APPMATHISRN Applied Mathematics2090-55722090-5564International Scholarly Research Network91093610.5402/2011/910936910936Research ArticleThe Stability Cone for a Difference Matrix Equation with Two DelaysIvanovS. A.1KipnisM. M.2MalyginaV. V.3KarpovE.MishurisG.QinY.1Department of Mathematical AnalysisChelyabinsk State Pedagogical University Chelyabinsk 454080Russiacspu.ru2Applied Mathematics DepartmentSouth Ural State UniversityChelyabinsk 454080Russiasusu.ac.ru3Faculty of Applied Mathematics and MechanicsPerm State Technical UniversityPerm 614990Russiapsu.ru201130052011201130032011250420112011Copyright © 2011 S. A. Ivanov et al.This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

We provide geometric algorithms for checking the stability of matrix difference equations xn=Axnm+Bxnk with two delays m,k such that the matrix ABBA is nilpotent. We give examples of how our results can be applied to the study of the stability of neural networks.

1. Introduction

The problem of the stability of the equationxn=axn-m+bxn-k with real coefficients a,b is basically solved .

The stability of matrix (1.1) with special 2×2 matrices a,b and m=1 was studied in [5, 6]. The case m=1, a=I, where I is the identity matrix, was studied in  without dimension restriction. In the paper  the dimension is also not bounded, and the results of  are generalized: it assumes that a=αI, α, 0α1. The representation of the solutions of (1.1) with commuting matrices a,b is given in  without considering a stability problem.

To the best of the authors' knowledge, the stability of (1.1) with complex coefficients a,b has not been studied yet.

In this paper we provide geometric algorithms for checking the stability of (1.1) with two delays m,k+, k>m1, for two cases: (1) a,b are complex numbers, (2) a,b are simultaneously triangularizable matrices. The results of this paper are based on the D-decomposition method (parameter plane method) [10, 11].

Matrices a,b commute in all the above articles, which implies the possibility of simultaneous triangularization . Therefore, our method can be applied to all of the above-mentioned cases. In the present paper the case m=1 is studied along with other values m+. The case m=1 is very important, so we separately examined it in detail in the paper .

The paper is organized as follows. In Section 2, we introduce the curve of D-decomposition and point out its key property of symmetry. In Section 3, we define the basic ovals and formulate their properties. In Section 4, we define a property of ρ-stability, which coincides with usual stability if ρ=1. Later in that section we solve a problem of geometric checking ρ-stability of (1.1) with positive real a and complex b. In Sections 5 and 6, we give a method of geometric checking the stability of (1.1) with complex coefficients and simultaneously triangularizable matrices, correspondingly. Finally, in Section 7, we employ our results to derive the stability conditions for neural nets.

2. <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M32"><mml:mrow><mml:mi>D</mml:mi></mml:mrow></mml:math></inline-formula>-Decomposition Curve for Given <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M33"><mml:mi>k</mml:mi><mml:mo>,</mml:mo><mml:mi>m</mml:mi><mml:mo>,</mml:mo><mml:mi>a</mml:mi><mml:mo>,</mml:mo><mml:mi>ρ</mml:mi></mml:math></inline-formula>

Consider the scalar variant of (1.1). The characteristic polynomial for (1.1) isf(λ)=λk-aλk-m-b. If k=dk1,m=dm1, d>1, then the trajectory of (1.1) splits into d independent trajectories, and degree of polynomial (2.1) gets smaller after the substitution λd=μ: f1(λ)=μk1-aμk1-m1-b. Therefore, often (but not always) we will assume that the delays m,k are relatively prime.

Definition 2.1.

D-decomposition curve for given k,m+, a, ρ+ is a curve on the complex plane of the variable b defined by the equationb(ω)=ρkexp(ikω)-|a|ρk-mexp(i(k-m)ω),        ωR.

Parameter ω moves along the interval of length 2π, the starting point of which is not fixed. We also call the curve (2.3) hodograph.

In this and the next sections we will consider only real positive values of a. Starting from Section 5 we will get rid of this restriction. Obviously, if we assume in (2.1) that b=b(ω) and a, a0, then (2.1) will have a root λ=ρexp(iω). Hodograph (2.3) splits the complex plane into the connected components. This decomposition is called the D-decomposition . If we put a, a0, and substitute any two internal points b1,b2 from one of the connected components of D-decomposition for coefficient b in polynomial (2.1), then the polynomials obtained will have equal number of roots inside the circle of radius ρ centred at the origin of the complex plane. In particular, if a, a0, ρ=1 and the substitution of some inner point of a component of D-decomposition into (1.1) gives a stable equation, then the substitution of any other internal point from that component also gives a stable equation.

Let us point out a key property of symmetry of hodograph (2.3).

Lemma 2.2 (symmetry).

If k,m are coprime, then hodograph (2.3) is invariant under the rotation by 2π/m.

Proof.

For coprime k,m there exist s,t+ such that ks-mt=1. From (2.3), (2.4) it follows that exp(i2πm)  b(ω)=b(ω+2πsm). Lemma 2.2 is proved.

