Part II: Noise sensitivies in systems with delays and multiple time scales Acknowle The auth NSF Grant sional Meas sity of Mich Nomencla 0.8 0.6 0.4 0.2 0 !0.2 !0.4 !0.6 !0.8 700 Wednesday, June 6, 2012 750 800 850 900 Ak ! p am ! a # % ! s d ! d & ! t f ! s '(t) ! t ( ! L Kc ! c k ! i M (t;z) ! u ) ! o p max ! m d p(i) ! d so ! n s(t) ! s t ! t u(m) ! p * bw ! o *n ! n x(t) ! r p z ! h + ! s Models of Balance • Applications: Human Postural Sway, Stick Balancing, Robotics • Mathematical Challenges: Nonlinear dynamics with delayed feedback, discontinuous control, noise • What are the contributing factors to stability, instability, balance, sway, other behaviors? Ideas from models in mechanics/ optics, bring to biological applications: balance, blood diseases, epidemiology Wednesday, June 6, 2012 • Applications: Human Postural Sway, Stick Balancing, Robotics 461 cm) attached at either end to a metal stand (see Holden et al. 1994). The stand rested on a rigid wooden platform (155 cm! 70 cm) that overlay the force plate and extended beyond its lateral edges. The touch device apparatus on one side of the platform was balanced by a comparable mass on the other side (see Fig. 1). This arrangement ensured that the Kistler force platform detected all forces applied to the touch device as well as all forces generated by the subject’s feet and that it did not interpret forces at the fingertip as changes in CFPx or CFPy. In this circumstance, the force applied by the finger to the touchbar and transmitted to the platform by the touchbar’s base will be exactly balanced by the resulting reaction force generated at the subject’s feet; consequently, there will be no spurious change in CFPx or CFPy as a result of the force at the fingertip. The horizontal bar was adjusted in height to allow individual subjects to assume a comfortable lateral arm position while touching the contact plate with their index finger. Two, dual-element, temperature-compensated strain gauges (Kulite Semiconductor, Type M(12) DGP-350-500) mounted on the metal bar transduced the lateral (FingerL) and vertical (FingerV) forces applied by the finger. The strain gauge signals were amplified and calibrated in units of force (newtons). A comparator could trigger an auditory tone when a specified threshold force was reached. Procedure Fig. 1 Schematic illustration of the experimental apparatus with a subject in the tandem Romberg stance signed an informed consent statement that had been approved by the Brandeis University Committee for the Protection of Human Subjects. Derivative of the Angle from vertical Apparatus and Measures Figure 1 depicts the test situation with a subject standing in the tandem Romberg position (heel-to-toe) on a force platform, touching a device used to measure the forces applied by the fingertip. The force platform (Kistler Model 9261A) measured the ground reaction forces at the feet. Center of foot pressure Medial-lateral (CFPx) and anterior-posterior (CFPy) coordinates of foot pressure were computed from the force components registered by piezoelectric crystals in the corners of the force platform. Wednesday, June 6, 2012 Head motion The subject stood with right foot behind left along the center of the anterior-posterior axis of the force platform. Adhesive tape was used to mark the position of the subject’s feet on the platform so that the same foot position could be repeated for the different conditions of the experiment. The touch bar was adjusted to a comfortable height (approximately waist level) and lateral distance for the subject to make contact with the right INDEX fingertip. The experimental conditions varied in terms of allowing vision (Vision) or having eyes closed (Dark) and type of fingertip contact: no contact (NoTouch) in which the subject’s arms hung passively, touch contact (Touch) in which the subject was limited to 1 N of applied force on the touch apparatus, and force contact (Force) during which the alarm was turned off and subjects could apply as much force as desired. One newton of applied force at the fingertip is mechanically inadequate to attenuate sway amplitude by more than 2.3% (Holden et al. 1994). The six experimental conditions included: Vision-NoTouch, Vision-Touch, VisionForce, Dark-NoTouch, Dark-Touch, and Dark-Force. One practice trial was given for each condition before the experiment began. Subjects started a trial by getting on the platform and looking straight ahead at a fixation target on a wall 2 m away that was covered with a black cloth. The subject’s peripheral visual field beyond 30° provided a rich, complex visual environment with many horizontal and vertical features. Subjects were told to let go of the safety railing surrounding them and take as much Asai, etal, 2009, PLOS1 Human postural sway Angle from vertical " − ! sec "" . ẍ = g tan 3 ystem. Finally, in Section 5 we conclude and point to some open problems. Simple model: inverted pendulum hus when the pendulum is at the inverted position, " = 0 and called “labyrinthus” in the inner ear which helps in the self-balancing ¨" = 0, equations .= Model the cart is not atouraeyesfixed position but moves with are closed. ome constant speed. The dynamics of the setup in Fig. 1 are governed by the second-order 2. MECHANICAL MODEL differential OF BALAN An alternate approach isConsider to the penisplacement θ of the pendulum, which canassume be as (see alsoinverted Refs. the written simplest that planar mechanical model[2,9,17]): of the inverted pen Its lowest point slides smoothly along the horizontal line. This mechan ! " ulum is 3m stabilized not by the application of forces at the possible model describing the “man-machine” system when somebody 3m 2 3g 3F 2 his fingertip, tries to point ofcos the stick in0. a way t 1 − by the θ̈ + application − and cos θdirect θ̇ onsin(2θ) sin θ move + thisatlowest θ= ase but of torque the pivot. In 4 8 its upper position. 2 LThe system has 2L(M + Mofc )freedom described by two p degrees z. The angle cp is detected together with its derivatives and the ho is case the model is veryandsimple, M. LANDRY, S. CAMPBELL, K. MORRIS, AND C. AGUILAR y l m θ O j i l g = -g j P F(t) i G. STP.