AIAA JOURNAL

Vol. 52, No. 3, March 2014

Unified Approach to Flutter Equations

Edward E. Meyer∗
Boreal Racing Shells, Seattle, Washington 98119
DOI: 10.2514/1.J052554
A general formulation of the flutter equation includes gyroscopics, viscous damping, active controls, Mach and
complex reduced-frequency-based unsteady aerodynamics, quasi-linear approximated structural nonlinearities, and
parameterized matrices. Solving the flutter equation usually involves a search for neutral stability points, the
boundary between stability and instability, followed by variations in the neutral stability point with various
parameters. A technique for solving the flutter equation under these conditions must necessarily be capable of solving
systems of nonlinear parameterized equations over a range of parameter values. Continuation methods are
specifically designed for solving such problems. A continuation method is presented that is capable of solving general
formulations of the flutter equation for neutral stability, parameter variations, optimization, model tuning, and
describing function flutter analyses, including an efficient technique for assessing limit cycle stability.
Downloaded by Anadolu University on May 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.J052554

Nomenclature of flutter. In the preliminary stages of design, various combinations of

A, B, C, Dc = control system matrices design parameters are considered. As the design becomes fixed,
a = sonic velocity flutter calculations are required at various combinations of fuel,
D = dynamic matrix payload, altitude, and other parameters that define flight conditions.
d = structural damping coefficient The flutter analyst is therefore concerned with two basic types of
J = Jacobian matrix calculations: a search for the lowest (critical) flutter speed under
J∶j = column j of matrix J a given set of conditions, and the variation of the critical flutter
M, K, G = mass, stiffness, gyroscopic matrices speed with various system parameters. Flutter equations commonly
M = V∕a, Mach number encountered in modern structures contain terms that make them
ns = order of the dynamic matrix difficult or impossible to treat with traditional approaches [1,2]. A
p = s∕V, complex reducedp frequency partial list includes active controls, structural nonlinearities, Mach







dependent aerodynamics, and gyroscopics. By treating the flutter
kxk = 2-norm of vector x
xt x
q = ρV 2 ∕2, dynamic pressure equation as a system of nonlinear algebraic equations, it is possible to
q = generalized coordinate amplitudes account for all of these terms and solve both neutral-stability and
Rn , Rm×n = the real n vectors and m-by-n matrices parameter-variation problems in a unified manner. Various types of
R , I  = real and imaginary parts of a complex variable parameter variations are possible; curves and contours are considered
s = σ  iω, characteristic exponent here, and surfaces are possible but not treated here.
t = tangent vector Continuation methods are a well known class of techniques for
V = velocity (true airspeed) the solution of systems of underdetermined nonlinear equations,
V, U, T = viscous damping, unsteady aero, and control- maintaining continuity over a specified range of the parameters,
system matrices making them ideally suited to solving flutter equations. Continuation
w = projection vector methods have been used previously [3,4] to solve flutter equations.
x = independent variables Here, a more general method is presented that allows for optimization
xij = ith iteration at the jth continuation step and model tuning solutions in addition to neutral stability and
xi = ith element of vector x parameter variations.
y = residual vector
ρ = fluid density
ρz = real vector comprising the real and imaginary II. Flutter Equations
parts of z The flutter equation considered here is a system of characteristic
σ = real part of the characteristic exponent equations following the assumption that the actual motion of the
Φ = transformation from generalized to physical structure (z) as a function of time is related to a set of generalized
coordinates coordinates (q) by
ω = oscillation frequency
zt
ΦRqest  (1)
Superscript
where s
σ  iω is the complex characteristic exponent, ω is the
t = transpose
frequency, i2
−1, and the real part σ determines the rate of growth
or decay of the oscillations and hence is of primary interest to the
I. Introduction flutter analyst. Φ is a transformation from the generalized coordinates
to physical degrees of freedom, often a set of vibration modes, and R
T HROUGHOUT the design of a modern aircraft, many
calculations are performed to ensure that the aircraft will be free
denotes the real part.
With this assumption, a general form of the flutter characteristic
equations can be written as
Received 11 January 2013; revision received 29 June 2013; accepted for
publication 6 August 2013; published online 31 January 2014. Copyright © s2 M  sG  sV  1  idK − qUp; M  Tq
Dq
0 (2)
2013 by the American Institute of Aeronautics and Astronautics, Inc. All
rights reserved. Copies of this paper may be made for personal or internal use,
on condition that the copier pay the $10.00 per-copy fee to the Copyright where M, K, G, V, U, and T are the (ns -by-ns ) mass, stiffness,
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923; include gyroscopic, viscous damping, unsteady aerodynamic, and control-
the code 1533-385X/14 and $10.00 in correspondence with the CCC. system matrices, respectively; d is the structural damping coefficient;
*Consultant. q
ρV 2 ∕2 is the dynamic pressure; p
s∕V is the complex
627
 628 MEYER