From now and further we will assume that -π<argzπ for any complex z, while Argz will be assumed as multivalued function, and the equality Argz=v will mean that one of the values of Argz equals to v.

The following lemma asserts that some part of the complex plane is free from points of the hodograph b(ω).

Lemma 2.3.

Let k,m be coprime, k>m>1, |a|<ρm. Let ω1[0;π/m] be the least positive root of the equation argb(ω)=π/m, and let 0<ω2ω1. Then, for any ω, from argb(ω)=argb(ω2), it follows that |b(ω)||b(ω2)|.

Proof.

The function |b(ω)| is 2π/m-periodic, increasing in [(2j-2)π/m,(2j-1)π/m] for any j and decreasing in [(2j-1)π/m,2jπ/m]. In addition, |b(ω+π/m)|=|b(-ω+π/m)| for any ω. Let us assume, in order to get a contradiction, that (2.6) and |b(ω)|<|b(ω2)| are true for some ω>0. Then there exists a positive integer s<m and δ such that ω=2πsm+δ,|δ|<ω2. From (2.3), (2.7) it follows that Argb(ω)=(k-m)ω+Arg(ρmexp(imω)-|a|)=2πksm+Arg(ρmexp(imδ)-|a|)+(k-m)δ=2πksm+Argb(δ). Since k,m are coprime and s<m, let us find natural numbers j,q such that ks=mj+q, m>q1. Then (2.8) implies Argb(ω)=2πqm+Argb(δ), which contradicts (2.6) and the inequality |argb(δ)|<π/m following from (2.7). Lemma 2.3 is proved.

3. Basic Ovals

For hodograph (2.3) the equality bω(0)=iρk-m(kρm-|a|(k-m)) takes place. If|a|<ρmkk-m, then let us look at a closed curve on the complex plane, which we call the basic oval. This curve is an image of an interval [-ω1,ω1] under the map b(ω) defined by (2.3). Here ω1(0,π/k] is the least positive root of the equation argb(ω)=π. We are also interested in those parts of hodograph (2.3) that can be obtained by the rotation of the basic oval by the angles 2πj/m, j,  0j<m (see Lemma 2.2). We also call them the basic ovals. Here is a formal definition.

Definition 3.1.

Let k,m be coprime, k>m1, j, 0j<m, and let (2.4), (3.1) hold. The basic oval Lj for (1.1) is a closed curve given by (2.3), where the variable ω runs from (-ω1+2πjs/m) to (ω1+2πjs/m), where ω1 is the least positive root of the equationargb(ω)=π.

From Lemma 2.2 and formula (2.5) it follows that all m basic ovals can be obtained from L0 by rotation by the angles 2πj/m, j=0,1,,(m-1).

Considering Definition 3.1, we get the following. For existence of the basic oval it is necessary that |a|<ρmk/(k-m). If m>1 and |a|>ρm, then the complex number 0 is outside any oval, and the intersection of all ovals is empty. If m=1, then for fixed k, ρ, |a|[0,ρmk/(k-m)) the basic oval L0 is unique. That is why the results related to the stability of (1.1) are different for m=1 and m>1.

The basic oval Lj decreases as |a| increases from 0, and it shrinks to the point b=-exp(2πj/m)ρk/(k-1) as |a| reaches ρmk/(k-m) (Figure 1 for m>1 and Figure 2 for m=1).

D-decomposition curve and three basic ovals for m=3,k=5: (a) |a|<ρm, (b) |a|=ρm, (c) ρm<|a|<ρmk/(k-m), and (d) |a|=ρmk/(k-m).

D-decomposition curve and the basic oval L0 for m=1,k=5: (a) |a|<ρ, (b) |a|=ρ, (c) ρ<|a|<ρk/(k-1), and (d) |a|=ρk/(k-1).

Lemma 3.2.

Let k,m be coprime, k>m1, j, 0j<m, a, and 0a<ρmk/(k-m). If the complex number b lies outside the basic oval Lj, then characteristic polynomial (2.1) has a root λ such that |λ|>ρ.

Proof.

Let us fix k,m,a,ρ,j, and let the complex number b lie outside the basic oval Lj. Having changed ρ to R>ρ in Definition 3.1 let us consider the system of ovals Lj(R). If R, then the ovals Lj(R) include a circle of an arbitrarily large radius. Therefore, there exists R0 such that the point b is inside the oval Lj(R0). The ovals Lj and Lj(R0) are homotopic, therefore, there exists R1(ρ,R0] such that b lies on the curve Lj(R1), which means the existence of a root λ of characteristic polynomial (2.1) such that |λ|=R1>ρ. Lemma 3.2 is proved.

4. Localization of Roots of Characteristic Polynomial (<xref ref-type="disp-formula" rid="EEq2">2.1</xref>) for Real Nonnegative <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M187"><mml:mrow><mml:mi>a</mml:mi></mml:mrow></mml:math></inline-formula> and Complex <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M188"><mml:mrow><mml:mi>b</mml:mi></mml:mrow></mml:math></inline-formula>

For the stability of (1.1) it is required that all the trajectories are bounded. But sometimes one needs to strenghten or weaken the stability requirement. It justifies the following definition.

