Decomposable and Essentially Univariate Mass-Action Systems: Extensions of the Deficiency One Theorem Authors: Abhishek Deshpande, Stefan Müller Published: 2026 ABSTRACT AbstractThe classical and extended deficiency one theorems by Feinberg apply to reaction networks with mass-action kinetics that have independent linkage classes or subnetworks, each with a deficiency of at most one and exactly one absorbing strong component. The theorems assume the existence of a positive equilibrium and guarantee the existence of a unique positive equilibrium in every stoichiometric compatibility class. In our work, we use the monomial dependency which extends the concept of deficiency. First, we provide a dependency one theorem for parametrized systems of polynomial equations that are essentially univariate and decomposable. As our main result, we present a corresponding theorem for mass-action systems, which permits subnetworks with arbitrary deficiency and arbitrary number of absorbing strong components. Finally, to complete the picture, we derive the extended deficiency one theorem as a special case of our more general dependency one theorem. Keywords: Reaction networks, Mass-action kinetics, Existence of a unique positive equilibrium, Parametrized systems of polynomial equations, Monomial dependency, Dependency one theorem Introduction Many systems in chemistry and biology (particularly in ecology and epidemiology) and also in economics and engineering are modeled as polynomial or power-law dynamical systems. These models can be formulated as reaction networks with mass-action or generalized mass-action kinetics. They capture complex dynamic behaviors, such as multistationarity, oscillations, and chaos, and the corresponding bifurcations. The most prominent results of (chemical) reaction network theory, as founded in the 1970’s by Horn, Jackson, and Feinberg, are the deficiency zero and one theorems (for mass-action systems) (Horn and Jackson 1972; Horn 1972; Feinberg 1972, 1987, 1995b). In essence, the concept of deficiency captures affine dependencies between complexes (representing the left- and right-hand sides of the reactions). The deficiency zero theorem assumes weak reversibility (and deficiency zero) and guarantees the existence of a unique positive equilibrium in every stoichiometric compatibility class (invariant subspace) and for all rate constants (system parameters); moreover, it ensures the asymptotic stability of this equilibrium. The deficiency one theorem has several assumptions. It applies to mass-action systems with independent linkage classes or subnetworks, each with a deficiency of at most one and exactly one absorbing strong component. Additionally, the theorem assumes the existence of a positive equilibrium (for given rate constants) and guarantees the existence of a unique positive equilibrium in every compatibility class (but does not address stability). Finally, weak reversibility ensures the existence of a positive equilibrium. Both, the deficiency zero and one theorems, have been difficult to improve upon. Only after 2010, the deficiency zero theorem has been extended, namely from mass-action to generalized mass-action kinetics (Müller and Regensburger 2012, 2014; Müller et al. 2016, 2019; Craciun et al. 2019; Boros et al. 2020; Müller and Regensburger 2024). In this work, we will extend the deficiency one theorem, even in the setting of mass-action kinetics. The deficiency one theorem was first fully stated in 1987 (Feinberg 1987), but was not proved until 1995 (Feinberg 1995a). Its extension from independent linkage classes to independent subnetworks was mentioned in (Feinberg 1987), and a proof was outlined in (Feinberg 1995a). Since then, the upper bound of one for both the deficiency and the number of absorbing strong components (per independent linkage class or subnetwork) has not been addressed. Only the assumption of existence (of a positive equilibrium) has been investigated further (Boros 2010, 2012, 2013a, b). In (Boros 2010, 2012), Boros provides an equivalent condition for the existence of a positive equilibrium for reaction networks that satisfy the assumptions of the deficiency one theorem. In (Boros 2013a), he characterizes single linkage-class, deficiency one mass-action systems for which a positive equilibrium exists for all or some rate constants. Finally, in (Boros 2013b), Boros proves the existence of a positive equilibrium within every stoichiometric compatibility class for weakly reversible, deficiency one mass-action systems. Using different methods, this result can be further extended to weakly reversible (not necessarily deficiency one) mass-action systems (Boros 2019). In this work, we extend the validity of the deficiency one theorem to a much broader class of mass-action systems. First, we consider independent subnetworks (instead of independent linkage classes) from the outset. Second, we do not assume that the subnetworks have a deficiency of at most one. Instead, we use the monomial dependency which generalizes the concept of deficiency. Specifically, dependency accounts for the fact that non-source vertices in a reaction network do not contribute monomials to the polynomial equations for the equilibria. As a consequence, the dependency is often smaller than the deficiency. Third, we do not assume that the subnetworks have exactly one absorbing strong component. We illustrate our results in a series of examples in Sect. 8. Here, we consider Example 2, given by the reaction network or, equivalently, by the “embedded” graph and assume mass-action kinetics. The network has one independent linkage class/subnetwork. Hence, its deficiency is δ=|V|-1-dim(S)=5-1-2=2, where V is the set of vertices and S is the stoichiometric subspace (the linear span of the differences of complexes). However, its dependency is d=|Vs|-1-dim(L)=4-1-2=1, where Vs is the set of source vertices and L is the monomial difference subspace (the linear span of the differences of source complexes). Moreover, this network has two absorbing strong components, one singleton (complex 0) and one non-singleton (with complexes X2 and 3X1). Clearly, it does not satisfy the conditions of the deficiency one theorem. Still, our dependency one results for mass-action systems, Theorem 14 and Corollary 15, can be applied. Technically, we treat positive equilibria of mass-action systems as parametrized systems of polynomial equations with classes and apply recent fundamental results for such systems (Müller and Regensburger 2026). Specifically, their solution set is essentially the solution set on the coefficient polytope (modulo an exponential fiber involving the monomial difference subspace L). We assume (i) the reaction network can be decomposed into independent subnetworks, (ii) the equation system is decomposable in the terminology of (Müller and Regensburger 2026), (iii) the subsystems (classes) have monomial dependency one and hence are univariate, and (iv) L=S (or L=K, where K is the kinetic subspace). Then, there exists a unique positive equilibrium in every stoichiometric (or kinetic) compatibility class. Note that, even if K=S, the subnetworks need not have exactly one absorbing strong component. We first provide a dependency one theorem for one class, Theorem 5, using the fundamental result from (Müller and Regensburger 2026), Theorem 2. Next, we extend it to decomposable systems (with several classes) and obtain Theorem 8. By applying the latter to reaction networks (and using Birch’s theorem), we arrive at a dependency one theorem for mass-action systems, Theorem 14. Finally, this allows us to provide a modular proof of the extended deficiency one theorem, Theorem 18, and hence of the classical deficiency one theorem. Organization of the Work In Sect. 2, we recall the relevant geometric objects corresponding to parametrized systems of generalized polynomial equations (with classes) including the coefficient polytope and the monomial dependency and difference subspaces. In Sect. 3, we provide sufficient conditions for the unique existence of a solution on the coefficient polytope, along with an equivalent condition for its existence. In Sect. 4, we introduce basic notions for reaction networks with mass-action kinetics, and in Sect. 5, we decompose a reaction network into subnetworks (as a pre-processing step) and finally into independent subnetworks. In Sect. 6, we present our main result, the dependency one theorem for mass-action systems, and in Sect. 7, we derive the extended deficiency one theorem as a special case. Finally, in Sect. 8, we provide three examples where the conditions of the deficiency one theorem are not satisfied, but the dependency one theorem can be applied (to conclude the unique existence of a positive equilibrium within every stoichiometric compatibility class for some or all rate constants). Notation For vectors x,y∈Rn, we denote their scalar product by x·y∈R and their component-wise (Hadamard) product by x∘y∈Rn. We denote the vector with all entries equal to one by 1n∈Rn and the identity matrix by Idn∈Rn×n. We denote the positive (non-negative) real numbers by R> (R≥). For x∈R>n and y∈Rn, we define the monomial xy=∏i=1n(xi)yi∈R>; and for Y=(y1,…,ym)∈Rn×m, we define the vector of monomials xY∈R>m via (xY)j=xj. For x∈Rn, we define ex=(ex1,ex2,…,exn)T∈R>n; and for x∈R>n, we define ln(x)=ln(x1),ln(x2),…,ln(xn)T∈Rn. Previous Results on Polynomial Systems In order to state Theorem 2 below for a parametrized system of generalized polynomial equations, we introduce geometric objects and auxiliary matrices as defined in (Müller and Regensburger 2023, 2026). Definition 1 For the positive variable vector x∈R>n, a coefficient matrix A∈Rn′×m, an exponent matrix B∈Rn×m, and a positive parameter vector c∈R>m, we define the parametrized system of generalized polynomial equations (i)We call C=kerA∩R>m the coefficient cone. Its closure C¯=kerA∩R≥m is a polyhedral cone, called an s-cone (subspace cone) in (Müller and Regensburger 2016). As a necessary condition for the existence of solutions, C must be non-empty. We call C=kerA∩R>m the coefficient cone. Its closure C¯=kerA∩R≥m is a polyhedral cone, called an s-cone (subspace cone) in (Müller and Regensburger 2016). As a necessary condition for the existence of solutions, C must be non-empty. (ii)We assume that A=A1…Aℓ∈Rn′×m with ℓ≥1 blocks Aj∈Rn′×mj (and hence m1+…+mℓ=m) such that the kernel of A is the direct product of the kernels of Aj, that is, kerA=kerA1×⋯×kerAℓ. Accordingly, B=B1…Bℓ∈Rn×m with ℓ blocks Bj∈Rn×mj and cT=(c1)T…(cℓ)T∈R>m with cj∈R>j. The decomposition of kerA induces a partition of the indices {1,…,m} into ℓ classes. In particular, the columns of B=(b1,…,bm) and hence the monomials xj, j=1,…,m are partitioned into classes. We assume that A=A1…Aℓ∈Rn′×m with ℓ≥1 blocks Aj∈Rn′×mj (and hence m1+…+mℓ=m) such that the kernel of A is the direct product of the kernels of Aj, that is, kerA=kerA1×⋯×kerAℓ. Accordingly, B=B1…Bℓ∈Rn×m with ℓ blocks Bj∈Rn×mj and cT=(c1)T…(cℓ)T∈R>m with cj∈R>j. The decomposition of kerA induces a partition of the indices {1,…,m} into ℓ classes. In particular, the columns of B=(b1,…,bm) and hence the monomials xj, j=1,…,m are partitioned into classes. (iii)We introduce the direct product Δ=Δm1-1×⋯×Δmℓ-1 of the standard simplices Δmj-1={y∈R≥j∣1mj·y=1} and define the bounded set P=C∩Δ. Clearly, P=P1×⋯×Pℓ with Pj=Cj∩Δmj-1. We call P the coefficient polytope. In fact, P is a polytope without boundary. Strictly speaking, only its closure P¯ is a polytope. We introduce the direct product Δ=Δm1-1×⋯×Δmℓ-1 of the standard simplices Δmj-1={y∈R≥j∣1mj·y=1} and define the bounded set P=C∩Δ. Clearly, P=P1×⋯×Pℓ with Pj=Cj∩Δmj-1. We call P the coefficient polytope. In fact, P is a polytope without boundary. Strictly speaking, only its closure P¯ is a polytope. (iv)Let Im=Idm-1-1m-1T∈Rm×(m-1), which can be seen as the incidence matrix of a star-shaped graph with vertices {1,…,m} and root m. We introduce the ℓ×ℓ block-diagonal (incidence) matrix I=Im10⋱0Imℓ∈Rm×(m-ℓ) with blocks Imj∈Rmj×(mj-1) and the “monomial difference” matrix M=BI∈Rn×(m-ℓ). Clearly, M=B1Im1…BℓImℓ is generated by taking the differences between the first mj-1 columns of Bj and its last column, for j=1,…,ℓ, and hence L=imM⊆Rn is the sum of the linear subspaces associated with the affine spans of the columns of B in the ℓ classes. We call L the monomial difference subspace. Further, we call d=dim(kerM) the monomial dependency. It can be determined as d=m-ℓ-dimL, cf. (Müller and Regensburger 2026, Proposition 1). Let Im=Idm-1-1m-1T∈Rm×(m-1), which can be seen as the incidence matrix of a star-shaped graph with vertices {1,…,m} and root m. We introduce the ℓ×ℓ block-diagonal (incidence) matrix I=Im10⋱0Imℓ∈Rm×(m-ℓ) with blocks Imj∈Rmj×(mj-1) and the “monomial difference” matrix M=BI∈Rn×(m-ℓ). Clearly, M=B1Im1…BℓImℓ is generated by taking the differences between the first mj-1 columns of Bj and its last column, for j=1,…,ℓ, and hence L=imM⊆Rn is the sum of the linear subspaces associated with the affine spans of the columns of B in the ℓ classes. We call L the monomial difference subspace. Further, we call d=dim(kerM) the monomial dependency. It can be determined as d=m-ℓ-dimL, cf. (Müller and Regensburger 2026, Proposition 1). (v)We introduce the ℓ×ℓ block-diagonal “Cayley” matrix J=1m1T0⋱01mℓT∈Rℓ×m with blocks 1mjT∈R1×mj and the matrix B=BJ∈R(n+ℓ)×m. We call D=kerB⊂Rm the monomial dependency subspace. It records affine dependencies between the columns of B within the ℓ classes. In fact, dimD=d, cf. (Müller and Regensburger 2026, Lemma 4). We introduce the ℓ×ℓ block-diagonal “Cayley” matrix J=1m1T0⋱01mℓT∈Rℓ×m with blocks 1mjT∈R1×mj and the matrix B=BJ∈R(n+ℓ)×m. We call D=kerB⊂Rm the monomial dependency subspace. It records affine dependencies between the columns of B within the ℓ classes. In fact, dimD=d, cf. (Müller and Regensburger 2026, Lemma 4). (vi)Finally, we introduce the “exponentiation” matrix E=IM∗∈Rm×n, where M∗∈R(m-ℓ)×n is a generalized inverse of M. Finally, we introduce the “exponentiation” matrix E=IM∗∈Rm×n, where M∗∈R(m-ℓ)×n is a generalized inverse of M. We can now state the main result of our previous work on parametrized systems of polynomial inequalities (Müller and Regensburger 2026), instantiated for equations, cf. (Müller and Regensburger 2023). Theorem 2 (Müller and Regensburger 2023, Theorem 1) Consider the parametrized system of generalized polynomial equations A(c∘xB)=0. The solution set Zc={x∈R>n∣A(c∘xB)=0} can be written as where is the solution set on the coefficient polytope P. Theorem 2 can be read as follows: In order to determine the solution set Zc, first determine the solution set on the coefficient polytope, Yc. Recall that the coefficient polytope P is determined by the coefficient matrix A, and the dependency subspace D is determined by the exponent matrix B (and the classes). To a solution y∈Yc, there corresponds the actual solution x=(y∘c-1)E∈Zc. In fact, if (and only if) dimLn and let S⊆Rn be a subspace. Then we have Theorem 3 was originally proved by Birch (Birch, 1963) and reproved by Horn and Jackson (Horn and Jackson 1972, Lemma 4B) in the context of reaction networks with mass-action kinetics. Motivated by applications, Birch’s theorem has been extended to cover generalized mass-action kinetics (Müller et al. 2019; Craciun et al. 2019). Example We consider a parametrized system of two non-overlapping trinomials in two variables, arising from a reaction network.1 Equivalently, where Clearly, the coefficient matrix A=A1A2∈R2×6 has two blocks A1,A2∈R2×3 such that kerA=kerA1×kerA2. Accordingly, the exponent matrix B=B1B2 has two blocks that is, there are two classes of monomials, 1, x1, x13 and x13, x12x2, x1x22, respectively. Further, c=c1c2 with c1=(k12,k21,k43)T and