reduced frequency; M
V∕a is the freestream Mach number; a is finite-difference approximations could be used, eliminating the effort
the sonic velocity; and D is the dynamic matrix. to code derivatives but introducing a possibly unacceptable approxi-
The matrices in this equation are, in general, not constant; rather, mation. A better solution, one that requires no extra coding yet gives
they are assumed to be functions of problem-dependent parameters the exact derivatives, is to write the code or transform it to use
using various matrix parameterizations. Nonlinear structural automatic differentiation ([7] p. 15). The control-system T is an
elements such as freeplay and nonlinear stiffness can be approxi- example of a matrix that is well suited to implementing in a custom
mated using describing functions ([5] p. 121), which conform to the computer code as it is often a complicated function of several
harmonic motion assumption [Eq. (1)]. parameters such as gains, phases, and time delays. The control-
Treating the flutter equation in the frequency domain has system matrix usually has more rows and columns than the mass
tremendous computational advantages over the time domain: the matrix. The extra rows and columns correspond to control-system
characteristic equation is a system of nonlinear algebraic equations degrees of freedom. T typically has the following form:
instead of a system of differential equations that must be integrated at  
each flight condition. In contrast, the algebraic equations can be T11 T12
solved over a continuous range of flight conditions. The disadvantage T
(5)
T21 T22
is that nonlinearities must be approximated to conform to the
harmonic motion assumption [Eq. (1)]. where T11 is the size of the mass matrix, T21 transforms generalized
coordinates to physical sensor motions, T22 contains control-system
equations, and T12 provides feedback to the structure from the
III. Parameterized Matrices control-system degrees of freedom. The extra control-system degrees
Various possibilities exist for creating matrices that are functions of freedom must be accounted for when forming the dynamic matrix.
If the controls equations are available in linear, time-invariant state
Downloaded by Anadolu University on May 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.J052554

of parameters, either those explicit in Eq. (2) or new problem-

dependent parameters. A few are discussed here, and others are easily variable form ([8] p. 172):
imagined. In the continuation method that follows, a requirement for
any type of matrix parameterization is the ability to compute x_ c
Axc  Bu ∈ Rnc
derivatives of the matrix with respect to its parameters. In addition,
y
Cxc  Dc u ∈ Rno
the matrix elements should be continuous with respect to the
parameters.
A simple parameterization is to set an element of the matrix to a where u is an ni vector of inputs, y is an no vector of outputs, xc is an
function. For example, nonlinear structural elements can be nc vector of state variables, and A, B, C, and Dc are real matrices,
approximated using the describing function method ([5] p. 121) by then it can be shown that
scaling a diagonal element of the stiffness matrix by a function of the
absolute value of the corresponding generalized coordinate. T11
−KEDc SΦc
Structural elements with freeplay can be modeled [6] by multiplying T12
−KEC
the diagonal of the stiffness matrix corresponding to the relevant
generalized coordinate by T21
BSΦc
1 T22
A − sI
π − 2 sin−1 u − sin2 sin−1 u if jqk j ≥ δ
cqk ; δ
π (3)
0 otherwise where Φc are the rows of the transformation matrix corresponding
to the control system inputs; and S is a diagonal matrix consisting
q



