Definition 4.1.

Equation (1.1) is said to be ρ-stable if for any of its solutions (xn) the sequence (|xn|/ρn) is bounded, and asymptotically ρ-stable if for any of its solutions (xn) one has limn|xn|/ρn=0.

If ρ=1, then the concept of (asymptotic) ρ-stability coincides with the concept of usual (asymptotic) stability. Evidently, (1.1) is ρ-stable, if there are no roots of polynomial (2.1) outside the circle of radius ρ centred at the origin, and there are no multiple roots of the polynomial on the boundary of circle. Equation (1.1) is asymptotically ρ-stable if and only if all the roots of its characteristic polynomial (2.1) lie inside the circle of radius ρ with the center at 0.

Let us call the equation (asymptotically) ρ-unstable if it is not (asymptotically) ρ-stable. As we noted in Section 2, if ρ=1, then the proportional change of both delays k,m in (1.1) has no influence on ρ-stability. It is not the case if ρ1. It is easy to see that the equationxn=axn-md+bxn-kd is asymptotically ρ-stable if and only if (1.1) is asymptotically ρd-stable. It implies the following important observation: proportional increase of both delays k,m with the conservation of the coefficients a,b in (1.1) preserves asymptotic ρ-stability if ρ>1 and may not preserve it if ρ<1.

Definition 4.2.

Let k,m be coprime, k>m1, a, and 0|a|<ρmk/(k-m). The stability domain D(k,m,a,ρ) is defined to be a set of all complex numbers b such that for any j(0j<m) the number b lies inside the basic oval Lj.

Under the same conditions if k,m are not coprime and d=gcd(k,m), let us put D(k,m,a,ρ)=D(k/d,m/d,a,ρd).

Theorems 4.35.2 will justify the name “stability domain” for D(k,m,a,ρ) a little later. Evidently, for coprime k,m such that k>m>1 the domain D(k,m,a,ρ) has the following properties. If 0|a|<ρm, then D(k,m,a,ρ) is the connected domain on the complex plane, containing 0, whose boundary is the D-decomposition curve (2.3). If |a|=ρm, then the domain D(k,m,a,ρ) degenerates into the point b=0. If ρm<|a|<ρmk/(k-m), then the domain D(k,m,a,ρ) is empty. If |a|ρmk/(k-m), then the domain D(k,m,a,ρ) is not defined in view of the fact that the basic ovals (Definition 3.1) are not defined.

If m=1, then the domain D(k,1,a,ρ) is a set of points lying inside the oval L0. If 0|a|<ρ, then D(k,1,a,ρ) includes 0. If ρ|a|<ρk/(k-1), then D(k,1,a,ρ) is nonempty and does not contain 0. If |a|=ρk/(k-1), then D(k,1,a,ρ) degenerates into the point b=-ρk/(k-1). Finally, if |a|>ρk/(k-1), then the domain D(k,1,a,ρ) is not defined.

The following theorems are based on the localization of roots of polynomial (2.1) with nonnegative a and complex b with respect to the circle of radius ρ centred at the origin.

Theorem 4.3.

Let k,m be coprime, k>m>1, a+, ρ>0.

If a>ρm, then for any b (1.1) is ρ-unstable.

If a=ρm, then for any complex b0 (1.1) is ρ-unstable; for b=0 it is ρ-stable (nonasymptotically).

If 0a<ρm, then (1.1) is asymptotically ρ-stable if and only if the complex number b lies inside the stability domain D(k,m,a,ρ).

If 0a<ρm, then (1.1) is ρ-stable if and only if the complex number b lies either inside or on the boundary of D(k,m,a,ρ).

Proof.

(1) Let a>ρm. Let us find R such that ρm<Rm<a<Rmkk-m. Taking into account the inequality a<Rmk/(k-m), let us consider m basic ovals Lj(R), j=0,1,,m-1 having ρ replaced by R in Definition 4.1. Since Rm<a, the system of ovals Lj(R) has no intersections. Hence, for any complex number b there exists j,    0j<m such that b lies outside the oval Lj(R). By Lemma 3.2 (1.1) is R-unstable. Since R>ρ, it is ρ-unstable. Statement 1 is proved.

(2) Let a=ρm. If b=0, then statement 2 of Theorem 4.3 is obvious. Let b0. If Reb0, then b lies outside the oval L0. If Reb0 and m is even, then b lies outside the oval Lm/2. If Reb0 and m is odd, then b lies either outside the oval L(m-1)/2 or outside the oval L(m+1)/2. In any case (1.1) is ρ-unstable by Lemma 3.2. Statement 2 is proved.

(3) Let 0a<ρm. Let the number b be inside the domain D(k,m,a,ρ). Then for any j(0j<m) the number b lies inside the oval Lj. By Lemma 2.3 the beam drawn on the complex plane from 0 to b does not intersect curve (2.3). Therefore, polynomial (2.1) has the same number of roots inside the circle of radius ρ for given b and for b=0. However if b=0, then all the roots of (2.1) lie inside the circle of radius ρ centred at 0. Therefore, (1.1) is asymptotically ρ-stable for given b.

