Approximations of the eigenvalues for the discrete and “continuous” operators ().
Analytic and numerical results of the Euler-Bernoulli beam model with a two-parameter family of boundary conditions have been presented. The co-diagonal matrix depending on two control parameters (k1 and k2) relates a two-dimensional input vector (the shear and the moment at the right end) and the observation vector (the time derivatives of displacement and the slope at the right end). The following results are contained in the paper. First, high accuracy numerical approximations for the eigenvalues of the discretized differential operator (the dynamics generator of the model) have been obtained. Second, the formula for the number of the deadbeat modes has been derived for the case when one control parameter, k1, is positive and another one, k2, is zero. It has been shown that the number of the deadbeat modes tends to infinity, as k1→1+ and k2=0. Third, the existence of double deadbeat modes and the asymptotic formula for such modes have been proven. Fourth, numerical results corroborating all analytic findings have been produced by using Chebyshev polynomial approximations for the continuous problem.
- matrix differential operator
- Chebyshev polynomials
- numerical scheme
- boundary control
The present paper is concerned with the spectral analysis and numerical investigation of the eigenvalues of the Euler-Bernoulli beam model. The beam is clamped at the left end and subject to linear feedback-type conditions with a non-dissipative feedback matrix [1, 2]. Depending on the boundary parameters and , the model can be either conservative, dissipative, or completely non-dissipative. We focus on the non-dissipative case, i.e., when the energy of a vibrating system is not a decreasing (or nonincreasing) function of time. In our approach, the initial-boundary value problem describing the beam dynamics is reduced to the first order in time evolution equation in the state Hilbert space . The evolution of the system is completely determined by the dynamics generator ,
Based on the results of [1, 2], the dynamics generator has a purely discrete spectrum, whose location on the complex plane is determined by the controls and . Having in mind the practical applications of the asymptotic formulas [3, 4, 5], we discuss the case of and , such that (see Proposition 2). As shown in , even though the operator is non-dissipative, for the case and (or and ), the entire set of eigenvalues is located in the closed upper half of the complex plane , which means that all eigenmodes are stable or neutrally stable. (We recall that to obtain an elastic mode from an eigenvalues of , one should multiply the eigenvalue by a factor ).
In the paper we address the question of accuracy of the asymptotic formulas for the eigenvalues.
We have also shown that there exists a sequence of values of the parameter , i.e., , such that for each there exist a finite number of deadbeat modes and each corresponds to a
Finally, we mention that the feedback control of beams is a well-studied area , with multiple applications to the control of robotic manipulators, long and slender aircraft wings, propeller blades, large space structure [7, 8], and the dynamics of carbon nanotubes . The analysis of a classical beam model with nonstandard feedback control law that originated in engineering literature [4, 10, 11, 12] may be of interest for both analysts and practitioners.
This paper is organized as follows. In Section 1 we formulate the initial-boundary value problem for the Euler-Bernoulli beam model. In Section 2, we reformulate the problem as an evolution equation in the Hilbert space of Cauchy data (the energy space). The dynamics generator , which is
|, , , , ,|
|, , , , ,|
|, , , ,|
|, , , ,|
1.1 The initial-boundary value problem for the Euler-Bernoulli beam model of a unit length
where is the transverse deflection, is the modulus of elasticity, is the area moment of inertia, is the linear density, and is the cross-sectional area of the beam.
Assuming that the beam is clamped at the left end and free at the right end , and applying Hamilton’s variational principle to the action functional defined by (1), we obtain the equation of motion
and the boundary conditions
where and are the moment and the shear, respectively :
where stands for transposition. The feedback control law is given by
where is the feedback matrix. We select
with being the control parameters. The feedback (6) can be written as
Notice that the choice of a feedback matrix defines whether the system is dissipative or not. Indeed, let be the energy of the system, defined by representation (1). Evaluating on the solutions of Eq. (2) satisfying the left-end conditions from Eqs. (3), we obtain
With the choice of as in Eq. (7), we have
Thus the system is not dissipative for all nonnegative values of and .