~N AND L. KOLL~R 200 M x x(t) Figure 1. Inverted pendulum system. Table 1 Notation. x(t) θ(t) M m L l P F (t) Displacement of the center of mass of the cart from point O Angle the pendulum makes with the top vertical Mass of the cart Mass of the pendulum Length of the pendulum Distance from the pivot to the center of mass of the pendulum (l = L/2) Pivot point of the pendulum Force applied to the cart experimental setup, the force F (t) on the cart is determined by them in acart, way that s 0 position be where F is the horizontal driving force applied to the andthe m upper = Mcpp /(M p + Mshould c ) is th niform pendulum Mp with respect to the sum of the mass of the cart Mc and the mass of 2¨ f the4 m! pendulum is L and g is""the gravitational acceleration. We adjust the time scale t√ " = T , !10" − mg! sin! control 3 f the pendulum by rescaling time, introducing a new dimensionless time tnew = told · F (t) = αV (t) − β ẋ(t), V (t) is the voltage in the motor driving the cart, and the second term represents al resistance in the motor. The voltage V (t) can be varied and is used to control the The values of the parameters for our experimental setup [21] are given in Table 2. ecise value of the viscous friction # is difficult to measure; however, we expect that it of a similar order of magnitude as β. Our work will thus consider # in the range [0, 15]. more convenient form of the equations is found by solving for ẍ and θ̈ from equations nd (1.2) and introducing the variables θ y = (y1 , y2 , y3 , y4 )T T = (x, θ, ẋ, θ̇) . here the more linear feedback control torque is Even simple model: � inverted pendulum w/ ˙. = q " + q " T control control 2 4 (torque at pivot T control the pivot) he linearization of Eq. !10" about " = 0 is very similar to Figure 1. Mechanical model of stick balancing. Wednesday, June 6, 2012 also explain the center manifold reduction and present the bifurcation diagram of the redu erted or“Reduced field model near the triple-zero eigenvalue bifurcation. In Section 3, we show that there a 4 ¨ to the essentials” balance model: ""give . examples of complex balancing mo ẍ =model, g tan 3 ! sec amics in the reduced and"in−Section 4 we em. Finally, in Section 5 we conclude and point to someor open as problems. a system Thus when the pendulum is at the inverted position, " = θ̈ − sin θ + F (θ, θ̇) cos θ = 0 !4" θ̇ = φ F ="˙ aθ(t + bcart θ̇(t −isτ )not at a fixed position but move = "¨ =−0,τ )the φ̇ = sin θ − F (θ, θ̇) cos θ Model equations istic some constant speed. F = aθ(t − τ ) + b θ̇(t − τ ) nThe!3" dynamics of the in Fig. 1approach are governedisby second-order equ Ansetup alternate tothe assume that differential the inverte lacement the pendulum, which can be written (see application also Refs. [2,9,17]): po- θ ofdulum is stabilized not by asthe of forces ! " of torque at the pi 3m base 3m 2the direct3 gapplication 3F 2 but by 1− cos θ θ̈ + θ̇ sin(2θ) − sin θ + cos θ = 0. 4 this case the 8 model is2 very L 2L(Mp + Mc ) simple, re F is the horizontal driving force applied to the cart, and m = Mp /(Mp + Mc ) is the r 2¨ to the sum of the mass of the cart Mc and the mass of the orm pendulum Mp4with respect m! " " " = T , − mg! sin! control 3 g is the gravitational acceleration. We adjust the time scale to t he pendulum is L and √ he!5" pendulum by rescaling time, introducing a new dimensionless time tnew = told · 3g conn the and • where the linear feedback control torque is Vertical Pendulum: need for control/ external force applied for stabilization Tcontrol = q2" + q4"˙ . Wednesday, June 6, 2012 cart becomes of Model equations ted 4 ¨(PD) Proportional /differential control: " − ! sec "" . ẍ = g tan The dynamics of the setup in Fig.3 1 are governed by the second-order differential e placement θ of the pendulum, which can be written as (see also Refs. [2,9,17]): Thus when the pendulum is at the inverted position, " = !4"! " 3m 3g 3m 2 3F ˙ ¨ 2 " =θ0,θ̈ the is not a θfixed positioncos but 1 − " =cos + cart − at sin θ̇ sin(2θ) + θ =moves 0. 4 8 2 L 2L(M + M ) p c stic some constant speed. !3" Mc ) is the ere F is the horizontal driving force applied to the cart, and m = M p /(M p + inverted An alternate approach is to assume that the F = aθ + b θ̇ iform pendulum Mp with respect to the sum of the mass of the cart Mc and the mass of th podulum is stabilized not by the application of forces a the pendulum is L and g is the gravitational acceleration. We adjust the time scale to √ base by time, the introducing direct application of torque at =the the pendulum by but rescaling a new dimensionless time tnew toldpiv · Derivative Proportional this case the model is very simple, or 4 2¨ m! " − mg! sin!"" = T , !5" onthe control 3 Aθ + B θ̇control torque is where the-Tlinear feedback control = Wednesday, June 6, 2012 ˙ Biological considerations: • Delay in the application of the control: neural transmission • On-off control: not active control all the time • Noise: Small random fluctuations can result in large changes in certain circumstances F = aθ(t − τ ) + bθ̇(t − τ ) Wednesday, June 6, 2012 2. The model. A nondimensional model of the hematopoietic production system is by Hematopoietic Stem cells Neutrophils (2) Platelets Erythrocytes dq dt dn dt dp dt dr dt (blood production) system: = −qb1 hq (q) + b1 µ1 q1 hq (q1 ) − q {b2 hn (n) + b3 hp (p) + b4 hr (r)} , = −γn n + an b2 qτnm hn (nτnm ), ! " = −γp p + ap b3 qτpm hp (pτpm ) − µ3 qτpsum hp (pτpsum ) , = −γr r + ar b4 {qτrm hr (rτrm ) − µ4 qτrsum hr (rτrsum )} , where q, n, r, and are nondimensional stem cells, neutrophils, Colijn,erythrocytes, Mackey, 2007and pla Fluctuations inpproliferation respectively. Subscripts indicate delays: q1 = q(t − 1) and so on. times in lower Theand functions hq , hn , hcell r , and hp are Hill functions given by population levels can drive transitions to large oscillations (disease): Weak attraction to healthy (steady) state (slow damping) 10 8 6 4 2 0 40 60 80 100 120 140 160 180 120 140 160 180 10 8 6 4 2 0 Wednesday, June 6, 2012 200 220 240 260 Mathematical Challenges: • Nonlinearity: non-uniqueness, complex dynamics • Delays: Delay Differential Equations (DDEs) instead of Ordinary Differential Equations (ODEs), functional/history dependence instead of initial conditions at a single point • On/off control: discontinuous equations rather than continuous or smooth behavior. • Noise: Stochastic DDEs (SDDEs), minimal theory, limited computational methods, additional complexities with nonlinearities References: Campbell, Milton, Ohira, Sieber, Krauskopf, Stepan, others Wednesday, June 6, 2012 Chatter in machine tool dynamics chip tool previous cut nonlinear force present cut dx = ydt dy = (−2κy − x + cj [x(t − τ ) − x(t)]j )dt 1 damped linear oscillator + 3 X η(1 + 3 X cj [x(t − τ ) − x(t)]j ) 1 Key parameters: variation in material x = fraction of variation from desired chip thickness Oscillations of tool position (about equilibrium) c1 = = material parameter, κ =cut damping (fixed) Is x(t) 0 stable? i.e. is a straight stable? τ = inversely proportional