ON STABILITY WITH RESPECT TO THE FIRST APPROXIMATION
N. A. IZOBOV
Submitted 1966 | SovietRxiv: ru-196601.44881 | Translated from Russian

Full Text

UDC 517.914.2

ON STABILITY WITH RESPECT TO THE FIRST APPROXIMATION

N. A. IZOBOV

We consider \(n\)-dimensional systems

\[ \frac{dx}{dt}=P(t)x, \tag{1} \]

\[ \frac{dy}{dt}=P(t)y+f(t,y) \tag{2} \]

with a matrix \(P(t)\) bounded and continuous for \(t \geqslant 0\), such that the characteristic exponents \(\lambda_i\) of system (1) are negative:

\[ \lambda_1 \leqslant \ldots \leqslant \lambda_n, \tag{3} \]

and the vector function \(f(t,y)\) satisfies the condition

\[ \|f(t,y)\| \leqslant g(t)\|y\|. \tag{4} \]

In [1], criteria are given for the asymptotic stability of the trivial solution of system (2)

\[ \left(\int_0^\infty g(t)\exp[(\sigma_\Gamma+\alpha)t]\,dt<+\infty,\ \alpha>0,\ \sigma_\Gamma \text{ is the Grobman irregularity coefficient of system (1)}\right) \]

for arbitrary \(f(t,y)\) satisfying (4).

In the present note, for systems (1) we introduce the notion of a quantity \(\gamma\), smaller than \(\sigma_\Gamma\) for a certain class of systems (and coinciding with \(\sigma_\Gamma\) for the remaining systems), but possessing a property analogous to the property of \(\sigma_\Gamma\) indicated above, and we establish one property of such systems.

\(1^\circ\). Let \(X(t)\) be a binormal matrix of solutions [2] of system (1); \(\lambda_i\) the characteristic exponent of the \(i\)-th column-solution of the matrix \(X(t)\); \(\delta_j\) the characteristic exponent of the \(j\)-th row of the matrix \(X^{-1}(t)\), \(\sigma_p \equiv \lambda_p+\delta_p\), \(p=1,\ldots,n\). In what follows we consider systems (1) for which

\[ \sigma_k=\max_p\{\sigma_p\}>\max_{p\ne k}\{\sigma_p\}=\sigma_l. \tag{5} \]

We note that \(\sigma_\Gamma=\sigma_k\) [2].

Let

\[ L_{ji}(U)\equiv \begin{cases} \displaystyle \delta_j+\lambda_i+\sum_{p=j}^{i-1}u_p, & j<i,\\[1.2em] \displaystyle \delta_j+\lambda_i-\sum_{p=i}^{j-1}u_p, & j>i. \end{cases} \]

where the point \(U(u_1,\ldots,u_{n-1})\) is a point of the Euclidean space \(E_{n-1}(u_n=0)\). By definition,

\[ \gamma(U)=\max_{j,i}\{L_{ji}(U)\}\qquad (j,i=1,\ldots,n) \]

(the maximum is taken over all pairs \((j,i)\) for which the corresponding \(L_{ji}(U)\) are defined, i.e., for which \(j\ne i\)). Let us specify a bounded domain \(G\subset E_{n-1}\). The point \(U(u_1,\ldots,u_{n-1})\in G\) if

\[ \lambda_j \leq \sum_{p=j}^{i-1} u_p \leq -\lambda_i,\qquad j<i. \tag{6} \]

Obviously, \(G\) is convex; \(G\) is nonempty, since the point \(U_\Gamma(\lambda_1-\lambda_2,\ldots,\lambda_{n-1}-\lambda_n)\in G\), and moreover is an interior point of it (not lying on the boundary). By definition,

\[ \gamma=\inf_{U\in G}\gamma(U). \]

Lemma.

\[ \gamma=\max\left\{\frac12(\sigma_k+\sigma_l),\ \sigma_k+\max_{p\ne k}\{\lambda_p\}\right\}. \]

Proof. We agree that

\[ \sum_{p=s}^{r}u_p=0,\qquad \text{if } r<s. \]

Compute \(\gamma(U_0)\):

\[ U_0(u_1^0,\ldots,u_{n-1}^0), \]

\[ u_i^0=\lambda_i-\lambda_{i+1},\qquad 1\leq i\leq k-2,\quad k+1\leq i\leq n-1, \tag{7} \]

\[ u_{k-1}^0=\lambda_{k-1}-\lambda_l+\frac12(\delta_k-\delta_l+\lambda_l-\lambda_k), \tag{8} \]

\[ u_k^0=\lambda_l-\lambda_{k+1}-\frac12(\delta_k-\delta_l+\lambda_l-\lambda_k). \tag{9} \]

Remark. If \(k=1\), then the coordinates of the point \(U_0\) are computed by formulas (7), (9), while if \(k=n\), by formulas (7), (8).