2. Operator form of the problem
In what follows, we incorporate the cross-sectional area into the density, write instead of , and also abbreviate . Let be the Hilbert space of two-component vector functions equipped with the following norm:
Assuming that are positive functions, we obtain that the closure of smooth functions satisfying will produce the energy space . Here , and the equality of function spaces is understood in the sense of a Hilbert-space isomorphism.
Problem (2) with conditions (3) can be represented as the time evolution problem:
where . The dynamics generator is given by the following matrix differential expression:
defined on the domain
For any , the adjoint operator  is given by
i.e., is defined by the same differential expression (14) on the domain described in Eq. (15), where and are replaced by and , respectively. It follows from Eq. (16) that is self-adjoint in and thus is the dynamics generator of the clamped-free beam model. For the reader’s convenience, we summarize the properties of from [1, 2] needed for the present work.
For each , the operator is a rank-two perturbation of the self-adjoint operator in the sense that the operators and exist and are related by the rule
where is a rank-two operator. A similar decomposition is valid for the adjoint operator, i.e.,
From now on, we assume that the structural parameters are constant. In the case of variable parameters, the spectral asymptotics will have the same leading terms and remainder terms depending on parameter smoothness.
The following asymptotic approximations for the eigenvalues (as ) of the operator hold:
If , then for one hasE20
If , then for one has
First of all, we address the question of accuracy of the asymptotic formulas (20) and (21). By its nature, formula (20) (as well as formula (21)) means that for any small , one can find a positive integer , such that all eigenvalues with satisfy the estimate
for the case when and
for the case when . The following important question holds:
The results of numerical simulations (see Tables 1 and 2 and Figures 1 and 2) show that the asymptotic formulas are indeed quite accurate. That is, if one places on the complex plane the numerically produced sets of the eigenvalues, then the theoretically predicted distribution of eigenvalues can be seen almost immediately. To obtain these results, we used the numerical procedure based on Chebyshev polynomial approximations [14, 15, 16], as outlined in Section 5.
In Figures 1 and 2, we represent the graphical distribution of the eigenvalues corresponding to the discretized operator (“numerical” eigenvalues) and the leading asymptotic terms from Eqs. (20) and (21) (“analytic” eigenvalues). In Tables 1 and 2, the numerical values of the corresponding graphical points on Figures 1 and 2 are listed. We have used the following notations: is the number of grid points on , and is the filtering parameter as described in Eq. (69). It can be easily seen from the graphs and tables that the two sets of data coincide almost immediately, i.e., the leading asymptotic terms in the approximations are very close to the numerically approximated eigenvalues.
Figure 3 shows the sub-domains of the -plane, which correspond to different intervals for the values of defined by Eq. (19). On the sub-domain with dark gray color such that , i.e., to evaluate the asymptotic approximation for the eigenvalues, one needs formula (20), while on the complementary sub-domain, one needs formula (21).
3. The deadbeat modes
An eigenvalue of the dynamics generator is called a deadbeat mode if . If the corresponding eigenfunction is , then the evolution problem (13) has a solution given in the form , which tends to zero without any oscillation.
As shown in paper , for the case when one of the control parameters is zero and the other one is positive, the entire set of the eigenvalues is located in the closed upper half plane. This result is not obvious since the operator is not dissipative; in fact, it requires a fairly nontrivial proof. However, due to this fact, we assume that any deadbeat mode can be given in the form , with . To deal with the deadbeat modes analytically, we rewrite the spectral equation in the form of an equivalent problem for an operator pencil  as
If and are an eigenvalue and eigenfunction of the pencil (24), then is also an eigenvalue of with the eigenfunction .
