11 M12 v1 b
= 1 , (4)
Lemma 1. Assume that L ̸= ∅ and that f1 and f2 M21 M22 v2 b2
are C 1 -smooth functions. Then, for every x(t0 ) = | {z
M
} | {z
v
} | {z
b
}
[xT T T
1 (t0 ) x2 (t0 )] ∈ L, the semi-explicit DAE form (2) has
a unique local solution satisfying the initial value given by where v = [v1T v2T ]T ∈ Rm , v1 ∈ Rm1 , and v2 ∈ Rm2 . Let
x(t0 ). T T T T T T
us define M1 = [M11 M21 ] and M2 = [M12 M22 ] .

Considering only consistent states x ∈ L, let Definition 2. [45] The linear equation M v = b is 1-full
ΦT (x(t0 ); θ) be the unique solution of system (2): with respect to v1 if Eq. (4) uniquely determines v1 for
any consistent vector b and known matrix M .
ΦT (x(t0 ); θ) := x(t0 + T )
Z t0 +T
Lemma 2. [45] The following statements are equivalent:
= x1 (t0 ) + f1 (x(t); θ)dt (3)
t0 1. the linear equation (4) is 1-full with respect to v1 ;
s.t. 0 = f2 (x1 , x2 ; θ), 2. rank(M1 ) = m1 and col(M1 ) ∩ col(M2 ) = ∅;
We explicitly denote in ΦT (x(t0 ); θ) the internal param- 3. rank(M ) = m1 + rank(M2 ).
eters θ of f1 and f2 (and, by equivalence, of the implicit
function F ), which are relevant for the identifiability anal- From Lemma 2, a local observability condition for non-
ysis. linear descriptor systems can be derived as follows [18].
We now formalize the local observability property for First, let us subsequently differentiate model (1):
descriptor systems as a generalization of the local (weak)
observability of nonsingular dynamical systems [47, 48]. 0 = F (x, ẋ), y = h(x),
0 = dF (x, ẋ)/dt, ẏ = dh(x)/dt,
Definition 1. The descriptor system (1), or the pair
0 = d2 F (x, ẋ)/dt2 , ÿ = d2 h(x)/dt2 , (5)
{F , h}, is locally observable at x0 if there exists a neigh-
borhood U ⊆ L of x0 such that, for every state (x′ ∈ U) ̸= .. ..
. .
x0 , h ◦ ΦT (x0 ; θ) ̸= h ◦ ΦT (x′ ; θ) for some finite time µ µ
interval t ∈ [t0 , t0 + T ]. 0 = d F (x, ẋ)/dt , y = dν h(x)/dtν .
(ν)
| {z } | {z }
0=F̄µ (x,ẋ,ẍ,...,x(µ+1) ) ȳ=H̄ν (x,ẋ,ẍ,...,x(ν) )
Assuming prior knowledge of the system model (1), a
descriptor system is locally observable around a state x0 if where µ ≥ 0 and ν ≥ 0 define respectively the number
x0 can be uniquely reconstructed from the measured sig- of times F and h are differentiated. According to Defini-
nal y(t) over a finite trajectory, as defined by the compo- tion 1, the descriptor system (1) is locally observable at
sition map h ◦ ΦT (x(t0 ); θ). The local weak observability x0 if there exists a local injective map from the system
property is related to the notion of smooth observability, model, the output signals, and its derivatives up to some
implying that there exists a smooth map from y and its order [i.e., Eqs. (5)] to the system state x0 . Consider-
successive derivatives ẏ, ÿ, . . . , y (ν) to the state vector x ing the Jacobian matrix of Eqs. (5) locally evaluated at
if the system is smoothly observable [18, 49]. x(t) = x0 , we have the following local linear map:
Remark 1. The observability property considered here 
∂ F̄µ (x, w)

∂ F̄µ (x, w)    
concerns only the reconstruction of consistent states x ∈  x 0
 ∂ H̄ ∂x ∂w = n×1 (6)

L. For the reconstruction of inconsistent states that do ν (x, w) ∂ H̄ν (x, w) w

ȳ
not satisfy the constraints in solution (3) (as described by ∂x ∂w
generalized solutions with impulse terms [50]), we refer to
| {z }
O(x,w)
the notion of impulse observability [2, 17].
where w = [ẋT ẍT . . . x(σ) T ]T , with order σ = max{µ +
2.1 Observability test 1, ν}, and O(x, w) is the observability matrix. Since F
and h are smooth functions, following the Constant Rank
The local observability of a nonlinear descriptor system Theorem, an algebraic observability condition can thus
can be verified through an algebraic rank condition [18, be derived for the unique reconstruction of x0 by testing
45], analogous to Kalman’s famous observability test. To under which conditions the linear set of equations (6) is
this end, we first define the notion of 1-fullness for a linear 1-full with respect to x:

3
Theorem 1. (Observability of nonlinear DAEs) [18]
Observability of 𝐸 = 𝐼! Observability of
The descriptor system (1), or the pair {F , h}, is locally
linear ODEs linear DAEs
observable at x0 ∈ L if (Kalman, 1959) (Hou, 1999)
 
∂ F̄µ (x, w)
𝒇 𝒙 = 𝐴𝒙 𝑭 𝒙, 𝒙̇ = 𝐴𝒙 − 𝐸𝒙̇
rank(O(x, w)) = n + rank 
 ∂w 
. (7) 𝒉 𝒙 = 𝐶𝒙 𝒉 𝒙 = 𝐶𝒙
∂ H̄ν (x, w) 
∂w
Observability of Observability of
holds at x = x0 for some µ and ν. nonlinear ODEs nonlinear DAEs
(Hermann and Krener, 1977) 𝑭 𝒙, 𝒙̇ = 𝒇 𝒙 − 𝒙̇ (Terreu, 1997)

Remark 2. The rank condition (7) depends on the num-

ber of times that functions F and h are differentiated Figure 1: Observability of descriptor systems and its special
in order to construct the observability matrix O(x, w). cases.
In practice, we iteratively increment µ and ν until the
image space of O stops growing for some σ, i.e., until
rank(O(x, ẋ, . . . , x(σ) )) = rank(O(x, ẋ, . . . , x(σ+1) )). 1. O(x, w) is nonsingular at x = x0 for some µ and ν;

Remark 3. Although here we focus on autonomous sys- 2. the Jacobian matrix

tems (1), the results can be directly generalized to control  0
Lf h(x)

systems F (x, ẋ) = Bu or F (x, ẋ, u) = 0, where u is the
1
input vector (control signal). Lf h(x)
∂  
Ψ(x) = ..
∂x 
 
. 
2.2 Relation to special cases ν
Lf h(x)
The observability test for nonlinear descriptor systems is full-column rank at x = x0 for some ν.
proposed in Ref. [18] was developed independently from
other well-known observability tests for linear descrip- Proof sketch. Condition 2 is a well-known condition in the
tor systems [17] and (linear and nonlinear) nonsingular literature for the observability of nonlinear nonsingular
systems [47, 51]. As such, a clear relationship between systems; proofs can be found in Refs. [47, 48, 52]. Here,
these tests is not yet available in the literature. As a we show that condition 1 follows directly from Theorem 1
first contribution of our paper, we derive and formalize for F (x, ẋ) = f (x) − ẋ. Furthermore, we prove that
the relationship between the aforementioned observability conditions 1 and 2 are equivalent. See Appendix 5 for
tests, showing that Theorem 1 is indeed a rigorous gener- details. □
alization of observability properties presented in previous Another important class of descriptor systems are rep-
work. These relationships are summarized in Fig. 1. resented by semi-explicit DAEs (2) with index 1 [53],
Nonsingular systems. The differential index of a sys- meaning that differentiating the constraints (2b) once
tem of DAEs is the number of times the DAEs have to leads to an ODE. Since f2 is smooth, an index-1 DAE
be differentiated in order to be expressed as an equiva- implies that Eq. (2b) is locally solvable for the algebraic
lent system of ODEs, also known as nonsingular systems. variable x2 . Following the Implicit Function Theorem, if
Thus, the well-studied class of nonsingular systems de- rank(∂f2 (x1 , x2 )/∂x2 ) = n2 , then there exists a locally
scribed by unique solution x2 = F2 (x1 ) which inserted into Eqs.
(2a) and (2c) yields
ẋ = f (x), (8a)
ẋ1 = f1 (x1 , F2 (x1 )) := F1 (x1 ), (9a)
y = h(x), (8b)
y = h(x1 , F2 (x1 )) := H(x1 ). (9b)
can be seen as a special form of descriptor systems (1)
with index 0, where F (x, ẋ) = f (x) − ẋ. An observabil- The local observability of an index-1 descriptor system
ity condition for nonsingular systems can thus be derived can, therefore, be verified using Proposition 1 for the
from Theorem 1: pair {F1 , H} at any state x0 ∈ L. From Definition 1,
if {F1 , H} is locally observable, then x1 can be locally
Proposition 1. (Observability of nonlinear ODEs) inferred from the measurement signal y and hence x2
The nonsingular system (8), or the pair {f , h} is locally can be obtained from its local solution F2 (x1 ). However,
observable at x0 if the following equivalent conditions are applying Proposition 1 requires the construction of the lo-
satisfied: cal map F2 , which can be significantly more cumbersome