c2=(k45,k56,k65)T. For the coefficient cone C=kerA∩R>6, we find C=C1×C2 with Cj=kerAj∩R>3, and for the coefficient polytope P=C∩Δ, where Δ=Δ2×Δ2 is a product of the standard simplices Δ2={y∈R≥3∣13·y=1}, we find with Pj=kerAj∩R>3∩Δ2 and dimP1=dimP2=1. By appending a Cayley matrix (rows of ones corresponding to the two blocks) to the exponent matrix B, we obtain which allows to determine the monomial dependency subspace D=kerB and the dependency d=dimD. In particular, d=dim(D)=2. In this example, the polynomial system is decomposable, that is, D=D1×D2 with D1=im2-31T and D2=im1-21T corresponding to the decomposition kerA=kerA1×kerA2. Clearly, d1=dimD1=d2=dimD2=1 and hence d=d1+d2. Finally, we obtain the monomial difference matrix M from B=B1B2 by taking the differences of the first two columns and the last column within the two blocks, This allows to determine the monomial difference subspace L=imM. In fact, L=R2. Dependency One Systems First, we consider systems with one class, second, we consider decomposable systems. One Class For one class, we consider d=dimP=1. (The case d=dimP=0 is trivial.) Definition 4 For a parametrized system of generalized polynomial equations A(c∘xB)=0 with kerA∩R>m≠∅, one class, one-dimensional coefficient polytope, and monomial dependency one, let y1,y2∈(kerA∩R≥m) be the two vertices of the coefficient polytope, let q=(y1-y2)∘(y1+y2)-1∈Rm, and assume that 1=q1≥⋯≥qm=-1 (after reordering of the index set {1,…,m}). Further, let I1,…,Iω⊂{1,…,m} be ω equivalence classes corresponding to equal (consecutive) components of q, and let q¯∈Rω with q¯i=qi′ for i′∈Ii be the vector of different q’s. Finally, let b∈Rm with imb=ker(B), and let b¯∈Rω with b¯i=∑i′∈Iibi′ be the vector of lumped b’s. Theorem 5 (d=1, one class) Let A(c∘xB)=0 be a parametrized system of generalized polynomial equations with kerA∩R>m≠∅, one class, one-dimensional coefficient polytope (dimP=1), and monomial dependency one (d=1). Then, |Yc|=1 for all c if ∑i′=1ib¯i′≥0 for all i=1,…,ω-1 (or “≤0” for all i) and b¯1·b¯ω<0. Proof Let y^=y1-y22, y¯=y1+y22>0, and hence q=y^∘y¯-1. Every y∈P can be written as y=y¯+ty^ with t∈(-1,1), and the binomial condition yb=cb for y∈P can be written as (y¯+ty^)b=cb for t∈(-1,1) or, after division by y¯, as f(t):=(1+tq)b=cby¯-b=:c∗. After considering equal components of q, Now, let b¯1·b¯ω<0, in particular, b¯1>0 and b¯ω<0. (The other case is analogous.) Clearly, q¯1=1 implies f(-1)→0, and q¯ω=-1 implies f(1)→∞. By continuity, there is a solution to f(t)=c∗ for all c∗ and hence for all c. That is, |Yc|≥1 for all c. Moreover, f′(t)=f(t)h(t) with Now, also let ∑i′=1ib¯i′≥0 for all i=1,…,ω-1 (or “≤0” for all i). Altogether, this implies f′(t)>0 (or f′(t)<0). That is, |Yc|≤1 for all c. □ Recently, existence of a unique, nondegenerate solution to a parametrized system of generalized polynomial equations (with arbitrary dependency and using one class) for all parameters) has been characterized using Hadamard’s Global Inversion Theorem (Deshpande and Müller 2024). Notably, existence (without uniqueness) of a solution on the coefficient polytope for all parameters can be characterized. Theorem 6 Let A(c∘xB)=0 be a parametrized system of generalized polynomial equations with kerA∩R>m≠∅, one class, one-dimensional coefficient polytope, and monomial dependency one. The following statements are equivalent: |Yc|≥1 for all c. |Yc|≥1 for all c. b¯1·b¯ω<0. b¯1·b¯ω<0. Proof See Appendix A. □ Decomposable Systems Definition 7 For a parametrized system of generalized polynomial equations A(c∘xB)=0 with kerA∩R>m≠∅ and ℓ classes, let Pj be the coefficient polytope and dj be the monomial dependency of the subsystem Aj(cj∘xBj)=0, j=1,…,ℓ. If dimPj=1, let yj,1,yj,2∈(kerAj∩R≥j) be the two vertices of Pj, let qj=(yj,1-yj,2)∘(yj,1+yj,2)-1∈Rmj, and assume that 1=q1j≥q2j≥⋯≥qjj=-1 (after reordering of the index set {1,…,mj}). Further, let I1j,I2j,…,Ijj⊂{1,…,mj} be ωj equivalence classes corresponding to equal (consecutive) components of qj. If dj=1, let bj∈Rmj with imbj=ker(Bj), and let b¯j∈Rωj with b¯ij=∑i′∈Iib′j be the vector of lumped bj’s. Theorem 8 (dj≤1) Let A(c∘xB)=0 be a parametrized system of generalized polynomial equations with ℓ classes that fulfills the following conditions: (i)kerA∩R>m≠∅. kerA∩R>m≠∅. (ii)d=d1+⋯+dℓ. d=d1+⋯+dℓ. (iii)For every (class) j=1,…,ℓ,dj=dimPj≤1.If dj=1, then∑i′=1ib¯′j≥0 for all i=1,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0. For every (class) j=1,…,ℓ,dj=dimPj≤1.If dj=1, then∑i′=1ib¯′j≥0 for all i=1,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0. dj=dimPj≤1. dj=dimPj≤1. If dj=1, then∑i′=1ib¯′j≥0 for all i=1,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0. If dj=1, then∑i′=1ib¯′j≥0 for all i=1,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0. ∑i′=1ib¯′j≥0 for all i=1,…,ωj-1 (or “≤0” for all i) ∑i′=1ib¯′j≥0 for all i=1,…,ωj-1 (or “≤0” for all i) b¯1j·b¯jj<0. b¯1j·b¯jj<0. Then, |Yc|=1 for all c. Proof By (i), the coefficient polytope P=P1×⋯×Pℓ is non-empty. By (ii) and (Müller and Regensburger 2026, Section 3.2), the system is decomposable. By (Müller and Regensburger 2026, Proposition 10), Yc=Yc,1×⋯×Yc,ℓ, that is, the solution set on the coefficient polytope is a direct product. In particular, |Yc|=|Yc,1|·⋯·|Yc,ℓ|. By the first item in (iii), dj=dimPj≤1. Case dj=dimPj=0: Pj is a point, and there is no binomial condition on Pj. Hence, |Yc,j|=1. Case dj=dimPj=1: By the second item in (iii), ∑i′=1ib¯′j≥0 for all i=1,…,ωj-1 (or “≤0” for all i) and b¯1j·b¯jj<0. By Theorem 5 for the subsystem Aj(cj∘xBj)=0, |Yc,j|=1. Altogether, |Yc|=|Yc,1|·⋯·|Yc,ℓ|=1, and all implications hold for all c. □ Reaction Networks We recall basic notions for reaction networks with mass-action kinetics (Adleman et al. 2014, Voit et al. 2015, Yu and Craciun 2018, Gunawardena 2003) following (Müller and Regensburger 2014). Note that we use index notation (rather than pure cardinality notation) in the definition of matrices, thereby avoiding arbitrary orderings of the underlying sets. A reaction network (G, y) is given by a simple directed graph G=(V,E) with a finite set of vertices V and edge set E⊆V×V together with an injective map y:V→Rn (a matrix Y∈Rn×V). Every vertex i∈V is labeled with a (stoichiometric) complex y(i)∈Rn, and every edge (i→i′)∈E represents a reaction y(i)→y(i′). The restriction of the map y (the matrix Y) to the set of source vertices Vs⊆V is denoted by ys:Vs→Rn (respectively, Ys∈Rn×Vs). If all components of the graph are strongly connected, the network is called weakly reversible. Remark In the classical definition of a reaction network, complexes (like reactions) are primary objects and correspond to the vertices (and edges) of the induced complex-reaction graph. Such reaction networks can also be represented as Euclidean-embedded graphs (Craciun 2015). See also (Craciun 2019; Craciun and Deshpande 2020). A mass-action system (Gk,y) is given by a reaction network (G, y) and positive edge labels k∈R>E. Every edge/reaction (i→i′)∈E is labeled with a rate constant ki→i′>0. The associated ODE system for the non-negative concentrations x∈R≥n (of n chemical species) is given by The sum ranges over all reactions, and every summand is a product of the reaction rate ki→i′xy(i), involving a monomial xy=∏j=1n(xj)yj determined by the stoichiometric complex of the reactant, and the reaction vector y(i′)-y(i) given by the stoichiometric complexes of product and reactant. Let IE∈{-1,0,1}V×E and IE,s∈{0,1}Vs×E be the incidence and source matrices of the digraph G, respectively, and Rk=IEdiag(k)(IE,s)T∈RV×Vs be the rectangular “Laplacian matrix”. (For details on the index notation, see Appendix B.1.) Further, recall the (source) complex matrices, Y∈Rn×V and Ys∈Rn×Vs, and let N=YIE∈Rn×E be the stoichiometric matrix. Then, the right-hand-side of (1) can be written in matrix form and decomposed into stoichiometric and graphical contributions, In the last