First compute \(L_{ki}(U_0)\); for this consider the case \(i>k\) (the case \(i<k\) is analogous). Since, according to (7),

\[ \sum_{p=k+1}^{i-1}u_p^0=\lambda_{k+1}-\lambda_i, \]

we have

\[ L_{ki}(U_0)=\delta_k+\lambda_i+u_k^0+\sum_{p=k+1}^{i-1}u_p^0 =\frac12(\sigma_k+\sigma_l),\qquad i>k. \]

In the case \(i<k\),

\[ L_{ki}(U_0)=\delta_k+\lambda_i-\sum_{p=i}^{k-2}u_p^0-u_{k-1}^0, \]

and therefore the relation

\[ L_{ki}(U_0)=\frac12(\sigma_k+\sigma_l),\qquad i\ne k, \]

is always valid.

Now estimate \(L_{jk}(U_0)\), \(j\ne k\):

\[ L_{jk}(U_0)=\delta_j+\lambda_k+\sum_{p=j}^{k-2}u_p^0-u_{k-1}^0 =\frac12(\sigma_j+\sigma_k)+\frac12(\sigma_j-\sigma_l),\qquad j<k, \]

\[ L_{jk}(U_0)=\delta_j+\lambda_k-u_k^0-\sum_{p=k+1}^{j-1}u_p^0 =\frac12(\sigma_j+\sigma_k)+\frac12(\sigma_j-\sigma_l),\qquad j>k. \]

Therefore, for any admissible \(j\),
\[ L_{jk}(U_0)\leq \frac12(\sigma_k+\sigma_l). \]
Obviously, it now suffices to estimate only those \(L_{ji}(U_0)\) for which \(j\ne k,\ i\ne k\). Compute
\[ \sum_{p=j}^{i-1}u_p^0 \]
(and similarly \(\sum_{p=i}^{j-1}u_p^0\)). Since \(j\ne k,\ i\ne k\), the sum \(\sum_{p=j}^{i-1}u_p^0\) either contains none of the terms \(u_{k-1}^0,\ u_k^0\), or contains the entire sum \(u_{k-1}^0+u_k^0\). Hence
\[ \sum_{p=j}^{i-1}u_p^0=\lambda_j-\lambda_i,\qquad \sum_{p=i}^{j-1}u_p^0=\lambda_i-\lambda_j, \]
and then
\[ L_{ji}(U_0)=\delta_j+\lambda_i+\lambda_j-\lambda_i=\sigma_j,\qquad j\ne k,\ i\ne k. \]

Finally,
\[ \gamma(U_0)=\frac12(\sigma_k+\sigma_l). \]
Let us show that \(\gamma=\gamma(U_0)\), if \(U_0\in G\). Suppose the contrary: there is a point
\[ U'(u_1',\ldots,u_{n-1}')\in G \]
for which \(\gamma(U')<\gamma(U_0)\).

Let \(l<k\) (the case \(k<l\) is analogous). Obviously,
\[ \sum_{p=l}^{k-1}u_p'\ne \frac12(\delta_k-\delta_l+\lambda_l-\lambda_k), \]
since otherwise we would have
\[ \gamma(U')\geq L_{lk}(U')=\frac12(\sigma_k+\sigma_l). \]
Denote
\[ \sum_{p=l}^{k-1}u_p'-\frac12(\delta_k-\delta_l+\lambda_l-\lambda_k)\equiv \alpha. \]
Then
\[ \gamma(U')\geq \max\{L_{kl}(U'),\,L_{lk}(U')\} =\max\{\gamma(U_0)-\alpha,\,\gamma(U_0)+\alpha\}. \]
We have obtained a contradiction. Finally,
\[ \gamma=\frac12(\sigma_k+\sigma_l),\qquad U_0\in G. \]