to rotation velocity Noise sources: δ = percentage variation of material parameter, η = δdw Material parameters (additive + multiplicative noise) Speed of rotation (delay) Wednesday, June 6, 2012 What about delays? Brief crash course in Delay Differential Equations (DDEs) � y (t) = −y(t − τ ), τ = 1 is the delay Need an initial history function: solution for the time interval [0, τ [ depends on initial history function Applied Delay Differential Equations y = 1 for the time interval [−1, 0[ Discontinuity in derivative at t=0, propagates as discontinuity in the second derivative at t=1, etc. Wednesday, June 6, 2012 http://site.ebrary.c Delays and cyclic/period behaviors Logistic equation: Populations � y(t − τ ) y (t) = ry(t) 1 − K � � Growth in birth rate, Decay due to overpopulation Delayed feedback in deaths from overpopulation Longer delays yield oscillations Lower graph has longer delay Wednesday, June 6, 2012 Carrying capacity = K: Bifurcation diagram for periodic behavior Logistic equation: Populations ations Bifurcation: fixed point stable for smaller delays, http://site.ebrary.com/lib/ubc/docPrint.action?encrypted=41e5c... For larger delays, branch indicates amplitude of Applied Delay Differential Equations oscillations http://site.ebrary.com/lib/ubc/docP Carrying capacity: K y Hopf bifurcation S U τ Wednesday, June 6, 2012 S which yieldsstability Linear 2x analysis: machine tool example dx � x (t) + κx (t) + x(t) = c (x(t − τ ) − x(t)) (x(s − τ ) − x(s)). + x = c − 2κ 1 1 s2 ds �� and linearized about x = 0, which yields (3.10) x(t) = e e Look characteristic equation for eigenvalues λt dx d2 x + x = c1 (x(s − τ ) − x(s)). − 2κ 2 ds ds −λτ + 1 = c1x(e − 1). λ2 − 2κλSubstituting the characteristic equation = eλt yields (3.11) λ2 − 2κλ + 1 = c1 (e−λτ − 1). termined by setting λ= iωdecay? in (3.11), dividing this equation into its λ = σ + iω Oscillations: grow or Look for The stability boundary is determined by setting λ = iω in (3.11), dividing this equation τ . These curves are shown d solving two equations for c 1 and σ = 0 these pure imaginary eigenvalues real and imaginary parts, and solving these two equations for c 1 and τ . These curves a in Figureparameters 3, in terms of theand non-dimensional parameters and the dimensional paramete non-dimensional the dimensional parameters. x=0 unstable (oscillations grow)The c1 c1 0.115 0.11 0.105 0.1 x=0 stable (oscillations decay) 40 6 Wednesday, June 6, 2012 3.7 x 10 45 50 55 ττ 60 65 70 75 ds2 − 2κ ds + x = c1 (x(s − τ ) − x(s)). Full nonlinear model: Substituting x = eλt yields the characteristic equation �� � λ2 − 2κλ + 1 = c1 (e−λτ − 1). The stability boundary is− determined setting λ = iω in (3.11), dividing this equati x (t) + κx (t) + x(t) = F(x(t τ ) −byx(t)) real and imaginary parts, and solving these two equations for c 1 and τ . These curves in Figure 3, in terms of the non-dimensional parameters and the dimensional parame x=0 unstable (oscillations grow) material c c Noise induced chatter parameter 1 1 0.115 0.11 Bifurcation scenario: Near stability boundaries x=0 stable (oscillations decay) 0.105 0.1 maximum amplitude 40 45 50 55 60 65 τ x delay 70 75 6 x 10 Hopf bifurcation: oscillatory large amplitude solutions for certain delay K nonlinear mode Smaller nonlinear oscillatory Hopf bifurcation solutions unstable Ω Figure 3: small amplitude Larger nonlinear oscillatory nonlinear mode (stable) solutions bi-stable with x=0 zero solution solution shape of the stability boundary implies that for certain control parameter (c ) t values of τ there is a larger rang 1.53.7 3.6 3.5 1 3.4 0 3.3 0.53.2 3.1 3 0 700 800 900 1000 1100 1200 rpm The stability !0.5 boundary for the zero solution of the linearized deterministic equation δ = 0; that is, there is a stable equilibrium at x = 0 for parameter values below these curve a series of curves, due to the periodicity of the functions in (3.11) with λ = iω. The top figu !1 of the non-dimensional parameters c and τ . The bottom figure sho stability boundary in terms 1 1.75 boundary for the dimensional parameters K0 (N/m ) and Ωrpm (rpm). Here Ωrpm = 60 · Ω/ for consistency with graphs given in previous references, for example [12]. !1.5 2200 1 Wednesday, June 6, 2012 2250 2300 2350 2400 2450 2500 which the steady state x = 0 is linearly stable. Since the delay τ is inversely proport the angular velocity of the rotating workpiece, then there may be certain preferred valu Variation in material parameters: noise source amplifies oscillations through CR x 0.8 0.6 0.4 0.2 0 Chatter sustained by noise in subthre !0.2 Computation of probability density for am !0.4 p(x) !0.6 !0.8 700 750 800 850 2.5 900 2 t p(x) 1.5 1 Probability density for x (x=0 is fixed point) 0.5 0 !1.5 !1 !0.5 0 x 0.5 1 x Wednesday, June 6, 2012 1.5 FIGURES SDDE: (additive noise) dx = y dt, dx = f (x(t), x(t − τ ); c1 ) dt + δdw τ ∼ τc + �2 τ2 or c1 = c1 c + �2 c1 2 T is a slow time T = �2 t, 0.13 material c parameter 1 Look for solution of the form 0.125 0.12 0.115 0.11 x = 0 stable 0.105 0.1 35 x̂(t) = A(T )cos bt + B(T )sin bt 40 45 50 55 τ 6 3.8 60 delay 65 x 10 Information from the deterministic problem: on t time scale 3.6 Derive equations for noisy amplitudes on T time scaleK1 0 @ dA dB 1 A = 0 @ ψA ψB 1 0 A dT + Σ @ Σ = constant matrix for added noise 3.4 3.2 dβA (T ) 3 700 A 800 dβB (T ) x = 0 stable 900 1000 1100 1200 Ωrpm Figure 1: The stability boundary for the zero solution for the linearized deterministic equatio Klosek, K. 2005 δ = 0. The equilibrium at x = 0 is stable for parameter values below these curves. The series o result of the periodicity of eλt in (2.10) for λ = iω. Find ψA , ψB and Σ by equating two equations for dx: Ito’s formula: change of variables from x to A, B Substitution in the SDDE (involves A(T ), A(T Wednesday, June 6, 2012 1 − �2 τ ) , etc.) Envelope equations for oscillations: ditional terms in (3.12) related to the multiplicative noise, but they are not included here, since K, 2006 he form of the equations for A and B is arean order of magnitude ey describe effects that smaller than those due to the additive noise δ " 1. dA ψA (A, B) = δ dT + $ σ11 dξ1 (T ) σ12 dB B)in detail in [19]. The dξB) σ21 form σ22 of ψA (A, B (A, 2 (Tis) are given The coefficients ψA and ψB ψ , (3.12) amplification ( ' ξj (t), j = 1, 2 independent standard the deterministic B(TFor − $parameters τ ) − B(T ) infactor A(T − $ τBrownian ) − A(T ) motions. + g (ω, κ) ψ(A, B) = c g (ω, κ) c 1 2 2 1 2 $2 $2 q = 0) stability region for x = 0, the amplitude (A, B) of the vibrations decay on the T + c12 [h1 (ω, κ)A(T ) + h2 (ω, κ)B(T )] 