step, we have introduced the kinetic matrix Γk=YRk∈Rn×Vs, in analogy to the stoichiometric matrix N=YIE. Remark Traditionally, one uses the source matrix IE,s′∈{0,1}V×E which involves all vertices (not just the source vertices), and one obtains with the (square) Laplacian matrix Lk∈RV×V. This formulation can be misleading since columns of Lk corresponding to non-source vertices are zero, and, after multiplication, non-source monomials do not appear in dxdt=YLkxY. Indeed, Rk arises from Lk by deleting zero columns, and imRk=imLk. The change over time lies in the kinetic subspace K=im(YRk) and further in the stoichiometric subspace stoichiometric subspace S=im(YIE), Hence, trajectories are confined to cosets of K and S, respectively, that is, x(t)∈x(0)+K⊆x(0)+S. For positive x′∈R>n, the sets (x′+K)⊆(x′+S)∩R>n are called kinetic and stoichiometric compatibility classes, respectively. Finally, we introduce non-negative integer characteristics of a graph or a reaction network. In particular, let l be the number of components of G (not to be confused with ℓ defined below), let t be the number of absorbing strong components, and let t′ be the number of non-singleton absorbing strong components. It is well known that which further implies Analogously, Most importantly, the (stoichiometric) deficiency is given by From the facts above, it can be easily shown that Mass-action systems with K≠S can be “pathological”. The assumption t=l in the classical deficiency one theorem (Feinberg 1987) rules out such systems. In the extended deficiency one theorem (Feinberg 1995a) (for independent subnetworks), this assumption is missing. For a summary of the notation introduced in this section, see Table 1(a). For an illustration, we continue the example from Sect. 2. Table 1Notation introduced in Sects. 4 and 5. (a) Matrices and subspaces for a reaction network (G, y) / a mass-action system (Gk,y) with underlying graph G=(V,E) and edge labels k. (b) Network decomposition: objects in (a) for subnetworks (Gj,yj) with Gj=(Vj,Ej) and combined objects for the full network (G, y). (c) Polytope and subspaces for the parametrized system of polynomial equations Γk∗xYs∗=0, arising from the decomposed dynamical system (4). See Definition 1 for Ac∘xB=0 with A=Γk∗, B=Ys∗, and c=1; analogously for the subsystems ΓkjxYsj=0symboldimensionnamerelation (a) reaction network IE V×E incidence matrix IE,s Vs×E source matrix Rk V×Vs (rectangular) Laplacian matrix Rk=IEdiag(k)(IE,s)T Y n×V complex matrix Ys n×Vs source complex matrix N n×E stoichiometric matrix N=YIE Γk n×Vs kinetic matrix Γk=YRk S stoichiometric subspace S=imN K kinetic subspace K=imΓk l number of connected components t number of absorbing strong comp δ deficiency δ=|V|-l-dim(S) (b) network decomposition ℓ number of (independent) subnetworks IEj,IE,sj,Rkj,Yj,⋯ objects in (a) for j=1,⋯,ℓ V⊔ disjoint union of vertex sets V⊔=V1⊔⋯⊔Vℓ Vs⊔ disjoint union of source vertex sets Vs⊔=Vs1⊔⋯⊔Vs2 IE∗ V⊔×E combined incidence matrix IE,s∗ Vs⊔×E combined source matrix Rk∗ V⊔×Vs⊔ combined (rect.) Laplacian matrix Rk∗=IE∗diag(k)(IE,s∗)T Y∗ n×V⊔ combined complex matrix Ys∗ n×Vs⊔ combined source complex matrix Γk∗ n×Vs⊔ combined kinetic matrix Γk∗=Y∗Rk∗ K∗ combined kinetic subspace K∗=imΓk∗ (c) polynomial equations P coefficient polytope L monomial difference subspace D monomial dependency subspace d monomial dependency d=dimD Pj,Lj,Dj,dj objects for j=1,⋯,ℓ Notation introduced in Sects. 4 and 5. (a) Matrices and subspaces for a reaction network (G, y) / a mass-action system (Gk,y) with underlying graph G=(V,E) and edge labels k. (b) Network decomposition: objects in (a) for subnetworks (Gj,yj) with Gj=(Vj,Ej) and combined objects for the full network (G, y). (c) Polytope and subspaces for the parametrized system of polynomial equations Γk∗xYs∗=0, arising from the decomposed dynamical system (4). See Definition 1 for Ac∘xB=0 with A=Γk∗, B=Ys∗, and c=1; analogously for the subsystems ΓkjxYsj=0 Example (continued) We return to the reaction network (with rate constants) or, equivalently, the “embedded” graph In our definition, the reaction network (G, y) is given by the graph G=(V,E) with vertex set V={1,2,3,4,5,6} and edge set E={12,21,43,45,56,65} (where we write ij short for i→j) and the map y:V→Rn with n=2 and y(1)=00,y(2)=10, y(3)=20,y(4)=30, y(5)=21,y(6)=12 (which labels vertices with complexes). The mass-action system (Gk,y) further labels edges with positive rate constants k=(k12,k21,k43,k45,k56,k65)T∈R>E. Note that the set of source vertices is Vs={1,2,4,5,6}. First, the V×E incidence and Vs×E source matrices (of the graph) are given by with resulting V×Vs (rectangular) Laplacian matrix Second, the n×V complex and n×Vs source complex matrices (arising from the underlying map) are given by Third, the n×E stoichiometric and n×Vs kinetic matrices are given by and They define the stoichiometric and kinetic subspaces, S=imN=R2 and K=imΓk=R2 (for all k). The network has l=2 components (with vertex sets {1,2} and {3,4,5,6}) and t=3 absorbing strong components (with vertex sets {1,2}, {3}, and {5,6}), in particular, t≠l. The resulting deficiency is δ=|V|-l-dimS=6-2-2=2. Further, we note that the deficiencies of the two components are δ1=2-1-1=0 and δ2=4-1-2=1, and hence δ≠δ1+δ2. Network Decomposition First, we decompose a reaction network/mass-action system and the associated ODE. Then, in Subsection 5.1, we formulate the resulting polynomial equations for positive equilibria. Only in Subsection 5.2, we assume independent subnetworks (in the sense of Feinberg (1987)) and address the decomposability of the polynomial equations (in the sense of Müller and Regensburger (2026)). In order to decompose the right-hand side of the ODE of a mass-action system, we proceed in three steps. (i) We assume that the edge set is partitioned into ℓ disjoint subsets, E=E1∪˙⋯∪˙Eℓ. (For the moment, the partition is arbitrary. In Subsection 5.2, it will arise from writing the kernel of N=YIE as a direct product.) The partition induces the subgraphs Gj=(Vj,Ej), j=1,…,ℓ, with vertex sets Vj, source vertex sets Vsj, and non-source vertex sets Vnsj. Note that the vertex sets need not be disjoint. Further note that absorbing strong components are either non-source vertices or non-singleton strong components. Let tj denote the number of absorbing strong components (of the subgraph Gj) and tj′ denote the number of non-singleton absorbing strong components. Clearly, tj=tj′+|Vnsj|. The subgraph Gj has incidence and source matrices The corresponding subnetwork (Gj,yj) has complex matrix Yj∈Rn×Vj and source complex matrix Ysj∈Rn×Vsj. (All matrices for a subgraph are submatrices of the corresponding matrices for the full graph.) Clearly, The vector of parameters k∈R>E is partitioned accordingly, with kj∈R>j, and the rectangular Laplacian matrix of the mass-action system (Gkj,yj) is given by (ii) In turn, we define “combined” block(-diagonal) matrices from the matrices of the subgraphs. Since the vertex sets need not be disjoint, we form the disjoint unions of (source) vertex sets, V⊔=V1⊔⋯⊔Vℓ and Vs⊔=Vs1⊔⋯⊔Vsℓ. We introduce the combined incidence and source matrices and the combined rectangular Laplacian matrix The (source) complex matrices can be combined accordingly, On the one hand, Y∗IE∗=YIE. On the other hand, Ys∗=YsIV,s∗ with a matrix IV,s∗∈{0,1}Vs×Vs⊔ that assigns to every source vertex of a subgraph the corresponding source vertex of the full graph. (iii) Now, we return to the ODE (3). Like the matrix YIE, the vector of monomials xYsIE,s∈R≥E has ℓ blocks and can be decomposed as Using the definitions above, we decompose the right-hand side of the ODE as Finally, we introduce the kinetic matrices and the combined kinetic matrix and summarize the decomposition as Before we proceed with the treatment of the resulting polynomial equations, we introduce linear subspaces associated with a subnetwork (Gj,yj) or a mass-action