To solve problem (24), we first redefine the spectral and control parameters to eliminate and from Eq. (24). We define , and by and , . Substituting these relations into Eq. (24) and eliminating the “tilde,” we obtain the following Sturm-Liouville eigenvalue problem:
The solution of Eq. (25) satisfying the left-end boundary conditions can be written in the form
For , the deadbeat modes do not exist.
For , there exist infinitely many deadbeat modes given explicitly byE29
For any , there exist a finite number of deadbeat modes. Each mode has the form , where is a root of the function
Let , and then satisfies the estimate
Hence as . (By we denote the greatest integer less than or equal to ).
we reduce Eq. (28) to the following form:
To prove Statement (3), we rewrite Eq. (32) in the form
The left-hand side of Eq. (33) is monotonically increasing, while the right-hand side is sinusoidal, with maximum and minimum , and period . So the graphs of the left- and right-hand side have intersections only on the interval . There are two intersections for each full period of the right-hand side that fits into the above interval (Figure 4). As it can be seen in Figure 4, one should add at least one more intersection for the first half-period after the full periods. Depending on the value of , the two graphs can have two intersections, one tangential intersection or no intersections on the second half-period. This leads to estimate (31).
A graphical illustration of the result of Theorem 1 is shown in Figure 5.
4. Structure of the deadbeat mode set
The main result on the existence and distribution of double roots of the function is presented in the statement below.
Eliminating from system given by (36) and (37), we obtain that the following equation has to be satisfied:
Now we show that when and , i.e.,
Eq. (38) does not have any solutions. Indeed, in the above range of , we have and . With such estimates we obtain that
which mean that Eq. (38) cannot be satisfied.
Now we consider the case when and , i.e.,
It is convenient to rewrite system given by (36) and (37) in the form , where , . If , then Eq. (38) generates the following equation for :
Let us show that for each , Eq. (40) has a unique solution. For we have
and for we have . Evaluating we have
Thus is a monotonically increasing function, such that , which means that has a unique root on .
Finally we show that the multiplicity of a multiple root cannot exceed 2. Using a contradiction argument, assume that there exists , such that in addition to Eqs. (36) and (37), one has , i.e., the multiplicity of is at least 3. The system and can be written as
Since , the second equation of (42) yields . Also, since is a multiple root, we must have . Then is in the third quadrant, which means that , as we have seen above. This contradicts our assumption that is a root of Eqs. (36) and (37).
Representation (44) implies that there exists , such that for all , we have . Taking into account that , we obtain that the root , , of the function is located on the interval . To find the location of this root more precisely , we use linear interpolation. Namely, substituting Eq. (43) into the expression for from Eq. (41) yields
Replacing by the linear function tangential to at the point , and finding the root of this function, we get
Having this approximation for , we immediately get
From the equation , we obtain the formula for as
First we observe that . Indeed, if , then we have
Since is purely imaginary, Eq. (51) is not valid. We define a new function:
One can readily check that satisfies the following boundary-value problem:
Eq. (54) is valid if and only if ; however, for a deadbeat mode, . The obtained contradiction means that for each double mode, there is one eigenfunction and one associate function.
4.1 Deadbeat mode behavior as
As approaches , the spectral branches are moving upward and toward the imaginary axis (Figures 6 and 7). As a result of this motion, eigenvalues approach the imaginary axis at different rates depending on whether approaches from above or below.
As follows from Table 3, the real parts of the eigenvalues decrease steadily as , to a point where the eigenvalue becomes a deadbeat mode. An increase in the number of deadbeat modes can be seen as , which is in agreement with Statement (3) of Theorem 1. One can see from Table 3 that there are pairs of modes such that the distance between them tends to zero as . (Compare modes no.5 and no.7 for , modes no.4 and no.7 for , modes no.5 and no.8 for , and modes no.4 and no.7 for ). Such behavior indicates convergence of the two simple deadbeat modes to a double mode, which is consistent with Theorem 2.
