Linear systems of ordinary differential equations (This is a draft and preliminary version of the lectures given by Prof. Colin Atkinson FRS on 21st, 22nd and 25th April 2008 at Tecnun) Introduction. .te cn un .es This chapter studies the solution of a system of ordinary differential equations. These kind of problems appear in many physical or chemical models where several variables depend upon the same independent one. For instance, Newton’s second law applied to a particle. m· d2 x1 = F1 (x1 , x2 , x3 , ẋ1 , ẋ2 , ẋ3 , t) dt2 m· d2 x2 = F2 (x1 , x2 , x3 , ẋ1 , ẋ2 , ẋ3 , t) dt2 m· d2 x3 = F3 (x1 , x2 , x3 , ẋ1 , ẋ2 , ẋ3 , t) dt2 (1) Let us consider this system of ordinary differential equations: dx = −3x + 3y dt (2) ww w dy = −xz + rx − y dt dx = xy − z dt where x, y, z are time-dependent variables, and r is a parameter. We find a system of three non-linear ODE’s. The interesting thing in this example is the strong dependence of the solution on the paramater r and the initial value conditions (t0 , x0 , y0 , z0 ). For the above reasons, the system presents as the chaos theory; a very known theory in Mathematics knows as the ”butterfly effect”1 . These equations were posed by Lorenz in the study of metheorology. 1 The phrase refers to the idea that a butterfly’s wings might create tiny changes in the atmosphere that may ultimately alter the path of a tornado or delay, accelerate or even prevent the occurrence of a tornado in a certain location. The flapping wing represents a small change in the initial condition of the system, which causes a chain of events leading to large-scale alterations of events. Had the butterfly not flapped its wings, the trajectory of the system might have been vastly different. While the butterfly does not cause the tornado, the flap of its wings is an essential part of the initial conditions resulting in a tornado. Recurrence, the approximate return of a system towards its initial conditions, together with sensitive dependence on initial conditions are the two main ingredients for chaotic motion. They have the practical consequence of making complex systems, such as the weather, difficult to predict past a certain time range (approximately a week in the case of weather). c 2008 Tecnun (University of Navarra) 1 There is an important relationship between the system of ODE and the ODE’s of any order superior to the first. It is a matter of fact, that an equation of order nth y (n) = F (t, y, y 0 , y 00 , . . . , y (n−1) ) (3) where y (n) = dn y/dtn . This can be converted in a system of n equations of first order. With this change of variables (4) x01 = F1 (t, x1 , x2 , . . . , xn ) x02 = F2 (t, x1 , x2 , . . . , xn ) ... x0n = Fn (t, x1 , x2 , . . . , xn ) (5) .te cn un .es or in general: x1 = y x2 = y 0 ⇔ x2 = x01 ... xn = y (n−1) ⇔ xn = x0n−1 x0n = F (t, x1 , x2 , . . . , xn ) ww w we reach n non-linear ODE’s of first order. There are three questions to be answered: (1) What about the existence of solutions? (2) What about uniqueness? (3) What is the sensitivity to the initial conditions? We are going to see in which cases we can assert that a system of ODE has a solution and this is unique. We must consider the following theorem Theorem. Let us assume that in a region A of the space (t, x1 , x2 , . . . , xn ), the functions F1 , F2 , . . . , Fn and ∂F1 ∂F1 ∂F1 , ,..., ∂x1 ∂x2 ∂xn ∂F2 ∂F2 ∂F2 , ,..., ∂x1 ∂x2 ∂xn ... ∂Fn ∂Fn ∂Fn , ,..., ∂x1 ∂x2 ∂xn (6) are continuous and such that the point (t0 , x01 , x02 , . . . , x0n ) is an interior point of A. Thus there exists an interval |t − t0 | < (7) (local argument) where there is a unique solution of the system given by eq. 5, x1 = Φ1 (t) x2 = Φ2 (t) . . . xn = Φn (t) (8) x02 = Φ2 (t0 ) . . . x0n = Φn (t0 ) (9) that fulfils the initial condition x01 = Φ1 (t0 ) c 2008 Tecnun (University of Navarra) 2 A way to prove this theorem is to use the Taylor’s expansion of the functions. Note.- We must point out that this is a sufficient condition theorem. So, weakening the conditions we can define a stronger expression of this theorem to get a unique solution. The systems are classified in the same manner as the ODE’s. They are linear and non-linear. If the functions F1 , F2 , . . . , Fn can be written as x0i = Pi1 (t) · x1 + Pi2 (t) · x2 + · · · + Pin (t) · xn + qi (t) (10) .te cn un .es with i = 1, 2, . . . , n, the system is called linear. If qi (t) are equal to zero for all i, this system is called linear and homogeneous; if not, non-homogeneous. For this kind of systems the theorem of existence and uniqueness is simpler and to some extent, more satisfactory. This theorem has global character. Notice that in the general case this theorem it is defined in the neighbourhood of the initial conditions, consequently giving a local character to the existence and uniqueness of the solution. Recall: If an equation is linear, it means we can add solutions together and still satisfy the differential equation, e.g. x01 = H(x1 ) and H is linear, then H(c1 · x1 + c2 · x2 ) = c1 · H(x1 ) + c2 · H(x2 ) (11) If x01 = H(x1 ) y x02 = H(x2 ), then (c1 x1 +c2 x2 )0 = H(c1 ·x1 +c2 ·x2 ) = c1 ·H(x1 )+c2 ·H(x2 ) = c1 ·x01 +c2 ·x02 (12) where c1 , c2 are constants. This is a nice property fulfilled when an equation is linear. Basic theory of linear systems of ODE’s ww w Let us consider a system of n linear differential equations of first order x01 = P11 (t) · x1 + P12 (t) · x1 + · · · + P1n (t) · xn + q1 (t) x01 = P21 (t) · x1 + P22 (t) · x1 + · · · + P2n (t) · xn + q2 (t) ... x01 = Pn1 · (t)x1 + Pn1 (t) · x1 + · · · + Pnn (t) · xn + qn (t) (13) We can write this as a matrix where x0 = P(t) · x + q(t) (14) x = (x1 x2 . . . xn )T q(t) = (q1 (t) q2 (t) . . . qn (t)) P(t) = P11 (t) P21 (t) .. . P12 (t) P22 (t) .. . T ... ... .. . P1n (t) P2n (t) .. . Pn1 (t) Pn2 (t) . . . Pnn (t) c 2008 Tecnun (University of Navarra) (15) 3 If q(t) = 0, we have an homogeneous system and eq. 14 becomes x0 = P(t) · x (16) .te cn un .es This notation emphasises the relationship between the linear systems of ODE’s and the first order linear differential equations Theorem. If x1 and x2 are solutions of eq. 16, so (c1 · x1 + c2 · x2 ) is solution as well. x01 = P(t) · x1 ⇒ (c1 · x1 + c2 · x2 )0 = P(t) · (c1 · x1 + c2 · x2 ) (17) 0 x2 = P(t) · x2 The question to be answered is: how many independent solutions of eq. 16 are there? Let us assume at the moment, if x1 , x2 , . . . , xn are solutions of the system, consider the matrix Ψ(t) –called fundamental matrix – given by Ψ(t) = (x1 x2 . . . xn ) (18) Its determinant will be x11 (t) x12 (t) . . . x21 (t) x22 (t) . . . |Ψ(t)| = .. .. .. . . . xn1 (t) xn2 (t) . . . = W (t) xnn (t) x1n (t) x2n (t) .. . (19) ww w where W (t) is called the wronskian of the system. These solutions will be linearly