system (Gkj,yj), j=1,…,ℓ. As above, we define the stoichiometric and kinetic subspaces Sj=im(YjIEj) and Kj=im(YjRkj), where Kj⊆Sj since imRkj⊆imIEj, and further the combined kinetic subspace K∗=im(Y∗Rk∗)=∑jKj. We record the following fact. Fact 9 Let (G, y) be a reaction network decomposed into ℓ subnetworks. Then, S=∑j=1ℓSj and K⊆K∗=∑j=1ℓKj. Proof Recall YIE=Y∗IE∗=Y1IE1…YℓIEℓ and Y∗Rk∗=Y1Rk1…YℓRkℓ. First, Second, K=im(YRk) and K∗=im(Y∗Rk∗), where and Since YIE=Y∗IE∗ and im(IE,s)T⊆im(IE,s∗)T, we have im(YRk)⊆im(Y∗Rk∗). □ Positive Equilibria Finally, we consider positive equilibria of the ODE (3), that is, x∈R>n such that dxdt=0. Clearly, we can specify them in the form A(c∘xB)=0 as that is, with A=N∈Rn×E, c=k∈R>E, B=YsIE,s∈Rn×E; equivalently, using the decomposition (4), we can specify them as that is, with A=Γk∗∈Rn×Vs⊔, c=1, and B=Ys∗∈Rn×Vs⊔. In Equation (5a), the coefficient matrix A=N (and hence the classes determined by its kernel) do not depend on the parameters k. However, there may be repeated monomials (within classes) giving rise to trivial dependencies. In Equation (5b), the coefficient matrix A=Γk∗ does depend on k, but repeated monomials are handled via the (rectangular) Laplacian matrix (which also eliminates non-source monomials). Hence, we will consider Equation (5b), but use classes arising from Equation (5a). Independent Subnetworks and Decomposability Let (G, y) with G=(V,E) be a reaction network. In the following, we assume that the partition of the edge set, E=E1∪˙⋯∪˙Eℓ, arises from a decomposition of kerN (as a direct product) or, equivalently, from a decomposition of S=imN (as a direct sum). That is, such that Equivalently, In the terminology of (Feinberg 1987), the resulting subnetworks (Gj,yj) are independent. In the setting of the deficiency one theorem, we will assume that the subgraphs Gj are connected (and say that the subnetworks are connected), see Section 7. Now, let (Gk,y) be the corresponding mass-action system with combined kinetic matrix By Proposition 10 below, Equivalently, In the terminology of Müller and Regensburger (2026) (for the polynomial equations Γk∗xYs∗=0), the decomposition of kerΓk∗⊆RVs⊔ induces a partition of the source vertex set Vs⊔=Vs1⊔⋯⊔Vsℓ into the ℓ classes Vsj. It remains to provide a formal argument. Proposition 10 Let (G, y) be a reaction network decomposed into ℓ independent subnetworks and (Gk,y) be the corresponding mass-action system. Then, (6a) implies (6b). Proof Consider ξ∈RVs⊔ with blocks ξj∈RVsj, j=1,…,ℓ, and assume Γk∗ξ=∑j=1ℓΓkjξj=0. Further, introduce αj=diag(kj)(IE,sj)Tξj∈Rj. Then, and hence ∑j=1ℓNjαj=0. By (6a), Njαj=Γkjξj=0. That is, (6b). □ To summarize, the decomposition of the kernel of the stoichiometric matrix implies a partition of the edge set (and a corresponding decomposition of the network into independent subnetworks) and further a decomposition of the kernel of the combined kinetic matrix (and a corresponding partition of the source vertex set into classes). An algorithm for network decomposition has been developed in Hernandez and la Cruz (2021) and applied to derive steady states analytically in (Hernandez et al. 2023). For the polynomial equations Γk∗xYs∗=0, the monomial dependency and difference subspaces, D and L, are given in Definition 1 (with B=Ys∗). Assuming ℓ classes, we introduce corresponding subspaces Dj and Lj, j=1,…,ℓ. In particular, Lj=im(YsjIEsj), where IEsj∈{-1,0,1}Vsj×Es is the incidence matrix of a star-shaped graph (Vsj,Es) (on the source vertices) with |Es|=|Vsj|-1. As defined in (Müller and Regensburger 2026, Section 3.2), the polynomial equations Γk∗xYs∗=0 are decomposable if D has the decomposition D=D1×⋯×Dℓ, corresponding to the decomposition of kerΓk∗; equivalently, d=d1+⋯+dℓ, where dj=dimDj. We provide another characterization. Proposition 11 Let (G, y) be a reaction network decomposed into ℓ subnetworks. Then d=d1+⋯+dℓ if and only if L=L1⊕⋯⊕Lℓ. Proof Recall d=|Vs⊔|-ℓ-dim(L) and dj=|Vsj|-1-dim(Lj). Clearly, Vs⊔=Vs1⊔⋯⊔Vsℓ implies |Vs⊔|=|Vs1|+⋯+|Vsℓ|. Hence, if and only if dim(L)=∑j=1ℓdim(Lj), that is, L=L1⊕⋯⊕Lℓ. □ For a summary of the notation introduced in this section, see Table 1(bc). For an illustration, we continue the example from Sects. 2 and 4. Example (continued) For the last time, we return to the reaction network (with rate constants) The decomposition of the kernel of the stoichiometric matrix N=N1N2 with as a direct product, that is, kerN=kerN1×kerN2 implies a partition of the edge set, E=E1∪˙E2 with E1={12,21,43} and E2={45,56,65}, and hence a decomposition of the reaction network (G, y) into ℓ=2 independent subnetworks. In the “embedded” graph, the subnetworks (G1,y1) and (G2,y2) are shown in red and green, respectively. Note that G1=(V1,E1) has vertex set V1={1,2,3,4}, whereas V2={4,5,6} for G2=(V2,E2), and hence V1∩V2={4}≠∅. Accordingly, we introduce the disjoint union of vertex sets which contains two “copies” of the element in the intersection, 4∈V1∩V2, and the disjoint union of the source vertex sets Vs⊔=Vs1⊔Vs2={1,2,41,42,5,6}. From the matrices for the subnetworks, we define the corresponding block(-diagonal) combined matrices: the combined V⊔×E incidence and Vs⊔×E source matrices, the combined V⊔×Vs⊔ (rectangular) Laplacian matrix, the combined n×V⊔ complex and n×Vs⊔ source complex matrices, and the combined n×Vs⊔ kinetic matrix, The latter defines the combined kinetic subspace, K∗=imΓk∗=R2 (for all k). Dependency One Mass-Action Systems We present a dependency one theorem for mass-action systems, extending the deficiency one theorems by Feinberg (Feinberg 1987, 1995a), cf. Theorem 18. In Theorem 14 below, we neither assume δi≤1, ti=1, nor K=S. Most importantly, we need to guarantee d=dimP (per class), in order to apply Theorem 8. Lemma 12 Let (G, y) be a reaction network with one subnetwork (the network itself) and let (Gk,y) be the corresponding mass-action system. Consider the equilibrium equation ΓkxYs=0 (with one class) and assume kerΓk∩R>s≠∅. Then, Proof On the one hand, dimP+1=dimC=dim(kerΓk)=dim(ker(YRk))=|Vs|-dim(im(YRk))=|Vs|-dimK. On the other hand, d+1=|Vs|-dimL. Hence, dimP=d if and only if dimK=dimL. □ As stated above, we do not explicitly restrict the number of absorbing strong components (per class), in particular, if they are singletons. However, we record that dimP≤1 (an at most one-dimensional coefficient polytope) implies t′≤2 (at most two non-singleton absorbing strong components). Fact 13 Let (G, y) be a reaction network with one subnetwork (the network itself) and let (Gk,y) be the corresponding mass-action system. Consider the equilibrium equation ΓkxYs=0 (with one class) and assume kerΓk∩R>s≠∅. Then, In fact, t′=2 only if there are two components (l=2) which are strongly connected. Proof Let dimP≤1. Then, t′≤2 since If t′=2, then kerΓk=kerRk is generated by two non-negative vectors with support on the two non-singleton absorbing strong components. Since the union of their supports is the set of all source vertices, there are two components, and they are strongly connected. □ Theorem 14 (dj≤1, mass-action systems) Let (Gk,y) be a mass-action system with ℓ independent subnetworks. Recall the kinetic and stoichiometric subspaces, K and S, respectively. For the equilibrium equation Γk∗xYs∗=0, recall the monomial dependency d and the monomial difference subspace L. Let (Gk,y) fulfill the following conditions: kerΓk∗∩R>Vs⊔≠∅. kerΓk∗∩R>Vs⊔≠∅. d=d1+⋯+dℓ. d=d1+⋯+dℓ. For every (class) j=1,…,ℓ,dj≤1 and dim(Kj)=dim(Lj).If dj=1, then∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0.(The components of the (lumped) dependency vector are ordered with respect to the vector q¯j∈Rωj of the polytope Pj, see Definition 7.) For every (class) j=1,…,ℓ,dj≤1 and dim(Kj)=dim(Lj).If dj=1, then∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0.