4
to handle compared to the direct test on the original de- for linear nonsingular systems [51, 54]. It follows directly
scriptor system’s equations proposed in Theorem 1. from Proposition 2 that:
Linear systems. If the pair {F , h} is given by linear Corollary 1. (Observability of linear ODEs) If E is
functions, then system (1) can be represented as nonsingular, then the following statements are equivalent:
E ẋ = Ax + Bu, (10a)
1′ ) the linear descriptor system (10) is observable;
y = Cx, (10b)
2′ ) rank(O′ ) = n, where A′ = E −1 A and
where A ∈ Rn×n , B ∈ Rn×p , and C ∈ Rq×n . The matrix
T
E ∈ Rn×n is possibly singular and we assume that the O′ = C T (CA′ )T . . . (C(A′ )n−1 )T

control signal u = 0 without loss of generality. Since we
restrict the observability property to the reconstruction is the Kalman’s observability matrix;
of consistent states x0 ∈ L (Remark 1), Definition 1 re-
λI − A′
 
duces to the notion of R-observability for linear descriptor 3′ ) rank = n, ∀λ ∈ C.
C
systems, which is defined globally as follows [2, 17]:
Definition 3. [2] The linear descriptor system (10) is Proof. If E is nonsingular, then the descriptor model
R-observable if, for any arbitrary x(t0 ) ∈ L, x(t0 ) can be (10) can be expressed as a linear system of ODEs: ẋ =
uniquely determined from the output signal y(t) for some A′ x + B ′ u, where A′ = E −1 A and B ′ = E −1 B. It thus
finite time interval t ∈ [t0 , t0 + T ]. follows that statement 3 of Proposition 2 reduces to state-
ment 3′ (also known as the Popov-Belevitch-Hautus test).
We now show that the R-observability condition follows Likewise, for a nonsingular E, statement 2 of Proposi-
as a special case of Theorem 1: tion 2 holds if and only if O is nonsingular (since O12
Proposition 2. (Observability of linear DAEs) The is full rank). −1
Moreover, O is nonsingular if and only if
′
following statements are equivalent: O 21 −O O
22 12 O11 = O is invertible, leading to statement
′
2.
1. the linear descriptor system (10) is R-observable;

O12

3 Identifiability of Descriptor Systems
2. rank(O) = n+rank , where the n(n+q)×n(n+
O22
1) observability matrix is defined by We now extend the observability condition presented in
  Theorem 1 to establish an algebraic identifiability con-
A −E 0 ... 0 0
 0 A −E . . . 0 0  dition for nonlinear descriptor systems. Let us consider
 again the autonomous descriptor system
 .. .. .. . .. 

..
. ..
  . . . . 


O11 O12  0 0 0 . . . A −E

 0 = F (x, ẋ; θ), (11a)
O= =
 ;
y = h(x; θ), (11b)
O21 O22  C 0 0 ... 0 0  
 0 C 0 . . . 0 0  where we explicitly denote θ ∈ Rp as the vector of un-


 . .. .. .. .. .. 
 .. . . . . .  known parameters sought to be identified. The notion of
0 0 0 ... C 0 parameter identifiability establishes the conditions under
  which, given a known model structure (11), the unknown
λE − A parameters θ can be uniquely determined from a mea-
3. rank = n, ∀λ ∈ C.
C surement signal y(t) recorded over t ∈ [t0 , t0 + T ]. This
is formally defined as follows.
Proof sketch. Statement 3 is a well-known sufficient
and necessary condition in the literature for the R- Definition 4. The descriptor system (11), or the pair
observability of linear descriptor systems; a complete {F , h}, is locally identifiable at x0 ∈ L with respect to
proof can be found in [2, Sec. 4.3 and 4.5]. Here, we parameter θ0 if there exists a neighborhood U ⊆ Rp of
show that statement 2 is sufficient for R-observability by θ0 such that, for every parameter (θ ′ ∈ U) ̸= θ0 , h ◦
applying Theorem 1 to linear functions. To prove the ne- ΦT (x0 ; θ0 ) ̸= h ◦ ΦT (x0 ; θ ′ ) for some finite time interval
cessity of statement 2, we show that statements 2 and 3 t ∈ [t0 , t0 + T ].
are equivalent. See Appendix 5 for details. □
If E is nonsingular, then L = Rn (all states are con- Comparing Definitions 1 and 4, it becomes evident that
sistent) and, hence, the notion of R-observability (Defini- the local identifiability problem can be treated as an ob-
tion 3) reduces to the well-known observability property servability problem if the parameters are represented as

5
state variables of the descriptor model—a framework pre- equations (11). This provides a practical and simple test
viously explored for nonsingular systems [32, 34]. To that can guide experimental design for system identifica-
this end, we first augment the descriptor system (11) tion, including optimal sensor placement (which variables
with additional state variables representing the param- to measure) and excitation conditions (which regions of
eters θ sought to be identified (assuming constant dy- the state space to explore for data collection in experi-
namics θ̇ = 0 so that θ(t) = θ, for all t ≥ 0). This yields ments). We explore these advantages in system analysis
the following augmented implicit model: with examples presented in Section 4.
An important underlying assumption in Theorems 1
0 = F (x, ẋ, θ), (12a) and 2 is that the observability and identifiability analyses
0 = θ̇, (12b) are evaluated locally in the neighborhood of a consistent
state (Remark 1). The definition of a consistent state in
y = h(x, θ). (12c)
this work (set L) requires an implicit model (1) to have an
An algebraic (local) identifiability condition can now be equivalent representation in semi-explicit form (2). This
derived for the augmented model (12) based on the 1- is often not a problem since DAEs in semi-explicit form
fullness property of linear systems (Lemma 2), analo- arise naturally in the modeling of many technological, bio-
gously to the observability test presented in Theorem 1. chemical, and environmental systems. However, if only
an implicit model is available, it may not be easy or even
Theorem 2. (Identifiability of nonlinear DAEs) Let possible to derive its corresponding semi-explicit repre-
the identifiability matrix be defined as sentation. Despite this challenge, if a consistent state can

∂ F̄µ (θ, z) ∂ F̄µ (θ, z)
 be determined (e.g., numerically), then the conditions in
∂θ ∂z Theorems 1 and 2 can be applied directly to the system’s
I(θ, z) =  , (13)
 
