This page explains, step by step, how to build a tracking controller for a mass-spring-damper system whose stiffness varies with a parameter. It goes from the physical model to the H∞ synthesis problem, with two interactive modules to build intuition about signal energy and system gain, then uses the method of [1] to derive sufficient matrix inequalities for this scalar parameter.
The page separates several levels of statement, which should not be confused:
Problem formulation
Sections 1 to 3 define the model, the LFT, the generalized plant and the weights.
Sufficient condition
Section 5 derives LMI conditions that suffice to certify the gain for the scalar parameter model, following the method of [1].
Numerical feasibility
The controller has a computed scaled gain of 0.8449. A second solver produced a storage matrix for γ=0.85, verified against the discrete-time dissipation inequality.
Stated guarantee
The certificate bounds the induced gain and establishes stability of the unsaturated discrete model for every parameter sequence satisfying ∣ρ∣≤1/γ≈1.17647.
Simulation
The simulation is an illustration in discrete time. A run does not prove stability or performance.
1Physical model & LFT
The system under study is a classical Mass-Spring-Damper (MSD) mechanical system. It consists of a mass m sliding on a frictionless horizontal surface, attached to a fixed wall via a spring with stiffness k(ρ(t)) and a damper with coefficient c. The system is subject to a control input u(t) and the goal of this wiki is to explain how to construct a tracking controller for this system using H∞ LPV-LFT control.
Figure 1: Schematic of the LPV Mass-Spring-Damper system.The force u(t) is applied to the mass and its position y(t) is measured from rest, both counted positive to the right.
The equation of motion is derived from Newton's Second Law:
my¨(t)+cy˙(t)+k(ρ(t))y(t)=u(t)
The stiffness k(ρ(t)) is parameter-dependent: k(ρ(t))=k0(1+αρ(t)). Physical parameters are presented in Table 1.
Table 1: Physical parameters of the LPV Mass-Spring-Damper system.
Parameter
Value
Unit
m (mass)
1
kg
c (damping)
2
N·s/m
k0 (nominal stiffness)
10
N/m
α (stiffness gain)
0.4
—
The system written just above is a Linear Parameter Varying (LPV) system: its dynamics depend on the time-varying parameter ρ(t), which sets the stiffness k(ρ(t)). To apply robust control tools, we must isolate the parameter variation from the nominal dynamics. This is done using a Linear Fractional Transformation (LFT), which creates a fictitious feedback loop with the parameter ρ. Here is the derivation of the LFT:
2Generalized LPV-LFT Model for synthesis
Now that we have the LPV-LFT system, we can proceed to the controller synthesis. Figure 2 illustrates a classical mixed-sensitivity synthesis scheme, a standard architecture in robust control design. It consists of three main components:
Plant P: Represents the Mass-Spring-Damper (MSD) system. It encapsulates both the nominal dynamics and the LFT decomposition of the varying stiffness.
Controller K: The controller we aim to synthesize. It also reads the parameter ρ(t), through an LFT loop of its own (Figure 2). Its internal architecture—derived to handle ρ(t)—will be detailed in Section 5.
Weighting Filters (We,Wu): Strictly design tools grafted onto the physical signals (the error e and the control signal u). They translate performance specifications into mathematical constraints.
Figure 2: The Synthesis Scheme.The same parameter ρ(t) is looped on the plant and on the controller.
By combining the LPV system model (Section 1) and the weighting filters presented in Figure 2, we construct the Generalized LPV Model. This is a single augmented state-space system that the optimizer 'sees'.
First, we need to define the two state models induced by the filters We and Wu:
The augmented state vector ξ concatenates the system states and the filter states:
ξ(t)=[x⊤(t)xe⊤(t)xu⊤(t)]⊤
To construct the augmented matrices, we must express the dynamics of the filters in terms of the global system inputs (wρ,r,u) and the augmented state vector ξ. Based on the synthesis scheme (Figure 2), the error signal is defined as e(t)=r(t)−y(t)=r(t)−Cyx(t).
Augmented State Dynamics We substitute the error equation into the state equation of the performance filter We. The control filter Wu depends only on the control input u. By assembling these with the nominal system dynamics, we obtain:
Completing the System Representation The state equation derived above governs how the system evolves internally. To fully characterize the Generalized Plant, we must now define how it interacts with its environment. This requires specifying the output equations for three distinct loops: the parameter variation loop, the performance evaluation loop (outputs of We and Wu), and the feedback loop to the controller. Table 2 provides a structured summary of these essential channels.
Table 2: Signal definitions for the Standard LPV Interconnection.
Channel
Description
Parametric (zρ→wρ)
Interaction with the varying parameter ρ(t). The signal zρ 'senses' the state affected by the parameter, and wρ reinjects the variation effect.
Performance (r→w)
Maps exogenous inputs (Reference r) to regulated outputs (Weighted Error e′, Weighted Control u′). This channel defines what we want to optimize.
Control (u→e)
Physical interface. u is the force applied by the actuator. The sensor measures the position y, and the controller receives the tracking error e=r−y.
Parametric Loop (zρ→wρ): The LFT feedback signal zρ depends exclusively on the physical system states x. However, to maintain a unified state-space representation for the generalized plant, we must express this output in terms of the augmented state vector ξ. This is done by zero-padding the matrix to account for the filter states:
Regulated Outputs w: The performance vector w consists of the weighted error e′ and the weighted control effort u′. We substitute the physical error expression into the output equation of We:
Measurement Output e: Finally, the controller observes the tracking error e. This observation equation acts as the feedback interface (we already have written it before):
e(t)=r(t)−Cyx(t)=Ce[−Cy00]ξ(t)+r(t)
By stacking these equations into matrix form, we identify the specific blocks of the generalized plant.
The full interconnection is described by the partitioned matrix:
We can now focus on the explicit definition of the filters We and Wu.
3Weighting Filters Design
The synthesis process relies on properly shaping the closed-loop response. Before solving the optimization problem, we must define the performance requirements (We and Wu) and formally state the input-output map (The Operator).
As we can see on Figure 2, we do not minimize the physical error directly. Instead, we use weighting filtersWe and Wu to shape the signals. These filters act as frequency-dependent penalties. The choice of these transfer functions dictates the trade-off between performance and robustness. For this example, we select:
We(s)=s+0.060.5s+3,Wu(s)=s+10000.2s+10
Justification of We (Performance): This filter shapes the sensitivity function S (the transfer from reference r to error e). The fundamental constraint is ∣S(jω)∣<1/∣We(jω)∣ and is represented in Figure 3. To ensure high precision, we focus on the behavior at zero frequency(s→0).
Frequency Zero implies Steady-State: A frequency of zero (s=0) corresponds to a signal that does not change over time (a constant). Therefore, the filter's gain at s=0 dictates the system's behavior in the steady-state regime (as t→∞).
Static Gain Calculation: Evaluating the filter at s=0 yields: We(0)=0+0.060.5(0)+3=0.063=50. This means the filter demands an amplification of 50 for constant signals.
Precision Guarantee (2% Error): Since the constraint forces the error sensitivity ∣S∣ to be less than the inverse of the weight, we have: ∣S(0)∣<We(0)1=501=0.02. So, for a constant reference and a frozen parameter, the steady-state error will never exceed 2%, as soon as the synthesis reaches the objective ∥WeS∥∞<1. For a time-varying ρ(t), this frequency-by-frequency reading is only an approximation: what the synthesis really bounds is the L2 gain of the whole weighted operator (Section 4).