0.12 (3.13) Compare probability density p(x) 0.1 scale. However, for δ != 0 the envelope of the vibrationsforhas standard deviation on the x(t)aand noisec1 2 τ ), B(T −$2 τ ), d similarly for ψB . Note that these coefficients depend on A(T ), B(T ) and A(T −$subthreshold x̂(t) = A(T )cos bt + B(T )sin bt driven oscillations of δ/$ (σ11 and σ22 are O(κ) while σ12 and σ21 are O(1)), indicating that2 the amplification c1 =time c1 c + � cT1 2 .this at is, they also depend on the delayed value of the envelope. On Here the slow scale Time scale of coherence resonance (CR): T (slow) 0.8 versely proportional to the square root of the proximity to the stability boundary. Figure 2 rresponds to a small delay. x(t) 5 0.10 4 0 0.4 graphs of the invariant density p(x) (for large time) for different values of c1 , showing the 3.5 0.2 3 p(x) ase in the variance as the parameter values approach the stability boundary. One can include 0 2.5 2 !0.2 Time scale of ional terms in (3.12) related to the multiplicative noise, but they are not included here, since 8 oscillations: t 1.5 !0.4 1 !0.6 0.5 t than those due to the additive describe effects thatBuckwar,K. are an order of magnitude smaller noise x , L’esperance, Soo, 2006 Wednesday, June 6, 2012 0.1 4.5 0.6 !0.8 0.11 700 750 800 850 900 0 !0.5 !0.4 !0.3 !0.2 !0.1 0 0.1 0.2 0.3 0.4 0.5 chip tool previous cut nonlinear force present cut dx = ydt dy = (−2κy − x + cj [x(t − τ ) − x(t)]j )dt 1 damped linear oscillator + 3 X η(1 + 3 X cj [x(t − τ ) − x(t)]j ) 1 Key parameters: variation in material x = fraction of variation from desired chip thickness Oscillations of tool position (about equilibrium) c1 = material parameter, κ = damping (fixed) Is x(t) = 0 stable? i.e. is a straight cut stable? τ = inversely proportional to rotation velocity Noise sources: δ = percentage variation of material parameter, η = δdw Material parameters (additive + multiplicative noise) Speed of rotation (delay) Wednesday, June 6, 2012 Results for variation of delay w/o variation of material parameters Ackno The NSF G sional sity of Nome Without CR: eigenvalues of discretized problem (Yilmaz, 2003) Ak am % d & f '(t) ( Kc k M (t;z) ) p max Sinusoidal variation of delay: Must be tuned precisely for damping (Namachchivaya, et al, 2003) p(i) so s(t) t u(m) * bw *n x(t) z + Refer Wednesday, June 6, 2012 Fig. 12 MRSSV hold duration, z, effect in chatter suppression #1$ Sl in x 0.4 Random switching of speed (delay) damps coherent oscillations −1 τ (w/o variation in material parameters) 0.2 x 0 !0.2 !0.4 0 1 2 3 4 5 6 7 8 9 10 5 6 7 8 9 10 s t 900 800 Ωrpm 700 600 500 0 1 2 3 4 t s • Noise-induced order via transients • Switching between exponential Figure 3: Top: The time series for (2.6) with δ = 0, τ = 66, c1 = .115 and q = 0 (dotted line), q = .25 (solid line). Bottom: Variation in Ωrpm corresponding to the time series in the top figure for q = .25. FIGURES damping and slow growth near a stability boundary 0.13 • Apparent controlled oscillations/quiescence is a sequence of transients Wednesday, June 6, 2012 0.125 0.12 c1 15 0.115 0.11 x = 0 stable 0.105 0.1 35 40 45 50 τ 6 3.8 3.6 x 10 55 60 65 70 ditional terms in (3.12) related to the multiplicative noise, but they are not included here, since Both types of noise: Envelope equations for oscillations he form of the equations for A and B is y describe effects that are an order of magnitude smaller than those due to the additive noise dA ψA (A, B) σ11 σ12 dξ1 (T ) δ = dT + , $ (A, The coefficients dBψA and ψB ψBare (A,given B) in detail in [19]. σ21 The σ22 form ofdξψ2A(T ) B) is δ " 1. ' ξj (t), j = 1, 2 independent c standard A(T −Brownian $2 τ ) − A(Tmotions. ) ψ(A, B) = c1 g1 (ω, κ) $2 (3.12) ( For parameters in the B(T − $2 τ ) − B(T ) + g2 (ω, κ) $2 deterministic q = 0) stability region for x = 0, the amplitude (A, B) of the vibrations decay on the T + c12 [h1 (ω, κ)A(T ) + h2 (ω, κ)B(T )] (3.13) scale. However, for δ != 0 the envelope of the vibrations has a standard on the K. deviation 2006 Time scale of coherence resonance (CR): T at is, they also dependof on oscillations: the delayed value of the envelope. On the slow time scale T this Time scale t versely proportional to the square root of the proximity to the stability boundary. Figure 2 Timeto ainterval responds small delay. for varying delay: L (discrete r.v.) d similarly for ψB . Note that these coefficients depend on A(T ), B(T ) and A(T −$2 τ ), B(T −$2 τ ), of δ/$ (σ11 and σ22 are O(κ) while σ12 and σ21 are O(1)), indicating that the amplification graphs of the invariant density p(x) (for large time) for different values of c1 , showing the ase in the variance as the parameter values approach the stability boundary. One can include Variance of resonant mode on intermediate scale: 8 Validthatfor large describe effects are L an sufficiently order of magnitude smaller than those due to the additive noise ional terms in (3.12) related to the multiplicative noise, but they are not included here, since Wednesday, June 6, 2012 19 CROSSING BOUNDARIES OF DELAY SYSTEMS IN LEUKEMIA Complicated stability regions not uncommon for DDE’s: Controlled container crane, Erneux, Kalmar-Nagy, 2008 FIGURES 1 −5 0.8 0.2 0.6 0.19 0.185 0.18 0.175 Immune response to leukemia treatment: (a) delays for cell division/interaction q0(ρ, jω) + q1(ρ,jω) 0.195 ρ(ω) x 10 0.4 0.13 0.2 0.125 0 0.12 −0.2 c1 −0.4 0.115 0.17 −0.6 0.11 0.165 −0.8 0.105 0.16 0 0.2 0.4 0.6 0.8 −1 0 1 ω x = 0 stable 0.002 0.10.004 35 0.006 6 3.8 x 10 0.008 40 ρ (b) 0.01 45 0.012 0.014 50 55 60 65 70 τ BIFURCATION ANALYSIS OF A MOD Delayed response for therapies: Fig. 4.1. (a) Graph of the stability crossing point ω(ρ) as a function of the time delay ρ based as a function of ρ. on the implicit relation (4.3). (b) Graph of q0 (ρ, jω(ρ)) + q1 (ρ, jω(ρ))e−jω(ρ)ρ 3.6 We are interested in when the expression equals 0, thus satisfying (4.2). 