∂ H̄ν (θ, z) ∂ H̄ν (θ, z)  equations in implicit form, not requiring conversion to
∂θ ∂z semi-explicit form.
where z = [xT ẋT . . . x(σ) T ]T and σ = max{µ + 1, ν}.
The descriptor system (12), or the pair {F , h}, is locally 3.1 Linear descriptor systems
identifiable at x0 ∈ L with respect to θ0 if
  Given that the observability test (Theorem 1) is appli-
∂ F̄µ (θ, z) cable both to (non)linear DAEs and ODEs, as proven in
rank(I(θ, z)) = p + rank 
 ∂z 
(14) Section 2.2, it follows that the identifiability test (Theo-
∂ H̄ν (θ, z)  rem 2) is also applicable to all these classes of systems. In
∂z what follows, we explore the structure of linear descrip-
holds at (x, θ) = (x0 , θ0 ) for some µ and ν. tor systems to determine sufficient conditions in which
the identifiability condition (14) is always locally satis-
Proof. See Appendix 5. fied, ensuring system identifiability.
Consider the linear descriptor system (10). In this sce-
p
Remark 4. The rank condition (14) is state dependent, nario, the parameters θ ∈ R sought to be identified
meaning that the system may be identifiable (or uniden- are typically a subset of the matrix elements Aij (i.e.,
tifiable) only in a subregion of the parameter space and θk = Aij if Aij is an unknown parameter, for k = 1, . . . , p
2
the set of consistent states (i.e., in a subset of Rp × L). and p ≤ n ). To this end, we model the parameters as
This is illustrated in detail in Section 4.1. state variables with constant dynamics [as in system (12)],
yielding
Remark 5. The observability condition (7) is sufficient E ẋ = A(θ)x, (15a)
and “almost” necessary for nonlinear descriptor systems.
From [55, Corollary 112], the observability condition is θ̇ = 0, (15b)
necessary on a dense subset of X , which implies that it y = Cx, (15c)
is also a necessary condition when restricted to L ⊆ X if
where we explicitly denote the dependence of A on the
L is dense. Thus, not satisfying condition (7) is a strong
parameters θ. The identifiability analysis of system (15)
indication on the unobservability of a system. The same
can thus be conducted by directly applying Theorem 2 on
argument can be extended to the identifiability test (14).
the functions F (x, ẋ) = Ax − E ẋ and h(x) = Cx. Since
Theorems 1 and 2 provide a clear advantage for the parameters θ are treated as state variables, the identi-
observability and identifiability analysis of nonlinear de- fiability of linear descriptor systems is typically a non-
scriptor systems: they establish algebraic conditions, linear problem due to the matrix-vector multiplication
which can be directly checked on the original system’s A(θ)x. Thus, Theorem 2 still only establishes a suffi-

6
cient condition for the identifiability of linear descriptor holds by construction. Hence, due to the equivalence of
systems. Nonetheless, the particular structure of the de- statements 2 and 3 in Lemma 2, if condition (19) is sat-
scriptor system (15) provides a concise form for the iden- isfied, then condition (14) is also satisfied.
tifiability matrix (13), which can be described by
  Corollary 3 highlights that full-state measurement is
I11 (θ, z) ∆1 ⊗ A − ∆2 ⊗ E not sufficient for the identifiability of general linear de-
I(θ, z) = , (16)
0q(ν+1)×p ∆1 ⊗ C scriptor systems. The identifiability will thus depend on
the structure of the system (encoded by matrices A and
where ∆1 = [Iσ 0σn×n ], ∆2 = [0σn×n Iσ ], and we as-
E) as well as the set of parameters to be identified. In
sumed that µ = ν. The matrix block I11 is defined as
particular, if the system is represented in semi-explicit
 
A(θ)x form      
In1 0 ẋ1 A11 A12 x1
 A(θ)ẋ 
∂  = (20)

I11 (θ, z) := .. . 0 0 ẋ2 A21 A22 x2
∂θ 

. 
(µ) further sufficient conditions can be derived depending on
A(θ)x
the set of parameters sought to be identified. Eq. (20)
By applying Theorem 2, we directly derive the following is a special form of the DAE system (2) for linear func-
identifiability test for linear descriptor systems: tions, where x1 ∈ Rn1 and x2 ∈ Rn2 are the differential
Corollary 2. (Identifiability of linear DAEs) Con- and algebraic variables, respectively. Matrices A12 and
sider the identifiability matrix I(θ, z) defined in Eq. (13). A21 describe interactions between differential and alge-
The descriptor system (15) is locally identifiable at x0 ∈ braic variables. Depending on the application, certain
L with respect to θ0 if submatrices may be known a priori and hence the set of
  unknown parameters can be restricted to the remaining
∆1 ⊗ A − ∆2 ⊗ E unknown matrices, e.g., θ = vec(A11 , A21 ). The follow-
rank(I(θ, z)) = p + rank (17)
∆1 ⊗ C ing result shows that some combinations of parameters
are identifiable if full-state measurement is available.
holds at (x, θ) = (x0 , θ0 ) for some µ = ν.
Assumption 1. The linear DAE (20) is index 1, that is,
−1
Remark 6. For applications in which all elements Aij A22 is nonsingular. This implies that x2 = −A22 A21 x1
are sought to be identified [i.e., θ = vec(A)], it follows and hence the DAEs can be reduced −1
to the ODEs ẋ1 =
that A x
c 1 , where A c = A 11 − A A A
12 22 21 .
 T
x ⊗ In xT Corollary 4. (Identifiability of semi-explicit linear
  
 ẋT ⊗ In   ẋT  DAEs) Suppose C = In and Assumption 1 holds. If
I11 (θ, z) =  ..  =  ..  ⊗ In . (18) A21 is full rank, then the descriptor system (20) is locally
   
 .   . 
identifiable at x0 ̸= 0 with respect to the set of parameters
(µ)T (x(µ) )T
x ⊗ In θ0 = vec(A11 , A22 ) or θ0 = vec(A12 , A21 ).
We now leverage the structure of the identifiability ma-
Proof. See Appendix 5.
trix (16) to derive some special, easier-to-check identifia-
bility conditions for linear descriptor systems under the Other parameter combinations like θ = vec(A11 , A12 )
assumption that full-state measurement is available (i.e., may also be identifiable in the case of full-state measure-
C = In ). ment depending on the sparsity of the matrix A. These
Corollary 3. (Identifiability of linear DAEs with scenarios are further explored numerically in Section 4.3.
full-state measurement) Suppose C = In . The de- The identifiability analysis of linear descriptor sys-
scriptor system (15) is locally identifiable at x0 ∈ L with tems can also be applied to linear nonsingular sys-
respect to θ0 if tems (i.e., when E = In ). For such cases, the fol-
lowing corollary shows that—contrariwise to descriptor
rank (I11 (θ, z)) = p (19) systems—full-state measurement is sufficient for the com-
plete identification of linear ODE systems (which is a well-
holds at (x, θ) = (x0 , θ0 ) for some µ.
known condition in the literature [31, 56]).
Proof. Recall that the identifiability matrices (13) and
(16) are equal for the linear functions F (x, ẋ) = Ax−E ẋ Corollary 5. (Identifiability of linear ODEs) Sup-
and h(x) = Cx. If C = In , then pose C = In and E = In . The descriptor system (15) is
    locally identifiable at x0 ̸= 0 with respect to θ = vec(A).
I11 (θ, z) ∆1 ⊗ A − ∆2 ⊗ E
col ∩ col =∅ Proof. See Appendix 5.
0q(ν+1)×p ∆1 ⊗ C

7
 a b c

Figure 2: Identifiable regions of the chemical reactor model for measurement signals given by: (a) y = x1 , (b) y = x2 , and (c)
y = x3 . The blue (red) colors correspond to states x in which the parameter Tc is identifiable (unidentifiable). The black solid
line represents the state trajectory x(t) starting at the initial condition x(t0 ) = [0.5 350 0.4995]T and converging to a limit cycle.
The system parameters were set to (c0 , T0 , Tc ) = (1, 350, 305), and (k1 , k2 , k3 , k4 , k5 ) = (1, 209.205, 2.0921, 8.7503·103 , 7.2·1010 ).
The tolerance for the numerical computation of the rank condition (14) is set to maxz σnϵ(∥I(θ, z)∥2 ), where ϵ(b) is the floating-
point relative accuracy of b. The numerical simulations are shown using the ode15s solver in MATLAB.