If b lies on the boundary of the domain D(k,m,a,ρ) or outside it, then b lies either on the boundary of one of the basic ovals Lj or outside one of them, and by Lemma 3.2 (1.1) is asymptotically ρ-unstable.

(4) If b lies outside the domain D(k,m,a,ρ), then the conclusion of statement 4 of Theorem 4.3 is a straightforward consequence of Lemma 3.2. If b lies inside D(k,m,a,ρ), then the conclusion of statement 4 of Theorem 4.3 is a straightforward consequence of statement 3 of Theorem 4.3. Let b lie on the boundary of D(k,m,a,ρ). Then for any root λ of polynomial (2.1) either |λ|<ρ or |λ|=ρ. In the latter case in view of the inequality 0a<ρm<ρmk/k-m we have dfdλ=λk-m-1(kλm-a(k-m))0, hence the root λ such that |λ|=ρ is simple. Theorem 4.3 is proved.

If in (1.1) the least delay m is equal to 1, then the situation is essentially different from the case m>1.

Theorem 4.4.

Let k>m=1, a+, ρ>0.

If aρk/(k-1), then for all complex numbers b (1.1) is ρ-unstable.

If 0a<ρk/(k-1), then (1.1) is asymptotically ρ-stable if and only if the complex number b lies inside the domain D(k,1,a,ρ).

If 0a<ρk/(k-1), then (1.1) is ρ-stable if and only if the complex number b lies inside D(k,1,a,ρ).

Proof.

(1) Let a>ρk/(k-m), and let b be a given complex number. Let us find R>ρ such that ρk/(k-m)<a<Rk/(k-m) and the point b is located outside the oval L0(R) obtained from Definition 3.1 by substituting R for ρ. By Lemma 3.2 there exists a complex root λ of polynomial (2.1) such that |λ|>R>ρ, so ρ-instability of (1.1) is proved. Let a=ρk/(k-1). Then the previous arguments also prove ρ-instability provided that b-ρk/(k-1). However, if b=-ρk/(k-1), then under the assumption that a=ρk/(k-1) the number λ=ρ is a multiple root of polynomial (2.1), and consequently, (1.1) is also ρ-unstable. Statement 1 of Theorem 4.4 is proved.

(2) Let 0a<ρk/(k-1). Since D(k,1,a,ρ) is the domain of inner points of the oval L0, it is connected. The function |b(ω)| (see (2.3)) increases as ω moves either from 0 to π or from 0 to (-π). Therefore, there are no points of hodograph (2.3) inside L0. To complete the proof of asymptotical ρ-stability of (1.1) at any point of L0 it is sufficient to prove that there exists at least one point b0 inside the oval L0 such that the equation is asymptotically ρ-stable for b=b0.

CASE 1. Let 0a<ρ. Then the point b=0 lies inside L0. If b=0, then polynomial (2.1) has the (k-1)-multiple root λ=0 and the simple root λ=a. This gives the asymptotic ρ-stability, in view of a<ρ.

CASE 2. Let ρa<ρk/(k-1). Let us consider the point b=ρk-aρk-1 at the boundary of L0 and consider characteristic polynomial (2.1) with given b: f2(λ)=λk-aλk-1-ρk+aρk-1. The equation f2(λ)=0 transforms into ((λρ)-1)((λρ)k-1-((aρ)-1)j=0k-2(λρ)k-j-2)=0. One of roots of (4.5) is equal to ρ, while others lie inside the circle of radius ρ centred at the origin in view of the inequality 0a-ρ<ρ/(k-1) .

Let us return to (2.1), and let us figure out in what direction the root λ=ρ moves as the coefficient b moves from the point b=ρk-aρk-1 toward the interior of L0 so that db, db<0. From (2.1) it follows that for λ=ρ we have dλdb=ρ-k+2kρ-a(k-1), and in view of a<ρk/(k-1) we get dλ/db>0. Therefore, db<0 implies λ<ρ. Consequently, there exist values of b inside L0 providing asymptotic ρ-stability of (1.1), therefore, for any value b inside L0 (1.1) is asymptotically ρ-stable.

(3) The proof of Statement 3 of Theorem 4.4 is analogous to the proof of Statement 4 of Theorem 4.3. Theorem 4.4 is proved.

5. Stability of (<xref ref-type="disp-formula" rid="EEq1">1.1</xref>) with Complex Coefficients <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M429"><mml:mi>a</mml:mi><mml:mo>,</mml:mo><mml:mi>b</mml:mi></mml:math></inline-formula>

Let us change the variables in (1.1) so that it has no influence on (asymptotic) ρ-stability:xn=ynexp(inmarga). Equation (1.1) changes toyn=αyn-m+βyn-k, whereα=|a|,β=bexp(-ikmarga). The characteristic polynomial for (5.2) has the formψ(μ)=μk-αμk-m-β. It is related to (2.1) by the change μ=λexp(-i(1/m)arga), that saves the absolute values of roots of the equation. It is important for us that new (5.2) has a real nonnegative coefficient at yn-m, in view of (5.3). This allows us to apply the results of the previous section. Therefore, from Theorems 4.3 and 4.4 we immediately derive the following theorems providing an answer to the question on the stability of (1.1) with complex coefficients a, b.