independent at each point t in an interval (α, β) if Example 1. W (t) 6= 0, ∀t ∈ (α, β) We will find it for a two by two system P11 (t) P12 (t) x0 = x P21 (t) P22 (t) (20) (21) whose solutions are x1 = (x11 (t) x21 (t))T y x2 = (x12 (t) x22 (t))T . They verify the system equations (see eq. 21) 0 x11 = P11 · x11 + P12 · x21 0 x1 = P(t) · x1 ⇔ x021 = P21 · x11 + P22 · x21 (22) 0 x = P · x + P · x 11 12 12 22 12 x02 = P(t) · x2 ⇔ x022 = P21 · x12 + P22 · x22 The fundamental matrix, Ψ(t), is x11 (t) x12 (t) Ψ(t) = x21 (t) x22 (t) c 2008 Tecnun (University of Navarra) (23) 4 and the wronskian, W (t), W (t) = x11 · x22 − x12 · x21 (24) The derivative of W (t) with respect to t is W 0 (t) = x011 · x22 + x11 · x022 − x012 · x21 − x12 · x021 and substituting x011 , x021 , x012 , x022 (25) as a function of x11 , x21 , x12 , x22 , we obtain W 0 (t) = (P11 + P22 ) · (x11 · x22 − x12 · x21 ) = (trace P(t)) · W (t) (26) This is the Abel’s formula. Solving the differential equation x (trace P(s))·ds .te cn un .es Rt W (t) = W (t0 ) · e t0 (27) As e is never zero, W (t) 6= 0 for all finite value of t if trace P(t) is integrable and W (t0 ) 6= 0. In n-dimensions the same happens. This generalises to n-dimensions to give Rt dW (trace = (P11 + · · · + Pnn ) · W (t) ⇒ W (t) = W (t0 ) · e t0 dt Example 2. A Sturm-Liouville equation has the form d dy a(t) · + b(t) · y = 0 dt dt This is a second order differential equation of the form P(s))·ds (28) dy d2 y + a1 (t) · + a0 (t) · y = 0 2 dt dt This kind of equations were studied in 18th/19th century and early 20th century, where a2 (t), a1 (t) and a0 (t) are linear in t. For instance, the vibration of a plate is given by these equations. We now write eq. (28) as a system ww w a2 (t) · x1 = y dy x2 = dt (29) (30) Then it gives x01 = x2 x02 = − (31) 0 b(t) a (t) · x2 − · x1 a(t) a(t) (32) We could use our theory of systems to get the connection –wronskian– between x1 and x2 independent solutions of the system. However, we consider eq. (28) directly assuming that y1 and y2 are two possible solutions. So d dy1 a(t) · + b(t) · y1 = 0 (33) dt dt d dy2 a(t) · + b(t) · y2 = 0 (34) dt dt c 2008 Tecnun (University of Navarra) 5 We now multiply y2 by eq. (33) and y1 by eq. (34) and substracting both expressions we get a(t) · (y100 · y2 ) + a0 (t) · (y10 · y2 − y20 · y1 ) = 0 (35) d y2 · y10 − y1 · y20 = y2 · y100 − y1 · y200 dt (36) but and W (t) = y2 · y10 − y1 · y20 a(t) · then .te cn un .es Then eq. (35) becomes d dW a(t) · W (t) = 0 + a0 (t) · W = 0 ⇒ dt dt h dW a0 (t) · dt =− ⇒ W (t) = W (t0 ) · exp W a(t) Z t t0 i −a0 (s) · ds a(s) Homogeneous linear system with constant coefficients. Let us consider the system x0 = A · x (37) where A is real , n × n constant matrix. As solution, we will try x = er·t · a (38) ww w where a is a constant vector. Then x0 = r · er·t · a (39) Then we have a solution provided that r · er·t · a = A · er·t · a ⇔ (A − r · I) · a = 0 (40) and for non-trivial solution (i.e. a 6= 0), r must satisfy |A − r · I| = 0 (41) Procedure. Finding eigenvalues, r1 , r2 , . . . , rn , solution of |A − r · I| = 0 and corresponding eigenvectors, a1 , a2 , . . . , an . Then if the n eigenvectors are linearly independent, we have a general solution x = c1 · er1 ·t · a1 + c2 · er2 ·t · a2 + · · · + cn · ern ·t · an c 2008 Tecnun (University of Navarra) (42) 6 where c1 , c2 , . . . , cn are arbitrary constants. xi = ai · eri ·t , with i = 1, 2, . . . , n. Then the a11 a12 . . . a1n a21 a22 . . . a2n W (t) = . .. .. .. .. . . . an1 an2 . . . ann Recall ai = (a1i a2i . . . ani ) and wronskian will be (r1 +r2 +···+rn )·t 6= 0 (43) e Example 3. .te cn un .es since a1 , a2 , . . . , an are linearly independent. In a large number of problems, getting the eigenvalues can be very difficult problem. 