of either 1, s, or s2, depending on whether the corresponding input
u
δ∕jqk j
δ∕ Rqk 2  Iqk 2 (4) is a displacement, velocity, or acceleration. E is a real matrix that
transforms the control system output y to generalized coordinates.
where δ is the freeplay, and qk is the generalized coordinate
associated with the nonlinear element. Note that this involves three IV. Neutral Stability
parameters: δ, and the real and imaginary parts of qk . The flutter The interest in the flutter equation is primarily in the behavior of
equations are no longer linear in the generalized coordinates, but the the real part of the characteristic exponent (s) with velocity; negative
motion is constrained by the harmonic assumption [Eq. (1)], which is values indicate decaying oscillations, positive values indicate
why this is often called a quasi-linear approximation. Similar growing oscillations, and zero indicates neutral stability. From
describing functions can approximate other types of structural Eq. (2), ns physically meaningful aeroelastic modes can be calculated
nonlinearities such as bilinear stiffness or quadratic damping. The use at each set of system parameters. Normally, the flutter analyst is
of describing functions to approximate nonstructural nonlinearities interested in only a few of these modes.
like control systems and unsteady aerodynamics needs further Continuity in these aeroelastic modes is important if the neutral-
investigation. stability point falls between two solution points and interpolation
A commonly used technique for parameterizing matrices is must be used to get the neutral-stability point. In addition, continuity
interpolation. Given a collection of matrices at various values of one is necessary in understanding the evolution of aeroelastic modes as
or more parameters numerous techniques are available for fitting each the velocity is increased.
element with, for example, cubic splines. Unsteady aerodynamic As an example of the results from such an analysis, Fig. 1 shows σ
matrices are often interpolated with respect to reduced frequency plotted against velocity for five aeroelastic modes, only one of which
(ω∕V), complex reduced frequency (p
s∕V), or complex reduced goes unstable (point a). This and all subsequent examples considered
frequency and Mach number. The mass matrix can be parameterized here are results from analyses of a simple cantilevered wing with five
by creating matrices at various mass distributions and interpolated vibration modes used for the generalized coordinate basis.
with respect to the relevant parameter, similarly for the stiffness
matrix.
A powerful way to parameterize a matrix is to use a custom V. Parameter Variations
computer code to generate or modify an existing matrix based on Once the lowest neutral-stability velocity (critical flutter speed)
various parameters. A matrix modified in this manner can be an has been found, interest usually turns to studying the variation in
almost arbitrary function of parameters. The problem with this kind critical flutter speed with various model parameters such as mass
of parameterization is computing derivatives of the matrix with and stiffness distributions, altitude, and control-system gain. Two
respect to its parameters. One solution is to write computer code to possible parameter variations are shown in Figs. 2 and 3. Figure 2
evaluate the derivatives, possibly an onerous task. Alternatively, shows the variation in critical flutter speed with a stiffness parameter
 MEYER 629

VI. Continuation Method

Continuation methods ([9] p. 13, [10,11]) are a class of techniques
for solving underdetermined systems of nonlinear equations:

fx
0 ∈ Rm
Real part of s ( σ )

(6)

a
0
x ∈ Rn (7)

where the number of equations m is fewer than the number of

independent variables n. Because there are fewer equations than
unknowns, there are an infinite number of solutions. Depending on
the difference n − m, the solution space consists of curves
(n − m
1), surfaces (n − m
2), and so on.
Continuation methods solve the system of equations so that
Velocity solutions form continuous curves, an important consideration for the
Fig. 1 Typical neutral-stability plot. solution of flutter equations where aeroelastic modes often become
close, making them difficult to keep separated.
There are a variety of continuation methods, most of which are
predictor–corrector schemes reminiscent of ordinary differential
Downloaded by Anadolu University on May 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.J052554

p1 , and Fig. 3 is a contour of constant flutter speed with p1 and a mass equation solvers. Many methods ([3], [9] p. 3) choose a particular
parameter p2. parameter to increment throughout the tracing of curves, but these
With parameter variations, it is even more important to track modes methods can fail if the curve has limit points in the continuation
in a continuous fashion. The alternative, to do a neutral-stability parameter. Locally parameterized continuation methods [11] avoid
analysis at each incremental value of the parameters would be this problem but require choosing the best parameter at each step.
prohibitively expensive. Moreover, the character of aeroelastic These methods constrain the corrector steps to be at constant values
modes can change rapidly, making it difficult to relate modes at of the continuation parameter. In contrast, pseudoarclength methods
different steps, which makes it necessary to do a neutral-stability take steps along the tangent to the curve and constrain the corrector
analysis tracking more than just the critical flutter mode at the steps to be perpendicular to the tangent, thus avoiding the problems
previous step. with continuation parameters. The method presented here is a
pseudoarclength method with minimum-norm corrector steps; the
corrector steps are computed in a way that gives the smallest step
instead of being constrained to be normal to the tangent.
Fundamental to continuation methods is the Jacobian matrix of
partial derivatives of f with respect to x:
a  
∂fi
Jx
f 0 x
∈ Rm×n (8)
∂xj

and a tangent vector t ∈ Rn with the following properties:

Velocity

Jt
0 tt t
1 (9)

b Starting from a known solution x0 , a curve is traced through a range of

the parameters. At step j  1, the solution is predicted from the
converged solution at step j and the tangent

x0j1
xj  αt (10)
p1
Fig. 2 Flutter speed as a function of p1 . where α is the step size, which must be chosen carefully. Too large a
step, and the corrector may not converge or may converge to the
wrong root; too small a step means extra work. Den Heijer and
Rheinboldt [12] studied the problem of computing step sizes, and
Rheinboldt [11] produced a continuation code using their algorithm,
taking into account the curvature and the rate and quality of corrector
convergence.
Starting with the prediction x0j1 , the corrector iterates to find the
solution xj1 using a Newton-like method computing corrections hi
satisfying

Jxi hi
−fxi  (11)

2
p

and setting
b
xi1
xi  hi (12)

Equation (11) is an underdetermined linear system that has an infinity

of solutions; the smallest such solution can be computed using a QR
p1 factorization of the transposed of Jacobian which splits it into an
Fig. 3 Constant flutter speed as a function of p1 and p2 . orthogonal matrix (Q) times an upper-triangular matrix R:
 630 MEYER

 
R these m − 2 components, three more parameters are needed to trace
Jt
QR
 Q1 Q2  1 (13)
0 curves. In addition to the three parameters, it is usually necessary to set
other parameters to constants. In the following examples, density ρ,
which when substituted into Eq. (11) yields the minimum-norm speed of sound a, and the structural damping coefficient d are held
solution ([13] p. 300) constant.
The first example is for a neutral stability analysis as in Fig. 1:
hi
−Q1 R−t i
1 fx  (14) 8 9
>
> V>
< > =
Corrector iterations continue using Eqs. (11–14) until suitable σ
x
∈ Rm1 (19)
convergence criteria are met, for example for some absolute tolerance >
> ω >
>
: ;
εa and relative tolerance εr : qr

kfk < ϵa and khk < ϵa  ϵr kxk (15) The variation in flutter speed with a parameter (p1 ) can be studied by
setting σ
0 and
The n; n − m matrix Q2 in Eq. (13), a byproduct of the QR 8 9
factorization, is an orthogonal matrix spanning the null space of the >
> V>
< > =
Jacobian, meaning ω
x
∈ Rm1 (20)
>
> p1 >>
JQ2
0 ∈ Rm×n−m : ;
qr
Qt2 Q2
I ∈ Rn−m×n−m (16)
Downloaded by Anadolu University on May 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.J052554

resulting in the parameter-variation analysis as shown in Fig. 2.

An arbitrary vector w is projected onto the null space with Another kind of parameter variation is possible with V held constant,
σ
0, and
t

Q2 Qt2 w ∈ Rn (17) 8 9
>
> ω>
< > =

tk
to satisfy Eq. (9). If n
m  1, Q2 p1
and normalized with t
t∕k x
∈ Rm1 (21)
>
> p >
consists of only one column, the projection only determines the sign : 2> ;
of the tangent, and a logical choice for the vector to project is the qr
tangent at the previous continuation step to keep the curve tracking in
the right direction. For the first step, any vector with the desired sign resulting in a constant-velocity contour, as shown in Fig. 3. Many more
will do. types of analyses are possible, provided the parameters included in x
and those held constant form a consistent set of equations.
An important consequence to including the generalized
VII. Flutter Equations as a System coordinates in the vector of independent variables is that the
of Nonlinear Real Equations continuation method is much more likely to maintain continuity
The complex flutter equation can be converted to an equivalent when tracking curves. The generalized coordinates participate in both
system of real equations of twice the size, opening up the vast amount the predictions and the corrections equal to any other parameters. A
of effort that has been put in to developing real continuation methods step-size algorithm that takes into account the curvature of the curve
([9] p. 146). Many of the parameters in the flutter equation are real, being tracked will therefore adjust the step size based on changes in
and no comparable theory has been developed for complex systems all parameters including the generalized coordinates. It may seem
dependent on real variables. The drawback to this conversion is a computationally more expensive to include the generalized
penalty in space (twice the computer memory) and time to factorize coordinates in the prediction and correction; that this is not the case
the Jacobian (again a factor of 2). can be seen by considering the operation count ([13] p. 299) for the
The flutter equation in a form suitable for solution using a most widely used solution techniques. For the traditional V-g
continuation method is a system of m
2ns real equations comprising technique, the major expense is an eigensolution of an ns ; ns 
the real and imaginary parts of the residual vector y
Dq: complex matrix [2]; for the determinant iteration [1], it is the
factorization of an ns ; ns  complex matrix; and for the continuation
fx
ρy ∈ Rm (18) method, it is the factorization of the 2ns ; 2ns  1 real Jacobian
matrix. Each of these require some multiple of n3s operations, but
where ρ is a notational convenience that creates a real vector of length because it uses the generalized coordinates, only the continuation
2k from a complex vector of length k: method can reliably maintain continuity in the solution curves.