Theorem 5.1.

Let k,m be coprime, k>m>1, a, ρ>0.

If |a|>ρm, then for any complex b (1.1) is ρ-unstable.

If |a|=ρm, then for any b0 (1.1) is ρ-unstable; for b=0 it is ρ-stable (nonasymptotically).

If |a|<ρm, then (1.1) is asymptotically ρ-stable if and only if the complex number β=bexp(-i(k/m)arga) lies inside the domain D(k,m,a,ρ).

If |a|<ρm, then (1.1) is ρ-stable if and only if the complex number β=bexp(-i(k/m)arga) lies either inside D(k,m,a,ρ) or on its boundary.

Theorem 5.2.

Let k>m=1, a, ρ>0.

If |a|ρk/(k-1), then for any complex b (1.1) is ρ-unstable.

If |a|<ρk/(k-1), then (1.1) is asymptotically ρ-stable if and only if the complex number β=bexp(-ikarga) lies inside the domain D(k,1,a,ρ).

If |a|<ρk/(k-1), then (1.1) is ρ-stable if and only if the complex number β=bexp(-ikarga) lies either inside D(k,1,a,ρ) or on its boundary.

Example 5.3.

Let m=6, a=0.8+0.9i, ρ=1.15 in (1.1). Let 6k12. For every given value k let us find all values of the complex coefficient b for which (1.1) is ρ-stable. The answer is demonstrated by Figure 4. Let us give some comments. First calculate |a|=1.204, arga0.844. If k=6, then to find ρ-stability domain one does not need to use Theorems 4.35.2. The domain is a circle given in Figure 4(a). Since 6,7 are coprime, then for k=7, by Theorem 5.1, (1.1) ρ-stable if and only if bexp(i(7/6)arga)D(7,6,a,ρ). The corresponding “curved hexagon” is shown in Figure 4(a). Similarly for k=11 the condition bexp(i(11/6)arga)D(11,6,a,ρ) is necessary and sufficient for asymptotic stability of (1.1). The corresponding “curved hexagon” is shown in Figure 4(b). For k=8 the stability criterion is the condition bexp(i(4/3)arga)D(4,3,a,ρ2). The corresponding “curved triangle” is shown in Figure 4(a). Similarly the “digon” exp(i(3/2)arga)D(3,2,a,ρ3) for k=9 is shown in Figure 4(a), and the “curved triangle” exp(i(5/3)arga)D(5,3,a,ρ2) for k=10 is shown in Figure 4(b). For k=12, according to Theorem 5.2, the stability criterion for (1.1) is bexp(i·2arga)D(2,1,a,ρ6) (Figure 4(b)). The corresponding “stability oval” is shown in Figure 4(a).

6. Stability Cones for Matrix Equation (<xref ref-type="disp-formula" rid="EEq1">1.1</xref>) with Simultaneously Triangularizable Matrices

Let us consider a matrix equationxn=A  xn-m+Bxn-k,x:+l; A,Bl×l. The characteristic equation for (6.1) isψ(λ)=det(Iλk-Aλk-m-B).

Definition 6.1.

Matrix equation (6.1) is called ρ-stable if for every solution (xn) the sequence (|xn|/ρn) is bounded. Equation (6.1) is called asymptotically ρ-stable if limn|xn|/ρn=0 holds for every solution (xn).

Obviously, matrix equation (6.1) is asymptotically ρ-stable if and only if all the roots of characteristic polynomial (6.2) lie inside the circle of radius ρ with the center at 0. We also observe that if at least one root of (6.2) lies outside the circle of radius ρ with the center at 0, then (6.1) is ρ-unstable.

In this paper we consider (6.1) only with triangularizable matrices A,B. It is known  that if the matrix AB-BA is nilpotent, then A,B can be simultaneously triangularized.

Definition 6.2.

If k>m>1, then the ρ-stability cone for given k,m,ρ is a set of points M=(u1,u2,u3)3 such that 0u31 and the intersection of the set with any plane u3=a  (0a1) is the stability domain D(k,m,a,ρ). If k>m=1, then the ρ-stability cone for given k,ρ is a set of points M=(u1,u2,u3)3 such that 0u3k/(k-1) and the intersection of the set with the plane u3=a(0ak/(k-1)) is the domain D(k,1,a,ρ).

Let us define a stability cone as the ρ-stability cone for ρ=1.

Returning to Figure 3, we can interpret the figures in Figure 3(a) as sections of the stability cone for k=5,m=1 at different heights u3=a, and the ones in Figure 3(b) as sections of the stability cone for k=5,m=2, and so on.

Stability domains D(k,m,a,ρ) for k=5, for different values of the coefficient a: (a) m=1, (b) m=2, (c) m=3, (d) m=4.

To Example 5.3. Stability domains for m=6, a=0.8+0.9i, ρ=1.19; (a) k=6,7,8,9, (b) k=10,11,12.