(The components of the (lumped) dependency vector are ordered with respect to the vector q¯j∈Rωj of the polytope Pj, see Definition 7.) dj≤1 and dim(Kj)=dim(Lj). dj≤1 and dim(Kj)=dim(Lj). If dj=1, then∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0.(The components of the (lumped) dependency vector are ordered with respect to the vector q¯j∈Rωj of the polytope Pj, see Definition 7.) If dj=1, then∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0. ∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i) ∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i) b¯1j·b¯jj<0. b¯1j·b¯jj<0. (The components of the (lumped) dependency vector are ordered with respect to the vector q¯j∈Rωj of the polytope Pj, see Definition 7.) L=K or (IVb) L=S. L=K or (IVb) L=S. Then, there exists a unique positive equilibrium within every (IVa) kinetic or (IVb) stoichiometric compatibility class. Proof The mass-action system (Gk,y) gives rise to the polynomial equations Γk∗xYs∗=0 with ℓ classes. In order to apply Theorem 8 (for Ac∘xB=0 with A=Γk∗, B=Ys∗, and c=1), we show that conditions (i), (ii), (iii) there follow from conditions (I), (II), (III) here. (i) By (I) with A=Γk∗. (ii) By (II). (iii) By (III), dim(Kj)=dim(Lj), and by Lemma 12, dj=dimPj (for j=1,…,ℓ). Regarding dj=1, (iii) and (III) are identical. By Theorem 8, the solution set on the coefficient polytope is a singleton, |Yc|=1, and by Theorem 2, the solution set is an exponential fiber, Zc=x∗∘e⊥ with x∗∈R>n. The set of positive equilibria within the kinetic compatibility class given by x′∈R>n is x∗∘e⊥∩(x′+K). By (IVa), K=L, and by Theorem 3, the intersection is a singleton. The set of positive equilibria within the stoichiometric compatibility class given by x′∈R>n is x∗∘e⊥∩(x′+S). By (IVb), L=S, and by Theorem 3, the intersection is a singleton. □ We now provide a simplified version of Theorem 14 that is closer to the (extended) deficiency one theorem (Theorem 18 in Sect. 7.1). In particular, we assume that the kinetic, stoichiometric, and monomial difference subspaces of the subnetworks (and hence of the full system) agree. Corollary 15 Let (Gk,y) be a mass-action system with ℓ independent subnetworks. Recall the kinetic and stoichiometric subspaces, K and S, respectively. For the equilibrium equation Γk∗xYs∗=0, recall the monomial dependency d and the monomial difference subspace L. Let (Gk,y) fulfill the following conditions: (I’) kerΓk∗∩R>Vs⊔≠∅. (I’) kerΓk∗∩R>Vs⊔≠∅. (III’) For every (independent subnetwork) j=1,…,ℓ,dj≤1 and Kj=Sj=Lj.If dj=1, then∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0.(The components of the (lumped) dependency vector b¯j∈Rωj are ordered with respect to the vector q¯j∈Rωj of the polytope Pj, see Definition 7.) (III’) For every (independent subnetwork) j=1,…,ℓ,dj≤1 and Kj=Sj=Lj.If dj=1, then∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0.(The components of the (lumped) dependency vector b¯j∈Rωj are ordered with respect to the vector q¯j∈Rωj of the polytope Pj, see Definition 7.) dj≤1 and Kj=Sj=Lj. dj≤1 and Kj=Sj=Lj. If dj=1, then∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0.(The components of the (lumped) dependency vector b¯j∈Rωj are ordered with respect to the vector q¯j∈Rωj of the polytope Pj, see Definition 7.) If dj=1, then∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i)b¯1j·b¯jj<0. ∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i) ∑i′=1ib¯′j≥0 for all i=1,2,…,ωj-1 (or “≤0” for all i) b¯1j·b¯jj<0. b¯1j·b¯jj<0. (The components of the (lumped) dependency vector b¯j∈Rωj are ordered with respect to the vector q¯j∈Rωj of the polytope Pj, see Definition 7.) Then, there exists a unique positive equilibrium within every stoichiometric compatibility class. Proof The mass-action system (Gk,y) gives rise to the polynomial equations Γk∗xYs∗=0 with ℓ classes (determined by the ℓ independent subnetworks), cf. Proposition 10. In order to apply Theorem 14, we show that conditions (I), (II), (III), (IVb) there follow from conditions (I’), (III’) here. (I) By (I’). (II) By (III’), Lj=Sj, and by the decomposition of kerN, S=S1⊕⋯⊕Sℓ. Hence, L=L1⊕⋯⊕Lℓ, and (II) by Proposition 11. (III) By (III’). (VIb) By (III’), see the argument for (II). □ Deficiency One Mass-Action Systems In the setting of the deficiency one theorem, we assume that the network is decomposed into ℓ independent subnetworks with one (absorbing) strong component. That is, the classes arise from kerN and they are assumed to be connected. This implies δ=δ1+⋯+δℓ, see Fact 17. The main result of this subsection is Lemma 16, a graph-theoretical argument used in the proof for the existence of a positive equilibrium in the case of weakly reversible networks, see Theorem 18 in Subsection 7.1 and its proof in Subsection 7.3. Given the directed graph G=(V,E) of a network, we introduce the bipartite graph G=(C,V,E) with the first vertex set C={1,…,ℓ} representing subnetworks of G, the second vertex set V⊆V representing “shared” vertices (appearing in more than one subnetwork), and the edge set E={(j,i)∈C×V∣i∈Vj}, connecting shared vertices with their subnetworks. Lemma 16 Let (G, y) with G=(V,E) be a reaction network decomposed into ℓ connected, independent subnetworks. Recall the (combined) incidence matrices, IE and IE∗, the number l of connected components, and the disjoint union V⊔ of source vertex sets. Then, |V⊔|-|V|=ℓ-l, that is, kerIE=kerIE∗. Proof Since we assume that subgraphs are connected, all vertices of a given subgraph derive from only one component of the original graph, and we can consider the decomposition of l=1 component into ℓ subgraphs. For the resulting bipartite graph G, |C|=ℓ and Since G is connected, |E|≥|C|+|V|-1, that is, V⊔-V≥ℓ-1. For clarity, note that a “double edge” (iji′) indicates that the shared vertices i,i′ reside in subgraph j and hence 0≠y(i′)-y(i)∈Sj. Now, assume V⊔-V>ℓ-1. Then, there is a cycle of n≥2 “double edges” with vertex sequence i1j1i2j2…injnin+1, where in+1=i1, i∗∈V⊆V, j∗∈C={1,…,ℓ}. Hence, contradicting S=S1⊕⋯⊕Sℓ. Hence, V⊔-V=ℓ-1. By combining the arguments for the l components of the original graph, V⊔-V=ℓ-l. Recall IE∈{-1,0,1}V×E and IE∗∈{-1,0,1}V⊔×E. Obviously, kerIE⊇kerIE∗. By the rank-nullity theorem, dim(kerIE)=|E|-dim(imIE)=|E|-dim(imIET)=|E|-|V|+dim(kerIET)=|E|-|V|+l as well as dim(kerIE∗)=|E|-|V⊔|+ℓ. Hence, dim(kerIE)-dim(kerIE∗)=|V⊔|-|V|-(ℓ-l). □ Fact 17 Let (G, y) be a reaction network decomposed into ℓ connected subnetworks. The following statements are equivalent. S=S1⊕⋯⊕Sℓ. S=S1⊕⋯⊕Sℓ. δ=δ1+⋯+δℓ and |V⊔|-|V|=ℓ-l. δ=δ1+⋯+δℓ and |V⊔|-|V|=ℓ-l. Proof 2⇒1. On the one hand, δ=|V|-l-dimS. On the other hand, δj=|Vj|-1-dimSj and hence ∑jδj=|V⊔|-ℓ-∑jdimSj. Altogether, 1⇒2. By Lemma 16, if S=S1⊕⋯⊕Sℓ, then |V⊔|-|V|=ℓ-l. By the displayed equation above, also δ=∑jδj. □ In the “classical” setting (ℓ=l), the connected subnetworks are the components of the graph (the linkage classes), and hence they are independent exactly when their deficiencies add up to the network deficiency. The (extended) deficiency one theorem We prove Theorem 18 below which already extends the “classical” deficiency one theorem (Feinberg 1987, Theorem 6.2.1) from components (“linkage classes”) to independent subnetworks, cf. (Feinberg 1987, Remark 6.2.D) and (Feinberg 1995a, Theorem A.1). Theorem 18 (δj≤1, independent subnetworks) Let (Gk,y) be a mass-action system with ℓ independent subnetworks. Recall the deficiency δ and the number t of absorbing strong components. Let (Gk,y) fulfill the following conditions: For every (independent subnetwork) j=1,…,ℓ, (i)δj≤1 and tj=1. δj≤1 and tj=1. If (ii) there exists a positive equilibrium, then there exists a unique positive equilibrium within every stoichiometric compatibility class. If the system is weakly reversible, then (ii) for all rate constants k. Remark (on the statement) Note that in the deficiency one theorems, the existence of a positive equilibrium is assumed. In our dependency one theorem, we only assume the necessary condition kerΓk∗∩R>Vs⊔≠∅. Indeed, in