100 K1 0.015 3.4 0.01 3.2 90 80 x = 0 stable 70 0.005 3 60 700 800 900 1000 1100 1200 1300 Ωrpm 0 ∼ τ ∼ τ 50 stable region 40 −0.005 Figure 1: The stability boundary for the zero solution for the linearized deterministic equation (2.6) with 30 δ = 0. The equilibrium at x = 0 is stable for parameter values below these curves. The series of curves is a −0.01of the periodicity of eλt in (2.10) for λ = iω. result 20 10 0 0 20 40 60 80 σ (a) Niculescu, et al, 2008 100 −0.015 −0.02 Colijn, Mackey, (b) 2007 −0.01 0 0.01 σ Fig. 4.2. Crossing curves with respect to σ and τ̃ when υ̃ = 0. (a) The zoomed-out plot shows the overall structure of the crossing curves. (b) The zoomed-in plot shows the small region of stability around the origin. Note that only nonnegative delays make sense. Wednesday, June 6, 2012 0.02 Hematopoietic system: dq dt dn Neutrophil dt ) dp Platelets dt dr dt Stem cells = −qb1 hq (q) + b1 µ1 q1 hq (q1 ) − q {b2 hn (n) + b3 hp (p) + b4 hr (r)} , = −γn n + an b2 qτnm hn (nτnm ), ! " = −γp p + ap b3 qτpm hp (pτpm ) − µ3 qτpsum hp (pτpsum ) , = −γr r + ar b4 {qτrm hr (rτrm ) − µ4 qτrsum hr (rτrsum )} , N ANALYSIS OF A MODEL OF HEMATOPOIESIS 383 here q, n, r, and p are nondimensional stem cells, neutrophils, erythrocytes, an 15 spectively. Subscripts indicate delays: q1 = q(t − 1) and so on. The functions hq , hn , hr , and hp are Hill functions given by 10 5 0 20 30 40 50 60 70 9 8 Steady-state and periodic oscillations in the decoupled stem cell compartment. Where the steadyshown as aJune thick line (left side), it is stable, and7where it is thin (right side), it is unstable. The Wednesday, 6, 2012 80 90 100 110 120 2. The model. A nondimensional model of the hematopoietic production system is by Hematopoietic Stem cells Neutrophils (2) Platelets Erythrocytes dq dt dn dt dp dt dr dt system: = −qb1 hq (q) + b1 µ1 q1 hq (q1 ) − q {b2 hn (n) + b3 hp (p) + b4 hr (r)} , = −γn n + an b2 qτnm hn (nτnm ), ! " = −γp p + ap b3 qτpm hp (pτpm ) − µ3 qτpsum hp (pτpsum ) , = −γr r + ar b4 {qτrm hr (rτrm ) − µ4 qτrsum hr (rτrsum )} , where q, n, r, and are nondimensional stem cells, neutrophils, Colijn,erythrocytes, Mackey, 2007and pla Fluctuations inpproliferation respectively. Subscripts indicate delays: q1 = q(t − 1) and so on. times in lower Theand functions hq , hn , hcell r , and hp are Hill functions given by population levels can drive transitions to large oscillations (disease): Weak attraction to healthy (steady) state (slow damping) 10 8 6 4 2 0 40 60 80 100 120 140 160 180 120 140 160 180 10 8 6 4 2 0 Wednesday, June 6, 2012 200 220 240 260 2. The model. A nondimensional model of the hematopoietic production system is given by Hematopoietic system: dq Stem cells dt dn dt (2) Platelets dp dt dr dt = −qb1 hq (q) + b1 µ1 q1 hq (q1 ) − q {b2 hn (n) + b3 hp (p) + b4 hr (r)} , = −γn n + an b2 qτnm hn (nτnm ), 15 ! " 10 = −γp p + ap b3 qτpm hp (pτpm ) − µ3 qτpsum hp (pτpsum ) , = −γr r + ar b4 {qτrm hr (rτrm ) − µ4 qτrsum hr (rτrsum )} , 5 0 20 30 40 50 60 70 80 90 100 110 120 9 where q, n, r, and p are nondimensional stem cells, neutrophils, erythrocytes, and platelets, respectively. Subscripts indicate q1 =OF q(tHEMATOPOIESIS − 1) and so on. BIFURCATION ANALYSIS OF delays: A MODEL The functions hq , hn , hr , and hp are Hill functions given by 8 7 Variable delay (production) 39 6 Platelet amplification 5 4 3 15 cell ( #): 0 50 100 150 10 5 0 30 delay 2 40 50 60 70 80 90 100 110 120 130 40 50 60 70 80 90 100 110 120 130 2.5 2 1.5 1 0.5 Colijn, Mackey, 2007 Wednesday, June 6, 2012 0 30 t Balance model: θ̇ φ̇ Stability diagram for PD control (w/delay), control always on 202 (b) = = φ sin θ − F (θ, θ̇) cos θ F = aθ(t − τ ) + bθ̇(t − τ ) G. ST&PAN AND L. KOLL~ Hopf bifurcation Pitchfork bifurcation (a) ” mg Figure 2. Stability chart.D-shaped region θ = 0 is stable in shaded area, Pitchfork bifurcation: stability of non-zero fixed point Hopf bifurcation, stability ofof oscillatory The corresponding stability chart in the plane the gain parameters behavior P, D for constant delays can be proved by checking the limit condition Pmin = Pm=(w) for the existence of any stability domain for P in (6). 71 < 72 < * ** is presented qualitatively in Figure 2. The encircled numbers show the number of Wednesday, June 6, 2012 se of the controllerp (solid ffiffiffiffiffiffiffiffiffiffiline), ffiffiffiffiffiffiffiffiffiffiffiand ffiffiffi c) plots the position of the target (dashed line) d c), DðxðtÞ,yðtÞÞ~ x2 ðtÞzy2 ðtÞ is the length of the position vector measured a ively, by the Qualisys motion capture system and the computer program. No ambig ment angle is small (see Figure 7d) and the movements of the fingertip and tip o What about other effects? On/off control, together with delays: Below the threshold: Control is off (growth) Above the threshold: Control is on (decay) bserved only when vertical with active control (‘‘act’’) taken only on System moves back and forth across the threshold certain thresholds [7,10,13,22,23,25–27 rtip, is associated with a small Balancing with Vibration generic model with ‘‘drift–and–act’’ con produces no changes in the F (x(t − τ )) with parametric excitation takes the form made by the fingertip. Taken that the skill enhancement is Toy model ertip and not to the effects of dx ~F ðxðt{tÞÞxðtÞzkxðtÞsin o–skeletal system. We suggest dt unexpected observation is to where k is a constant, f is the forcing f ced position is not a simple delay, xðMilton, tÞ,xðt{tÞetare, the lex bounded time–dependent al, respectively, 2009 variable at times t and t{t, and jðtÞ d of attraction whose size is of uently, for sufficiently large noise with variance g2 . The feedback funct If fixed point is and not the reached, periodic behavior he basin of attraction, step–like shape shown incontinues Figure 9a. Mode Calculate conditions to periodic solution g, any mechanism that biasesthat lead successfully employed, for example, to Figure 9. Effects of parametric excitation on the dynamics of a simple ‘‘drift and act’’ controller. a) Graphical representation of a simple realization of the feedback function that produces a limit cycle oscillation in (2) in the absence of parametric excitation and noisy perturbations, where F ðxðt{tÞÞ~ðazbÞ{b=ð1zexpðQðxðt{tÞ{ThÞ and a~0:18, b~{0:20, Q~500, and Th~1. The displacement from the upright position, xðt{tÞ, grows when xðt{tÞvTh and decreases when xðt{tÞwTh. b) Periodic parametric excitation is turned on at the ;. The effect is to decrease the amplitude of the limit cycle oscillation. Parameters are f ~2 and k~0:14. doi:10.1371/journal.pone.0007427.g009 Wednesday, June h i 6, 2012 What about other effects? On/off control: Observations indicate control on only for large enough variation from upright position Asai, etal, 2009, PLOS1 Figure 8. Sample of human postural sway, collected from a subject in quiet standing for 120 s. Left upper pane Left lower panel: trajectory in the phase plane; Right panel: power spectral density of the angular sway. doi:10.1371/journal.pone.0006169.g008 Effects of delays: Delays can naturally introduce oscillations Noise-free DDE simulations of Model 4 for various sets of P-D-a parameters show that the stable regions in the P-D-a parameter space of Model 4 are the same as those of Model 3, with the difference that the attractor of Model 3 is an asymptotically stable equilibrium