Justification of Wu (Control Effort & Robustness): This filter shapes the Control Sensitivity function KS (the transfer from reference r to control input u). The constraint is ∣KS(jω)∣<1/∣Wu(jω)∣ and is represented in Figure 3. We analyze the inverse of the weight (the ceiling) at the frequency extremes:
Low Frequencies (s→0): Permissiveness. At steady-state, Wu(0)=100010=0.01⇒Wu(0)1=100. This sets a high ceiling (100) for the control action. It allows the controller to generate large static forces (up to 100 times the reference amplitude) if necessary to compress the spring and eliminate the static error. We do not want to restrict the controller's authority here.
High Frequencies (s→∞): Restriction (Roll-off). At high frequencies, Wu(s)∼s0.2s=0.2 as s→∞, hence Wu(∞)1=5. This sets a low ceiling (5), twenty times lower than the first one. At these frequencies the loop no longer has any effect and S≈1, so KS≈K: a high-frequency ceiling of 5 (14 dB) limits the control action produced by sensor noise.
Between the two: why the ceiling only falls above 50 rad/s. The weight keeps its low value up to its zero, at 50 rad/s, and reaches its high-frequency value beyond its pole, at 1000 rad/s. This leaves room to the controller where the mass needs it. Tracking is asked up to about 3.5 rad/s, the frequency at which ∣We(jω)∣=1. There the loop still follows the reference, so that ∣T∣=∣PKS∣ stays close to 1 and ∣KS∣=∣T∣/∣P∣ is close to 1/∣P(jω)∣=∣k0−mω2+jcω∣. This quantity grows like mω2: moving a mass faster takes a force that grows with the square of the frequency, and the loop only lets go of the reference about a decade further. So ∣KS∣ must first rise before it can fall. For the controller of this page, with the parameter frozen at ρ=0, it reaches 36 at 10 rad/s and about 48 near 20 rad/s, then falls to 0.34 at 1000 rad/s. With the earlier filter (2s+0.2)/(s+10), the ceiling for ∣KS∣ is about 0.71 at 10 rad/s and tends to 0.5 at high frequencies. For the earlier weights and ρ=0, the smallest possible weighted gain at 1.23 rad/s already exceeds 1.63 in the continuous model, regardless of the controller. This rules out γ<1 even at the nominal stiffness.
Figure 3: Frequency response of the weighting filters.
4L2-norm, H∞-norm and H∞-synthesis
In the first three sections, all the objects related to the design of an H∞ controller have been defined and motivated. But to be honest, it is perfectly normal for you to wonder whatH∞actually is and what its purpose is. This section is therefore dedicated to this topic.
The H∞-norm is a norm based on signal energy (L2-norm). Let's firstly define this L2-norm.
The L2-norm of a signal u(t) can be infinite, e.g. u(t)=sin(t) or u(t)=cos(t) have infinite energy. A necessary condition for a uniformly continuous signal to have a finite L2-norm is that its value tends to zero as t goes to infinity. This is not a sufficient condition, e.g. u(t)=t+11 has an infinite L2-norm but tends to zero (but not fast enough) as t goes to infinity.
The purpose of this section is not to delve into the theoretical depths of the L2-norm. However, the module below is available for those who wish to gain a practical sense of this metric. Use it to experiment with different signals and see firsthand which ones possess finite energy. It displays the norm ∥u∥L2, that is the square root of the integral above, and draws the integral itself as a reference line.
Signal L2 Norm Explorer
Enter a function of time 't' (e.g., sin(t)*exp(-0.5*t)) to visualize its energy. The blue area represents the squared magnitude.
t (s)
u(t)
∣u(t)∣2
∫0t∣u(τ)∣2dτ (accumulated energy)
The plot above is computed directly. To get the exact and numeric value of the norm, SymPy and SciPy run in your browser: this downloads about 90 MB from cdn.jsdelivr.net, once.
Find the Worst-Case Input
Below is a stable linear system (unknown to you). Try to find an input signal u(t)=sin(ωt) that maximizes the gain (output norm / input norm). For a sine input, in steady state, this ratio is exactly ∣H(jω)∣, which is what the module plots. If you maximize this ratio, you have found the H∞ norm of the system.
Input u(t) and output y(t)
Time (s)
Input u(t)
Output y(t)
Gain ∥y∥/∥u∥=∣H(jω)∣ at the frequencies explored
ω (rad/s)
Gain
u(t)=sin(ωt)
Current gain ratio
7.0711
Your max found
7.0711
To mathematically formulate the problem, we define the Closed-Loop Operator Tzw. This operator maps the exogenous input r to the regulated performance outputs w:
w=Tzw(P,K,ρ)r
Figure 4: The closed-loop operator Tzw: the whole dashed box, seen as a single system.The same measured parameter ρ(t) is looped on the plant at the top and on the controller at the bottom, as in [1].
Figure 4 shows what is inside this operator. Once the plant P, the controller K and the parameter ρ are connected, the signals zρ,wρ,e,u,zθ and wθ stay inside the loop. Seen from outside, only the reference r enters and only the weighted signals w leave. That is all Tzw is: the map between the two.
Its gain compares the size of what leaves with the size of what enters. Keeping this gain below γ means that, for every non-zero r of finite energy, from zero initial conditions and for every admissible ρ(t):
∥w∥L2<γ∥r∥L2
Since w gathers the weighted error e′ and the weighted control u′, its energy is the sum of theirs, ∥w∥L22=∥e′∥L22+∥u′∥L22. A small gain therefore means a small tracking error and a restrained control effort, in the frequency ranges emphasized by We and Wu.
This operator depends on the Generalized Plant P, the Controller K and the parameter ρ. The H∞ synthesis problem is formulated as follows:
Let's read this line slowly, because γ appears twice, and that is no accident. On the right, γ is the performance level: the smaller it is, the better. Under the sup,1/γ is the range of ρ covered by the guarantee. Both are linked by the small-gain idea: if what surrounds ρ has a gain below γ and if ∣ρ∣≤1/γ, a signal going around the loop through ρ comes back smaller than it left, so the loop stays stable. Lowering γ therefore improves the performance and widens the range of ρ at the same time. Here the unsaturated discrete model is certified at γ=0.85, so ∣ρ∣≤1/γ≈1.17647 covers the design range ∣ρ∣≤1.
How such a controller is found, and why the result can be trusted, is the subject of Section 5.
5Gain-scheduled synthesis: the method of Apkarian and Gahinet, with its proofs
5.1The plan
Section 4 ended on a problem statement: find a controller K(ρ) and the smallest γ such that the gain of Tzw stays below γ for every ρ(t) such that ∣ρ(t)∣≤1/γ. Stated like this, the problem cannot be handed to a solver: the unknown is a whole dynamical system, and the condition must hold for infinitely many functions ρ(t). Following the method of Apkarian and Gahinet [1], we derive sufficient matrix inequalities for the scalar parameter in this example. The derivation uses completion of squares and changes of coordinates to separate the controller from the certificate.
Step 1. Take ρ out: the closed loop is one constant system, looped on ρ.
Step 2. State the small-gain theorem with a change of units called a scaling: a single inequality, written (D) below, guarantees stability and performance for every admissible ρ(t).
Step 3. Turn (D) into a matrix inequality: this is the bounded real lemma.
Step 4. Eliminate the controller from this inequality: this is the projection lemma.
Step 5. Read what remains: the Linear Matrix Inequalities for this scalar parameter, which are convex.
Step 6. Rebuild the controller from their solution.