Let us now consider the case when \(U\in {}_0G\). Let the coordinates of the point
\[ U_1(u_1^{(1)},\ldots,u_{n-1}^{(1)}) \]
be determined as follows:
\[ u_i^{(1)}=\lambda_i-\lambda_{i+1},\qquad 1\leq i\leq k-2,\quad k+1\leq i\leq n-1, \tag{7'} \]
\[ u_{k-1}^{(1)}=\lambda_{k-1}-\lambda_k-\max_{p\ne k}\{\lambda_p\}, \tag{8'} \]
\[ u_k^{(1)}=\lambda_k-\lambda_{k+1}+\max_{p\ne k}\{\lambda_p\}. \tag{9'} \]

Here the preceding remark should be taken into account. Let us verify that \(U_1\in G\).

First of all, \(\lambda_i \leqslant u_i^{(1)} \leqslant -\lambda_{i+1}\); this follows from the negativity of \(\lambda_i\) and from the inequalities

\[ -\lambda_{k+1}+\max_{p\ne k}\{\lambda_p\} \geqslant 0,\quad k<n, \]

\[ \lambda_{k-1}-\max_{p\ne k}\{\lambda_p\}\leqslant 0,\quad k>1. \]

Therefore it now suffices to consider all possible pairs \(\{j,i\}\) for which \(j<i-1\). Obviously, for \(\{j,k\}\), \(\{k,i\}\) we have, respectively,

\[ \sum_{p=j}^{k-1} u_p^{(1)} =\lambda_j-\lambda_k-\max_{p\ne k}\{\lambda_p\}, \tag{10} \]

\[ \sum_{p=k}^{i-1} u_p^{(1)} =\lambda_k-\lambda_i+\max_{p\ne k}\{\lambda_p\} \tag{11} \]

(the relations (10) and (11) are also valid for \(j=k-1\), \(i=k+1\), respectively). But
\(\lambda_j-\max_{p\ne k}\{\lambda_p\}\leqslant 0\) for \(j<k\),
\(\max_{p\ne k}\{\lambda_p\}-\lambda_i\geqslant 0\) for \(i>k\). Therefore (10) and (11) satisfy conditions (6). For all remaining \(\{j,i\}\), \(j<i-1\),

\[ \sum_{p=j}^{i-1}u_p^{(1)}=\lambda_j-\lambda_i, \]

and the fulfillment of (6) is obvious.

Thus, \(U_1\in G\). Let us show that \(U_1\) is a boundary point of \(G\). Let \(\lambda_s=\max_{p\ne k}\{\lambda_p\}\). If \(s<k\), then from (10) we have

\[ \sum_{p=s}^{k-1}u_p^{(1)} =\lambda_s-\lambda_k-\max_{p\ne k}\{\lambda_p\} =-\lambda_k, \tag{12} \]

whereas if \(s>k\), then from (11)

\[ \sum_{p=k}^{s-1}u_p^{(1)} =\lambda_k-\lambda_s+\max_{p\ne k}\{\lambda_p\} =\lambda_k. \tag{13} \]

Comparing (12) and (13) with (6), we obtain what is required.

Let us compute \(\gamma(U_1)\). It is easy to see that

\[ L_{ki}(U_1)=\sigma_k+\max_{p\ne k}\{\lambda_p\}. \]

Let \(\Pi\) be the set of points of the line passing through the points \(U_\Gamma\) and \(U_0\). It is not hard to verify that the point \(U_1\in\Pi\). Obviously, \(u_n=L_{ji}(U)\), \(U\in\Pi\), is the equation of a line in the \(n\)-dimensional space \(E_n\), and since

\[ L_{ki}(U_\Gamma)>L_{ki}(U_0),\quad L_{ki}(U_\Gamma)>L_{ki}(U_1), \]

and the point \(U_1\in G\), the point \(U_1\) lies between the points \(U_\Gamma\) and \(U_0\) on the line \(\Pi\) (otherwise \(L_{ki}(U_1)>L_{ki}(U_\Gamma)\)) and divides the “segment” \(U_\Gamma U_0\) in some positive ratio \(\tau\). Using the inequalities obtained earlier and the obvious inequalities