point and that of Model 4 is a limit cycle. This fact provides a common basis for understanding the noisy dynamics of Wednesday, June 6, 2012 both Models 3 and 4. As shown in Fig. 7, the power spectra of How do we analyze these? mgh<0.79 for the right panel) and commo (D = 10 Nms/rad, a = 20.4 s21, s = 0.2 the switching parameter a, the optimal va of Model 3 was about 60% of mgh regard the dark band was located at P=mgh&0 values of P used for Fig. 9 left and right a and larger than the optimal value of P. A we also explain the center manifold reduction and present the bifurcation diagram of the vector field model near the triple-zero eigenvalue bifurcation. In Section 3, we show that th dynamics in the reduced model, and in Section 4 we give examples of complex balancin system. Finally, in Section 5 we conclude and point to some open problems. Feedback control in balance models: 2. Model equations PD control with delay or θ̇ φ̇ = = φ sin θ − F (θ, θ̇) cos θ F = aθ(t − τ ) + bθ̇(t − τ ) The dynamics of the setup in Fig. 1 are governed by the second-order differentia displacement θ of the pendulum, which can be written as (see also Refs. [2,9,17]): ! " 3m 3m 2 3g 3F 1− cos2 θ θ̈ + θ̇ sin(2θ) − sin θ + cos θ = 0. 4 8 2L 2L(Mp + Mc ) where F is the horizontal driving force applied to the cart, and m = Mp /(Mp + Mc ) is uniform pendulum Mp with respect to the sum of the mass of the cart Mc and the mass o of the pendulum is L and g is the gravitational acceleration. We adjust the time scale Asai, etal, 2009, PLOS1 of the pendulum by rescaling time, introducing a new dimensionless time tnew = told · State-dependent on/off control: control is on in state drifting away from origin q q OFF s e ON e Ws W Fig. 1. Sketch of the inverted pendulum of length L and mass Mp on a cart of mass Mc at horizontal position δ. such that it is positive in the position shown in the sketch. ON Wednesday, June 6, 2012 OFF or field model near the triple-zero eigenvalue bifurcation. In Section 3, we show that there are complex symmetric amics in the reduced model, and in Section 4 we give examples of complex balancing motion in the full control em. Finally, in Section 5 we conclude and point to some open problems. Back to the balance models: PD control with delay θ̈ − sin θ + F (θ, θ̇) cos θ = 0 On/Off Control : F = aθ(t τ )1 + bθ̇(t − τ )second-order differential equation he dynamics of the setup − in Fig. are governed by the for the angular Note delay in Model equations or lacement θ of the pendulum, which can be written as (see also Refs. [2,9,17]): ! " 3m 3m 2 3g 3F 1− cos2 θ θ̈ + θ̇ sin(2θ) − sin θ + cos θ = 0. 4 8 2L 2L(Mp + Mc ) switching from on/off, (1) system enters off of the massbefore of the re F is the horizontal driving force applied to the cart, and m = M /(M + M ) is the ratio region control orm pendulum M with respect to the sum of the mass of the cart M and the mass of the pendulum. The length Less control: he pendulum is L and g is zig-zag the gravitational acceleration. We adjust the time scale to intrinsic time scale off switched √ √ theis he pendulum by rescaling time, introducing a new dimensionless time t = t · 3g/ 2L. This converts p p p c behavior Intermittent Postural Control new old ON ON OFF c . Sketch of the inverted pendulum of length L and mass Mp that it is positive in the position shown in the sketch. Stronger control: OFF on a cart of mass M at horizontal position δ. The inclination angle θ is defined yields spiral c on/off switch Wednesday, 6, 2012 gure 3. TypicalJune solutions of Model 3 in the phase plane. In each plane, the initial state at t = 0 is represented by the thick curve segment Asai, etal, 2009, PLOS1 Periodic behavior: θ̈ − sin θ + F (θ, θ̇) cos θ = 0 F = aθ(t − τ ) + bθ̇(t − τ ) PD control with delay Zig-zag orbits - periodic solutions away from origin, moving back and forth from on/off θ̇ " (!0,"0) t=0 $ t = !# ON ! t = Tint t = Tint + # (!1,"1) OFF % 1 Figure 6: The forward orbit, Γ, of a point on Σ1 . The primary purpose of this figure is to illustrate some labels used in expansions given the text. Unlike elsewhere in this Note, t = 0 corresponds to a switching point instead of a point on Σ1 . Wednesday, June 6, 2012 Periodic behavior: zig-zag orbits θ̈ − sin θ + F (θ, θ̇) cos θ = 0 F = aθ(t − τ ) + bθ̇(t − τ ) Calculating Zig-zag orbits: •Calculate solution in OFF region (no delays in equation) •Calculate solution in ON region, using off solution as (Asymptotic analysis) initial history, ON " θ̇ = φ •Approximate with small $ delays (next page) •Does the value at end of % on/off switch match with OFF initial condition in off position? Figure 6: The forward orbit, Γ, of a point on Σ . The primary purpose o is to illustrate some labels used in expansions given the text. Unlike elsew If yes, then periodic (!0,"0) t=0 t = !# t = Tint t = Tint + # (!1,"1) 1 Note, t = 0 corresponds to a switching point instead of a point on Σ1 . Wednesday, June 6, 2012 and assuming, without loss of generality, that this point corresponds to t = Some calculations: F = aθ(t − τ ) + bθ̇(t − τ ) θ̈ − sin θ + F (θ, θ̇) cos θ = 0 " ON (!0,"0) t=0 $ t = !# OFF ! t = Tint % t = Tint + # (!1,"1) 1 where The forward orbit, Γ, of a point on Σ1 . The primary purpose of this figure strate some labels used in expansions given the text. Unlike elsewhere in this 3.2 Series in s instead and τ of a point on Σ . = 0 corresponds to a expansions switching point 1 β4 = − bθ0 cos(θ0 ) , 2 1 0 2<(θ0θ)0, < π2 , β̂5 = − abθ0 cos # 2 2 1 − <0 )s. ≤ 0 , β6 = − b sin(θ0 ) cos(θ π 6 # $ 1 (θ20 ) < 0 . 1 β̂6 = β6 + abθF 0 cos (θ0 ) = − b cos(θ0 ) sin(θ0 ) − aθ0 cos(θ0 ) . 6 6 (63) (64)(29) 21 and ensures that the control is sufficiently strong. e last inequality refers (21) It may be showntothat n that φ(−τ ) = sθ(−τ )% for the orbit 2Γ, the value φ0 is completely determined ξ1 τ + ξ2 sτ + ξ3 τ + O(|s, τ |3) , φ(τ ) − sθ(τ ) ≤ 0 T = , (67)it is int alue of θ0 , so we first derive relation between φ0 and θ0 . From (4), 3 ˆ 2 ξ1 τ + ξthe 2 sτ + ξ3 τ + O(|s, τ | ) , φ(τ ) − sθ(τ ) ≥ 0 rified that21 for t ∈ [−τ, 0], 6 ully this isn’t too confusing in that we’ve been thinking of initial conditions as points on Σ 1 . hat it is somewhat superfluous to also write φ(t) because φ(t) = θ̇(t). Wednesday, June 6, 2012 13 "3 . 