5.2Step 1: one constant system, looped on the parameter
In Figure 4, ρ(t) appears twice: on the plant, through wρ=ρzρ, and on the controller, through wθ=ρzθ. Stack the two channels into zΔ=(zθ,zρ) and wΔ=(wθ,wρ). The two loops then read as a single one:
wΔ(t)=ρ(t)zΔ(t)
Everything else inside the dashed box, that is P and K connected through e and u, contains no ρ: it is a linear time-invariant system, which we call M. Its state gathers the states of the plant and of the controller, xcl=(ξ,xK). Its inputs are ν=(wΔ,r) and its outputs η=(zΔ,w):
Two facts about M matter for what follows. First, the whole difficulty of a time-varying system is now concentrated in one multiplication by a number, ρ(t), of which we only know that ∣ρ(t)∣≤1/γ. Second, M depends on the unknown controller in the simplest possible way. Gather the nine matrices of the controller into a single one, Ω. Substituting the equations of K into those of P gives:
The matrices with the index 0 or Ω contain only the matrices of the generalized plant of Section 2, zeros and identities. The controller appears once and to the first power: the closed loop is affine in Ω. This can be seen on one line of the computation: the plant state obeys ξ˙=(A+BuDK11Ce)ξ+BuCK1xK+…, where each controller matrix is multiplied by plant matrices, never by another controller matrix.
Let us write these matrices down, because Step 4 will use them. The controller reads three things, stacked into one vector: its own state xK, the error e and its parameter channel wθ. It produces three things: the derivative of its state, the control u and zθ. In one line:
x˙Kuzθ=ΩxKewθ
What the controller reads is a linear function of the state xcl=(ξ,xK) and of the inputs ν=(wθ,wρ,r) of M, because the error is e=Ceξ+r (Section 2):
What it produces enters M in two places. In the state equation, the derivative of its state is the lower half of x˙cl, and the control acts on the plant through Bu. In the output equation, zθ is the first block of η=(zθ,zρ,w), and the control reaches w through Dwu:
BΩ=[0IBu000],DΩ=00000Dwu100
BΩ=[0IBu000]DΩ=00000Dwu100
The rest does not go through the controller. It is the plant alone, as if the controller produced nothing:
Check one block to see how the pieces fit: the upper-left block of A0+BΩΩCΩ is A+BuDK11Ce, the matrix that multiplies ξ in the line above.
5.3Step 2: the small-gain theorem and its scaling
The idea first. Consider the loop made of two elements: M, seen from wΔ to zΔ, and the multiplication by ρ. If M never amplifies a signal by more than γ and ρ never by more than 1/γ, a signal that goes once around the loop comes back no larger than it left. If one of the two bounds is strict, it comes back strictly smaller, and nothing is left after many turns: the loop is stable. This is the small-gain theorem. Its strength is that it asks nothing of ρ(t) but its size: the parameter may vary as fast as it likes. Figure 5 lets you try it.
Amplification by M
0.90 γ
Amplification by ρ
0.90 /γ
Product over one turn
0.81
At each turn the signal comes back smaller: it dies out, and the loop is stable.
Figure 5: A signal going around the loop.Each bar is the largest size the signal can have after one more turn, its size at the start being 1. One turn multiplies it by at most the amplification of M times that of ρ. The first is counted in multiples of γ and the second in multiples of 1/γ: only their product matters.
Its weakness is its pessimism, which a scaling reduces. The size of zΔ depends on the units in which its components are measured and on the way they are combined. Let us measure zΔ and wΔ through the same symmetric positive definite matrix L: the size of zΔ becomes zΔ⊤LzΔ instead of zΔ⊤zΔ. Since ρ(t) multiplies every component by the same number, it shrinks the new size exactly as it shrank the old one. For ∣ρ∣≤1/γ:
wΔ⊤LwΔ=ρ2zΔ⊤LzΔ≤γ21zΔ⊤LzΔ(S)
wΔ⊤LwΔ=ρ2zΔ⊤LzΔ≤γ21zΔ⊤LzΔ(S)
Figure 6: The scaling in a plane.The large ellipse gathers the vectors zΔ of size 1 when the size is measured through L. Multiplying by ρ shrinks every direction by the same factor: the image of the ellipse is an ellipse of the same shape, ∣ρ∣ times smaller, and it carries wΔ=ρzΔ. This is what (S) says.
The gain of M, on the other hand, changes with L. We are free to pick the L for which M looks smallest: one more unknown, and a far less pessimistic test. With several parameters, L must keep the property used above, that of passing through the parameter block, which restricts its shape [1]. With the single parameter of this page, every symmetric positive definite L qualifies.
An energy that can only go down: the idea of Lyapunov. The small-gain idea of Figure 5 follows a signal around the loop. To turn it into a proof, the method calls on a second idea, older and more general, due to Aleksandr Lyapunov (1892): to show that a motion dies out, there is no need to compute it. It is enough to find a quantity, attached to the state, that measures how far the system is from rest, and to check that it can only decrease.
Think of a marble in a bowl (Figure 7). As it rolls up the opposite side, its height rises again. Its kinetic and potential energies exchange, while friction dissipates their sum. In Figure 7 the height represents the stored energy of the system, so it decreases along the trajectory. A Lyapunov function is the same object for a dynamical system: a function V of the state with two properties.
It is positive away from rest.V(0)=0, and V(x)>0 for every other state x: rest is the bottom of the bowl.
It decreases along every motion.dV/dt<0 as long as the state is not zero: the marble only goes down.
Then the state never leaves the region enclosed by the level curve of V on which it starts, and if V decreases fast enough, the state tends to rest; part (a) of the proof of Result 1, below, writes this argument out in full. For a mechanical system, the natural candidate is the energy, hence the name used on this page, stored energy. For the linear systems of this page, the candidates are the quadratic forms V(x)=x⊤Xx, with X symmetric positive definite: their level curves are ellipses, and their graph is a bowl.
Figure 7: A Lyapunov function as a bowl.The surface is the graph of the energy V of the mass alone, built in the example below (Figure 9), above the plane of its states (y,v/ω0). In blue, a motion of the mass while ρ(t) oscillates over the whole range ∣ρ∣≤1.17; in copper, the same motion lifted to the height V: it can only go down.
Why such a detour? For a frozen value of ρ, the mass is a linear system with constant coefficients, and the eigenvalues of its matrix decide stability: here, the mass is stable for every frozen ρ>−2.5, where the stiffness is positive. But eigenvalues only describe a frozen system. When ρ(t) varies, a family of systems that are all stable can produce a motion that grows, if one switches between them at the right moments. Figure 8 shows it on the mass alone. With ρ frozen at −2 or at 6, the mass comes back to rest. Now take the soft spring, ρ=−2, while the mass moves away from rest, and the stiff one, ρ=6, while it comes back. Each switch to the stiff spring happens at the end of a swing, far from rest, where it multiplies the energy held by the spring; each switch back happens as the mass passes through rest, where the spring holds none. The damper cannot keep up: every swing is wider than the previous one, as when a child pumps a swing.
Figure 8: Stable at every frozen value, unstable when switched.The mass alone, from the same start. With ρ frozen at −2 or at 6, it comes back to rest. With ρ switched between these two values at the instants shown below, every swing is about 40% wider than the previous one.
No eigenvalue sees this. A Lyapunov function does, provided it does not depend on ρ: its derivative at an instant involves only the current value of ρ, so a decrease checked at each value of a range holds for every way of moving through that range, however fast. The energy of the next example does this for ∣ρ∣<1.5; the switching of Figure 8 leaves that range.