\[ L_{ji}(U_0)\leqslant \frac12(\sigma_k+\sigma_l),\quad L_{ji}(U_\Gamma)\leqslant \sigma_i,\quad j\ne k, \]

we have

\[ L_{ji}(U_1)=\frac{L_{ji}(U_\Gamma)+\tau L_{ji}(U_0)}{1+\tau} \leq \frac{\frac{1}{2}(\sigma_k+\sigma_l)+\tau\cdot\frac{1}{2}(\sigma_k+\sigma_l)}{1+\tau} = \frac{1}{2}(\sigma_k+\sigma_l). \tag{14} \]

Since \(U_1\) is an interior point of the segment \(U_\Gamma U_0\), from \(L_{ki}(U_\Gamma)>L_{ki}(U_0)\) it follows that

\[ L_{ki}(U_1)>L_{ki}(U_0) \tag{15} \]

(recall that the point \(\overline U_1\bigl(u_1^{(1)},\ldots,u_{n-1}^{(1)},L_{ki}(U_1)\bigr)\) lies on the line passing through the points \(\overline U_\Gamma\bigl(\lambda_1-\lambda_2,\ldots,\lambda_{n-1}-\lambda_n,L_{ki}(U_\Gamma)\bigr)\) and \(\overline U_0\bigl(u_1^0,\ldots,u_{n-1}^0,L_{ki}(U_0)\bigr)\)), and hence, taking (14) into account,

\[ L_{ji}(U_1)<\sigma_k+\max_{p\ne k}\{\lambda_p\},\qquad j\ne k. \]

Thus, \(\gamma(U_1)=\sigma_k+\max_{p\ne k}\{\lambda_p\}\). By virtue of (12) or (13) (the point \(U_1\) is a boundary point) and (6), \(\gamma=\gamma(U_1)\), for for any point \(U\in G\)

\[ L_{ks}(U)=\gamma(U_1)+\varepsilon(U)\geq \gamma(U_1), \]

since

\[ \varepsilon(U)= \begin{cases} \displaystyle \sum_{p=k}^{s-1} u_p-\lambda_k, & k<s,\\[1.2em] \displaystyle -\sum_{p=s}^{k-1} u_p-\lambda_k, & k>s, \end{cases} \]

is a nonnegative quantity.

Finally,

\[ \gamma= \begin{cases} \displaystyle \frac{1}{2}(\sigma_k+\sigma_l), & \text{if } U_0\in G,\\[1.2em] \displaystyle \sigma_k+\max_{p\ne k}\{\lambda_p\}, & \text{if } U_0\notin G. \end{cases} \tag{*} \]

If \(U_0\notin G\), then from inequality (15) it follows that

\[ \sigma_k+\max_{p\ne k}\{\lambda_p\}>\frac{1}{2}(\sigma_k+\sigma_l). \]

Therefore, to prove the lemma it is sufficient to show that in the case \(U_0\in G\) the inequality

\[ \frac{1}{2}(\sigma_k+\sigma_l)\geq \sigma_k+\max_{p\ne k}\{\lambda_p\} \tag{16} \]

holds. From the inequalities

\[ L_{ki}(U_\Gamma)>L_{ki}(U_1),\qquad L_{ki}(U_\Gamma)>L_{ki}(U_0) \]

and from the fact that the point \(U_1\) is a boundary point of the domain \(G\), it follows that the point \(U_0\) is a point of the segment \(U_\Gamma U_1\). Hence, as above, \(L_{ki}(U_0)\geq L_{ki}(U_1)\), i.e. (16) holds. On the basis of (*) we have the assertion of the lemma.

Corollary 1. The quantity \(\gamma\) does not depend on the choice of the binormal matrix \(X(t)\).

The proof follows from Theorem 2 of [2] and the expression for \(\gamma\) established by the lemma.

Corollary 2. \(\sigma_l < \gamma < \sigma_\Gamma\).

Corollary 3. If system (1) is of second order, then \(\gamma\) does not exceed the Perron number \(\sigma_\Pi\) [2].

Let us note that in the case \(\sigma_l=\sigma_k\), \(\gamma\) coincides with Grobman’s irregularity coefficient.