The stability cones for m>1 are the intersections of m conical surfaces formed by the basic ovals as the parameter a changes from 0 to k/(k-m) (Figure 5).

The stability cone as an intersection of three surfaces formed by the basic ovals, m=3,k=4.

Let us consider the simple case of a diagonal systemyn-diag(a11,,all)yn-m-diag(b11,,bll)yn-k=0 with complex entries ajj,bjj, 1jl. Let us construct the points Mj=(u1j,u2j,u3j)   (1jl) in 3 in the following way:u1j=Re(bjjexp(-ikmargajj)),u2j=Im(bjjexp(-ikmargajj)),u3j=|ajj|.

It follows from the definition of the ρ-stability cone and from Theorems 5.15.2 that (6.3) is asymptotically ρ-stable if and only if all the points Mj(1jl) lie inside the ρ-stability cone for given k,m. All the points Mj with u3j=0, u1j2+u2j2<ρk are considered as inner points of the ρ-stability cone.

The natural extension of the the class of diagonal systems is that of systems with simultaneously triangularizable matrices. The following theorem is our main result.

Theorem 6.3.

Let k>m1, let the numbers k,m be coprime, and ρ>0. Let A,B,Sl×l, and S-1AS=AT, and S-1BS=BT, where AT and BT are lower triangle matrices with elements aj  s,bj  s  (1j,  sl). Let one construct the points Mj=(u1j,u2j,u3j), (1jl) by the formulas (cf. (6.4)) u1j=|bjj|cos(argbjj-kmargajj),u2j=|bjj|sin(argbjj-kmargajj),u3j=|ajj|. Then (6.1) is ρ-asymptotically stable if and only if all the points Mj(1jl) lie inside the ρ-stability cone for the given k,m,ρ.

If some point Mj lies outside the ρ-stability cone, then (6.1) is ρ-unstable.

Proof.

Let us make the change yn=Sxn. Then (6.1) transforms to the following one: yn=ATyn-m+BTyn-k. The characteristic polynomial for (6.6) has the form ψ(λ)=j=1l(λk-ajjλk-m-bjj). It coincides with the characteristic polynomial of diagonal system (6.3). Therefore, from statement 3 of Theorem 5.1 (for m>1) and from statement 2 of Theorem 5.2 (for m=1) we obtain asymptotic ρ-stability if all the points Mj lie inside the ρ-stability cone. Similarly from statement 4 of Theorem 5.1 (for m>1) and statement 3 of Theorem 5.2 (for m=1) we obtain ρ-instability of (6.1) if some point Mj lies outside the cone. Theorem 6.3 is proved.

7. Applications to Neural Networks

Let us apply the results of the previous sections to the problem of the stability of discrete neural networks similar to continuous networks studied in [15, 16]. Let us consider a ring configuration of l neurons (Figure 6) interchanging signals with the neighboring neurons.

A neural net.

Let yn(j) be a signal of the j-th neuron at the n-th moment of time. Let us suppose that the neuron reaction on its state, as well as on that of the previous neuron, is m-units delayed, and reaction on the next neuron is k-units delayed. The neuron chain is closed, and the first neuron is next to the l-th one. Let us assume that the neurons interchange the signals according to the equationsyn(1)=f(yn-m(1))+g(yn-m(l))+h(yn-k(2)),yn(2)=f(yn-m(2))+g(yn-m(1))+h(yn-k(3)),yn(l)=f(yn-m(l))+g(yn-m(l-1))+h(yn-k(1)), where f,g,h are sufficiently smooth real-valued functions of a real variable. Let us assume that there is a real number y* such that the stationary sequences yn1=y*,,ynl=y* form a solution of (7.1). Let us introduce the variables xn(j)=yn(j)-y* and the vector xn=(xn(1),,xn(l))T, and let us linearize system (7.1) in new variables about zero. We get (6.1) with the circulant  matricesA=(α00ββα000βα0000α),B=(0γ0000γ00000γ000). Here α=df(y*)/dy, β=dg(y*)/dy, γ=dh(y*)/dy. Let us introduce a matrix P (a lines permutation operator):P=(0100001000001000). Then B=γP, A=αI+βPl-1, therefore, diagonalization P generates simultaneous diagonalization of A,B. The eigenvalues of P are 1, ɛ, ɛ2,,ɛl-1, where ɛ=exp(i2π/l). Therefore,AT=αI+βdiag(1,ɛl-1,ɛl-2,,ɛ),BT=γdiag(1,ɛ,ɛ2,,ɛl-1). Granting (7.4), by Theorem 6.3, we can build points Mj=(u1j,u2j,u3j) in 3 for system (6.1), (7.2):u1j=γcos(2π(j-1)l-(km)arg(α+βexp(-2πi(j-1)l))),u2j=γsin(2π(j-1)l-(km)arg(α+βexp(-2πi(j-1)l))),u3j=|α+βexp(-2πi(j-1)l)|. We get the following consequence of Theorem 6.3.

Corollary 7.1.

If for every j(1jl) the point Mj=(u1j,u2j,u3j) defined by formulas (7.5) lies inside the ρ-stability cone for given k,m,ρ, then system (6.1), (7.2) is asymptotically ρ-stable. If at least one point Mj lies outside the ρ-stability cone for given k,m,ρ, then system (6.1), (7.2) is ρ-unstable.