Where does V come from? For the mass alone, physics suggests it, and one correction is enough, as the example shows. For the closed loop, nobody can guess it: the matrix X becomes an unknown, which the synthesis looks for together with the controller (Steps 3 to 6). Once it is found, it acts as a certificate: the guarantee follows from a few lines of algebra, without simulating anything.
See it first on the mass alone. The statement below rests on one idea: an energy stored in the system, which can only go down. Before stating it for the closed loop, which has eight states, let us look at it where it can be drawn: the mass of Section 1 without control, u=0. Two states are enough, the position y and the velocity v=y˙, and the motion obeys:
y˙mv˙=v=−cv−ky,k=k0(1+αρ(t))
y˙mv˙k=v=−cv−ky=k0(1+αρ(t))
A natural candidate is the mechanical energy at the nominal stiffness, V0=21k0y2+21mv2. Differentiate it along the motion and replace y˙ and v˙ by the two equations:
The first term is the power lost in the damper: it never has the wrong sign. But it vanishes whenever v=0, and the second term has no sign at all as soon as k≠k0: when the stiffness differs from its nominal value, this energy can grow. V0 is not enough. Add to it a term that couples the position and the velocity:
V=21k0y2+21mv2+2cyv(V)
It is still an energy, positive as soon as the state is not zero, as long as the damper is not too strong. Complete the square in v:
V=2m(v+2mcy)2+21(k0−4mc2)y2
V=2m(v+2mcy)2=+21(k0−4mc2)y2
Both squares have a positive coefficient when c2<4mk0: this is the condition for the mass alone to oscillate before it comes to rest, and it holds for the values of Table 1. The derivative of V along the motion is computed like that of V0:
The first square is never negative. The bracket is therefore positive for every non-zero state exactly when the coefficient of the second square is positive, that is when κ2<c2k/m: if the bracket vanished, the second square would give y=0 and the first one v=0. Replace κ by its expression and expand: the terms in c2(k−k0)/m cancel, and what remains bounds the distance between the stiffness and its nominal value:
(k−k0)2<mc2(k0−4mc2)
Since k−k0=k0αρ, divide by k02 and introduce the damping ratio ζ=c/(2mk0): the condition reads α∣ρ∣<2ζ1−ζ2. The damper sets the room left to the parameter. Without damping there is none: a stiffness that changes at the right moments feeds energy into the mass, as a child pumps a swing.
With the values of Table 1, ζ≈0.32 and the condition is ∣ρ∣<1.5, which contains the range ∣ρ∣≤1.17 of this page. So a single energy, with no ρ in it, decreases along every motion, whatever ρ(t) does in that range and however fast: at each instant, dV/dt only sees the current value of k. Over that range the bracket is even at least ε(y2+v2) for some ε>0, because its smallest eigenvalue depends continuously on ρ and is positive at every point of a closed interval: the energy falls at least in proportion to the squared size of the state. This is exactly what Result 1 asks of the closed loop.
Figure 9 shows this computation. In the plane of the states, the level curves of an energy are ellipses, and a motion is a curve. The energy decreases when the motion crosses the ellipses inwards, that is when the velocity of the state makes an obtuse angle with the outward normal: dV/dt is the product of the two. Choose the mechanical energy, give ρ a value other than zero, and watch the ellipse carried by the state swell at times. Choose the energy (V): it only shrinks, even when the stiffness jumps at the worst moments.
V(t)/V(0), logarithmic scale · t (s)
V(t)/V(0)
9.1% of the starting energy is left
exp(−βt)
95% guaranteed bound
dV/dt
−0.036 it decreases
ρ
0.00
Along this motion, this energy never grew.
EnergyV=21k0y2+21mv2+2cyv
Parameter
Figure 9: The stored energy of the mass alone.The plane of the states (y,v) of the mass, with the velocity divided by the natural pulsation ω0=k0/m, so that both axes are lengths drawn at the same scale. In blue, the motion and its velocity; in green, the outward normal to the level curve. In copper, the level curve of the start: the state never leaves it when the energy only decreases. The ellipse in ink is the level curve carried by the state. On the right, the energy left over time and, in copper, the bound V(0)exp(−βt) that the proof of Result 1 guarantees.Click in the plane to launch the mass from another state.
The closed loop of this page has eight states, four for the generalized plant and four for the controller: its ellipses become ellipsoids in a space of dimension eight, which cannot be drawn, but nothing else changes. One difference of substance remains: for the mass alone we could guess the energy, while for the closed loop the matrix X is an unknown that the synthesis must find, together with the controller.
Here is the precise statement. It uses an energy stored in the closed loop, V=xcl⊤Xxcl, where X is symmetric positive definite: V is positive as soon as the state is not zero.
Where the left-hand side of (D) comes from. At first sight, (D) falls from the sky. It is built in three moves, from the two things we want: performance and stability.
First move: performance is a balance. We want (b), ∥w∥L2<γ∥r∥L2. Square it and divide by γ>0:
γ1∫0∞w⊤wdt<γ∫0∞r2dt
Read it as a balance of energy, kept in a currency where what enters through r counts γ times its energy and what leaves through w counts 1/γ times its energy: the loop must give back less than it receives. Splitting γ2 into γ on one side and 1/γ on the other changes nothing to the inequality; it gives the two ports of the loop the same form, as the third move will show.
Second move: from a balance over all times to a balance at each instant. A balance over all times cannot be checked instant by instant, because energy received now may leave later. In the meantime, the loop stores it, in a stored energy V of the kind described above. At each instant, what is stored plus what leaves must not exceed what enters:
dtdV+γ1w⊤w−γr2≤0
Integrated from rest, it gives V(T)+γ1∫0Tw⊤wdt≤γ∫0Tr2dt, and since V(T)≥0, the balance over all times follows. With r=0, it says dV/dt≤0: the same V is a Lyapunov function, and stability comes with performance. Up to the terms in ε, this is the inequality (D′) of the proof below.
Third move: cut the loop at the parameter. This inequality concerns the closed loop, in which ρ(t) is at work, and ρ(t) is not known when the controller is designed. So cut the loop where ρ acts, and look at M alone. For M,wΔ is one more input that it does not choose, and zΔ one more output. Treat this port like the other one, measuring sizes with the scaling L of (S): count γ1zΔ⊤LzΔ for what M sends to the parameter, and γwΔ⊤LwΔ for what comes back. The balance of M with its two ports is the left-hand side of (D), term for term: what M stores, then what goes to the parameter net of what comes back, then what leaves through w net of what enters through r:
dtdV+(γ1zΔ⊤LzΔ−γwΔ⊤LwΔ)+(γ1w⊤w−γr2)
dtdV+(γ1zΔ⊤LzΔ−γwΔ⊤LwΔ)+(γ1w⊤w−γr2)
Asking this balance for every wΔ, and not only for those that ρ would send, removes ρ from the condition: (D) is a property of M alone, of its matrices, which a computer can test (Step 3). This has a cost: M must behave even for inputs that no admissible ρ would produce, and this is the caution discussed in Section 5.8. And it loses nothing that matters, because of (S): when the loop is closed with ∣ρ∣≤1/γ, what comes back from the parameter is at most γ2ρ2≤1 times what it received, counted in the currency of its port. The net flow to the parameter is then positive or zero: the parameter can take energy from M, never give it. Dropping this flow from (D) leaves exactly the balance of the second move. This is the small-gain theorem, read with energies: a block that gives back less than it receives cannot feed the loop.
The terms in ε only make each inequality strict, with a margin: the proof uses them to obtain an exponential decrease in (a) and a strict gain in (b).