ρz
 Rz1  Iz1  : : : Rzk  Izk  t ∈ R2k A. Jacobian Matrix

The Jacobian matrix needed for the continuation method serves
The vector of independent variables x ∈ Rn comprises the real and two purposes: iterating to find a solution at each continuation step
imaginary parts of elements of the generalized-coordinate amplitudes using Newton’s method, and computing the tangent vector for the
(q), flutter equation parameters such as V, ω, σ, and problem- prediction of the next step. It is therefore important to compute the
dependent parameters. Depending on which parameters are included Jacobian reasonably accurately both to ensure continuity and to
in x, a variety of solutions are possible. Three examples are shown here, speed convergence.
all of which have one more independent variable than equations The Jacobian is an m (
2ns )-by-n real matrix, much of which is
resulting in curves. Each of these examples and some subsequent the dynamic matrix due to the partial of the residual with respect to the
examples include a reduced set of generalized coordinates: generalized coordinates. Column j of the Jacobian matrix is
qr
 ρq1 t : : : ρqk−1 t ρqk1 t : : : ρqns t t ∈ Rm−2 J∶j
ρyj 
where qr is a real vector comprising the real and imaginary parts of the ∂y ∂D ∂q
yj

qD (22)
generalized coordinates, excluding the kth element, which is held ∂xj ∂xj ∂xj
constant for two reasons: to normalize the generalized coordinates, and
to prevent convergence to the trivial solution q
0. The index k is The second term on the right-hand side is zero unless xj is the real or
usually chosen as the largest generalized coordinate component. With imaginary part of a generalized coordinate. If it is the real part of
 MEYER 631

generalized coordinate k, then the term is the kth column of D, and if VIII. Guided Tangent
it is the imaginary part, then it is i times the kth column of D, Normally, to track curves with a continuation method, the number
where i2
−1. of unknowns must be exactly one greater than the number of
For the set of variables in the example of Eq. (19), equations; two greater yields surfaces, three yields volumes, and so
  on. Likewise, the tangent is a vector for curves, a plane for surfaces,
∂y q ∂U qs ∂U and so on for higher dimensions. By restricting the tangent to being a
y1

ρVU  − 2 q
∂V a ∂M V ∂p vector in the higher-dimensional spaces, it is possible, and useful, to
  trace a curve. The method used to compute the tangent vector
∂y q ∂U [Eq. (17)] allows for this restriction simply by choosing the projected
y2

2sM  G  V  q
∂σ V ∂p vector w.
 
∂y iq ∂U The dimension of the Jacobian null space (columns in Q2 ) is
y3

2isM  iG  V  q (23) n − m. If this difference is greater than 1, there are an infinite number
∂ω V ∂p
of tangent vectors possible instead of just one, which means that there
is more freedom to guide the tangent in particular directions by
choosing the vector, w in Eq. (17), that is projected onto the null
space. A few examples will illustrate the technique.
B. Start Points
Finding start points for parameter-variation analyses is straight- A. Optimization
forward; they are taken from previous neutral stability or parameter- If the projection vector is one of the n coordinate directions, that is
variation analyses. For example, the start point for the curve in Fig. 2 if it is all zeros except for one of the elements, say the kth element,
Downloaded by Anadolu University on May 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.J052554