2 a2 θ02 "2 ∆H = # . ζ2 ζ3 ζ5 ζ̂4 1 b sin3 (θ0 ) cos(θ0 ) 2 "3 . sin(θ0 ) − aθ0 cos(θ0 ) 1 sin2 (θ0 ) − 3aθ0 sin(θ0 ) cos(θ0 ) + a2 θ02 cos2 (θ0 ) b cos(θ0 ) . ξˆ3 = ! "2 2 sin(θ0 ) − aθ0 cos(θ0 ) Also, it is useful to introduce (72) ξ˜1 = ξ1 + 1 = −aθ0 cos(θ0 ) . sin(θ0 ) − aθ0 cos(θ0 ) Then 2 2 2 3 4 ζ1 sτ + ζ2 τ + ζ3 s τ + ζ4 sτ + ζ5 τ + O(|s, τ | ) , Tint ≤ τ ζ1 sτ + ζ2 τ 2 + ζ3 s2 τ + ζ̂4 sτ 2 + ζ̂5 τ 3 + O(|s, τ |4) , Tint ≥ τ , (73) ∆H = # ζ1 sτ + ζ2 τ 2 + ζ3 s2 τ + ζ4 sτ 2 + ζ5 τ 3 + O(|s, τ |4) , Tint ≤ τ ζ1 sτ + ζ2 τ 2 + ζ3 s2 τ + ζ̂4 sτ 2 + ζ̂5 τ 3 + O(|s, τ |4) , Tint ≥ τ , where a2 θ03 cos2 (θ0 ) , sin(θ0 ) − aθ0 cos(θ0 ) sin(θ0 ) − 21 aθ0 cos(θ0 ) , = a2 θ02 cos2 (θ0 ) sin(θ0 ) − aθ0 cos(θ0 ) sin(θ0 ) − 21 aθ0 cos(θ0 ) = 2abθ03 cos2 (θ0 ) ! "2 , sin(θ0 ) − aθ0 cos(θ0 ) ζ4 = ξ3 = ! (71) where ζ1 = ξ2 2 cos (θ0 ) − sin(θ0 ) , sin(θ0 ) − aθ0 cos(θ0 ) −bθ0 sin(θ0 ) cos(θ0 ) = ! "2 , sin(θ0 ) − aθ0 cos(θ0 ) ξ1 = (70) sin(θ0 ) − aθ0 cos(θ0 ) Then Continuity of Tint may be verified at φ(τ ) − sθ(τ ) = 0 as follows: When φ(τ ) − sθ(τ ) = 0, by (59), 0) 0) 2 a = 2 tan(θ −bs− b sin(θ ). With this value of a, ξ1 +ξ2 s+ξ3 τ = ξ1 +ξ2 s+ ξ̂3 τ = 1+O(|s, τ |2 ) θ0 2θ0 τ +O(|s, ζ̂5 1= τ byτ | definition. 1 as expected, since here T int θ(t) = θ0 + φ0 t + sin(θ0 )t2 + φ0 cos(θ0 )t3 + O(t4 )16 . (32) 2 1 b sin3 (θ0 ) cos(θ0 ) 2 −aθ0 cos(θ0 ) . ξ˜1 = ξ1 + 1 = sin(θ0 ) − aθ0 cos(θ0 ) (65)(30) (66)(31) (69) Also, it is useful to introduce 1 2 mary goal motivating ourabθanalysis in the β̂0 = − 0 cos (θ0 ) , is the determination of the change (60) 6 nian, (5), of Γ between1 t = 02 and t = Tint + τ ; this will indicate whether the abθ0 cos (θ0 ) , (61) β̂2 = 2away from the origin. pproaching or moving " 1! is point we makeβ3the assumptions sin(θ0 ) − = following aθ0 cos(θ0 ) , on a, b, s and θ0 and as we progress (62) 2 the analysis the role of each assumption should become clear: 1 ξ2 1 sin (θ0 ) − 3aθ0 sin(θ0 ) cos(θ0 ) + b cos(θ0 ) ξˆ3 = ! 2 sin(θ0 ) − aθ0 cos(θ0 ) 15 %of generality, & that this point corresponds to t = 0 . We let ming, without loss θ0 + %sθ0 + sin(θ0 )τ &t + %12 sin(θ0 )t2& + O(|s, τ, t|4 ) , t ∈ [−τ, 0] 2 3 te the smallest positive intersects Σ1 .τ,Then sin(θ0 )τatt which + β3 + βΓ t|4 ) , thet next ∈ [0, τ ]switching θ0 + sθ0 +time 4 s &t +%β6 t + O(|s, & % θ(t) = 3 + sθ + sin(θ )τ + β̂ τ 2 t + β + β s + β̂ τ t2 + β̂ t3 + O(|s, τ, t|4 ) , + β̂ τ θ 3 4 5 6 0 0 0 0 2 r at the point t ∈ [τ, Tint + τ ] ! " (59) (28) (θ1 , φ1 ) = θ(Tint + τ ), φ(Tint + τ ) . where (68) ξ3 = ! Below we will need higher order terms of (47). To prevent calculations and statements becoming too messy we find it useful to perform series expansions in both s and τ . Thus we rewrite (46) as: where − sin(θ0 ) , sin(θ0 ) − aθ0 cos(θ0 ) −bθ0 sin(θ0 ) cos(θ0 ) = ! "2 , sin(θ0 ) − aθ0 cos(θ0 ) ξ1 = sin 2abθ02 cos2 (θ0 ) 3 (θ0 ) − 11 2 4 aθ0 sin (θ0 ) cos(θ0 ) + 2a2 θ02 sin(θ0 ) cos2 (θ0 ) − 12 a3 θ03 cos3 (θ0 ) , "3 ! sin(θ0 ) − aθ0 cos(θ0 ) a2 θ03 cos2 (θ0 ) ζ1 = , (74) sin(θ0 ) − aθ0 cos(θ0 ) sin(θ0 ) − 21 aθ0 cos(θ0 ) , ζ2 = a2 θ02 cos2 (θ0 ) (75) sin(θ0 ) − aθ0 cos(θ0 ) sin(θ0 ) − 21 aθ0 cos(θ0 ) ζ3 = 2abθ03 cos2 (θ0 ) ! "2 , (76) sin(θ0 ) − aθ0 cos(θ0 ) ζ4 = 2abθ02 cos2 (θ0 ) (77) sin3 (θ0 ) − 11 2 4 aθ0 sin (θ0 ) cos(θ0 ) + ! 2a2 θ02 sin(θ0 ) cos2 (θ0 ) − 12 a3 θ03 "3 sin(θ0 ) − aθ0 cos(θ0 ) 1 sin3 (θ0 ) − 6aθ0 sin2 (θ0 ) cos(θ0 ) + 8a2 θ02 sin(θ0 ) cos2 (θ0 ) − 2 "3 ! 1 sin3 (θ0 ) − 6aθ0 sin2 (θ0 ) cos(θ0 ) + 8a2 θ02 sin(θ0 ) cos2 (θ0 ) − 3a3 θ03 cos3 (θ0ζ)5 = 6 abθ0 sin(θ0 ) cos (θ0 ) 2 = (78), abθ0 sin(θ0 ) cos (θ0 ) ! "3 sin(θ0 ) − aθ0 cos(θ0 ) 6 sin(θ0 ) − aθ0 cos(θ0 ) sin2 (θ0 ) − 43 aθ0 sin(θ0 ) cos(θ0 ) + 14 a2 θ02 cos2 (θ0 ) 3 1 2 2 2 2 , ζ̂4 = 2abθ02 cos2 (θ0 ) ! "2 sin (θ ) − aθ sin(θ ) cos(θ ) + 0 0 0 0 2 2 4 4 a θ0 cos (θ0 ) = 2abθ0 cos (θ0 ) , (79) sin(θ0 ) − aθ0 cos(θ0 ) "2 ! sin(θ0 ) − aθ0 cos(θ0 ) 1 sin3 (θ0 ) + 7aθ0 sin2 (θ0 ) cos(θ0 ) − 9a2 θ02 sin(θ0 ) cos2 (θ0 ) + 3a3 θ03 c abθ0 cos2 (θ0 ) ζ̂5 = "2 ! 3 2 2 θ 2 sin(θ ) cos 2 (θ ) + 3a3 θ 3 cos3 (θ ) (θ ) + 7aθ sin (θ ) cos(θ ) − 9a 1 sin 6 0 0 0 0 0 0 0 0 0 sin(θ0 ) − aθ0 cos(θ0 ) = . (80) abθ0 cos2 (θ0 ) "2 ! 6 sin(θ0 ) − aθ0 cos(θ0 ) 18 18 Back to the balance models: (w/ cart) 5 4 PD control with delay !!/s super DIB F =3 aθ(t − τ ) + bθ̇(t − τ ) SN (!0,"0) " 2 ON t=0 $ t = !# PF 1 OFF 0 0.5 1 1.5 a 2 ! t = Tint % t = Tint + # (!1,"1) 1 zig-zag orbits 2.5 Figure 6: The forward orbit, Γ, of a point on Σ1 . The primary purpose of this figure is to illustrate some labels used in expansions given the text. Unlike elsewhere in this Note, t = 0 corresponds to a switching point instead of a point on Σ1 . B HC 1 U " 0.5 0 0.5 super DIB PF 1 and assuming, without loss of generality, that this point corresponds to t = 015 . We let Tint denote the smallest positive time at which Γ intersects Σ1 . Then the next switching will occur at the point ! " (28) (θ1 , φ1 ) = θ(Tint + τ ), φ(Tint + τ ) . U S sub DIB 1.5 a The primary goal motivating our analysis is the determination of the change in the Hamiltonian, (5), of Γ between t = 0 and t = Tint + τ ; this will indicate whether the orbit is approaching or moving away from the origin. At this point we make the following assumptions on a, b, s and θ0 and as we progress through the analysis the role of each assumption should become clear: Zigzag orbits S 2 0 < θ0 < π2 , # − π2 < s ≤ 0 , 2.5 Bifurcation diagram: control a vs. θ F (θ0 ) < 0 . Wednesday, June 6, 2012 (29) (30) (31) where the last inequality refers to (21) and ensures that the control is sufficiently strong. Given that φ(−τ ) = sθ(−τ ) for the orbit Γ, the value φ0 is completely determined Simpson, K., Li, J. Nonl Sci, to appear A A 7 7 HC delay 6 5 HC 6 sub DIB sub DIB State diagram: control vs. delay 5 4 !!/s !!/s4 3 super super DIB DIB 3 SN 2 2 1 0 0.5 B B a $ 1 " 0.5 1.5 1 2 a HC ! HC t = Tint U S t = Tint + # (!1,"1) 0.5 2 1.5 control 2.5 2.5 t=0 t = !# 1 PF 1 (!0,"0) " zig-zag orbits PF 1 0 0.5 " SN % S U 1 ure 6: The forward orbit, Γ, of a point on Σ1 . The primary purpose ofsub this figure super super sub o illustrate some labels used in expansionsPF given the text. Unlike elsewhere DIB DIBin this DIB 0 to a switching point instead ofPF e, t = 0 corresponds a pointDIB on Σ1 . 0 0.5 1 1.5 0.5 1 1.5 a a assuming, without loss of generality, that this point corresponds to t = 015 . We let denote the smallest positive time at which Γ intersects Σ1 . Then the next switching occur at the point ! " (28) (θ1 , φ1 ) = θ(Tint + τ ), φ(Tint + τ ) . Wednesday, June 6, 2012 S 2 2 2.5 2.5 θ Bifurcation diagram: control a vs. Panel A is a bifurcation set of (4)-(6) when s = −0.01, b = 2 and G(θ) = 4: ure 4: Panel A is a bifurcation set of (4)-(6) when s = −0.01, b = 2 and G(θ) = Mathematical Challenges: A 7 6 5 • Nonlinearity: non-uniqueness, complex dynamics • Delays: Delay Differential Equations (DDEs) HC sub DIB instead of Ordinary Differential Equations (ODEs), functional/history dependence instead of initial conditions at a single point 4 !!