Figure 10 keeps these accounts on the closed loop of this page, with the matrices X and L computed for its controller (Section 5.8). The reference is a pulse of finite energy: r=0.2 m for 2 s, then 0. Push the parameter beyond 1/γ to see it give energy back, and the guarantee fall.
γ∫0tr2dτ
100% entered through r
γ1∫0tw⊤wdτ
92.2% left through w
∫0t(γ1zΔ⊤LzΔ−γwΔ⊤LwΔ)dτ
1.9% went to ρ, net
V(t)
1.6% stored
Dissipated by M
4.3%
At every instant, ∣ρ(t)∣≤1/γ: the parameter only receives, and what leaves through w stays below what entered through r. This is the conclusion of Result 1 (b).M itself always dissipates: (D) is a property of M alone.
Parameter
Figure 10: The energy balance of the loop of this page.The loop of the simulation, cut at the parameter, computed every millisecond with the discrete-time form of (D) (Section 5.8). On the left, the block M: its stored energy fills the tank, and what it dissipates leaves at the bottom; the width of each arrow follows the flow it carries. On the right, the accounts since the start, as shares of what entered through r. The two bars always have the same length: this is the balance itself.
Proof. Close the loop on the parameter: wΔ=ρzΔ with ∣ρ∣≤1/γ. (D) holds whatever the inputs of M, hence for this one too, and (S) applies. Multiply (S) by γ>0, which keeps the direction of the inequality:
The first inequality, written on two lines, is (D) itself. The second drops the bracket, which is positive or zero and carries a minus sign: the right-hand side can only grow. The third drops −ε∥wΔ∥22, which is negative or zero, for the same reason. Keeping only the first and the last member:
dtdV+γ1w⊤w−γr2≤−ε(∥xcl∥22+r2)(D′)
dtdV+γ1w⊤w−γr2≤−ε(∥xcl∥22+r2)(D′)
(a) Take r=0. In (D′), the term γ1w⊤w is positive or zero: dropping it from the left-hand side keeps the inequality true.
dtdV≤dtdV+γ1w⊤w≤−ε∥xcl∥22
To compare ∥xcl∥22 with V, let λmin and λmax be the smallest and the largest eigenvalue of X: both are positive, since X is symmetric positive definite. In an orthonormal basis of eigenvectors of X,V is the sum of the squared coordinates of xcl, each weighted by an eigenvalue, and ∥xcl∥22 is the same sum without the weights. Hence the double bound below, which Figure 11 draws: a level curve of V lies between two circles.
λmin∥xcl∥22≤V≤λmax∥xcl∥22
Figure 11: A level curve of the energy, V(x)=V(0), between two circles.The largest circle it contains has the radius rmin=V(0)/λmax; the smallest circle around it has the radius rmax=V(0)/λmin. A state that stays inside the ellipse stays inside the large circle.
The bound on the right gives ∥xcl∥22≥V/λmax. With β=ε/λmax>0:
dtdV≤−ε∥xcl∥22≤−βV
The derivative of exp(βt)V(t), which is exp(βt)(dtdV+βV), is therefore negative or zero: this product never grows, and exp(βt)V(t)≤V(0) for every t≥0. With the bound on the left:
The state tends to zero exponentially, wherever it starts: this is (a). The factor λmax/λmin has a geometric meaning: the state never leaves the ellipse of its start (Figure 9), so it never leaves the circle drawn around that ellipse (Figure 11), but inside it may first move away from the origin. The rate β is a guarantee, not a prediction: on Figure 9, the energy falls much faster than the copper bound.
(b) Start from zero initial conditions, xcl(0)=0, so that V(0)=0, and take a non-zero reference r of finite energy. We may assume ε<γ: if (D) holds for some ε, it holds for every smaller ε, since its right-hand side only grows. In (D′), the term −ε∥xcl∥22 is negative or zero: dropping it from the right-hand side keeps the inequality true. This is where the state leaves the computation:
dtdV+γ1w⊤w−γr2≤−ε(∥xcl∥22+r2)≤−εr2
dtdV+γ1w⊤w−γr2≤−ε(∥xcl∥22+r2)≤−εr2
Move γr2 to the right-hand side:
dtdV+γ1w⊤w≤(γ−ε)r2
This holds at every instant, so it also holds for the integrals from 0 to T. The integral of dV/dt is V(T)−V(0)=V(T):
V(T)+γ1∫0Tw⊤wdt≤(γ−ε)∫0Tr2dt
V(T)+γ1∫0Tw⊤wdt≤(γ−ε)∫0Tr2dt
V(T) is positive or zero, since X is positive definite: dropping it from the left-hand side keeps the inequality true. Multiply then by γ>0. Finally, since r2≥0 and γ−ε>0, the right-hand side can only grow if its integral runs over all times, which gives ∥r∥L22:
∫0Tw⊤wdt≤γ(γ−ε)∫0Tr2dt≤γ(γ−ε)∥r∥L22
The right-hand side no longer depends on T. The left-hand side grows with T and stays below this bound: it has a limit, which is ∥w∥L22 by the definition above, and this limit satisfies the same bound. Since r is not zero, ∥r∥L2>0, and since ε>0:
∥w∥L22≤γ(γ−ε)∥r∥L22<γ2∥r∥L22
Taking square roots gives ∥w∥L2<γ∥r∥L2: this is (b), and the proof is complete.
Three remarks. The proof never used the way ρ varies, only ∣ρ∣≤1/γ at each instant: the guarantee holds for arbitrarily fast variations. Then, γ plays the two roles announced in Section 4: it bounds the gain from r to w, and 1/γ bounds the parameter. Finally, (D) also ensures that the loop on ρ has a solution at every instant, although zΔ depends instantly on wΔ and conversely: taken with a zero state and a zero reference, (D) says that the instantaneous part of M has a gain below γ in the units of L, so that no non-zero signal can sustain itself instantly around the loop.
5.4Step 3: from trajectories to a matrix inequality
(D) must hold along every trajectory of M. To turn it into a condition that a computer can test, write it with the two stacked signals of Step 1, the inputs ν=(wΔ,r) and the outputs η=(zΔ,w). Two matrices keep the accounts of the scaling: Λν=diag(L,1) weighs the inputs and Λη=diag(L,I2) the outputs. Each one measures the parameter channel through L and leaves the other signals as they are; they only differ in size, because r is one number and w=(e′,u′) is two. With them, (D) reads:
dtdV+γ1η⊤Ληη−γν⊤Λνν≤−ε(∥xcl∥22+∥ν∥22)
dtdV+γ1η⊤Ληη−γν⊤Λνν≤−ε(∥xcl∥22+∥ν∥22)
At a given instant, a trajectory of M can go through any state xcl with any input ν, and the rest is imposed by the equations of M:x˙cl=Aclxcl+Bclν and η=Cclxcl+Dclν. So (D) is a condition on all pairs (xcl,ν), and each term of its left-hand side is a quadratic expression of such a pair. Take them one at a time. First the stored energy, V=xcl⊤Xxcl:
The second line uses two facts: a number is equal to its transpose, and X is symmetric, so that x˙cl⊤Xxcl=xcl⊤Xx˙cl. The last line uses them again to write 2xcl⊤XAclxcl as two halves, one of them transposed. Then the outputs, replaced by their expression:
The term of the inputs needs no rewriting. Stack the state and the inputs into one vector, q=(xcl,ν). The three terms then add up to a single quadratic form, q⊤Φq, where the symmetric matrix Φ gathers them block by block:
The first matrix comes from the stored energy and from the inputs, the product from the outputs. Since ∥xcl∥22+∥ν∥22=∥q∥22, (D) says that q⊤Φq≤−ε∥q∥22 for every q.