is the neutral-stability point in Fig. 1 (point a). It may be necessary to then that component of the resulting tangent will be maximal. That is,
interpolate results from the previous analysis to obtain the precise moving along this tangent will have the greatest change in xk of all
start point, for example if the neutral-stability point (σ
0) lies possible tangent vectors, and at each continuation step the curve will
between two solution points. head toward maximally increasing or decreasing xk , depending on
Start points for parameter-variation analyses can also be obtained the sign of the projection vector.
from previous parameter-variation analyses. For example, a start Returning to the example of Eq. (21), to optimize flutter speed with
point for the flutter speed contour of Fig. 3 could be obtained from the respect to the two parameters p1 and p2 , let
analysis of Fig. 2 by interpolating to get the desired flutter speed 8 9
(point b). >
> V> >
Start points for neutral-stability analyses are more problematic. >
> > >
>
<ω> =
Generally, the task is to solve Eq. (2) at zero (or small) velocity, giving x
p1 ∈ Rm2 (27)
rise to the following nonlinear eigenvalue problem: >
> >
>
> p2 >>
>
>
: qr >;
Dsq
0 (24)
and if the projection vector is all zeros but for 1.0 in the element
Solutions to Eq. (24) are eigenpairs si ; qi  that can be used to form corresponding to velocity:
start points using Eq. (19). A straightforward technique due to Ruhe
[14] solves Eq. (24) with a series of linear, generalized eigenvalue w
 1 0 : : : 0 t ∈ Rm2 (28)
problems starting with an initial guess s0 for the first eigenvalue.
1) Solve the complex generalized eigenvalue problem Equation (17) will drive the continuation curve toward maximally
increasing or decreasing velocity (depending on the sign); thus, it is a
∂D j method for optimizing the model flutter speed with respect to model
Dsj q
−λ s q parameters. Only two parameters p1 and p2 were used in this
∂s
example, but any number greater than one could be used.
for the smallest (absolute value) λ.
2) Set sj1
sj  λ. B. Contour Start Points
3) If λ > ϵjsj1 j, then repeat. Figure 3 is a constant-velocity contour of two parameters p1 and
Then choose the second smallest λ, set s0
sj1  λ, and repeat p2 . More such contours may be traced, giving an outline of the cone-
the process for the next eigenvalue. This algorithm is based on a shaped p1 − p2 − V surface shown in Fig. 4. A problem arises when
Taylor series expansion of Eq. (24), keeping only two terms. If searching for start points for these contours in that, if a parameter
instead three terms are retained,

∂Ds λ2 ∂2 Ds
Ds  λ ≈ Ds  λ   ::: (25)
∂s 2 ∂s2

the eigenvalue problem becomes a polynomial eigenvalue problem

 
∂D j 1 ∂2 Dsj 
Dsj   λ s   λ2 q
0
p2

(26)
∂s 2 ∂s2

that takes more effort to solve but should result in faster convergence.
In fact, in the absence of control-systems, gyroscopics, and viscous
damping, at zero velocity and s0
0, Eq. (26) becomes the usual
free-vibration eigenvalue problem λ2 M  Kq
0, and all
eigenpairs are valid start points with no iteration required. Note
also that s0 should be small but not zero with the first algorithm p1
because ∂D∕∂s0 is zero in this case. Fig. 4 Constant velocity contours.
 632 MEYER

variation as in Fig. 2 is used to find start points, the velocities may not oscillation (LCO) because the oscillations, although still not growing
extend the full range desired for the contours. What is needed is a or decaying, are at a finite amplitude. More important is what happens
curve that traverses the surface toward increasing (or decreasing) to the limit cycle if the amplitude is changed slightly; if the limit cycle
contour values: in this case, velocity. That is exactly what the is stable, then the amplitude will return to the limit cycle amplitude,
optimization technique does using a set of independent variables as in but if it is an unstable limit cycle, then the amplitude will increase or
Eq. (27) and a projection vector as in Eq. (28). Figure 5 shows this decrease to the closest stable limit cycle. Determining whether a limit
optimization curve with the contours and the p1 variation curve (from cycle is stable or unstable requires the partial derivative of σ with
Fig. 2). Clearly, the p1 variation curve is capable of providing start respect to the nonlinear amplitude [6], information that is available
points for only three contours; the guided tangent curve can provide from the Jacobian.
start points for them all. Approximating nonlinearities with a quasi-linear technique like
describing functions complicates finding the lowest neutral-stability
C. Tuning point by the fact that the generalized coordinates cannot be
The guided tangent technique can also be used to tune model normalized arbitrarily because the flutter equation is no longer linear
parameters to force an aeroelastic mode to match certain criteria, for in the generalized coordinates. As in Eq. (19), one generalized
example to match test results. If x^ is a set of parameter values, a subset coordinate could be left out of x and the critical flutter speed found
of the problem parameters (x), by adjusting the rest of the problem with this one generalized coordinate amplitude held constant.
parameters it may be possible to match the subset of parameter Beyond this, it is also desirable to study the variation in flutter speed
values. The idea is to form a projection vector that is the difference with nonlinear amplitude, perhaps with the goal of finding the
between the current parameter values and the target values. For combination of nonlinear amplitudes that gives the lowest flutter
example, to match ω ^ and p^ 2 in Eq. (27), the following projection speed. This can be done by replacing σ with the real part of the
Downloaded by Anadolu University on May 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.J052554