\(2^\circ\). Together with systems (1), (2) we consider the system

\[ \frac{dy}{dt}=[P(t)+Q(t)]y . \tag{2'} \]

Denote

\[ \overline{\omega}[r]\equiv \overline{\lim}_{t\to+\infty}\frac{1}{t}\ln\|r(t)\|, \]

where \(r(t)\) is a scalar, a vector, or a matrix. The following is true.

Theorem 1. If \(\overline{\omega}[Q]<-\gamma\), then the characteristic exponents of system \((2')\) are negative.

Proof. Denote

\[ H(t)=((h_{ji}(t)))\equiv X^{-1}(t)Q(t)X(t), \]

where \(j\) is the row number; \(\overline{\omega}[Q]+\gamma=-\varepsilon,\ \varepsilon>0\). Obviously,

\[ \overline{\omega}[h_{ji}]\leq \delta_j+\lambda_i-\gamma-\varepsilon \quad (j,i=1,\ldots,n). \tag{17} \]

We shall establish for \(\overline{\omega}[h_{kk}]\) a sharper estimate; for this purpose compute \(h_{kk}\):

\[ h_{kk}=\sum_{s=1}^{n}\left(\sum_{i=1}^{n}x'_{ki}q_{is}\right)x_{sk} =\sum_{s=1}^{n}\left(\sum_{i=1}^{n}x'_{ki}x_{sk}q_{is}\right), \tag{18} \]

where \(X=((x_{ji}))\), \(X^{-1}=((x'_{ji}))\), \(Q=((q_{ji}))\), and \(j\) is the row number. The estimate for \(\overline{\omega}[h_{kk}]\) needed by us will be obtained if an estimate of the quantity \(\overline{\omega}[x'_{ki}x_{sk}]\) is given in a suitable way. Multiplying the \(s\)-th row of the matrix \(X(t)\) by the \(i\)-th column of its inverse matrix \(X^{-1}(t)\), we have

\[ \sum_{p=1}^{n}x_{sp}x'_{pi}=\delta_{si}, \]

where \(\delta_{si}\) is the Kronecker symbol; whence

\[ x_{sk}x'_{ki}=\delta_{si}-\sum_{p\ne k}x_{sp}x'_{pi}. \tag{19} \]

But

\[ \overline{\omega}[x_{sp}x'_{pi}]\leq \lambda_p+\delta_p=\sigma_p \]

for arbitrary \(s\) and \(i\). Therefore from (19) we have \((\sigma_p\geq 0)\)

\[ \overline{\omega}[x_{sk}x'_{ki}]\leq \max_{p\ne k}\{\sigma_p\}=\sigma_l, \]

and from (18), finally,

\[ \overline{\omega}[h_{kk}]\leq \sigma_l-\gamma-\varepsilon. \tag{20} \]

By the definition of \(\gamma\), there is a point \(U_m\in G\) such that \(\gamma=\gamma(U_m)\). In view of the convexity of the domain \(G\) and the existence of its interior point \(U_\Gamma\), in any neighborhood of the point \(U_m\) there are interior points of the domain \(G\). The function \(\gamma(U)\) is continuous; therefore, for the chosen \(\varepsilon\) there is an interior point \(U_\varepsilon(u_1^\varepsilon,\ldots,u_{n-1}^\varepsilon)\) of the domain \(G\) such that

\[ |\gamma(U_\varepsilon)-\gamma|<\frac{\varepsilon}{2}. \tag{21} \]

(if \(U_m\) is an interior point, i.e., \(U_m\) coincides with \(U_0\), then the point \(U_0\) may be taken as \(U_\varepsilon\)).