A symmetric matrix is called negative definite, written Φ≺0, when q⊤Φq<0 for every non-zero q; all its eigenvalues are then negative. This is the same requirement as (D). If (D) holds for some ε>0, then q⊤Φq is negative for every non-zero q. Conversely, if Φ≺0, the computation made with the eigenvalues of X in the proof of Result 1 gives q⊤Φq≤λmax∥q∥22, where λmax is now the largest eigenvalue of Φ, a negative number: (D) holds with ε=−λmax.
One difficulty is left in Φ: the matrices Ccl and Dcl appear twice in the same product. They contain the unknown controller, which would thus appear squared. A classical tool unfolds this product.
Figure 12: The quadratic form of the block matrix.It is drawn for a fixed a, as a function of b, in the case where b is a single number. Because H≺0, it is a parabola that opens downwards. Its top is reached at b=−H−1G⊤a, and the height of that top, in copper, is a⊤(F−GH−1G⊤)a: the whole parabola is below zero exactly when its top is.
Apply the tool to Φ, read from right to left. Take for F the first matrix of Φ, and:
G=[Ccl⊤ΛηDcl⊤Λη],H=−γΛη
H≺0 because γ>0 and L≻0. And F−GH−1G⊤ is exactly Φ, because the three copies of Λη reduce to one:
So Φ≺0 exactly when the block matrix built from F,G and H is negative definite. In that larger matrix, Ccl and Dcl appear once each: nothing is multiplied by itself any more.
The test has a frequency-domain interpretation: the H∞ norm of M is below γ once its parameter channel is measured through L. For a linear time-invariant system and L=I, the bounded real lemma is in fact an equivalence: such an X exists if and only if the system is stable with an H∞ norm below γ. We have proved the direction that gives guarantees, from the inequality to the gain. The converse, which says that this test loses nothing for a system without parameter, is a deeper theorem, due to Kalman, Yakubovich and Popov; it is presented in [2], we admit it, and the guarantee of this page does not rest on it.
5.5Step 4: eliminating the controller
The unknowns of (B) are X,L and the controller Ω, hidden in the four matrices of the closed loop. The trouble comes from products such as XAcl: they multiply two unknowns, X and Ω. A set described by such products is in general not convex, and a solver can get stuck in it. The way out of [1] is to remove Ω from the problem altogether. First, see where Ω sits in (B). Replace the closed-loop matrices by their expression of Step 1: in each block, the part without Ω and the part with Ω separate.
The four parts with Ω have the same factor on the left of Ω along each block row, and the same factor on its right along each block column. They are therefore the blocks of one product, placed where (B) has them:
The other blocks of (B) that contain Ω are the transposes of these, since (B) is symmetric: they form V⊤Ω⊤U. Call Ψ what is left, that is the matrix of (B) written for Ω=0, with A0,B0,C0 and D0. Then (B) is:
Ψ+U⊤ΩV+V⊤Ω⊤U≺0(E)
The question becomes: Ψ,U and V being given, when does some Ω satisfy (E)? The answer does not involve Ω.
Why these conditions are necessary. Multiply (E) on the left by a⊤ and on the right by a. If Ua=0, the two terms that contain Ω vanish, whatever Ω is, and a⊤Ψa<0 remains. The same happens if Va=0. In words: in the directions it cannot reach, the controller cannot help, so the inequality must already hold there without it.
Why they are sufficient: the idea, in a plane. This is the remarkable part: nothing else is required. Take first the smallest case that shows everything: a plane, with one direction that U does not see and one direction that V does not see. Call α and β the coordinates of a vector along these two directions. Then Ua only depends on β and Va only on α, so the two terms in Ω add up to a multiple of the product αβ, and (E) reads:
ψ11α2+2(ψ12+ω)αβ+ψ22β2<0∀(α,β)≠(0,0)
ψ11α2+2(ψ12+ω)αβ+ψ22β2<0∀(α,β)≠(0,0)
Here ψ11,ψ12 and ψ22 are the entries of Ψ in these coordinates, and ω is a number that Ω sets as it likes. The first condition, read with β=0, says ψ11<0; the second, read with α=0, says ψ22<0. Choose ω=−ψ12: the middle term disappears, and what is left, ψ11α2+ψ22β2, is negative as soon as α or β is not zero. The controller cannot change the form along the two axes, but it tilts it everywhere else, and it can cancel the coupling between them. Figure 13 lets you do it.
ψ11α2+2(ψ12+ω)αβ+ψ22β2
ω
0.00
ψ12+ω
1.60
The form is positive or zero in the copper directions: (E) fails.
On the two axes, the form is always ψ11α2 or ψ22β2.
Figure 13: The projection lemma in a plane.The quadratic form ψ11α2+2(ψ12+ω)αβ+ψ22β2, with ψ11<0 and ψ22<0. In ink, one of its level curves, where it is equal to −1. In copper, the directions in which it is positive or zero: the inequality fails there. On the two axes the form does not depend on ω: the blue axis is the direction that U does not see, the green axis the one that V does not see.
The general case: four families of directions. In a larger space, a direction can be seen by U, by V, by both or by neither. Choose coordinates that sort them. Take a basis of the vectors that both U and V send to zero; complete it into a basis of the vectors that U sends to zero; complete the first one, separately, into a basis of the vectors that V sends to zero; then complete the whole into a basis of the space. Every vector now has four pieces, a=(a1,a2,a3,a4):a1 is seen neither by U nor by V,a2 is seen by V only, a3 by U only, and a4 by both. In these coordinates:
U=[00U3U4],V=[0V20V4]
U=[00U3U4]V=[0V20V4]
Two facts make the rest work. First, the three families built before the last completion are independent together: by the dimension formula for the sum of two subspaces, there are exactly as many vectors in them as the dimension of the space they span. Second, the columns of [U3U4] are independent: if U3a3+U4a4=0, the vector (0,0,a3,a4) is sent to zero by U, so it is a combination of the first two families, which is only possible if a3 and a4 are zero. The same holds for [V2V4]. Now compute the term in Ω:
The four matrices K32,K34,K42 and K44 can take any values we want. Indeed, gather them into one matrix K, and write U~=[U3U4] and V~=[V2V4]: we need U~⊤ΩV~=K. Since the columns of U~ are independent, U~⊤U~ has an inverse, and likewise for V~. The following choice works, as one checks by multiplying it by U~⊤ on the left and by V~ on the right:
Ω=U~(U~⊤U~)−1K(V~⊤V~)−1V~⊤
So (E), in these coordinates, is the matrix of Ψ cut into the same four pieces, with freely chosen matrices added in seven places:
The first row and the first column carry no free matrix: the piece that neither U nor V sees is out of reach. The two conditions of the lemma, written in these coordinates, say that Ψ is negative definite on the pieces (1,2), which U sends to zero, and on the pieces (1,3), which V sends to zero. In particular Ψ11≺0. Make three choices, as in the plane: cancel what can be cancelled.
Look at its first three pieces, and call N that block. Apply Tool 1 to it with Ψ11≺0 in the role of the negative block (Tool 1 was written with that block in the lower-right corner; changing the order of the pieces puts any diagonal block there). N≺0 exactly when the matrix below is negative definite, and the choice of K32 has emptied its off-diagonal blocks:
[Ψ22−Ψ12⊤Ψ11−1Ψ1200Ψ33−Ψ13⊤Ψ11−1Ψ13]
Its two diagonal blocks are negative definite: by Tool 1 again, the first one says that Ψ is negative definite on the pieces (1,2), and the second one on the pieces (1,3). These are the two conditions of the lemma. So N≺0.