0 x = 1 ·x 1 1 4 (44) Consider x = a · er·t . Then we require (A − r · I) · a11 a21 1 − r |A − r · I| = 4 0 0 1 =0 1 − r = (1 − r)2 − 4 = 0 ⇒ (1 − r)2 = 4 ⇒ 1 − r = ±2 r = 3, −1. So the eigenvalues are 3 and −1. With r = 3 −2 1 a 0 · 1 = 4 −2 a2 0 ww w So −2 · a1 + a2 = 0 implies that a1 = 1 and a2 = 2. Hence, the eigenvector will be 1 2 With r = −1 2 4 1 a 0 · 1 = 2 a2 0 So 2 · a1 + a2 = 0 implies that a1 = 1 and a2 = −2. Therefore, the eigenvector will be 1 −2 The general solution is 1 1 3·t x = C1 · · e + C2 · · e−t 2 −2 (45) The equation (45) is a family of solutions since C1 and C2 are arbitrary (i.e., if x1 and x2 are known at t = 0, we can solve eq. (45) to get C1 and C2 for c 2008 Tecnun (University of Navarra) 7 .te cn un .es Figure 1: Phase plane of example 3 specific solution). Note: eq. (44) does not involve time explicitly. It could be written as dx1 x1 + x2 = (46) dx2 4 · x1 + x2 So we can study the problem in 2 − d space (x1 , x2 ). This is often called the phase plane. Note that eq. (46) defines dx1 /dx2 uniquely except for points where top and bottom terms of the quotient are zero simultaneously. dx1 dt a11 = a dx 21 2 dt a12 a22 x1 x2 ⇔ ww w dx1 dt = a11 · x1 + a12 · x2 (47) dx2 = a21 · x1 + a22 · x2 dt dx2 a21 · x1 + a22 · x2 = dx1 a11 · x1 + a12 · x2 a11 · x1 + a12 · x2 = 0 a21 · x1 + a22 · x2 = 0 (48) (49) In general, what about A? T 1. If A is hermitian (i.e., AH = A = A, where A is the complex conjugate of the matrix), then the eigenvalues are real and we can find n linearly-independent eigenvectors. 2. A is non-hermitian. We have the following possibilities (2.a.) n real and disttinct eigenvalues and n independent eigenvectors. (2.b.) Complex eigenvalues. (2.c.) Repeated eigenvalues. c 2008 Tecnun (University of Navarra) 8 Example 4. 0 x = 1 4 1 1 x (50) Put x = er·t · a to get 1 =0 1−r 1−r 4 (51) we obtain r1 = −1 y r2 = 3. We need the eigenvectors r1 = −1 ⇒ a1 = (1 − 2)T r2 = 3 ⇒ a2 = (1 2)T .te cn un .es (52) We have two solutions x1 = 1 −2 e −t x2 = 1 2 e3t (53) To find the general solution, we construct it via a linear combination –by adding these two linearly independent solutions together– 1 1 x = c1 · e−t · + c2 · e3·t · (54) −2 2 c1 and c2 are arbitrary constants to be determined by intial conditions or other conditions on x. ww w We can plot (see Figure 2) the family of the solutions in the (x1 , x2 ) plane with an arrow to signify the direction of time (means time increasing). Our solution in components looks like x1 =c1 · e3·t + c2 · e−t x2 =2 · c1 · e3·t − 2 · c2 · e−t we can study the motion of a pendulum, we can use systems in order to study position and velocity. Suppose c2 ≡ 0 and we are interested in the point x1 = x2 = 0 and note x1 /x2 = 1/2 with c1 6= 0. We only reach (0, 0) if t → −∞ since we need e3·t → 0. If c1 ≡ 0 and c2 6= 0, then x1 /x2 = −1/2. To get (0, 0) we need t → +∞. So (0, 0) is a special point. I can represent the time by an arrow x1 x2 x1 t → +∞ ⇒ x2 t → −∞ ⇒ → c1 · e−t ⇔ x2 → −2 · x1 → −2 · c1 · e−t → c2 · e3·t ⇔ x2 → 2 · x1 → 2 · c2 · e3·t (55) Point (0, 0) is a saddle point. Thinking of the future, we can observe that detA = −3 < 0. c 2008 Tecnun (University of Navarra) 9 .te cn un .es Figure 2: Phase plane of Example 4 Example 5. Let us consider this problem. √ −3 2 √ x0 = ·x 2 −2 In order to solve this problem, we obtain the eigenvalues √ −3 − r 2 √ =0 2 −2 − r (56) (57) and r1 = −4 y r2 = −1. The eigenvectors are ww w √ r1 = −4 ⇒ a1 = (− √2 1)T r2 = −1 ⇒ a2 = (1 2)T they are linearly independent. Hence the solutions are √ 1 − 2 −4·t √ x1 = e x2 = e−t 1 2 (58) (59) And the general solution is x = c1 · e−4·t · √ 1 − 2 + c2 · e−t · √ 1 2 Plotting the paths on the phase plane (x1 , x2 ) √ x1 → c2√ · e−4·t t → −∞ ⇒ −4t ⇔ x2 → − 2 · x1 x → − 2 · c · e 2 2 x1 → 0 t → +∞ ⇒ ⇔ (x1 , x2 ) → (0, 0) x2 → 0 (60) (61) Point (0, 0) is a stable node. Thinking of the future again, detA = 2 > 0, traceA = −5 < 0 and detA < (traceA)2 /4. c 2008 Tecnun (University of Navarra) 10 .te cn un .es Figure 3: Phase plane of Example 5 Example 6. Let us consider another 0 x= 1 1 where A is a hermitian matrix −r 1 1 system 1 1 0 1 x 1 0 1 −r 1 1 1 −r =0 (62) (63) we obtain r1 = −1 (double) and r2 = 2. The eigenvectors are ww w r1 = −1 ⇒ a1 = (1 − 1 0)T , a2 = (1 1 − 2)T r2 = 2 ⇒ a3 = (1 1 1)T They are linearly independent. Hence, the solutions are 1 1 1 x1 = −1 · e−t x2 = 1 · e−t x3 = 1 · e2·t 1 0 −2 And the general solution is 1 1 1 x = c1 · e−t · −1 + c2 · e−t · 1 + c3 · e2·t · 1 0 −2 1 (64) (65) (66) Complex eigenvalues. If A is non hermitian (but real) and has complex eigenvalues then the determinant |A − r · I| = 0 takes complex conjugates eigenvalues |A − r1 · I| = 0 c 2008 Tecnun (University of Navarra) (67) 11 As A and I are real, if we work out the conjugate of eq. 67, we obtain |A − r1 · I| = 0 (68) this means that if r1 is an eigenvalue, its complex conjugate is eigenvalue as well. Therefore the eigenvectors will be complex conjugates, too: (A − r1 · I) · x1 = 0 ⇔ (A − r1 · I) · x1 = 0 (69) The solution associated to r1 is .te cn un .es x1 = er1 ·t · a1 = e(λ1 +i·µ1 )·t · (u1 + i · v1 ) = = eλ1 ·t · (cos(µ1 · t) + i · sin(µ1 · t)) · (u1 + i · v1 ) = = eλ1 ·t · (u1 · cos(µ1 · t) − v1 · sin(µ1 · t)) + i · eλ1 ·t (u1 · sin(µ1 · t) + v1 · cos(µ1 · t)) (70) And the other one x1 = eλ1 ·t ·(u1 ·cos(µ1 ·t)−v1 ·sin(µ1 ·t))−i·eλ1 ·t (u1 ·sin(µ1 ·t)+v1 ·cos(µ1 ·t)) (71) We are looking for real solution to this system. We know that a linear combination of these solutions will be a solution as well. Hence, we can take the real and the imaginary parts of these ones x = c1 ·eλ1 ·t (u1 ·cos(µ1 ·t)−v1 ·sin(µ1 ·t))+c2 ·eλ1 ·t (u1 ·sin(µ1 ·t)+v1 ·cos(µ1 ·t)) (72) Example 7. . 0 ww w x = −1/2 −1 where A is not symmetric. Solving −1/2 − r −1 1 −1/2 x (73) 1 =0 −1/2 − r (74) we get r = −1/2 ± i. The eigenvector is 1 a1 = i (75) Hence the solution is 1 1 (−1/2+i)·t x1 = e = e−t/2 · (cos(t) + i · sin(t)) i i (76) Then x1 = cos(t) + i · sin(t) − sin(t) + i · cos(t) ·e−t/2 = cos(t) − sin(t) ·e−t/2 +i· sin(t) cos(t) ·e−t/2 (77) c 2008 Tecnun (University of Navarra) 12 The general solution is x = c1 · e−t/2 · cos(t) − sin(t) + c2 · e−t/2 · sin(t) cos(t) (78) .te cn un .es Plotting this family of curves on the phase plane (x1 , x2 ) Figure 4: Phase plane of Example 7 t → +∞ ⇒ x1 → 0 ⇔ (x1 , x2 ) → (0, 0) x2 → 0 (79) Let us study the points (x 0)T