Analyzing Table 4, one can see that the eigenvalues get closer to the imaginary axis as . However the rate at which their real parts approach zero is significantly lower than in the case . Even at , the eigenvalue closest to the imaginary axis has a real part of about , which means that it is not a deadbeat mode (see Statement (1) of Theorem 1).
The eigenvalues near the imaginary axis approach the same double deadbeat modes in both cases when (see Statement (2) of Theorem 1). In conclusion, one can claim that the eigenvalues are indeed approaching the imaginary axis; however, the rate of this approach is different for and . In the former case, an eigenvalue’s distance from the imaginary axis decreases very slowly; in the latter case, the eigenvalues quickly “jump” on the imaginary axis and turn into deadbeat modes.
5. Outline of the numerical scheme
Recall that the th Chebyshev polynomial of the first kind is defined by
The cardinal functions, , and the Chebyshev-Gauss-Lobatto (CGL) grid points are defined as follows:
where coefficients are such that and for . The main property of cardinal functions is (using the Kronecker delta). The family forms a basis in the space of polynomials of degree , i.e., if is such polynomial, then and can be written in the forms
If and , then , where is the Chebyshev derivative matrix with the elements
5.1 Discretization of
where , equipped with the norm
We approximate the action of on the finite-dimensional subspace of polynomials of degree at most . Using the CGL grid and the cardinal functions, we substitute for and their truncated expansions:
Let and be -dim vectors and be a -dim vector defined by
Let be the finite-dimensional approximation of the differential operator . The discretized operator induced by can be given by
where is the identity matrix and is the derivative matrix (58).
5.2 Incorporating the boundary conditions
Discretization of the boundary conditions in the domain description (60) yields
Let be auxiliary row-vectors
is called the boundary operator. Let be the kernel of , i.e., . We have to identify all eigenvalues of the operator , when its domain is restricted to . It is clear that is isomorphic to with . Let be the matrix consisting of column vectors that form an orthonormal basis in . It is clear that is the identity matrix on and is the identity matrix on . The following result holds: if is an eigenvalue of the operator , and the corresponding eigenvector satisfies Eq. (67), then the same is an eigenvalue of the matrix . However, the inverse statement is not necessarily true. Indeed, we observe that is the identity in , which is not equivalent to the identity in . Assume now that is an eigenvalue of with corresponding eigenvector . If , we have
but which indicates that fake eigenvalues may exist.
5.3 Filtering of spurious eigenvalues
In order to decide which eigenvalues of should be discarded, we impose the following condition. Let be the spectrum of and be the set of its eigenfunctions. We construct the set of “trusted” eigenvalues [14, 15], for some filtering precision, as
where is a discrete approximation to the integral norm defined in Eq. (61). (The subscript is short for Chebyshev). Using the CGL quadrature, we obtain the following formula for the norm of a vector defined as in Eq. (63):
In this work we have considered the spectral properties of the Euler-Bernoulli beam model with special feedback-type boundary conditions. The dynamics generator of the model is a non-self-adjoint matrix differential operator acting in a Hilbert space of two-component Cauchy data. This operator has been approximated by a “discrete” operator using Chebyshev polynomial approximation. We have shown that the eigenvalues of the main operator can be approximated by the eigenvalues of its discrete counterpart with high accuracy. This means that the leading asymptotic terms in formulas (20) and (21) can be used by practitioners who need the elastic modes.
Further results deal with existence and formulas of the deadbeat modes. It has been shown that for the case when one control parameter, , is such that and the other one , the number of deadbeat modes approaches infinity. The formula for the rate at which the number of the deadbeat modes tends to infinity has been derived. It has also been established that there exists a sequence of the values of parameter , such that the corresponding deadbeat mode has a multiplicity 2, which yields the existence of the associate mode shapes for the operator . The formulas for the double deadbeat modes and asymptotics for the sequence as have been derived.
Partial support of the National Science Foundation award DMS-1810826 is highly appreciated by the first author.