The last piece remains. Call Z the column made of Ψ14 above two zero blocks: the whole matrix has N and Z on its first row of blocks, Z⊤ and Ψ44+K44+K44⊤ on the second. By Tool 1 with N≺0, it is negative definite exactly when Ψ44+K44+K44⊤−Z⊤N−1Z≺0. Choose:
K44=21(Z⊤N−1Z−Ψ44−I)
This matrix is symmetric, so K44+K44⊤ is twice it, and the expression above is equal to −I, which is negative definite. With these four choices (E) holds, and the formula given above turns them into a matrix Ω: the lemma is proved. If one of the four families of directions is empty, its blocks are simply absent, and the same argument applies.
The second condition, on the plant alone. Apply the lemma to (E), starting with V. A vector on which (B) acts has three blocks, a=(xcl,ν,σ): the state, the inputs, and a third block σ of the size of the outputs, the one that Tool 1 added in Step 3. And Va=CΩxcl+EΩν is what the controller reads (Step 1). So Va=0 means xK=0,e=Ceξ+r=0 and wθ=0: the controller reads nothing.
On these vectors, undo Tool 1. Its proof never used that the first block ranges over the whole space: for a fixed q=(xcl,ν), the largest value of a⊤Ψa over σ is q⊤Φ0q, where Φ0 is the matrix Φ of Step 3 written for Ω=0. The second condition is therefore: q⊤Φ0q<0 for every non-zero q such that the controller reads nothing. By Step 3, q⊤Φ0q is the left-hand side of (D) for the loop without controller. Compute it. With xK=0 and wθ=0, the matrices of Step 1 give:
The state and its derivative both have a zero lower half. In dV/dt=2xcl⊤Xx˙cl, only the upper-left block of X takes part: call it S, it weighs the plant state. Likewise the parameter channel is reduced to the plant channel, zΔ=(0,zρ) and wΔ=(0,wρ), so only the last diagonal entry of L takes part: call it L3. The second condition of the lemma is: for every non-zero (ξ,wρ,r) such that Ceξ+r=0,
(i) is the energy balance (D) of the plant alone, without control, with the stored energy ξ⊤Sξ, on the trajectories that the controller does not see.
The first condition, after a change of variables.U contains the unknowns X and Λη, so one more step is needed. Multiply (E) on both sides by the symmetric invertible matrix Π=diag(X−1,Λν−1,Λη−1). A matrix stays negative definite under such a multiplication, because b⊤ΠMΠb is the quadratic form of M at the vector Πb, which runs over all non-zero vectors when b does. The term in Ω keeps its shape, with UΠ and VΠ in place of U and V, and:
UΠ=[BΩ⊤XX−10DΩ⊤ΛηΛη−1]=[BΩ⊤0DΩ⊤]
It is free of unknowns. The lemma, applied to the multiplied inequality, gives the condition: b⊤ΠΨΠb<0 for every non-zero b such that UΠb=0. Write b=(pcl,sν,sη), with pcl=(p,pK) cut like the state and sη=(sθ,sρ,sw) cut like the outputs. With the matrices of Step 1:
UΠb=BΩ⊤pcl+DΩ⊤sη=pKBu⊤p+Dwu⊤swsθ
UΠb=BΩ⊤pcl+DΩ⊤sη=pKBu⊤p+Dwu⊤swsθ
So UΠb=0 means pK=0,sθ=0 and Bu⊤p+Dwu⊤sw=0: these are the directions on which the control u has no effect. The multiplied matrix is computed block by block, each product of a matrix by its inverse reducing to the identity:
It has the shape of Ψ, with the roles of the inputs and of the outputs exchanged and every matrix of the plant transposed. This time the middle block sν is free, and Tool 1 removes it as it removed σ above: for fixed pcl and sη, the largest value over sν replaces the blocks that involve sν by γ1t⊤Λν−1t, with t=B0⊤pcl+D0⊤sη. With pK=0 and sθ=0, the matrices of Step 1 give:
As before, only one block of each unknown is left: the upper-left block of X−1, which we call R, and the last diagonal entry of L−1, which we call J3. The first condition of the lemma is: for every non-zero (p,sρ,sw) such that Bu⊤p+Dwu⊤sw=0,
(ii) is the same balance as (i), for the plant with transposed matrices and on the directions in which the control cannot act. Written with matrices, (i) and (ii) can be restricted to these subspaces using matrices NS and NR whose columns span the two sets of vectors.
5.6Step 5: the linear matrix inequalities, and why they are convex
The controller has disappeared, and with it most of X and of L: (i) and (ii) only involve S and L3, taken from X and L, and R and J3, taken from their inverses. One question remains. Can any two matrices S and R be the corners of a positive definite matrix and of its inverse? No, and the exact condition is simple.
Proof, from the controller to the conditions. Suppose that a controller, X≻0 and L≻0 satisfy (B), which is (E). By Tool 2, the two conditions of the lemma hold, and Step 4 has written them as (i) and (ii), with S and L3 read in X and L, and R and J3 read in their inverses. Tool 3 gives the first of the two conditions above. The same tool, applied to L and its inverse, whose lower-right corners are the numbers L3 and J3, gives the second one.
Proof, from the conditions to the controller. Suppose that R,S,L3 and J3 satisfy the four conditions. First, R and S are positive definite. Indeed the quadratic form of the first block matrix at a pair (a,b) is a⊤Ra+2a⊤b+b⊤Sb, and it is never negative. Taking b=0 shows R⪰0. If Ra=0 for some a≠0, the pair (a,−ta) would give −2t∥a∥22+t2a⊤Sa, which is negative for a small t>0: impossible. So no eigenvalue of R is zero, and R≻0. The same argument applies to S, and to the numbers L3 and J3, which are positive.
Tool 3 then builds a matrix X≻0 whose corner is S and whose inverse has the corner R, and a matrix L≻0 whose corner is L3 and whose inverse has the corner J3. For L, the block that the tool adds is the part of the scaling that weighs the parameter channel of the controller. The controller of this page has a single such channel, like the mass: this block is one number, and L is twice as large as its corner, as in Tool 3. With these X and L, (i) and (ii) are exactly the two conditions of Tool 2, since Step 4 has shown that these conditions only involve the four corners. So some Ω satisfies (E), which is (B): the controller exists.
Why these conditions are convex. Count what is left: the unknowns are R,S,L3 and J3. Two symmetric matrices of size n hold n(n+1) numbers; with n=4 states for the generalized plant, that is 22 unknown numbers in all. For a fixed γ, each of them appears to the first power in every condition. Take two sets of values that satisfy the conditions and mix them, with weights 1−t and t between 0 and 1. The left-hand side of (i), for a given (ξ,wρ,r), is then the same mixture of the two left-hand sides:
g1<0,g2<0⟹(1−t)g1+tg2<0
g1<0,g2<0⟹(1−t)g1+tg2<0
where g1 and g2 are the values of the left-hand side for the two sets. A mixture of two negative numbers is negative, so (i) holds for the mixture. The same is true of (ii), and of the last two conditions, where a mixture of two forms that are never negative is never negative. The set of solutions is therefore convex: with two of its points it contains the whole segment between them. Conditions of this kind are called Linear Matrix Inequalities (LMIs). At a fixed performance level, these inequalities form a convex feasibility problem. When a solver returns a candidate, its matrices can be substituted into the inequalities to check feasibility.
Figure 14 shows the last condition, the only one that can be drawn: with numbers, it says L3J3≥1. It also shows what the method has gained. In (B), the scaling and its inverse were tied by an equality, which for numbers would read L3J3=1: a curve, and a curve of this kind is not convex. Tool 3 has replaced the curve by the region above it.