vector, used to guide the tangent with Eq. (17), will tend toward the excluded generalized coordinate; the imaginary part remains
target parameter values: excluded and set to a constant value:
8 9 8 9 8 9
>
> 0 > > >
> 0 > > >
>V > >
>
> >
> >
> >
> >
< >
=
>
> ω^ >
> >
> ω >
> ω
>
> >
> >
> >
> x
∈ Rm1 (30)
>
< 0 > = > < 0 > = > Rqk  >
>
> >
w
p
^ 2 − p 2 ∈ Rm2 (29) : qr > ;
>
> 0 > > >
> 0 > >
>
> > > >
>
> . >>
> >
> . >
> .. >>
> .. >
>
> > > > > >
: ; > > : > ; Tracing a curve with these variables gives the variation in flutter speed
0 0 and frequency with generalized coordinate amplitudes. Figure 6
shows a typical variation in velocity with nonlinear amplitude. At one
Of course, there may be no set of parameter values that will cause the particular speed, arrows show a stable limit cycle on the upper branch
mode to match the target, but this procedure will bring the values as and an unstable limit cycle on the lower branch. At this speed,
close as possible, and intuitively, including more parameters in x perturbations of amplitude less than a decay to zero, but with
would make for a closer match. amplitudes greater than a, the oscillations will grow to b and remain
at this stable LCO.
At each solution point, it is possible to determine if the curve
IX. Nonlinear Flutter represents a stable or unstable limit cycle from the tangent vector of a
modified Jacobian. Partial derivatives of variables on a curve can be
Nonlinearities can be approximated using describing functions
obtained from the tangent vector simply by dividing the appropriate
that replace matrix elements with terms that are in some sense
elements; thus, for example, if V were replaced in Eq. (30) with σ and
equivalent for a given amplitude of oscillation, resulting in flutter
the nonlinear amplitude is ζ
Rqk , then
equations that are nonlinear in the generalized coordinate amplitudes.
Because the flutter equations are already nonlinear in other
∂σ t
parameters, this does not change the solution technique.
1 (31)
A basic assumption made with linear flutter equations is that the ∂ζ t2k1
motion is infinitesimally small. A neutral stability condition found
with linear flutter equations means that, given an infinitesimally If this quantity is positive, then the limit cycle is unstable; otherwise,
small perturbation, the structure will oscillate with sinusoidal motion it is stable. All that is needed is the tangent corresponding to the
that neither decays nor grows. In contrast, a neutral stability condition modified set of variables, and that can be obtained by doing a rank-
found with nonlinear flutter equations represents a limit cycle one update to the Jacobian:

b
stable LCO
Amplitude
p2

p1 variation
unstable LCO

a
optimization curve

p1 Velocity
Fig. 5 Contours with guided tangent. Fig. 6 Flutter speed variation with nonlinear amplitude.
 MEYER 633

J^
J  uvt (32) effective way of including most terms encountered in modern flutter
analyses and (almost) guarantees continuous curves. A predictor–
where corrector continuation method using a QR factorization allows for
more unknowns than equations and can be used to guide curves toward
 
∂y optimal or target sets of model parameters.
u
ρ − J∶1 ∈ Rm Several simple examples illustrate how these techniques might be
∂σ
v
 1 0 : : : 0 t ∈ Rn (33) used as well as the tremendous flexibility gained by this formulation
of the flutter equation. Although the techniques were illustrated by a
effectively replacing column one of the Jacobian with the partial of very small model, they have been used routinely on models
the residual with respect to σ. It is not necessary to actually make this approaching 300 degrees of freedom.
replacement; only the vectors u and v are of interest because, with
them, the QR factorization of the Jacobian can be updated ([13]
p. 334), and from that the desired tangent for Eq. (31) is Q2 [Eq. (13)]. References
Thus, the determination of limit cycle stability requires little
[1] Hassig, H. J., “An Approximate True Damping Solution of the Flutter
additional effort. Equation by Determinant Iteration,” Journal of Aircraft, Vol. 8, No. 11,
1971, pp. 885–889.
X. Speeding-Up Curve Tracking doi:10.2514/3.44311
[2] Rodden, W. P., “Handbook for Aeroelastic Analysis,” MSC/NASTRAN
An advantage in tracking each aeroelastic mode separately as the Ver. 65, Santa Ana, CA, Nov. 1987.
continuation method does is that it is almost trivial to parallelize the [3] Cardani, C., and Mantegazza, P., “Continuation and Direct Solution of
computation because the tracking of each mode is entirely inde-
Downloaded by Anadolu University on May 10, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.J052554