Remark 7. Corollaries 4 and 5 provide sufficient condi- lenges compared to the special case of linear nonsingular
tions for the local identifiability of linear systems around systems. Sections 4.1 and 4.2 numerically evaluate the
any consistent state x ∈ L. The only exception is the identifiability condition presented in Theorem 2, while
origin (x, ẋ, . . . , x(σ) ) = 0, where it follows trivially from Section 4.3 illustrates the results derived for linear sys-
Eq. (18) that [x ẋ ẍ . . . x(µ) ]T ⊗ In = 0 and hence condi- tems proposed in Corollaries 2–5.
tion (19) is not satisfied. This illustrates why parameter
identification of a linear system is not possible when using 4.1 Chemical reactor
data collected from a system operating on an equilibrium
point; that is, the system dynamics were not “excited” Consider an exothermic reactor system with a single first-
r (t)
during the data collection (as commonly referred in sys- order reaction (A −−→ B) and generated heat removed
tem identification). through an external cooling circuit. The chemical reactor
is modeled by [57]

4 Numerical Examples ẋ1 = k1 (c0 − x1 ) − x3 (21a)

ẋ2 = k1 (T0 − x2 ) + k2 x3 + k3 (Tc − x2 ), (21b)
This section demonstrates the identifiability analysis re- 0 = x3 − k5 exp{−k4 /x2 }x1 , (21c)
sults on several descriptor systems, showing its applica-
bility to experimental design and method validation in where x = [x1 x2 x3 ]T ∈ R3 is the state vector, x1 is the
system identification. Section 4.1 presents a tutorial iden- concentration of reactant A, x2 is the temperature, and
tifiability analysis for a simple index-1 chemical reactor x3 is the reaction rate per unit volume (algebraic vari-
model, showing step-by-step the computation of the pro- able). Eq. (21c) arises from a conservation law1 (Arrhe-
posed algebraic rank condition. This example highlights nius equation) that imposes the dependence of reaction
the parameter identifiability dependence on the choice of rates on temperature. The parameters are: inlet feed re-
measured variables, also known as the sensor placement actant concentration c0 and feed temperature T0 , coolant
problem. Moving up to an index-3 pendulum model, temperature Tc , and other constants ki . Fig. 2a shows
Section 4.2 investigates the dependence of the parameter that the chemical reactor dynamics converge to a limit
identifiability on the system state and experimental con- cycle for the considered set of parameters.
ditions (e.g., if the dataset is informative enough), shed- Let the coolant temperature be the parameter sought
ding some light on the limitations of recent tools devel- to be identified (i.e., θ = Tc ). The parameter estima-
oped for parameter identification in DAEs [28]. Finally, 1 This descriptor model can be directly converted to ODEs and
Section 4.3 shows that even linear descriptor systems in then numerically integrated with conventional solvers (e.g., 4th-
which all state variables are independently measured may order Runge-Kutta). However, numerical solvers for DAEs enforce
be unidentifiable, posing additional identification chal- the constraint (21c) to be preserved in simulations like Fig. 2.

8
tion depends on the choice of observable (i.e., which very close to unidentifiable regions. If the numerical rank
state variable xi is measured by a sensor). To evalu- tolerances are not appropriately chosen, the rank condi-
ate the appropriate sensor data for parameter estima- tion (14) may be numerically satisfied for ill-conditioned
tion, we now test which observables make the system matrices I(θ, z) even though the system is practically
identifiable. As a tutorial, we analyze step-by-step the unidentifiable.
identifiability of the descriptor model (21) considering After conducting the same identifiability analysis for
the output signal y(t) = x1 (t), t ∈ [0, 10] s. First, we y(t) = x2 (t) and y(t) = x3 (t), Fig. 2b,c shows that the
augment model (21) by representing θ(t) = Tc (t) as an system is locally identifiable everywhere around the limit
additional state variable with constant dynamics, i.e., cycle solution when these other types of measurement sig-
θ̇(t) = 0. Functions H̄ν and F̄µ are then constructed nals (sensors) are considered. This is a particularly bene-
up to order µ = ν = n − 1 = 2: H̄ν = [x1 ẋ1 ẍ1 ]T and ficial result for this sensor placement problem since tem-
T d2 F T T
F̄µ = [(F )T ( dF
dt ) ( dt2 ) ] , where
perature sensors (y = x2 ) are very affordable and practi-
  cal to install in chemical reactors.
k1 (c0 − x1 ) − x3 − ẋ1
F = k1 (T0 − x2 ) + k2 x3 n + k3 (θo− x2 ) − ẋ2  ,

x3 − k5 exp − xk42 x1
4.2 Pendulum equation

−k1 ẋ1 − ẋ3 − ẍ1
 The motion of an undamped pendulum is represented in
dF Cartesian coordinates as an index-3 DAE system:
=  −k1 ẋ2 + k2 ẋn3 + ko3 (θ̇ − ẋ2 ) − ẍ2  ,
 
dt
−ẋ3 − k5 exp − xk42 ẋ1 + k4 xx1 ẋ2 2 ẋ1 = x3 , (22a)
2
ẋ2 = x4 , (22b)
and d2 F /dt2 is omitted for the sake of brevity. Note that x1 x5
F̄µ and H̄ν are functions of...θ, x, and its time derivatives. ẋ3 = , (22c)
m
By assumption, θ̇ = θ̈ = θ = 0, ∀t. After building the x2 x5
ẋ4 = − g, (22d)
identifiability matrix, condition (14) can then be evalu- m
ated to test if the system is identifiable at any consistent 0 = x21 + x22 − L2 , (22e)
state x0 ∈ L ⊂ R3 and parameter value θ0 ∈ R1 .
where (x1 , x2 ) denote the Cartesian coordinates of the
Fig. 2 shows the identifiable regions in the state pendulum endpoint in the plane, (x3 , x4 ) are the corre-
space depending on the observable choice. For the out- sponding velocities, and x5 is the tension per unit length
put y(t) = x1 (t) (Fig. 2a), the limit cycle solution in the pendulum arm. The parameters are: mass m, grav-
ΦT (x(t0 ); θ) predominantly lies in an unidentifiable re- itational force g, and pendulum-arm length L. Assume
gion. At unidentifiable states x0 , measured trajectories that only the pendulum angle is measured, i.e.,
h ◦ ΦT (x0 ; θ) sufficiently close to x0 are not guaranteed  
to be distinguishable from a trajectory h◦ΦT (x0 ; θ′ ) cor- x1
y = h(x) = arctan − . (23)
responding to some distinct parameter θ′ ̸= θ lying in a x2
neighborhood U of the true parameter value θ. In other We now test the identifiability of the pendulum’s indi-
words, there may exist some (θ′ ∈ U) ̸= θ such that vidual parameters (θ = m, g, or L) as well as all its
h ◦ ΦT (x0 ; θ) = h ◦ ΦT (x0 ; θ′ ). The predominance of possible combinations (e.g., θ = [g L]). The parameter
unidentifiable regions in the state space implies the pos- identifiability is evaluated for all (consistent) states in
sible existence of non-unique solutions for θ in the neigh- the simulated trajectory presented in Fig. 3a. Figure 3b
borhood of ΦT (x(t0 ); θ). This directly impairs the esti- shows that the system is locally identifiable with respect
mation accuracy of θ if the dataset mostly include mea- to all individual and pairwise combination of parameters
surement data collected in these unidentifiable states. for the entire set of consistent states (x1 , x2 ) ∈ L. On
Despite the predominance of unidentifiable regions in the other hand, the set of all parameters θ = [m g L] is
Fig. 2a, parameter identifiability is achieved periodi- unidentifiable2 for almost every x ∈ L (Fig. 3c), high-
cally in a small region of the limit cycle around x(t) = lighting that—in practice—we expect that at least one of
[0.25 360 0.5]T and other regions containing part of the the parameters must be known a priori before methods
transient response. Theoretically, parameter estimation for parameter estimation can be successfully applied.
could be possible by using data collected only at the The above theoretical analysis of the parameters iden-
identifiable states. However, this may not be possible 2
in practice if the identifiable region is too small as in ever,The identifiability condition is formally only sufficient. How-
given that L is dense, it follows from Remark 5 that the con-
Fig. 2a, either due to limited data availability in these dition is also necessary except for a local set of parameters with
regions or ill-conditioned identifiability matrices at states Lesbegue dimension zero.