Figure 14: The condition L3J3≥1.The shaded region is convex: the segment between two of its points stays inside (in blue). The curve L3J3=1 is not: the segment between two of its points leaves it (in copper).
And γ? It multiplies L3 and J3 in (i) and (ii), so it is not one more unknown of the same kind. But the conditions get easier as γ grows: in (D), the terms in 1/γ decrease and the terms in −γ decrease too when γ increases, so the matrices and the controller that work for one value of γ work for every larger one. The smallest γ can therefore be approached by trying values and halving the interval at each trial (Figure 15). The value found in this way is the best one of this framework.
Figure 15: Searching the smallest γ.Each trial solves the convex conditions for one value of γ: a full dot when they have a solution, an empty one when they have none. Since a solution for one value is a solution for every larger one, each trial halves the interval that contains the smallest γ.
5.7Step 6: rebuilding the controller
A solution (R,S,L3,J3) is not yet a controller. Two steps remain, and both are constructive: the proofs above already contain their formulas.
First, complete the matrices. The proof of Tool 3 gives X from R and S:
N=(S−R−1)1/2,X=[SNNI]
N=(S−R−1)1/2X=[SNNI]
The same formula, with numbers, gives L from L3 and J3. This is where the order of the controller is decided: the lower-right block of X weighs the state of the controller, and it has the size of S. The controller therefore has as many states as the generalized plant, 4 here.
Then, solve for the controller. Look at (E) again with X and L now known: Ψ,U and V are arrays of numbers, and the only unknown, Ω, appears to the first power. (E) is an LMI in Ω, and Tool 2 guarantees that it has a solution. Its proof even builds one, by the four choices of Step 4 and the formula that turns them into Ω. A solver may return another solution of (E), for instance one whose entries are smaller: any of them will do. Reading Ω block by block gives the nine constant matrices of the controller (Step 1).
5.8The controller obtained, and what its guarantee covers
Note on Controller Structure: It is important to note that the controller structure (LPV-LFT) is not an arbitrary choice. Following the methodology of Apkarian & Gahinet [1], assigning the controller the same parameter dependence (LFT) as the plant allows the non-convex optimization problem to be transformed into a convex system of Linear Matrix Inequalities (LMIs). At a fixed γ, feasibility is convex. Bisection can approach the infimum of the levels admitted by these sufficient conditions. For the controller used here, the numerical search and the subsequent certificate establish an achievable γ. With a frozen parameter and without actuator saturation, the discrete closed loop remains stable beyond ∣ρ∣=1.17, down to about ρ≈−2.5, where the stiffness vanishes (k=0). This frozen-parameter calculation provides no guarantee for arbitrary parameter variations outside ∣ρ∣≤1/γ.
Where does this caution come from? From Step 2. A single matrix X and a single matrix L serve for every value of ρ(t) in the range and for every speed of variation. A parameter known to vary slowly, or over a smaller range, could allow a smaller γ: the method does not use that information. In exchange, it gives a controller and a guarantee by solving convex problems only.
What is inside K(ρ)? As promised in Section 2, let's open the controller. Look at the bottom of Figure 4: K is a linear time-invariant block, which never changes, looped on ρ(t) exactly like the plant at the top. Since ρ(t) is measured in real time, the controller can use its current value at every instant: this is what makes it gain-scheduled. Here it is in discrete time, as it runs in the simulation (the index θ comes from [1], where the parameter block is called Θ):
The matrices AK,…,DKθθ are constant: the LMIs give them once and for all, offline, and the browser never solves anything. The controller has 4 states, as many as the generalized plant (2 for the mass, 1 for each filter). Its loop on ρ has a single channel, like that of the mass, hence the block ρ(t) at the bottom of Figure 4. When ρ changes, no matrix is recomputed: only the signal going around this loop changes, and the output u adapts with it. One practical detail: wθ appears on both sides of its own equation, through zθ. This algebraic loop is solved at every step, before computing u:
wθ[k]=1−ρ[k]DKθθρ[k](CKθxK[k]+DKθ1e[k])
wθ[k]=1−ρ[k]DKθθρ[k]=(CKθxK[k]+DKθ1e[k])
You may have noticed that we just switched to discrete time, while the rest of this page, proofs included, is written in continuous time. The numerical certificate applies to the unsaturated discrete model with Ts=1 ms. The mass uses an explicit Euler step. The synthesis discretizes the two weighting filters with a zero-order hold.
Nothing essential changes in discrete time, and [1] treats both cases. The stored energy is compared between two successive instants instead of being differentiated, and (D) becomes:
Result 1 is proved in the same way, line for line. The scaling inequality (S) holds at each instant, so the two terms of the parameter channel are dropped as before, and what is left is:
(a) Take r=0. With the largest eigenvalue λmax of X, as in continuous time, ∥xcl[k]∥22≥V[k]/λmax; and we may suppose ε≤λmax, since a smaller ε is always allowed:
V[k+1]≤V[k]−ε∥xcl[k]∥22≤(1−λmaxε)V[k]
At each step the energy is multiplied by a number that lies between 0 and 1 and is not 1: it decreases at least like a geometric sequence, and the state tends to zero. (b) Start from a zero state and add the inequalities from k=0 to k=N−1. The differences of V cancel in pairs, and V[0]=0:
V[N]+γ1k=0∑N−1w[k]⊤w[k]≤(γ−ε)k=0∑N−1r[k]2
V[N]+γ1k=0∑N−1w[k]⊤w[k]≤(γ−ε)k=0∑N−1r[k]2
V[N] is positive or zero and is dropped; the rest of the argument is the one of continuous time, with the sum of the squares of a sequence in place of the integral, and the gain from r to w is below γ. Steps 3 to 6 keep their arguments too. One thing changes in Step 3: V[k+1] is the quadratic form of X at xcl[k+1], which contains the closed-loop matrices twice. The matrix inequality gets one more block, and one more Schur complement separates the two copies before the controller is eliminated.
How the controller of this page was computed. The synthesis was run in MATLAB on the sampled model, with a single scheduling channel for the controller, as for the mass. For each scaling L, MATLAB solves the scaled H∞ synthesis problem. A direct search over L from several starting points found a best level of 0.829. The controller is computed 2% above this level, which gives better-conditioned matrices, then checked apart from its synthesis. Its scaled gain is 0.8449. At γ=0.85, a second solver, MOSEK, returns matrices X and L that satisfy the discrete-time form of (D) above, and the inequality is checked on the numbers returned. In balanced coordinates, the margin at 0.85 is about 3.9×10−10. The resulting parameter bound is ∣ρ∣≤1/γ, with 1/γ≈1.17647. The figures use 1.17 as a conservative limit. Figure 10 uses these X and L.
By minimizing γ in the synthesis step, we are essentially 'squashing' this worst-case gain ratio for all parameter trajectories ρ(t) that stay in the range ∣ρ∣≤1/γ.
References
[1]
P. Apkarian and P. Gahinet, “A convex characterization of gain-scheduled H∞ controllers,” IEEE Transactions on Automatic Control, vol. 40, no. 5, pp. 853–864, May 1995.doi (opens in a new tab)
[2]
P. Gahinet and P. Apkarian, “A linear matrix inequality approach to H∞ control,” International Journal of Robust and Nonlinear Control, vol. 4, no. 4, pp. 421–448, 1994.doi (opens in a new tab)
An opportunity, a research collaboration, or a project to develop? Tell me about the context and what you would like to discuss.