In equation \((2')\), written in matrix form, we make the change of variables

\[ Y(t)=X(t) \begin{pmatrix} e^{-\alpha_1 t} & 0\\ \cdot & \\ & \cdot & \\ 0 & & e^{-\alpha_n t} \end{pmatrix} Z(t), \tag{22} \]

where \(Z(t)\) is the fundamental matrix of the system

\[ \frac{dZ}{dt} = \begin{pmatrix} \alpha_1 & 0\\ \cdot & \\ & \cdot & \\ 0 & & \alpha_n \end{pmatrix} Z+B(t)Z, \tag{23} \]

\[ B(t)\equiv \begin{pmatrix} e^{\alpha_1 t} & 0\\ \cdot & \\ & \cdot & \\ 0 & & e^{\alpha_n t} \end{pmatrix} X^{-1}(t)Q(t)X(t) \begin{pmatrix} e^{-\alpha_1 t} & 0\\ \cdot & \\ & \cdot & \\ 0 & & e^{-\alpha_n t} \end{pmatrix}, \]

and \(\alpha_i\), \((i=1,\ldots,n)\), are determined from the following linear system of equations:

\[ \alpha_n=0, \]

\[ \alpha_i-\alpha_{i+1}=u_i^\varepsilon \quad (i=1,\ldots,n-1). \]

Let us now estimate \(\bar\omega[B]\). From the definition of \(B(t)\) it follows that

\[ b_{ji}(t)=h_{ji}(t)\exp[(\alpha_j-\alpha_i)t]. \]

Hence

\[ \bar\omega[b_{ji}]\leqslant (\delta_j+\lambda_i+\alpha_j-\alpha_i)-\gamma-\varepsilon, \qquad j\ne i. \]

Let \(j<i\); then

\[ \alpha_j-\alpha_i = \sum_{p=j}^{i-1}(\alpha_p-\alpha_{p+1}) = \sum_{p=j}^{i-1}u_p^\varepsilon, \]

and therefore

\[ \delta_j+\lambda_i+(\alpha_j-\alpha_i)=L_{ji}(U_\varepsilon). \]

An analogous result is obtained also in the case \(j>i\). Thus,

\[ \bar\omega[b_{ji}]\leqslant L_{ji}(U_\varepsilon)-\gamma-\varepsilon, \qquad j\ne i, \]

i.e., on the basis of (21) and the definition of \(\gamma(U)\), we have

\[ \bar\omega[b_{ji}] \leqslant \gamma(U_\varepsilon)-\gamma-\varepsilon \leqslant |\gamma(U_\varepsilon)-\gamma|-\varepsilon < -\frac{\varepsilon}{2}, \qquad j\ne i. \]

According to Corollary 2 of the lemma and estimate (20),

\[ \bar\omega[b_{ii}]\leqslant \sigma_1-\gamma-\varepsilon \quad (i=1,\ldots,n). \]

Thus, always \(\bar\omega[B]<0\). Therefore, according to [3], the totality of characteristic exponents of system (23) is the set \(\{\alpha_1,\ldots,\alpha_n\}\).

Let \(Z(t)=((Z_{ji}(t)))\) be a normal matrix of solutions of system (23), such that the characteristic exponent of the \(i\)-th solution-column is equal to \(\alpha_i\). Denote

\[ S(t) \equiv \begin{pmatrix} e^{-\alpha_1 t} & 0\\ \cdot & \\ & \cdot & \\ 0 & & e^{-\alpha_n t} \end{pmatrix} Z(t). \]

Then \(s_{ji}(t)=z_{ji}(t)e^{-\alpha_j t}\), and therefore \(\bar\omega[s_{ji}]\leqslant \alpha_i-\alpha_j\). From our substitution

\[ y_{ji}(t)=\sum_{p=1}^{n} x_{jp}s_{pi}, \qquad Y(t)=\bigl((y_{ji}(t))\bigr) \]

it follows, taking into account the preceding estimate, that

\[ \bar\omega[x_{jp}s_{pi}]\leqslant \lambda_p+\alpha_i-\alpha_p. \]

If \(i<p\), then

\[ \alpha_i-\alpha_p=\sum_{r=i}^{p-1}(\alpha_r-\alpha_{r+1}) =\sum_{r=i}^{p-1}u_r^\varepsilon, \]

and therefore

\[ \bar\omega[x_{jp}s_{pi}]\leqslant \lambda_p+\sum_{r=i}^{p-1}u_r^\varepsilon . \]

But the point \(U_\varepsilon\) is an interior point of the domain \(G\); therefore from (6) we have

\[ \beta_{ip}\equiv \lambda_p+\sum_{r=i}^{p-1}u_r^\varepsilon<0,\qquad i<p. \]