9
 identifiable
a 15
b 0
c 0
unidentifiable

10

-3 -3
0

-5

-10
-6 -6
-15
0 5 10 15 -6 -3 0 3 6 -6 -3 0 3 6

Figure 3: (a) State evolution of the pendulum equation as a function of time. (b, c) Identifiable (blue) and unidentifiable
(red) regions in the phase plane (x1 , x2 ) depending on the parameter set θ sought to be identified: (b) parameter sets m, g, L,
[m g], [g L], and [m L] are all (locally) identifiable everywhere in the phase space; (c) parameter set [m g L] is unidentifiable
almost everywhere in the phase space. The parameters are set in the simulation as L = 6.25, g = 9.81, and m = 0.3.

a θ = {m} θ = {g} θ = {L} b θ = {m, g, L}

0.4 0.4

11
7
estimates

estimates

estimates

estimates
10
0.3 0.3
6
9

8 5
0.2 0.2
100 300 500 100 300 500 100 300 500 100 300 500
data size data size data size data size

θ = {m, g} θ = {g, L} θ = {m, L} 11

estimates
0.4
10
11
7
estimates

estimates
estimates

9
10
0.3
6 8
9 100 300 500
data size
8 5
0.2
100 300 500 100 300 500 100 300 500
data size data size data size 7
estimates

0.4
6
11
7
estimates

estimates

estimates

10 5
0.3
6 100 300 500
9 data size

8 5 mass
0.2
100 300 500 100 300 500 100 300 500 gravity
data size data size data size
length

Figure 4: Parameter estimation of the pendulum model using a prediction-error method for nonlinear DAE identification.
Boxplots of the estimated parameter values for the sets of parameters predicted to be: (a) identifiable versus (b) unidentifiable.
The x-axis shows the number of data points used for the parameter estimation. Each box comprise the estimated values for
a 100 independent experiments. The dashed lines indicate the true values of the parameters. In each experiment, the mean
square error between a measured noisy trajectory and a trajectory predicted by the model is minimized over the model
parameters. The minimization is performed using the Levenberg-Marquardt algorithm, using the same settings as in the
paper proposing the method [29]. The initial estimates for the parameters are set to m′ = 0.1, g ′ = 8.21, and L′ = 5.41.

10
 a b

differential variable
algebraic variable

c d

Figure 5: Identifiability of linear descriptor systems for different model structures and parameter sets. The network structure
of the underlying model is shown on the left panels: each node represents a (differential or algebraic) variable and an edge
points from node j to i if Aij ̸= 0. The parameters (represented by edges) sought to be identified in each case are highlighted
in green. The identifiable (unidentifiable) states in the simulated trajectory are indicated in blue (red) on the right panels.
The initial conditions of differential variables x1 (t0 ) were randomly drawn from a normal distribution. (a) Identifiability of
submatrices A12 and A21 in a dense DAE model. (b) Identifiability of matrix A in a dense DAE model. (c) Identifiability of
matrix A in a sparse DAE model. (d) Identifiability of matrix A in a dense ODE model.

tifiability is supported by computational results using a variables are independently measured, i.e., C = I4 . We
prediction-error method developed for parameter estima- choose a random system matrix such that the system is
tion of general nonlinear descriptor systems [28, 29]. Fig- asymptotically stable:
ure 4 shows that identifiable parameter sets can indeed be  
−1.72 0.20 0.02 −0.23
accurately estimated using the prediction-error method;  0.15 −1.30 −0.31 −0.21 
the parameter estimates converge to the true values as the A=  0.51 −0.87 −1.50 −0.21  .

number of data samples increases. On the other hand, the
−1.05 −0.20 −0.23 −1.34
estimation of the parameter set θ = [m g L], which is pre-
dicted to be unidentifiable by our theory, diverged from
the true values. Although such biased estimation in Fig.
4b could be mistakenly attributed to numerical issues, Figure 5 shows the identifiability analysis of linear sys-
poor data, or lack of scalability of prediciton-error meth-tems for different choices of parameters and model struc-
ods to larger sets of parameters, our analysis highlights tures. Note that A22 is invertible and hence the descriptor
that the failed parameter estimation is actually a direct system is index 1 (Assumption 1). Since rank(A21 ) = 2,
consequence of the identifiability of the model structure. it follows from Corollary 4 that the set of parameters
θ1 = vec(A12 , A21 ) is identifiable. As predicted by the
theory, Fig. 5a shows that the system is indeed locally
4.3 Linear DAE system
identifiable with respect to θ1 at the states x(t) in the
Consider the linear descriptor system (15) of order n = 4. simulated trajectory. In these simulations, the identifi-
Let E = diag(I2 , 02×2 ), leading to n1 = 2 differential able states are determined by evaluating the rank condi-
and n2 = 2 algebraic variables. Assume that all state tion (17) at state x(t). Only when the system converges

11
to equilibrium (x(t) → 0) that identifiability is lost3 , as implementations of the examples discussed in Section 4.
pointed in Remark 7. Similar results can be verified for We expect that the presented results will find a wide range
the identifiability of θ = vec(A11 , A22 ). of DAE applications beyond the examples considered in
For the parameter set θ2 = vec(A), Fig. 5b shows this work, such as singular systems, power systems, multi-
that the system is not identifiable. This demonstrates body robotics, biological regulatory networks, and many
that, for a generic descriptor system, full-state measure- others. In particular, by treating feedback control laws
ment is not enough for identifiability of the system matrix u = k(x) as algebraic equations (2b), the proposed iden-
A. Prior knowledge of the system is thus essential to as- tifiability test may also provide insight in the limitations
certain identifiability. One alternative is to reduce the of closed-loop identification [62, 63].
set of parameters to be identified (as in Fig. 5a), but There are still many challenges for the data-driven
this indirectly assumes that all other parameters are al- modeling of descriptor systems. First, it still remains un-
ready known a priori. Another approach is to explore clear which types of model structures and measurement
the model sparsity and prior knowledge of zero entries in functions are sufficient to guarantee the system identi-
matrix A; the interconnection structure is often reliably fiability, even in the linear case as explored in Section
known in many network systems. Consider, for exam- 4.3. We expect that methods based on structural net-
ple, the sparse model in which the following matrix en- work inference [33], graph-theoretical conditions for ob-
tries are zero: {a12 , a13 , a23 , a24 , a31 , a34 , a42 }. Figure 5c servability [52,58,64], or rank properties of the Kronecker
illustrates that identifiability of the sparse system with product [65] might be able to address these problems.
respect to θ2 can indeed be achieved. Beyond the results Second, the identifiability analysis is grounded on the as-
presented in Corollary 4, we expect that generic identifia- sumption that the parameters can be modeled as time-
bility conditions based on model sparsity can be derived, constant state variables. However, whether the identi-
although this is a topic of future research. fiability of a DAE system is promoted or inhibited by
If E = In (leading to an ODE system of order 4), it time-dependent parameters (i.e., θ̇ = f (t)) has yet to be
follows from Corollary 5 that the system is identifiable explored. Finally, our results can be directly extended to
with respect to θ2 . This statement is verified numeri- discrete-time systems and feedback systems (Remark 3),
cally in Fig. 5d. The contrast between Figs. 5b and d but generalizations to broader classes of systems such as
demonstrates, even in the linear setting, DAEs are fun- stochastic DAEs [66] or partial DAEs [67] remain an open
damentally more challenging to identify from data than problem.
ODEs.