y (0 y)T . Substituting into eq. 73 ww w x = (x 0)T ⇒ x0 = (−x/2 − x)T ⇒ dy =2 dx (80) dy −1 x = (0 y) ⇒ x = (y − y/2) ⇒ = dx 2 T 0 T Point (0, 0) is a stable spiral point. Thinking of the future once again, det A = 5/4 > 0, traceA = −1 < 0 and detA > (traceA)2 /4. Repeated eigenvalues. Let us consider the system given by eq. 37, with A a real non-symmetric matrix. Let us assume that one of its eigenvalues, r1 , is repeated (m1 = m > 1), and there are not m linearly independent eigenvectors, 1 ≤ k < m. We should obtain (m − k) additional linearly independent solutions. How do we amend the procedure to deal with cases when there are no m linearly independent eigenvectors? For example, the exponential of a square matrix: eAt = I + At + An tn A2 t2 + ··· + + ... 2! n! c 2008 Tecnun (University of Navarra) (81) 13 We recall the Cayley–Hamilton theorem. If f (λ) is the characteristic poynomial of a matrix A, f (λ) = det (A − λ · I) (82) This theorem says that f (A) = 0. For instance, f (λ) = ri2 − (trace · A)ri + det A = ri2 − p · ri + q = 0 (83) Hence, by Cayley-Hamilton theorem, f (A) = A2 − p · A + q · I = 0 .te cn un .es (84) using this theorem, any expansion function like eA can be reduced to polynomials on A. This is very useful theorem in problems of materials and continuum mechanics. A2 = p · A − q·I A3 = p · A2 − q · A = p2 − q · A − p · q · I A4 = p · A3 − q · A2 = p · p2 − 2q · A − (p2 − q) · q · I ... (85) So, we can use this theorem to get a finite expansion of eq. 81. Differentiating with respect to t that expression A3 t2 An tn−1 deAt = A + A2 t + + ··· + + ··· = dt (2! (n − 1)! (86) An−1 tn−1 = A I + At + · · · + + ... (n − 1)! ww w Hence x=e At v= = AeAt A2 t2 An tn I + At + + ··· + + ... 2! n! v (87) where v is a constant vector, then dx = AeAt v = Ax dt (88) Av = A · ai = ri · ai A2 · v = ri · A · ai = ri2 · ai ... An v = rin · ai (89) So it looks all right. Hence, Hence, substituing into eq. 87 rin · tn ri2 · t2 + ··· + + . . . · ai = eri ·t · ai x = 1 + ri · t + 2! n! c 2008 Tecnun (University of Navarra) (90) 14 Note that e(A+B)·t = eA·t eB·t ⇔ A · B = B · A (91) eAt v = e(A−λI)·t eλI·t · v, ∀λ (92) Let us consider I can do this for all λ since (A − λ · I)(λ · I) = λ · A − λ2 · I = (λ · I) · (A − λ · I) (93) hence .te cn un .es Let us come back to eq. 87 λ 2 · I2 · t 2 λn In · tn eλ·I·t · v = I + λ · I · t + + ··· + + . . . v = eλ·t · v (94) 2! n! eAt v = eλ·t e(A−λ·I)·t · v A·t (95) But x = e · v is a solution of the system. Moreover, let us assume that v = ai where ai is an eigenvector of ri , then t2 x = eri ·t I + t(A − ri · I) + (A − ri · I)2 + . . . · ai = 2! (96) t2 ri ·t 2 =e ai + t · (A − ri · I) · ai + · (A − ri · I) · ai + . . . = 2! But, as ai is eigenvector, it verifies ww w (A − ri · I) · ai = 0 = (A − ri · I)2 · ai = (A − ri · I)3 · ai = . . . (97) Then eq. 96 becomes x = eri ·t · ai (98) We can look for another vector v such that (A − ri · I) · v 6= 0 y (A − ri · I)2 · v = 0 = (A − ri · I)3 · v = . . . (99) Lemma. Let us assume that the characteristic polynomial of A, non-hermitian, of order n, have repeated roots (r1 , r2 , . . . , rk , 1 ≤ k < n), of order (m1 , m2 , . . . , mk (m1 + m2 + · · · + mk = n), respectively such that f (λ) = (λ − r1 )m1 (λ − r2 )m2 . . . (λ − rk )mk (100) if A has only nj < mj eigenvectors of the eigenvalue rj (i.e., (A − rj · I) · v = 0 has nj independent solutions), then (A − rj · I)2 · v = 0 has at least nj + 1 independent solutions. In general, if (A − rj · I)m · v = 0 has got nj < mj independent solutions, (A − rj · I)m+1 · v = 0 has at least nj + 1 independent solutions c 2008 Tecnun (University of Navarra) 15 Example 8. Let it be the system 1 x0 = 0 1 1 ·x (101) In order to solve, we take 1−r 0 1 =0 1−r (102) .te cn un .es the root is r1 = 1, double, with only one eigenvector obtained from (A − r1 I) · a1 = 0 r1 = 1 ⇒ a1 = (1 0)T (103) We cannot solve the second solution we need. So, we have 1 x1 = · et 0 (104) We must determine another vector a2 such that 2 (A − r1 · I) · a2 6= 0 (105) (A − r1 · I)2 · a2 = 0 = (A − r1 · I)3 · a2 = . . . (106) 3 ww w As (A − r1 I) = (A − r1 I) = 0, any vector a2 verifies eq. 106. However, according to the inequality given by eq. 105, a2 cannot be a linearly dependent vector of a1 (eq. 103). So 0 1 0 1 a2 = (0 1)T ⇒ · = 6= 0 (107) 0 0 1 0 From eq. 96, we obtain t2 t 2 x2 = e · a2 + t · (A − r1 · I) · a2 + · (A − r1 · I) · a2 + . . . = 2! 0 1 t t t =e · +t +0 =e · 1 0 1 (108) The general solution is t x = c1 · e · 1 0 t + c2 · e · t 1 (109) Thinking of next paragraph, detA = 1 > 0, traceA = 2 > 0 and detA = (traceA)2 /4. c 2008 Tecnun (University of Navarra) 16 .te cn un .es Figure 5: Phase plane of Example 8 Résumé of the study of a system of two homogeneous equations with constant coefficients. Let us consider the system x0 = a · x + b · y y0 = c · x + d · y (110) f (r) = r2 − p · r + q (111) its characteristic polynomial is ww w where p = a + d (the trace) y q = a · d − b · c (the determinant). Its eigenvalues are p p ± p2 − 4 · q r= (112) 2 The study of eigenvalues and eigenvectors is very useful in order to classify the critical point (0, 0) and to know the trajectories on the phase plane. 1. q < 0. The eigenvalues are positive and negative, respectively. Saddle point. 2. q > 0. 2.a. p2 > 4q. Both eigenvalues are either positive or negative. - Stable node, if p < 0. - Unstable node, if p > 0. 2.b. p2 < 4 · q. The eigenvalues are complex conjugates. - Stable spiral, if p < 0. - Unstable spiral, if p > 0. - Center, if p = 0. 2.c. p2 = 4 · q. The eigenvalue is a double root of the characteristic polynomial. - Stable node, if p < 0 and there is only one eigenvector. - Unstable node, if p > 0 and there is only one eigenvector. c 2008 Tecnun (University of Navarra) 17 - Sink point, if p < 0 and there are two independent eigenvectors. - Source point, if p > 0 and there are two independent eigenvectors. 3. q = 0. It means that the matrix rank is 1 and therefore, one row can be obtained multipliying the other one by a constant (c/a = d/b = k). Then, (0, 0) is not an isolated critical point. There is a line y = −a · x/b, b 6= 0, of critical points. The trajectories on the phase plane are y = k ·x+E, E being a constant. - If p > 0, paths start at the critical points and go to the infinity. - If p < 0, conversely, the trajectories end at the critical points. .te cn un .es It is very convenient and useful to know the plane trace-determinant (p, q). Fundamental matrix of a system Let us consider the homogeneos system given by eq. 16 and let x1 , x2 , . . . , xn be a set of its independent solutions. We know that we can build the general solution via a linear combination of them. We denote fundamental matrix, Ψ(t), a matrix whose columns are the solution vectors