If, however, \(i>p\), then

\[ \lambda_p+\alpha_i-\alpha_p =\lambda_p-(\alpha_p-\alpha_i) =\lambda_p-\sum_{r=p}^{i-1}u_r^\varepsilon . \]

And again from (6) (the point \(U_\varepsilon\) is interior) we have

\[ \beta_{ip}\equiv \lambda_p-\sum_{r=p}^{i-1}u_r^\varepsilon<0,\qquad i>p. \]

In the case \(p=i\), \(\bar\omega[x_{ji}s_{ii}]\leqslant \lambda_i\). Denote

\[ \beta_i=\max\{\max_{p\ne i}\{\beta_{ip}\},\lambda_i\}. \]

Then \(\lambda_i'\leqslant\beta_i\), where \(\lambda_i'\) is the characteristic exponent of the \(i\)-th column-solution of the fundamental matrix of solutions \(Y(t)\). Obviously, \(\beta_i<0\) \((i=1,\ldots,n)\). The theorem is proved.

Let

\[ \beta(\varepsilon)=\max_i\{\beta_i\}, \]

and let \(y(t)\) be an arbitrary solution of system (2′).

Corollary 1. For any \(Q(t)\), \(\bar\omega[Q]+\gamma=-\varepsilon\), \(\varepsilon>0\), the inequality \(\bar\omega[y]\leqslant\beta(\varepsilon)<0\) holds.

Corollary 2. If there exists an \(\varepsilon>0\) such that

\[ \int_{0}^{\infty}\|Q(t)\|e^{(\gamma+\varepsilon)t}\,dt<+\infty, \tag{24} \]

then \(\bar\omega[y]\leqslant\beta(\varepsilon)\) for all \(Q(t)\) satisfying (24).

The proof follows from [3].

Theorem 2. If there exists a positive \(\varepsilon\) such that

\[ \int_{0}^{\infty} g(t)e^{(\gamma+\varepsilon)t}\,dt<+\infty, \]

then the trivial solution of system (2) is asymptotically stable.

Proof. According to the Bylov–Grobman principle of linear inclusion [4], every nontrivial solution \(y(t)\) of system (2) is a solution of a linear system

\[ \frac{dv}{dt}=P(t)v+Q_y(t)v \]

of the form \((2')\), where the matrix \(Q_y(t)\), depending on \(y\), has the property

\[ \|Q_y(t)\|\leq g(t) \tag{25} \]

for all \(y\). Then, according to the substitution (22),

\[ y(t)=X(t) \begin{pmatrix} e^{-a_1t} & 0\\ \cdot & \cdot\\ 0 & e^{-a_nt} \end{pmatrix} z_y(t), \tag{22'} \]

where \(z_y(t)\) is a solution of system (23) with the corresponding matrix \(B_y(t)\). By the well-known Gronwall lemma,

\[ \|z_y(t)\|\leq \|z_y(0)\| \exp\left\{\int_0^t \|B_y(\tau)\|\,d\tau\right\} e^{\left[\max_p\{a_p\}\right]t},\qquad t\geq 0. \]

Let

\[ a_s=\max_p\{a_p\}. \]

From the definition of the matrix \(B_y(t)\), taking (21) into account, it follows that, for a given \(\varepsilon\), there is a positive constant \(b=b(\varepsilon)\), depending only on \(\varepsilon\), such that for every \(y(t)\), by virtue of (25),

\[ \|B_y(t)\|\leq b g(t)e^{(\gamma+\varepsilon)t}. \]

Therefore

\[ \|z_y(t)\|\leq N\|z_y(0)\|e^{a_s t},\qquad N=\exp\left[b\int_0^\infty g(t)e^{(\gamma+\varepsilon)t}\,dt\right]. \]

From the substitution \((22')\), for any \(\varepsilon_1>0\) there is an \(M(\varepsilon_1)>0\) such that

\[ \|y(t)\|\leq M\|z_y(t)\| \exp\left[\left(\max_p\{\lambda_p-a_p\}+\varepsilon_1\right)t\right]. \]

But

\[ \max_p\{\lambda_p-a_p\}+a_s\leq \beta(\varepsilon),\qquad \|z_y(0)\|\leq \|X^{-1}(0)\|\,\|y(0)\|. \]

Finally,

\[ \|y(t)\|\leq MN\|X^{-1}(0)\|\,\|y(0)\|e^{[\beta(\varepsilon)+\varepsilon_1]t}. \]

From this estimate, choosing \(\varepsilon_1\) so that \(\beta(\varepsilon)+\varepsilon_1<0\), the validity of the theorem follows.