Acknowledgement
5 Conclusion
The large adoption of Kalman’s controllability and ob- The authors thank Mohamed R.-H. Abdalmoaty and
servability tests can be attributed to its simple for- Håkan Hjalmarsson for insightful discussions. FL was
mulation and easy implementation involving only ma- supported by the Luxembourg National Research Fund
trix multiplications and prior knowledge of the sys- (FNR), grant CORE19/13684479/DynCell, and is now a
tem model. These tests can be intuitively examined Quilvest Research Fellow under a Quilvest donation. RB
based on the system structure [52, 58, 59] and have is supported by the Swedish Research Council under con-
since been extended to more general classes of sys- tracts 2019-04956 and 2016-06079 (the research environ-
tems [16, 47, 60, 61]. Our results are motivated by a ment NewLEADS) and by Digital Futures.
similar goal: to derive an identifiability condition that
can be straightforwardly tested without excessive ma-
nipulation of system equations. The proposed identifi- Appendix: Proofs
ability test (Theorem 2) requires only prior knowledge
of the system’s implicit function F (x, ẋ), measurement PROOF OF PROPOSITION 1. Sufficiency of con-
function h(x), and their successive derivatives (which dition 1. Let O(x, w) be the corresponding observabil-
can be obtained via symbolic computation). Codes ity matrix for a nonsingular system F (x, ẋ) = f (x) − ẋ
for the identifiability test of general descriptor systems and measurement function h(x). Note that ∂ F̄µ /∂w is a
{F , h} are available in GitHub (https://github.com/ block lower triangular matrix with diagonal matrices −In .
montanariarthur/IdentifiabilityDAE), along with Thus, it follows from condition (7) in Theorem 1 that the
3 Theoretically, identifiability is lost exactly at x(t) = 0. In prac-
pair {F , h} is observable if rank(O(x, w)) = n + n(µ + 1)
tice, the rank condition (17) holds only up to some numerical rank
or, equivalently, O(x, w) is invertible.
tolerance as x(t) approaches zero (Fig. 5a,c,d). Equivalence between conditions 1 and 2. Since ∂ F̄µ /∂w

12
is invertible, then O(x, w) is invertible if and only if (5). Subsequent differentiation of Eq. (12) yields
−1
F̄µ (x, . . . , x(µ+1) , θ)
   
0

′ ∂ H̄ν ∂ H̄ν ∂ F̄µ ∂ F̄µ
O (x, w) = − Θ  = 0 , (24)
∂x ∂w ∂w ∂x

H̄ν (x, . . . , x(ν) , θ) ȳ
is invertible. We now show that O′ (x, w) = Ψ(x), there-
T
fore implying that conditions 1 and 2 are equivalent. We where Θ = [θ̇ T θ̈ T . . . θ (µ+1) ]T . We define the function
show this, without loss of generality, for the univariate Ψ(θ, . . . , θ (σ) , x, . . . , x(σ) ) := [F̄µT ΘT H̄νT ]T . The Jaco-
case f (x) and h(x), where n = 1. After some matrix bian matrix of DΨ is given by
manipulations, it follows that
∂ F̄µ ∂ F̄µ ∂ F̄µ
 
 ∂h 
2
∂x  ∂θ ∂(x, ẋ, . . . , x(σ) ) ∂(θ̇, θ̈, . . . , θ (σ) ) 
∂ h ∂h ∂f
 

∂x2 ẋ + ∂x ∂x 
  ∂Θ ∂Θ ∂Θ 
DΨ = 
O′ = 
 2  
(σ)

3 2 2 ∂f
 ∂∂xh3 ẋ2 + ∂∂xh2 ẍ + 2 ∂∂xh2 ∂x ẋ + ∂h ∂f ∂h ∂ 2
f
+ ∂x ∂x2 ẋ

 ∂θ ∂(x, ẋ, . . . , x ) ∂( θ̇, θ̈, . . . , θ (σ) ) 
∂x ∂x
 
   ∂ H̄ν ∂ H̄ν ∂ H̄ν 
..
. ∂θ (σ)
∂(x, ẋ, . . . , x ) ∂(θ̇, θ̈, . . . , θ ) (σ)

∂h
   
∂x
 0
Lf h(x)
 ∂ F̄µ ∂ F̄µ
∂ ∂h

0n(µ+1)×pσ 
  2∂x ∂x f
  L1f h(x)  ∂θ ∂z
 ∂  =

0 0 ∆

=  ∂ ∂ h2 f 2 + ∂h ∂f f  = L2 h(x) = Ψ(x), p(µ+1)×p p(µ+1)×n(σ+1)
     
∂x ∂x ∂x ∂x ∂x  f   ∂ H̄ν ∂ H̄ ν


..
 .. 0q(ν+1)×pσ
. . ∂θ ∂z
where ∆ = [Ip(µ+1)×p(µ+1) 0p(µ+1)×p(σ−µ−1) ].
where we have substituted ẋ = f (x) and ẍ = ∂f∂x (x)
f (x)
Now, consider the following linear map of Eq. (24) eval-
and the equality holds by induction for all higher-order
uated locally around (θ, z, w):
terms Lνf h(x), ν ≥ 3. □
   
θ 0n(µ+1)×1
PROOF OF PROPOSITION 2. Sufficiency of DΨ  z  =  0p(µ+1)×1  , (25)
statement 2. The sufficiency of statement 2 follows di- w ȳ
rectly from Theorem 1 by computing the observability
T
matrix O(x, w) corresponding to the linear functions where w = [θ̇ T θ̈ T . . . θ (σ) ]T . By Lemma 2, the lin-
F (x, ẋ) = Ax − E ẋ and h(x) = Cx up to order σ = n. ear system (25) is 1-full with respect to θ if and only if
In the linear case, condition (7) becomes globally valid rank(∂Ψ/∂θ) = p and col(∂Ψ/∂θ) ∩ col(∂Ψ/∂(z, w)) =
since O(x, w) = O, which has constant rank for every ∅. Clearly, ∂Ψ/∂θ is full column rank if and only if
state x ∈ L. It follows from Lemma 2 that if statement 2  
∂ F̄µ
holds, then O is 1-full with respect to x. Therefore, any
arbitrary x(t0 ) ∈ L can be uniquely reconstructed from rank  ∂θ  = p. (26)
 
∂ H̄µ
y(t) (and its successive derivatives up to order n−1) and,
∂θ
by Definition 3, the system is R-observable.
Moreover, note that col(∂Ψ/∂θ) ∩ col(∂Ψ/∂w) is empty
Necessity of statement 2. Since statement 3 is sufficient for any system {F , h}. Therefore, it follows that the 1-
and necessary condition for R-observability, we now prove fullness of system (25) with respect to θ is equivalent to
the necessity of statement 2 by showing that statement 3 Eq. (26) and col(∂Ψ/∂θ)∩col(∂Ψ/∂z) = ∅, which in turn
entails statement 2. It follows from statement 2 that are equivalent to condition (14) according to Lemma 2.
 
A

O
 Recall that Ψ : M 7→ N is assumed to be smooth,
rank(O) = rank + rank 12 where M ⊂ R(σ+1)(n+p) and N ⊂ R(µ+1)(n+p)+nν . Let
C O22
  Tp M denote the tangent space of M around a point
O12 p = [θ0T z0T w0T ]T ∈ M. Following the Constant Rank
= n + rank .
O22 Theorem, there exist open neighborhoods U around p
T T T
Observe that rank[A C ] = n is equivalent to state- and V around Ψ(p), as well as some diffeomorphisms
ment 3 for λ = 0, which concludes the proof. □ ϕ : Tp M 7→ U and ψ : TΨ(p) N 7→ V, such that
Ψ(U) ⊆ V and the Jacobian DΨ : Tp M 7→ TΨ(p) is
−1
PROOF OF THEOREM 2. Consider the extended given by DΨ = ψ ◦ Ψ ◦ ϕ.
system (12) and recall that µ and ν respectively define The proof now follows from the fact that, if condi-
the higher-order derivatives of functions F and H in Eq. tion (14) holds, then the system (25) is 1-full with respect

13
to θ. This implies that θ can be uniquely determined for the identifiability of (20) follows from Corollary 3 given
any consistent measurement ȳ and known matrix DΨ. that
Finally, the equivalence DΨ = ψ −1 ◦ Ψ ◦ ϕ shows that θ 
∂Γ12
 
∂Γ21

can also be uniquely determined from ȳ and the system rank(I11 ) = rank + rank
∂ vec(A12 ) ∂ vec(A21 )
model Ψ. Therefore, the system is locally identifiable −1 0 1 µ


around a neighborhood U of (θ0 , x0 ) and its successive = rank A22 A21 [Ac Ac . . . Ac ] ⊗ x1 · n1
+ rank [A0c A1c . . . Aµc ] ⊗ x1 · n2


derivatives up to order σ. □
PROOF OF COROLLARY 4. Consider the general = 2n1 n2 = p. □
identifiability analysis where θ = vec(A11 , A12 , A21 , A22 ).
Denote PROOF OF COROLLARY 5. Since E = In , it fol-
lows that x(µ) = Aµ x. Thus,
Aij (θ)xj
   
A(θ)x
rank (I11 (θ, z)) = rank [A0 A1 . . . An−1 ] ⊗ x rank(In )


 A(θ)ẋ   Aij (θ)ẋj 
Γ= ..  and Γij =  .. .
   