Theorem 18, the existence assumption can be replaced by the necessary condition. A different (more involved) condition was given by (Boros 2013b). Remark (on the proof) A detailed proof of the classical theorem is given in (Feinberg 1995a, Sections 5–8 and Appendix B). For an outline of the proof of the extension (to independent subnetworks or “direct partitions”), see (Feinberg 1995a, Appendix A). In (Feinberg (1995a), Sections 5 and 6), Feinberg shows that the set of positive equilibria has the form x∗∘e⊥ (a particular solution x∗ times the exponentiation of the orthogonal complement of the stoichiometric subspace S), by reformulating the problem using a monotonic function. Our approach is based on a geometric framework for parametrized systems of polynomial equations and inequalities (Müller and Regensburger 2023, 2026). Within this framework, we establish the existence of a point y∗ on the coefficient polytope (where y∗ is in one-to-one correspondence with x∗∘e⊥). As Feinberg, we use the “second salt theorem” (a network argument) and a version of “Birch’s theorem”. In the case of weakly reversible networks, we replace the proof for the existence of a positive equilibrium in (Feinberg 1995a, Section 8) by a graph-theoretical argument (see the last paragraph in the proof of Theorem 18, which uses Lemma 16). In Subsection 7.3, we will show that the deficiency one theorem follows from our more general dependency one results, Theorem 14 or its Corollary 15. Indeed, we exhibit how the network conditions in Theorem 18 ensure the conditions in Corollary 15. First, we show that the kinetic, stoichiometric, and monomial difference subspaces of the subnetworks agree. Lemma 19 Let (Gk,y) be a mass-action system with t=1 (one absorbing strong component). Recall the number t′ of non-singleton absorbing strong components and the deficiency δ. Further, recall the kinetic and stoichiometric subspaces, K and S, respectively. For the equilibrium equation ΓkxYs=0, recall the dependency d and the monomial difference subspace L, and assume that kerΓk∩R>s≠∅. Then, K=L=S. K=L=S. d=δ+t′-1. d=δ+t′-1. Proof 1. Since t=l=1, imRk=imIE and hence K=im(YRk)=im(YIE)=S. If t′=1 (the absorbing strong component is not a singleton), then Vs=V and L=S. If t′=0 (the absorbing strong component is a singleton), then V=Vs∪˙{i∗} and E=Es∪˙E∗, where Es is the set of reactions between source vertices (and hence E∗ is the set of reactions with target vertex i∗). Now, note that S is the linear subspace associated with the affine hull of the complexes, aff(Y)=aff(y(i)∣i∈V), and L is the linear subspace associated with the affine hull of the source complexes, aff(Ys)=aff(y(i)∣i∈Vs). We will show y(i∗)∈aff(Ys) which implies aff(Ys)=aff(Y) and hence L=S. Since kerΓk∩R>s≠∅, there is ξ∈R>s such that Γkξ=0, and hence there is α=diag(k)(IE,s)Tξ∈R>E such that YIEα=0. That is, YIEsαEs+YIE∗αE∗=0, using IE=IEsIE∗ and α=αEsαE∗. Clearly, YIE∗αE∗=-YIEsαEs∈L. Explicitly, where α¯=∑(i→i∗)∈E∗αi→i∗, βi→i∗=αi→i∗/α¯, and ∑(i→i∗)∈E∗βi→i∗=1. That is, y(i∗)=y+δy/α¯, where y∈aff(Ys) and δy∈L. Hence, also y(i∗)∈aff(Ys). 2. Recall d=|Vs|-1-dimL and δ=|V|-1-dimS. By statement 1, L=S and hence d-δ=|Vs|-|V|=t-t′=1-t′. Hence d=δ+t′-1. □ Regarding dimensions, the assumption t=1 has two consequences: if t′=1 (non-singleton absorbing strong component), then d=δ; however, if t′=0 (singleton absorbing strong component), then d=δ-1. In particular, δ=1 and t′=0 imply d=0. Second, to show that the sign conditions (for the dependency vector) in Corollary 15 are fulfilled, we provide a variant of Feinberg’s “second salt theorem” (Feinberg 2019). The Second Salt Theorem We continue our study of subnetworks (with δ≤1, t=1, and ker(YRk)∩R>s≠∅). By Lemma 12 and Lemma 19, d=dimP≤1. We consider the nontrivial case d=dimP=1 which has δ=t′=1, again by Lemma 19. Let y1,y2∈ker(YRk)∩R≥s be the vertices of P. Since t′=1, there is a non-negative y^∈P¯ with Rky^=0 and support on the (non-singleton) absorbing strong component. (It lies in the interior of the coefficient polytope, if the network is weakly reversible, and on the boundary, otherwise.) Since δ=1, there is a positive y¯∈P with Rky¯=β∈RV and imβ=kerY∩imRk. (It lies in the interior, possibly after adding y^.) That is, Without loss of generality, we assume that the vertices y1, y2 of the coefficient polytope (a line segment) and the vectors y^, y¯ are ordered as y1⋯y^⋯y¯⋯y2, where y^=y1 if the network is not weakly reversible. Now, y^=λ^y1+(1-λ^)y2 for some λ^∈(0,1] and y¯=λ¯y1+(1-λ¯)y2 for some λ¯∈(0,1), where λ^>λ¯. Let It is easy to see2 that the order of the entries of q^ agrees with the order of the entries of specified in Definition 4. Finally, we introduce the scaled rectangular graph Laplacian and we rewrite Equations (7) as Now, we state Feinberg’s “second salt theorem” (Feinberg 2019) for the rectangular (rather than the square) graph Laplacian. Lemma 20 Let Gk=(V,E,k) be a labeled simple digraph with one component. Let T={1,…,t}⊆V={1,…,m} be (the vertices of) an absorbing strong component, let q^∈R≥s be the corresponding vector in the kernel of the rectangular graph Laplacian, that is, Rkq^=0 and suppq^=T, and assume q^1≥⋯≥q^t>q^t+1=⋯=q^m=0. Further, let β=Rk1Vs∈RV. Then, If iq^i+1, then ∑i′=1iβi′>0. Finally, ∑i′=1tβi′=0 if and only if T=V. Proof See Appendix B.2. □ Proof of the extended deficiency one theorem (for independent subnetworks) Proof of Theorem 18 The mass-action system (Gk,y) gives rise to the polynomial equations Γk∗xYs∗=0 with ℓ classes (determined by the ℓ connected, independent subnetworks), cf. Proposition 10. In order to apply Corollary 15, we show that conditions (I’), (III’) there follow from conditions (i), (ii) here. (I’)If (ii), there exists a positive equilibrium, x∈R>n with Γk∗xYs∗=0, and hence (I’). This further implies kerΓkj∩R>Vsj≠∅ for the mass-action systems (Gkj,yk), j=1,…,ℓ, having tj=1 by (i). Hence, Lemma 19 applies to the subsystems. If (ii), there exists a positive equilibrium, x∈R>n with Γk∗xYs∗=0, and hence (I’). This further implies kerΓkj∩R>Vsj≠∅ for the mass-action systems (Gkj,yk), j=1,…,ℓ, having tj=1 by (i). Hence, Lemma 19 applies to the subsystems. (III’)By Lemma 19, Kj=Sj=Lj and dj≤1, j=1,…,ℓ. It remains to consider dj=1 in detail, implying δj=1 and tj′=1; in particular, all vertices are source vertices, that is, Vsj=Vj. By the argument before Lemma 20, there are β∈Rj with imβ=kerYj∩imRkj and q^∈R≥Vsj such that Rk¯jq^=0 and Rk¯j1=β (for a certain k¯∈R>Vsj), cf. Equation (8). In particular, suppq^=T, where T⊆Vj denotes (the vertices of) the absorbing strong component. Let t=|T| and m=|Vj| and assume q^1≥⋯≥q^t>q^t+1=⋯=q^m=0. By Lemma 20, ∑i′=1iβi′≥0,fori=1,…,t. Further, if iq^i+1, then ∑i′=1iβi′>0. Finally, ∑i′=1tβi′=0 if and only if T=Vj. The polynomial system ΓkjxYsj=0 is of the form A(c∘xB)=0 with B=Ysj=Yj, and the dependency vector b∈Rj in Corollary 15 is given by imb=Dk=kerB1T=kerYj1T. Since imRkj=imIEj=ker1T, equivalently imb=kerYj∩imRkj. That is, we can choose b=β and the (in)equalities for β guaranteed by Lemma 20 also hold for b. Now, recall that I1,…,Iω⊂Vj denote ω equivalence classes corresponding to equal (consecutive) components of q^ and that b¯∈Rω with b¯i=∑i′∈Iibi′ is the vector of lumped b’s. Since ∑i′=1ibi′≥0 for i=1,…,t, we get ∑i′=1ib¯i′>0 for i=1,2,…,ω-1. (If T⊂Vj and hence Iω={t+1,…,m}, then ∑i′=1ω-1b¯i′=∑i′=1tbi′>0.) Finally, since ∑i=1ωb¯i=0, we get b¯ω<0 and hence b¯1·b¯ω<0. By Lemma 19, Kj=Sj=Lj and dj≤1, j=1,…,ℓ. It remains to consider dj=1 in detail, implying δj=1 and tj′=1; in particular, all vertices are source vertices, that is, Vsj=Vj. By the argument before Lemma 20, there are β∈Rj with imβ=kerYj∩imRkj and q^∈R≥Vsj such that Rk¯jq^=0 and Rk¯j1=β (for a certain k¯∈R>Vsj), cf. Equation (8). In particular, suppq^=T, where T⊆Vj denotes (the vertices of) the absorbing strong component. Let t=|T| and m=|Vj| and assume q^1≥⋯≥q^t>q^t+1=⋯=q^m=0. By