/s super DIB 3 SN 2 1 • On/off control: discontinuous equations rather 0 0.5 B PF than continuous or smooth behavior. a Biophysical model: impact of noise? U U S S a 1 1.5 2 2.5 HC 1 " 0.5 0 0.5 super DIB PF 1 sub DIB 1.5 2 2.5 Bifurcation diagram: control a vs. : Wednesday, Panel June A 6,is2012 a bifurcation set of (4)-(6) when s = −0.01, b = 2 and G(θ) = θ Balance model w/ noise: small random fluctuation A For larger control: oscillations near the origin: weak attraction to the origin 7 0.2 HC 6 0.15 OFF 0.1 5 ON 0.05 4 " sub DIB !!/s 0 super DIB !0.05 3 ON !0.1 SN OFF 2 !0.15 !0.2 1 !0.5 !0.4 !0.3 PF !0.1 !0.2 0 0.5 0 ! 0.1 1 0.2 0.3 0.4 0.5 1.5 2 a B For smaller control, noise can cause transition to zigzag orbit 2.5 HC U 1 " 0.5 0 0.5 super DIB PF 1 U S sub DIB 1.5 S 2 2.5 a Bifurcation diagram: control a vs. Wednesday, June 6, 2012 θ Model 3. In this m intermittently switched mechanism, which div four regions separated and the ordinate axis: Balance model w/ noise: Asai, 2009, PLOS1 Experimental observations ( ( fP (hD )~P hD fD (h_ D )~D h_ D fP (hD )~0 fD (h_ D )~0 , , i:e:, PD c i:e:, PD co Note that the phase D.0 is infinite dimens system at time t is Therefore the h{h_ pl Nevertheless, with kee refer to the h{h_ plane PD-on regions corresp phase plane, augment second quadrants, resp the switching paramet areas of the phase plan between 50% to 100 Figure 1. Characterization of the 4 control models in the phase a?{?, the PD-off plane (h_ vs: h). In Models 1 and 2 the control is active in the whole becomes to ample of human postural sway, collected from a subject in quiet for 1203 s. upper panel: swayidentical sequen plane.standing The shaded areas in Models andLeft 4 identify the areas where Angular the condition for the cont control switched off. anel: trajectory in the phase plane; Right panel: power spectral density ofis the angular sway. Fig. 3: it describes a doi:10.1371/journal.pone.0006169.g001 1.2 journal.pone.0006169.g008 breaking the stability 0.2 equilibrium would be Noise drives Control off near origin A C transitions !0.028 OFF to larger ON values of D, a, and U mgh<0.79 for the right panel) and common e DDE simulations of Model 4 for various sets of P-D-a 0.8 zig-zag(D = 10 Nms/rad, a = 20.4 s21, s = 0.2 Nm). For this value show that the stable regions in the P-D-a parameter !0.03 ! Model 4 are the sameSas those of Model" 3, with the even the switching parameter a, the optimal value of P for the stabi orbits, hat the attractor of Model 3 is an asymptotically stable of Model 3 was about 60% of mgh regardless the value of D, i for 0.4 point and that of Model 4 is a limit cycle. This factlarger the dark band was located at P=mgh&0:6 in Fig. 6C. Thus t !0.032 U the noisy dynamicscontrol common basis for understanding of values of P used for Fig. 9 left and right are, respectively, smal ls 3 and 4. As shown in Fig. 7, the power spectra of and larger thanON the optimal valueOFF of P. As in these two exampl values S the two power nd 04, if affected by noise, exhibit law the showed 0.3 0.31 0.32 PSD 0.33 0.34the two power law scaling for smaller values of 0.15 PLoS ONE | www.plosone.org 3 0.1 0.05 " 0 !0.05 U a physiological sway movements. mes that are typical of 1 1.2 1.4 1.6 B power law scaling factor at the low ar, 0.08 the first egime of Model 4 changes depending on the values of " 0.04 Wednesday, 6, 2012 s as these parameters determine the noise June intensity !0.1 and! it was more like a second order system and similar to the PS of Model 2 with or without a resonance for large values of P. Figure 10 shows, for Model 4, the dependence of the scal factor a of the power spectrum of the noisy sway at the l !0.15 !0.2 !0.5 !0.4 !0.3 !0.2 !0.1 0 ! 0.1 0.2 0.3 0.4 0.5 duced bifurcation was analyzed in [24] (see also [28, 29]) for a general time-delaye ecewise-smooth system comprised of ordinary differential equations on each side Bursting-like Experimental e switching manifold.observations In that paper itoscillations was proved that in a neighbourhood of t A 1 SN 0.8 ! 0.6 0.4 sub DIB 0.2 PF 0 1 1.2 B a Text 1.4 1.6 C 0.32 0.04 " ! 0.28 0 0.24 400 420 440 460 time 480 500 !0.04 0.24 0.28 0.32 ! gure 8: Panel A is a bifurcation diagram of (4)-(6) when τ = 0.3, s = −0.1, b = nd G(θ) = cos(θ). The meanings of the abbreviations and colours are the same as 6, 2012 g.Wednesday, 4-B.JunePanel B shows a partial time series of an orbit with transients decayed wh Biophysical models of balance: • Human postural sway, stick balancing • Nonlinearity: non-uniqueness, complex dynamics • Bifurcations in context of delays and on/off control • Identify sensitivies to control conditions and noise spiral, along the dotted curve the orbit falls onto W s and limits upon the origin without again crossing either of the switching manifolds. The dash-dot curve is a Hopf-like bifurcation of (4) at which a symmetric spiral periodic orbit is born. Note that one may choose the control parameters such that zigzag orbits approach the origin and spiral orbits head away from the origin (as in Fig. 1) and vice-versa. We have not been able to identify an intersection between the dashed and dotted curves of Fig. 9 for any τ and s, nor have been able to show that such an intersection cannot occur. Such an intersection could permit for the existence of an orbit that repeatedly switches between zigzag and spiral motion. 4 Discts control: regions of zigzags/spirals depend on control coeffs spiral out or zigzag in spiral out or zigzag out spiral in or zigzag out b2 1 spiral in or zigzag in zigzag to spiral in spiral to zigzag out D-shaped region (cts control) spiral to zigzag in 1 Wednesday, June 6, 2012 zigzag to spiral out 3 1.5 2 a 2.5 3 3.5 Figure 9: A bifurcation set of the linearization, (56), with (5)-(6), τ = 0.5 and s = 0. Noise driven ZZ-spiral, amplified spirals, stabilized transients 0.4 0.4 0.2 0.2 ON 0 0 ï0.2 ï0.2 ï0.4 OFF ï0.4 ï0.2 0 0.2 0.4 0.6 ï0.4 0.4 0.4 0.2 0.2 0 0 ï0.2 ï0.2 ï0.4 ï0.4 Wednesday, June 6, 2012 ï0.4 ï0.2 0 0.2 0.4 0.6 ï0.4 ï0.2 0 0.2 0.4 0.6 ï0.4 ï0.2 0 0.2 0.4 0.6 Lots of mathematical challenges: • New analytical and computational methods needed • Nonlinear models with delay: complex behaviors • Piecewise continuous nonlinear systems: recently receiving more attention • Stochastic modeling for systems with delay and discontinuities: open problems analytically and computationally • Act and wait: apply control for times just beyond the delay time for stabilization, some suggestions of sensitivies to these optimal control strategies Wednesday, June 6, 2012