 .   .  = n2 .
(µ)
A(θ)x(µ) Aij (θ)xj As in the proof of Corollary 4, the second equality fol-
lows from the fact that rank([A0 A1 . . . An−1 ]) = n and
As in Eq. (18), it follows that
  rank(x) = 1 for x ̸= 0. Given that p = n2 (size of θ), the
∂Γ ∂Γ ∂Γ ∂Γ identifiability of system (15) follows from Corollary 3. □
I11 =
∂ vec(A11 ) ∂ vec(A12 ) ∂ vec(A21 ) ∂ vec(A22 )
 
∂Γ11 ∂Γ12 References
0 0
 ∂ vec(A11 ) ∂ vec(A12 )
= ,

∂Γ21 ∂Γ22 [1] P. Kunkel, V. Mehrmann, Differential-Algebraic Equa-
0 0
∂ vec(A21 ) ∂ vec(A22 ) tions: Analysis and Numerical Solution, European Math-
∂Γij ematical Society, 2006.
(µ)
where = [xj ẋj ẍj . . . xj ]T ⊗ Ini . [2] G.-R. Duan, Analysis and Design of Descriptor Linear
∂ vec(Aij )
Systems, Springer, 2010.
For the case of θ = vec(A11 , A22 ), it holds that
    [3] L. Petzold, Differential/algebraic equations are not
∂Γ11 ∂Γ22 ODE’s, SIAM Journal on Scientific and Statistical Com-
rank(I11 ) = rank + rank
∂ vec(A11 ) ∂ vec(A22 ) puting 3 (1982) 367–384.
[4] H. J. Dudley, L. Lu, Z. J. Ren, D. M. Bortz, Sensitivity
 
(µ)
= rank [x1 ẋ1 ẍ1 . . . x1 ]T rank(In1 ) and bifurcation analysis of a differential-algebraic equa-
tion model for a microbial electrolysis cell, SIAM Journal
 
(µ)
+ rank [x2 ẋ2 ẍ2 . . . x2 ]T rank(In2 ) on Applied Dynamical Systems 18 (2) (2019) 709–728.
= rank [A0c A1c . . . Aµc ] ⊗ x1 · n1


[5] S. Nabavi, A. Chakrabortty, Structured identification of
reduced-order models of power systems in a differential-
+ rank A−1 0 1 µ


22 A21 [Ac Ac . . . Ac ] ⊗ x1 · n2 algebraic form, IEEE Transactions on Power Systems
= rank [A0c A1c . . . Aµc ] · rank(x1 ) · n1


32 (1) (2016) 198–207.
+ rank A−1 0 µ

[6] D. Pogorelov, Differential-algebraic equations in multi-
22 A21 [Ac . . . Ac ] · rank(x1 ) · n2
body system modeling, Numerical Algorithms 19 (1998)
= n21 + n22 , 183–194.
where we have applied in the equalities above the rela- [7] D. Del Vecchio, J.-J. E. Slotine, A contraction theory
tion rank(A ⊗ B) = rank(A) rank(B). In the third equal- approach to singularly perturbed systems, IEEE Trans-
actions on Automatic Control 58 (3) (2012) 752–757.
ity, Assumption 1 implies that ẋ1 = Ac x1 and x2 =
A−1 [8] F. Dorfler, F. Bullo, Kron reduction of graphs with ap-
22 A21 x1 . In the fourth equality, note that rank(x1 ) = 1
plications to electrical networks, IEEE Transactions on
for all x1 ̸= 0. Moreover, since A22 is invertible due to
Circuits and Systems I: Regular Papers 60 (1) (2012)
Assumption 1 and rank(A21 ) = min{n2 , n1 } is satisfied,
150–163.
it follows that
[9] T. N. Chang, E. J. Davison, Decentralized control of de-
rank(A−1 0 1 n−1
22 A21 [Ac Ac . . . Ac ]) = n2 . scriptor systems, IEEE Transactions on Automatic Con-
trol 46 (10) (2001) 1589–1595.
Since p = n21 + n22 (size of θ), the identifiability of sys- [10] P. Kunkel, V. Mehrmann, Analysis of over-and under-
tem (20) with respect to θ = vec(A11 , A12 ) follows from determined nonlinear differential-algebraic systems with
Corollary 3. application to nonlinear control problems, Mathematics
Analogously, for the case of θ = vec(A12 , A21 ) ∈ Rp , of Control, Signals and Systems 14 (2001) 233–256.

14
[11] S. Campbell, P. Kunkel, Solving higher index DAE op- [26] J. Schoukens, L. Ljung, Nonlinear system identification:
timal control problems, Numerical Algebra, Control and A user-oriented road map, IEEE Control Systems Maga-
Optimization 6 (4) (2016) 447–472. zine 39 (6) (2019) 28–99.
[12] M. Nadeem, M. Bahavarnia, A. F. Taha, On wide-area [27] M. Gerdin, T. B. Schön, T. Glad, F. Gustafsson,
control of solar-integrated DAE models of power grids, L. Ljung, On parameter and state estimation for linear
American Control Conference (2023) 4495–4500. differential-algebraic equations, Automatica 43 (3) (2007)
[13] M. Gerdin, J. Sjöberg, Nonlinear stochastic differential- 416–425.
algebraic equations with application to particle filtering, [28] M. R.-H. Abdalmoaty, O. Eriksson, R. Bereza, D. Bro-
IEEE Conference on Decision and Control (2006) 6630– man, H. Hjalmarsson, Identification of non-linear
6635. differential-algebraic equation models with process dis-
[14] R. K. Mandela, R. Rengaswamy, S. Narasimhan, L. N. turbances, IEEE Conference on Decision and Control
Sridhar, Recursive state estimation techniques for nonlin- (2021) 2300–2305.
ear differential algebraic systems, Chemical Engineering [29] R. Bereza, O. Eriksson, M. R.-H. Abdalmoaty, D. Bro-
Science 65 (16) (2010) 4548–4556. man, H. Hjalmarsson, Stochastic approximation for iden-
[15] S. A. Nugroho, A. F. Taha, N. Gatsis, J. Zhao, Observers tification of non-linear differential-algebraic equations
for differential algebraic equation models of power net- with process disturbances, IEEE Conference on Decision
works: Jointly estimating dynamic and algebraic states, and Control (2022) 6712–6717.
IEEE Transactions on Control of Network Systems 9 (3) [30] R. Bellman, K. J. Åström, On structural identifiability,
(2022) 1531–1543. Mathematical Biosciences 7 (3-4) (1970) 329–339.
[16] S. L. Campbell, N. K. Nichols, W. J. Terrell, Duality, ob-
[31] E. Walter, G. Le Cardinal, P. Bertrand, On the identifi-
servability, and controllability for linear time-varying de-
ability of linear state systems, Mathematical Biosciences
scriptor systems, Circuits, Systems and Signal Processing
31 (1-2) (1976) 131–141.
10 (4) (1991) 455–470.
[17] M. Hou, P. C. Müller, Causal observability of descrip- [32] E. Tunali, T.-J. Tarn, New results for identifiability
tor systems, IEEE Transactions on Automatic Control of nonlinear systems, IEEE Transactions on Automatic
44 (1999) 158–163. Control 32 (2) (1987) 146–154.