Let us proceed to the problem of stability of a neural network with a large number of neurons. The points Mj=(u1j,u2j,u3j) defined by (7.5) lie on the closed curveu1(t)=γcos(t-(km)arg(α+βexp(-it))),u2(t)=γsin(t-(km)arg(α+βexp(-it))),u3(t)=|α+βexp(-it)|,0t2π.

If l, then the points Mj are dense in the curve (7.6). We get the following consequence of Theorem 6.3.

Corollary 7.2.

Let one consider system (6.1), (7.2) with l×l matrices A,B. If any point of the curve (7.6) lies inside the ρ-stability cone for given k,m,ρ, then system (6.1), (7.2) is asymptotically ρ-stable for any l. If at least one point of the curve (7.6) lies outside the ρ-stability cone for given k,m,ρ, then there exists l0 such that system (6.1), (7.2) is ρ-unstable for any l>l0.

Example 7.3.

Let us consider the ring of neurons shown in Figure 6. Put k=4, m=3, ρ=1, β=0.1, γ=0.4. Let us pose a question: what are the values of α>0 for which the system of two neurons described by (6.1), (7.2) is stable?

For applications of Corollaries 7.1 and 7.2 we construct the curves (7.6) for six values of α (Figures 7(a) and 7(b)). Assuming l=3, we construct the points M1, M2, M3 in each of the six curves. Then we construct the stability cone for given k,m,ρ (Figure 7(b)). It is the ρ-stability cone for ρ=1. In Figure 7(b) we see that two curves (7.6) corresponding to the values α=0.1,  α=0.3 are hidden inside the cone. The point M1 corresponding to the values α=0.5 is on the surface of the cone, while all other points of the curve (7.6) for α=0.5 lie inside the cone. Therefore, according to Corollaries 7.1 and 7.2, if α<0.5, then system (6.1), (7.2) is asymptotically stable for any l2. In our interpretation it means that the neuron configuration in Figure 6 is stable for any number of neurons. If α>0.5, then all the curves (7.6) lie entirely or partially outside the cone. In view of Corollaries 7.1 and 7.2 system (6.1), (7.2) is unstable. In our interpretation it means that even the configuration of two neurons is unstable.

To Example 7.3. (a) The curves (7.6) for α=0.1+rh,  h=0.2,  r=0,1,,5. (b) The stability cone covers the curves (7.6) for α<0.5. If α=0.5, then the point M1 is on the surface of the cone. If α>0.5, then all the curves (7.6) are outside the cone either entirely or partially.

Now let us consider Example 7.3 without the condition ρ=1. Under assumptions of Example 7.3, for any value α there exists a number ρ0(α) such that system (6.1), (7.2) is asymptotically ρ-stable if ρ<ρ0(α) and ρ-unstable if ρ>ρ0(α). Corollaries 7.1 and 7.2 allow us to find ρ0(α) by means of construction of different ρ-stability cones. Table 1 shows how ρ0 depends on α. The value L(α)=lnρ0(α) is a Lyapunov exponent  for system (6.1), (7.2).

ρ0 versus α in Example 7.3.

 α 0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 ρ0(α) 0.87 0.938 1 1.056 1.108 1.156 1.201 1.243
Example 7.4.

Let us consider the neuron chain shown in Figure 6. Let us fix the parameters: k=4, m=3, ρ=1, β=0.1, α=0.5. In this example we demonstrate how the change of the parameter γ in system (6.1), (7.2) changes the mutual location of the curve (7.6) and the stability cone. In Figure 8(a) we show the curves (7.6) corresponding to the values γ=rh, h=0.1, r=0,1,,5, and the points M1, M2, M3 for l=3. The upper part of the stability cone corresponding to u3>0.7 is removed. In Figure 8(b), one third of the lateral surface of the stability cone is removed too. Figure 8 demonstrates that the curves (7.6) lie inside the cone if 0γ<0.4. Therefore, if 0γ<0.4, then system (6.1), (7.2) is asymptotically stable for any l2. If γ=0.4, then the point M1=(u1,u2,u3) defined by (7.5) lies on the cone surface. If γ>0.4, then the point M1 lies outside the cone, and this shows the instability of system (6.1), (7.2). In our interpretation, for 0γ<0.4, the neuron chain is stable no matter how many neurons are in the chain, and for γ>0.4 it is unstable even if it consists only of two neurons.

To Example 7.4. The stability cone and the curves (7.6) for γ=rh,  h=0.1,  r=0,1,,5 in two projections.

8. Conclusion

The condition A+B<1 is sufficient for the asymptotic stability of matrix (6.1) , and it does not require simultaneous triangularization of the matrices A,B. There are sufficient conditions for stability of nonautonomous scalar difference equations in .

Stability cones for differential matrix equations ẋ=Ax+Bx(t-τ) with one delay τ are introduced in the paper  and for some integrodifferential equations in the paper .