On the basis of Corollary 3 of the lemma, in the case \(n=2\) the following is true.

Corollary. If \(\bar\omega[Q]<-\sigma_{\Pi}\), then for every solution \(y(t)\) of system \((2')\), \(\omega[y]<0\).

An even stronger result is true, namely the following.

Remark. Under the conditions of the corollary to Theorem 2,

\[ \bar\omega[y]\leq \max_p\{\lambda_p\}. \]

Proof. Obviously, consideration is required only in the case

\[ \sigma_{\Pi}<\sigma_{\Gamma}. \]

But in this case

\[ \sigma_{\Pi}=\max_p\{\lambda_p+\delta_{3-p}\}\qquad (p=1,2), \]

and the validity of the remark is obvious if one makes the substitution

\[ y(t)=X(t)z(t) \]

in equation \((2')\). Using estimate (20), we see that \(z(t)\)—

solution of a system having a linear asymptotic equilibrium ([5], p. 74).

Let us note, in connection with the last remark, that by changing the domain \(G\) in a suitable way, one can obtain certain modifications of the quantity \(\gamma\) possessing definite properties.

\(3^\circ\). In conclusion, let us clarify one property of the quantity \(\gamma_0=\dfrac{1}{2}(\sigma_k+\sigma_l)\).

The following assertion is valid.

If \(\bar\omega[Q]<-\gamma_0\), then the set of characteristic exponents \(\{\lambda'_1,\ldots,\lambda'_n\}\) of system \((2')\), written in a certain order, satisfies the condition

\[ \lambda'_i \leq \lambda_i+\frac{1}{2}(\sigma_k-\sigma_l). \]

Proof. In equation \((2')\) we make the substitution (22), where the \(\alpha_i\) are now defined as follows:

\[ \alpha_i=\frac{1}{2}(\delta_n-\delta_i+\lambda_i-\lambda_n)\quad (i=1,\ldots,n). \]

On the basis of (17), (20), and the definition of \(\alpha_i\), we have \(\bar\omega[B]<0\). Therefore the set of characteristic exponents of system (22), as before, is \(\{\alpha_1,\ldots,\alpha_n\}\). Using the estimate for \(s_{pi}\) proved in Theorem 1, the inequality
\[ \bar\omega[x_{jp}s_{pi}]\leq \lambda_p+\alpha_i-\alpha_p =\lambda_i+\frac{1}{2}(\sigma_p-\sigma_i) \]
is valid for every \(j\) and fixed \(i\). Hence from the substitution \(y_{ji}(t)=\sum_{p=1}^{n} x_{jp}s_{pi}\) it follows that

\[ \bar\omega[y_{ji}]\leq \lambda_i+\frac{1}{2}(\sigma_k-\sigma_i), \]

i.e., there exists a fundamental matrix of solutions \(Y(t)\) such that the characteristic exponent of its \(i\)-th column does not exceed the quantity
\[ \lambda_i+\frac{1}{2}(\sigma_k-\sigma_i). \]
This implies the validity of the assertion.

Corollary. If \(\bar\omega[Q]<-\gamma_0\), then system \((2')\) has at least one nontrivial solution whose characteristic exponent does not exceed \(\lambda_k\).

Everything stated above can be reformulated in the corresponding way also for system (2).

References

  1. D. M. Grobman, Mat. sb., 30 (72), 1, 1952, pp. 121–166.
  2. R. E. Vinograd, DAN SSSR, 119, No. 3, 1958.
  3. D. M. Grobman, DAN SSSR, 86, No. 1, 19–22, 1952.
  4. B. F. Bylov, D. M. Grobman, UMN, 17, issue 3 (105), 1962.
  5. L. Cesari, Asymptotic Behavior and Stability of Solutions of Ordinary Differential Equations. Mir Publishers, Moscow, 1964.

Received by the editors
November 17, 1965

Belorussian State University
named after V. I. Lenin

Submission history

ON STABILITY WITH RESPECT TO THE FIRST APPROXIMATION