[18] W. J. Terrell, Observability of nonlinear differential al- [33] J. Gonçalves, S. Warnick, Necessary and sufficient condi-
gebraic systems, Circuits, Systems and Signal Processing tions for dynamical structure reconstruction of LTI net-
16 (1997) 271–285. works, IEEE Transactions on Automatic Control 53 (7)
(2008) 1670–1674.
[19] K. Sato, Algebraic controllability and observability of
nonlinear differential algebraic systems with geometric [34] A. F. Villaverde, A. Barreiro, A. Papachristodoulou,
index one, SICE Journal of Control, Measurement, and Structural identifiability of dynamic systems biology
System Integration 7 (5) (2014) 283–290. models, PLoS Computational Biology 12 (10) (2016)
e1005153.
[20] Y. Yuan, G.-B. Stan, S. Warnick, J. Goncalves, Robust
dynamical network structure reconstruction, Automatica [35] A. Raue, C. Kreutz, T. Maiwald, J. Bachmann,
47 (6) (2011) 1230–1235. M. Schilling, U. Klingmüller, J. Timmer, Structural and
[21] N. M. Mangan, S. L. Brunton, J. L. Proctor, J. N. Kutz, practical identifiability analysis of partially observed dy-
Inferring biological networks by sparse identification of namical models by exploiting the profile likelihood, Bioin-
nonlinear dynamics, IEEE Transactions on Molecular, formatics 25 (15) (2009) 1923–1929.
Biological and Multi-Scale Communications 2 (1) (2016) [36] J. D. Stigter, J. Molenaar, A fast algorithm to assess local
52–63. structural identifiability, Automatica 58 (2015) 118–124.
[22] A. Aalto, L. Viitasaari, P. Ilmonen, L. Mombaerts, [37] D. Joubert, J. Stigter, J. Molenaar, Assessing the role of
J. Gonçalves, Gene regulatory network inference from initial conditions in the local structural identifiability of
sparsely sampled noisy data, Nature Communications 11 large dynamic models, Scientific Reports 11 (1) (2021)
(2020) 3493. 16902.
[23] Y. Yuan, X. Tang, W. Zhou, W. Pan, X. Li, H.-T. Zhang, [38] N. Evangelou, N. J. Wichrowski, G. A. Kevrekidis, F. Di-
H. Ding, J. Goncalves, Data driven discovery of cyber etrich, M. Kooshkbaghi, S. McFann, I. G. Kevrekidis, On
physical systems, Nature Communications 10 (1) (2019) the parameter combinations that matter and on those
4894. that do not: data-driven studies of parameter (non) iden-
[24] T. B. Schön, A. Wills, B. Ninness, System identifica- tifiability, PNAS Nexus 1 (4) (2022) pgac154.
tion of nonlinear state-space models, Automatica 47 (1) [39] M. Gerdin, Identification and estimation for models de-
(2011) 39–49. scribed by differential-algebraic equations, Ph.D. thesis,
[25] L. A. Aguirre, A bird’s eye view of nonlinear system iden- Department of Electrical Engineering, Linköping Univer-
tification, arXiv:1907.06803 (2019). sity (2006).

15
[40] A. Ben-Zvi, P. J. McLellan, K. McAuley, Identifiability [57] W. H. Ray, Advanced Process Control, New York:
of non-linear differential algebraic systems via a lineariza- McGraw-Hill, 1981.
tion approach, The Canadian Journal of Chemical Engi- [58] C. T. Lin, Structural Controllability, IEEE Transactions
neering 84 (5) (2006) 590–596. on Automatic Control 19 (3) (1974) 201–208.
[41] M. Gerdin, Using DAE solvers to examine local identifia- [59] T. Boukhobza, F. Hamelin, D. Sauter, Observability of
bility for linear and nonlinear systems, IFAC Proceedings structured linear systems in descriptor form: A graph-
Volumes 39 (1) (2006) 808–813. theoretic approach, Automatica 42 (4) (2006) 629–635.
[42] L. Ljung, T. Glad, On global identifiability for arbitrary [60] H. F. Chen, On stochastic observability and controllabil-
model parametrizations, Automatica 30 (2) (1994) 265– ity, Automatica 16 (2) (1980) 179–190.
276.
[61] A. N. Montanari, L. Freitas, D. Proverbio, J. Gonçalves,
[43] M. Ashyraliyev, Y. Fomekong-Nanfack, J. A. Kaandorp, Functional observability and subspace reconstruction in
J. G. Blom, Systems biology: parameter estimation for nonlinear systems, Physical Review Research 4 (2022)
biochemical models, The FEBS Journal 276 (4) (2009) 043195.
886–902.
[62] P. Van den Hof, Closed-loop issues in system identifica-
[44] L. Chen, X. Chen, J. Qian, Z. Yao, A new method for tion, Annual Reviews in Control 22 (1998) 173–186.
parameter identifiability of DAE systems, IEEE Inter-
[63] U. Forssell, L. Ljung, Closed-loop identification revisited,
national Conference on Control and Automation (2013)
Automatica 35 (7) (1999) 1215–1241.
280–283.
[64] A. N. Montanari, C. Duan, A. E. Motter, Target con-
[45] S. L. Campbell, E. Griepentrog, Solvability of general dif-
trollability and target observability of structured network
ferential algebraic equations, SIAM Journal on Scientific
systems, IEEE Control Systems Letters 7 (2023) 3060–
Computing 16 (1995) 257–270.
3065.
[46] C. C. Pantelides, The consistent initialization of
[65] Y. Tian, Some rank equalities and inequalities for Kro-
differential-algebraic systems, SIAM Journal on Scientific
necker products of matrices, Linear and Multilinear Al-
and Statistical Computing 9 (1988) 213–231.
gebra 53 (6) (2005) 445–454.
[47] R. Hermann, A. J. Krener, Nonlinear controllability and [66] R. Winkler, Stochastic differential algebraic equations of
observability, IEEE Transactions on Automatic Control index 1 and applications in circuit simulation, Journal of
22 (5) (1977) 728–740. Computational and Applied Mathematics 163 (2) (2004)
[48] M. Vidyasagar, Nonlinear Systems Analysis, 2nd Edition, 435–463.
Prentice Hall, 1978. [67] B. Benhammouda, H. Vazquez-Leal, Analytical solutions
[49] S. L. Campbell, W. J. Terrell, Observability of linear for systems of partial differential-algebraic equations,
time-varying descriptor systems, SIAM Journal on Ma- SpringerPlus 3 (1) (2014) 137.
trix Analysis and Applications 12 (1991) 484–496.
[50] Y. Chen, S. Trenn, Impulse-free jump solutions of nonlin-
ear differential–algebraic equations, Nonlinear Analysis:
Hybrid Systems 46 (2022) 101238.
[51] R. Kalman, On the general theory of control systems, IRE
Transactions on Automatic Control 4 (3) (1959) 110–110.
[52] A. N. Montanari, L. A. Aguirre, Observability of net-
work systems: A critical review of recent results, Journal
of Control, Automation and Electrical Systems 31 (6)
(2020) 1348–1374.
[53] R. Winkler, Stochastic differential algebraic equations of
index 1 and applications in circuit simulation, Journal
of Computational and Applied Mathematics 163 (2004)
435–463.
[54] C.-T. Chen, Linear System Theory and Design, 3rd Edi-
tion, Oxford University Press, 1999.
[55] M. Vidyasagar, Nonlinear Systems Analysis, Prentice
Hall, Englewood Cliffs, NJ, USA, 1993.
[56] J. Delforge, A sufficient condition for identifiability of a
linear system, Mathematical Biosciences 61 (1) (1982)
17–28.

16