Lemma 20, ∑i′=1iβi′≥0,fori=1,…,t. Further, if iq^i+1, then ∑i′=1iβi′>0. Finally, ∑i′=1tβi′=0 if and only if T=Vj. The polynomial system ΓkjxYsj=0 is of the form A(c∘xB)=0 with B=Ysj=Yj, and the dependency vector b∈Rj in Corollary 15 is given by imb=Dk=kerB1T=kerYj1T. Since imRkj=imIEj=ker1T, equivalently imb=kerYj∩imRkj. That is, we can choose b=β and the (in)equalities for β guaranteed by Lemma 20 also hold for b. Now, recall that I1,…,Iω⊂Vj denote ω equivalence classes corresponding to equal (consecutive) components of q^ and that b¯∈Rω with b¯i=∑i′∈Iibi′ is the vector of lumped b’s. Since ∑i′=1ibi′≥0 for i=1,…,t, we get ∑i′=1ib¯i′>0 for i=1,2,…,ω-1. (If T⊂Vj and hence Iω={t+1,…,m}, then ∑i′=1ω-1b¯i′=∑i′=1tbi′>0.) Finally, since ∑i=1ωb¯i=0, we get b¯ω<0 and hence b¯1·b¯ω<0. By Corollary 15, there exists a unique positive equilibrium within every stoichiometric compatibility class. It remains to show that weak reversibility implies (ii). By Lemma 16, kerIE=kerIE∗, and hence the cycles of the graph G agree with the cycles of the subgraphs Gj, j=1,…,ℓ. Hence, if G is weakly reversible, then every Gj is weakly reversible and kerRkj∩R>s≠∅. Since Γkj=YjRkj, also kerΓkj∩R>s≠∅, and since Γk∗=Γk1…Γkℓ, also kerΓk∗∩R>Vs⊔≠∅. That is, (ii). □ Examples Example 1 Deficiency two network with one (singleton) absorbing strong component as shown in Fig. 1. Reaction network (as an embedded graph) with 5 vertices and 4 source vertices and hence δ=5-1-2=2, but d=4-1-2=1 The ODE associated with the mass-action system is given by and steady state can be written as ΓkxYs=0 with Γk=k14k23-k35k42k14-k2300 and Ys=02310101. We show that conditions (I), (II), (III), and (IVb) of Theorem 14 are satisfied. (I) kerΓk∩R>04 is generated by y1=(1k14,1k23,2k35,0)T and y2=(0,0,1k35,1k42)T and hence kerΓk∩R>04≠∅. As a consequence, q=(1,1,q3,-1)T with q3∈(-1,1). (The vertices of the graph G have been labeled such that the components of q are ordered.) (II) Satisfied trivially. (III) Clearly, d=4-1-2=1. Since dim(K)=dim(imΓk)=2, we get K=L. Finally, the (lumped) dependency vectors are given by b=(1,3,-1,-3)T and b¯=(4,-1,-3)T and hence ∑i′=1ib¯i′≥0 for i=1,2 as well as b¯1·b¯3=4·(-3)<0. (IVb) Since L=R2 holds for the monomial difference subspace (and L⊆S), we get L=S. Using Theorem 14, we find that there exists a unique positive equilibrium in every stoichiometric compatibility class, for all rate constants. Example 2 Deficiency two network with two absorbing strong components (one singleton and one non-singleton) as shown in Fig. 2. Reaction network (as an embedded graph) with 5 vertices and 4 source vertices and hence δ=5-1-2=2, but d=4-1-2=1 The ODE associated with the mass-action system is given by and steady state can be written as ΓkxYs=0 with Γk=-k15-k233k34-3k43k12-k21-k34k43 and Ys=11030110. We show that conditions (I), (II), (III), and (IVb) of Theorem 14 are satisfied. (I) kerΓk is generated by y1=(3k21+k23,3k12-k15,k12k23+k15k21k34,0)T and y2=(0,0,1k34,1k43)T and hence kerΓk∩R>04≠∅ iff 3k12-k15>0. In this case, q=(1,1,q3,-1)T with q3∈(-1,1). (The vertices of the graph G have been labeled such that the components of q are ordered.) (II) Satisfied trivially. (III) Clearly, d=4-1-2=1. Since dim(K)=dim(imΓk)=2, we get K=L. Finally, the (lumped) dependency vectors are given by b=(1,2,-2,-1)T and b¯=(3,-2,-1)T and hence ∑i′=1ib¯i′≥0 for i=1,2 as well as b¯1·b¯3=3·(-1)<0. (IVb) Since L=R2 holds for the monomial difference subspace (and L⊆S), we get L=S. Using Theorem 14, we find that there exists a unique positive equilibrium in every stoichiometric compatibility class iff 3k12-k15>0. Example 3 Deficiency two network with two (singleton) absorbing strong components as shown in Fig. 3. Reaction network (as an embedded graph) with 5 vertices and 3 source vertices and hence δ=5-1-2=2, but d=3-1-2=0 The ODE associated with the mass-action system is given by and steady state can be written as ΓkxYs=0 with Γk=-k21+k23-k32+k34k45k21k34-k45 and Ys=123001. We show that conditions (I), (II), (III), and (IVb) of Theorem 14 are satisfied. (I) kerΓk is generated by the vector (k32-2k34k21k32-2k21k34+k23k34,k23k21k32-2k21k34+k23k34,1k45)T and hence kerΓk∩R>04≠∅ iff k32-2k34>0. (II) Satisfied trivially. (III) Clearly, d=3-1-2=0. Since dim(K)=dim(imΓk)=2, we get K=L. (IVb) Since L=R2 holds for the monomial difference subspace (and L⊆S), we get L=S. Using Theorem 14, we find that there exists a unique positive equilibrium in every stoichiometric compatibility class iff k32-2k34>0. Acknowledgements This research was funded in whole, or in part, by the Austrian Science Fund (FWF), grant DOI 10.55776/PAT3748324 to SM. For open access purposes, the authors have applied a CC BY public copyright license to any author-accepted manuscript version arising from this submission. Appendix A Proof of Theorem 6 Proof In the proof of Theorem 5, we have shown that b¯1·b¯ω<0 implies |Yc|≥1 for all c. We now show the other direction, |Yc|≥1 for all c implies b¯1·b¯ω<0. Recall that determining Yc is equivalent to solving f(t)=c∗ for t∈(-1,1), where and 1=q¯1>⋯>q¯ω=-1. More explicitly, f(t)=(1+t)b¯1·⋯·(1-t)b¯ω>0. Hence, |Yc|≥1 for all c is equivalent to |{t∈(-1,1)∣f(t)=c∗}|≥1 for all c∗>0. Now, for all δ>0, the function f attains positive minimum and maximum values on the compact interval [-1+δ,1-δ], and f being surjective onto (0,∞) requires limt→-1f(t)=0 and limt→1f(t)=∞ (or vice versa). This is equivalent to b¯1>0 and b¯ω<0 (or vice versa), that is b¯1·b¯ω<0. □ B Reaction networks B.1 Index notation We explicitly state the incidence and source matrices as well as the (rectangular) Laplacian matrix of a labeled simple digraph (V, E, k), using index notation. (i)The incidence matrix IE∈R>V×E is given by (IE)i,j→j′=-1,ifi=j,1,ifi=j′,0,otherwise. The incidence matrix IE∈R>V×E is given by (IE)i,j→j′=-1,ifi=j,1,ifi=j′,0,otherwise. (ii)The source matrix IE,s∈R>Vs×E is given by (IE,s)i,j→j′=1,ifi=j,0,otherwise. The source matrix IE,s∈R>Vs×E is given by (IE,s)i,j→j′=1,ifi=j,0,otherwise. (iii)The rectangular Laplacian matrix Rk∈RV×Vs is given by (Rk)i,j=kj→i,if(j→i)∈E,-∑(i→i′)∈Eki→i′,ifi=j,0,otherwise. It can be decomposed as Rk=IEdiag(k)(IE,s)T. The rectangular Laplacian matrix Rk∈RV×Vs is given by (Rk)i,j=kj→i,if(j→i)∈E,-∑(i→i′)∈Eki→i′,ifi=j,0,otherwise. It can be decomposed as Rk=IEdiag(k)(IE,s)T. B.2 Proof of the second salt theorem We prove the second salt theorem for rectangular (rather than square) Laplacian matrices. Proof of Lemma 20 Let T′={1,…,t′}⊆T. The equations Rkq^=0 and Rk1Vs=β imply If T′⊂T, then the first equation and the order on q^ imply and the second equation and q^t′≥q^t′+1 further imply In particular, ∑i∈T′βi>0 if q^t′>q^t′+1. If T′=T, then (T→V\T)=∅, since T is an absorbing strong component, and The sum is zero if and only if (V\T→T)=∅, that is, T=V. □ Author Contributions A.D. and S.M. contributed equally. Funding Open access funding provided by University of Vienna. Data Availability No datasets were generated or analyzed during the current study. Declarations Conflict of interest The authors declare no conflict of interest. Footnotes The reaction network (with rate constants) defines a mass-action system with associated ODE Reaction networks are formally introduced in Sect. 4, where we will continue the example. For two vectors y1,y2∈R≥m with y1+y2,λ¯y1+(1-λ¯)y2∈R>m, consider the maximal minors of the matrices (y1y2),(y1-y2y1+y2),(y^y¯)∈Rm×2. Explicitly, let(y1y2)=a1a2b1b2⋮⋮,(y1-y2y1+y2)=a1-a2a1+a2b1-b2b1+b2⋮⋮,(y^y¯)=λ^a1+(1-λ^)a2λ¯a1+(1-λ¯)a2λ^b1+(1-λ^)b2λ¯b1+(1-λ¯)b2⋮⋮.If λ^>λ¯, thensignλ^a1+(1-λ^)a2λ¯a1+(1-λ¯)a2-λ^b1+(1-λ^)b2λ¯b1+(1-λ¯)b2=signa1-a2a1+a2-b1-b2b1+b2=signa1a2b1b2. Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Change history 6/28/2026 In this article, Label ``1B.1'' should be B.1. References Associated Data Data Availability Statement No datasets were generated or analyzed during the current study.