of eq. 18. The determinant of this matrix is not zero (eq. 19). This determinant is called wronskian (eq. 20). Let us assume that we are looking for a solution x such that x(t0 ) = x0 . Then x0 = c1 · x1 + c2 · x2 + · · · + cn · xn = Ψ(t0 ) · c (113) where c = (c1 c2 . . . cn )T . As |Ψ(t)| 6= 0, ∀t, c can be obtained using the inverse matrix of Ψ(t0 ): c = Ψ−1 (t0 ) · x0 (114) and the solution will be x = Ψ(t) · Ψ−1 (t0 ) · x0 (115) ww w This matrix is very useful when Ψ(t0 ) = I (116) This special set of solutions builds the matrix Φ(t). It verifies x = Φ(t) · x0 (117) −1 We obtain that Φ(t) = Ψ(t) · Ψ (t0 ). Moreover, with constant coefficients: (1.) eA·t (eq. 81) is a fundamental matrix of fundamental solutions since it verifies eq. 86 and eA·0 = I. (2.) If we know two fundamental matrices of the system, Ψ1 and Ψ2 , there is always a constant matrix C such that Ψ2 = Ψ1 · C, since each column of Ψ2 can be obtained by a linear combination of the columns of Ψ1 . (3.) It can be shown that eA·t = Ψ(t) · Ψ−1 (t0 ) (see eq. 115). According to paragraphs 1 and 2, there exists a matrix C such that eAt = Ψ(t) · C (118) I = Ψ(0) · C ⇒ C = Ψ−1 (0) (119) This expression at t = 0 c 2008 Tecnun (University of Navarra) 18 Nonhomogeneous systems. We have x0 = P(t) · x + q(t) (120) 0 We assume that we have solved x = P(t) · x. We actually have a procedure for P(t) = A, a constant matrix. Consider special cases (1) .te cn un .es If P(t) = A and A has n independent eigenvectors, the procedure is to build the matrix T with the eigenvectors of A: T = (a1 a2 . . . an ) (121) Then, we change variables x = T · y with x0 = T · y0 . Going back to the system x0 = T · y0 = A · x + q = A · T · y + q (122) As T is built with n eigenvectors, this is a regular matrix (det T 6= 0), so we can work out T−1 , the inverse of T. Hence, from eq. 114 y0 = T−1 · A · T · y + T−1 · q = D · y + h (123) where D is the diagonal matrix of the eigenvalues of A. Therefore, yi0 (t) = ri · yi (t) + hi (t), ∀i = 1, 2, . . . , n Hence ri t · e−ri ·t · hi (t) · dt + ci · eri ·t , ∀i = 1, 2, . . . , n ww w yi (t) = e Z (124) (125) After obtaining yi , we can get x = T · y. This method is only possible if there are n linearly independent eigenvectors, i.e., A is a diagonalizable constant matrix. Because we can reduce the matrix to its diagonal form, the above procedure works. In cases where we do not have the n independent eigenvectors, the matrix A can only be reduced to its Jordan canonical form. (2) Variation of the parameters. Let us consider the system of eq. 14, and we know the solution of the associated homogeneous system (eq. 16). Then we can build the fundamental matrix of the system, Ψ(t) (eq. 18), whose columns are the linearly independent of the homogeneous system solutions. We are looking for a solution like x = Ψ(t) · u(t) c 2008 Tecnun (University of Navarra) (126) 19 where u(t) is a vector to be determined such that eq. 118 is a solution of eq. 14). Substituting Ψ0 (t) · u(t) + Ψ(t) · u0 (t) = P(t) · Ψ(t) · u(t) + q(t) (127) as we know that Ψ0 (t) = P(t) · Ψ(t), Ψ(t)u0 (t) = q(t) ⇒ u(t) = Z Ψ−1 (t)q(t)dt + c (128) ww w .te cn un .es where c is a constant vector and there exists Ψ−1 (t) since the n columns of matrix Ψ(t) are linearly independent (eq. 20). The general solution is Z x = Ψ(t) · Ψ−1 (t) · q(t) · dt + Ψ(t) · c (129) c 2008 Tecnun (University of Navarra) 20