There are images of the stability domains in the space of parameters of scalar differential equations ẋ=ax(t-τ1)+bx(t-τ2) with delays τ1,τ2 [25, 26] and scalar difference equations xn=xn-1+axn-m+bxn-k with delays k,m . The results of the papers [25, 26] imply that there is no simple complete description of the stability domains for these equations.

Acknowledgments

The authors are indebted to K. Chudinov, A. Makarov, and D. Scheglov for the very useful comments.

NikolaevYu.The geometry of D-decomposition of a two-dimensional plane of arbitrary coefficients of the characteristic polynomial of a discrete systemAutomation and Remote Control200465121904191410.1023/B:AURC.0000049876.54417.e62128191DannanF. M.The asymptotic stability of x(n+k)+ax(n)+bx(nl)=0Journal of Difference Equations and Applications2004106589599206041410.1080/10236190410001685058ZBL1054.39003KipnisM. M.NigmatulinR. M.Stability of trinomial linear difference equations with two delaysAutomation and Remote Control200465111710172310.1023/B:AURC.0000047886.46498.792114854ZBL1080.39013ChengS. S.HuangS. Y.Alternate derivations of the stability region of a difference equation with two delaysApplied Mathematics E-Notes200992252532550493ZBL1204.39015KaslikE.BalintSt.Bifurcation analysis for a two-dimensional delayed discrete-time Hopfield neural networkChaos, Solitons & Fractals200734412451253233127910.1016/j.chaos.2006.03.107ZBL1142.39300MatsunagaH.Stability regions for a class of delay difference systemsDifferences and Differential Equations200442Providence, RI, USAAmerican Mathematical Society273283Fields Institute Communications2089617ZBL1067.39012LevitskayaI. S.A note on the stability oval for xn+1=xn+AxnkJournal of Difference Equations and Applications2005118701705215664810.1080/10236190512331333851ZBL1080.39014KaslikE.Stability results for a class of difference systems with delayAdvances in Difference Equations20092009139384922588199ZBL1187.39025DiblikJ.KhusainovD. Ya.Representation of solutions of discrete delayed system x(k+1)=Ax(k)+Bx(km)+f(k) with commutative matricesJournal of Mathematical Analysis and Applications200331816376221087210.1016/j.jmaa.2005.05.021ZBL1094.39002GryazinaE. N.PolyakB. T.Stability regions in the parameter space: D-decomposition revisitedAutomatica2006421132610.1016/j.automatica.2005.08.0102183067ZBL1121.93028ŠiljakD.Parameter space methods for robust control design: a guided tourIEEE Transactions on Automatic Control1989347674688100066110.1109/9.29394ZBL0687.93033HornR.JohnsonC.Matrix Theory1986Cambridge, UKCambridge University PressKipnisM.MalyginaV.The stability cone for a matrix delay difference equationInternational Journal of Mathematics and Mathematical Sciences2011201115860326CohnA.Über die Anzahl der Wurzeln einer algebraischen Gleichung in einem KreiseMathematische Zeitschrift1922141110148154454310.1007/BF01215894CampbellS. A.YuanY.BungayS.Equivariant Hopf bifurcation in a ring of identical cells with delayed couplingNonlinearity200518628272847217696110.1088/0951-7715/18/6/022YuanY.CampbellS. A.Stability and synchronization of a ring of identical cells with delayed couplingJournal of Dynamics and Differential Equations2004163709744210916310.1007/s10884-004-6114-yZBL1071.34079DavisP. J.Circulant Matrices1994New York, NY, USAAMS ChelseaKaplanJ.YorkeJ.PeitgenH. O.WaltherH. O.Chaotic behavior of multidimensional difference equationsFunctional Differential Equations and Approximation of Fixed Points2006730Berlin, GermanySpringer204227Lecture Notes in MathematicsKipnisM. M.KomissarovaD. A.Stability of delay difference systemAdvances in Difference Equations2006200692249006314109KulikovA. Yu.Stability of a linear nonautonomous difference equation with bounded delaysRussian Mathematics20105411182610.3103/S1066369X10110022KulikovA. Yu.MalyginaV. V.On the stability of nonautonomous difference equations with several delaysRussian Mathematics2008523152310.3103/S1066369X0803002X2444077MalyginaV. V.KulikovA. Yu.On the stability of semi-autonomous difference equationsRussian Mathematics20115551927KhokhlovaT. N.KipnisM. M.MalyginaV. V.The stability cone for a delay differential matrix equationApplied Mathematics Letters201124574274510.1016/j.aml.2010.12.020SabatulinaT. L.MalyginaV. V.Several stability tests for linear autonomous differential equations with distributed delayRussian Mathematics2007516526010.3103/S1066369X070600722380833KipnisM. M.LevitskayaI. S.Stability of delay difference and differential equations: similarities and distinctionsDifference Equations, Special functions and Orthogonal Polynomials2007Hackensack, NJ, USAWorld Scientific31532410.1142/9789812770752_00272451180ZBL1127.39017LevitskayaI. S.Stability domain of a linear differential equation with two delaysComputers & Mathematics with Applications2006511153159219561410.1016/j.camwa.2005.05.011ZBL1111.34054