the Flutter Equation,” Computers and Structures, Vol. 8, No. 2, 1978,

pendent of the tracking of other modes, and so they may be done in pp. 185–192.
parallel. Modern multicore, multi-CPU, and GPU architectures offer doi:10.1016/0045-7949(78)90021-4
great potential for speeding up flutter analyses using continuation [4] Meyer, E. E., “Application of a New Continuation Method to Flutter
methods. Equations,” 29th AIAA Structures, Structural Dynamics, and Materials
Conference, AIAA Paper 1988-2350, April 1988.
Much of the computational expense in the continuation method is
[5] Šiljak, D. D., Nonlinear Systems, Wiley, New York, 1969.
in factorizing the Jacobian. Two things can be done to mitigate this [6] Gordon, J. T., Meyer, E. E., and Minogue, R. L., “Nonlinear Stability
expense: a modified Newton-type corrector, and updating the Analysis of Control Surface Flutter with Freeplay Effects,” Journal of
Jacobian factorization. Modified Newton’s method simply uses a Aircraft, Vol. 45, No. 6, 2008, pp. 1904–1916.
Jacobian for multiple corrector iterations; more important is the doi:10.2514/1.31901
reduction in factorizations. Convergence for a modified Newton’s [7] Griewank, A., and Walther, A., Evaluating Derivatives: Principles and
method is slower, but whether that is compensated by the Techniques of Automatic Differentiation, Vol. 105, Frontiers in Applied
factorization savings is problem-dependent. Mathematics, Society for Industrial and Applied Mathematics,
Updating the Jacobian factorization is an improvement on the Philadelphia, 2008.
[8] Brogan, W. L., Modern Control Theory, 3rd ed., Prentice–Hall, Upper
modified Newton’s method. The Jacobian is evaluated and factorized
Saddle River, NJ, 1991.
at the predicted point and used for the first corrector iteration. For [9] Allgower, E. L., and Georg, K., Numerical Continuation Methods,
subsequent corrector iterations, the QR factors are updated using the Springer–Verlag, New York, 1990, pp. 3, 13.
changes in x and f. Broyden’s update formula ([9] p. 61) is a rank-one [10] Rheinboldt, W. C., “Numerical Analysis of Continuation Methods for
update to the Jacobian: Nonlinear Structural Problems,” Computers and Structures, Vol. 13,
Nos. 1–3, 1981, pp. 103–113.
Jxi1  ≈ Jxi   uvt (34) doi:10.1016/0045-7949(81)90114-0
[11] Rheinboldt, W. C., “A Locally Parameterized Continuation Process,”
ACM Transactions on Mathematical Software, Vol. 9, No. 2, 1983,
where
pp. 215–235.
doi:10.1145/357456.357460
u
fxi1 ∕khi k ∈ Rm [12] Den Heijer, C., and Rheinboldt, W. C., “On Steplength Algorithms for a
Class of Continuation Methods,” SIAM Journal on Numerical Analysis,
v
hi ∕khi k ∈ Rn (35)
Vol. 18, No. 5, 1981, pp. 925–947.
doi:10.1137/0718066
Given u and v, the QR factorization of Jt xi  can be updated ([13] [13] Golub, G. H., and Van Loan, C. F., Matrix Computations, 4th ed., Johns
p. 334) and used for corrector step i  1 in Eq. (14). Hopkins Univ. Press, Baltimore, MD, 2013.
[14] Ruhe, A., “Algorithms for the Nonlinear Eigenvalue Problem,” SIAM
Journal on Numerical Analysis, Vol. 10, No. 4, 1973, pp. 674–689.
XI. Conclusions doi:10.1137/0710059
Formulating the flutter equation as a system of real nonlinear
equations in which all variables, including the generalized coordinates, W. Silva
are treated equally and solving with a continuation method is a very Associate Editor