Convolution Integrals — Question 9

PDF ↗

Question 9

All functions are causal (zero for t<0t<0). Use the one-sided Laplace transform and (f*g)(t)=∫0tf(t−u)g(u)du(f*g)(t)=\int_0^t f(t-u)g(u)\,du. Write Ha(t)=0H_a(t)=0 for t<at<a and 11 for t≥at\ge a. Values at isolated endpoints do not affect an ordinary integral.

Two normalized stages kα(t)=αe−αtk_\alpha(t)=\alpha e^{-\alpha t} and kβ(t)=βe−βtk_\beta(t)=\beta e^{-\beta t}, with α,β>0\alpha,\beta>0, are followed by a delay d≥0d\ge 0: r(t)=Hd(t)(kα*kβ)(t−d).r(t)=H_d(t)(k_\alpha*k_\beta)(t-d). The observed response is positive precisely for t>1t>1, has mass 11, mean μ=5/2\mu=5/2 and variance v=5/4v=5/4. For a nonnegative unit-mass kernel pp, μp=∫0∞tp(t)dt\mu_p=\int_0^\infty tp(t)\,dt and vp=∫0∞(t−μp)2p(t)dtv_p=\int_0^\infty(t-\mu_p)^2p(t)\,dt.

Tasks

  1. Prove that convolution adds the means and variances of two nonnegative unit-mass kernels with finite second moments.

  2. Compute the mean and variance of kαk_\alpha, and explain the effect of a delay on mass, mean and variance.

  3. Recover dd and the two rates from the observations. State precisely what is and is not uniquely identifiable.

  4. For general prescribed delay dd, mean μ\mu and variance vv, give necessary and sufficient conditions for two finite positive rates to exist. Interpret the equality case and the excluded upper boundary.

Original worksheet page 1: question and worked solution for 4-9-009
Show solutionHide solution

Question 9 – Solution

Strategy. Work with reciprocal rates. Convolution moment identities reduce identification to a quadratic equation.

Step 1: Prove the moment identities. In the convolution integral use t=u+wt=u+w, u,w≥0u,w\ge 0. Nonnegativity permits interchange, and finite second moments make the following values finite: ∫(p*q)=(∫p)(∫q)=1,μp*q=∬(u+w)p(u)q(w)dudw=μp+μq,∫t2(p*q)(t)dt=∫u2p(u)du+2μpμq+∫w2q(w)dw.\begin{aligned} \int(p*q)&=(\int p)(\int q)=1,\\ \mu_{p*q}&=\iint(u+w)p(u)q(w)\,du\,dw=\mu_p+\mu_q,\\ \int t^2(p*q)(t)\,dt&=\int u^2p(u)\,du+ 2\mu_p\mu_q+\int w^2q(w)\,dw. \end{aligned} Subtracting (μp+μq)2(\mu_p+\mu_q)^2 from the last line proves vp*q=vp+vq\boxed{v_{p*q}=v_p+v_q}.

Step 2: Calculate the stage statistics. Integration by parts gives mass 11, first moment 1/α1/\alpha and second moment 2/α22/\alpha^2 for kαk_\alpha. Its variance is therefore 1/α21/\alpha^2. A delay replaces time by t=d+ut=d+u: mass stays 11, the mean increases by dd, and the centered second moment is unchanged. Hence μ=d+1α+1β,v=1α2+1β2.\mu=d+\frac 1\alpha+\frac 1\beta,\qquad v=\frac 1{\alpha^2}+\frac 1{\beta^2}.

Step 3: Recover the parameters. Both stage kernels are positive for positive times, so their convolution is positive at every positive time and zero at zero. The observed onset thus forces d=1d=1. Put x=1/αx=1/\alpha, z=1/βz=1/\beta. Then x+z=32,x2+z2=54,xz=12,(X−x)(X−z)=X2−32X+12.x+z=\tfrac 32,\quad x^2+z^2=\tfrac 54,\quad xz=\tfrac 12,\quad (X-x)(X-z)=X^2-\tfrac 32X+\tfrac 12. Its roots are 11 and 1/21/2, giving {α,β}={1,2}\boxed{\{\alpha,\beta\}=\{1,2\}}. The delay and unordered rate pair are unique. Stage order cannot be identified even from the entire output because convolution is commutative.

Step 4: Give the full feasibility condition. Write m=μ−dm=\mu-d. Positive x,zx,z require and are possible exactly when m>0,m22≤v<m2.\boxed{m>0,\qquad \frac{m^2}{2}\le v<m^2.} Indeed xz=(m2−v)/2>0xz=(m^2-v)/2>0 and (x−z)2=2v−m2≥0(x-z)^2=2v-m^2\ge 0; conversely these conditions make (m±2v−m2)/2(m\pm\sqrt{2v-m^2})/2 real and strictly positive. At the lower equality the rates coincide. At the upper boundary one reciprocal rate would be zero, requiring an infinite rate, excluded here.

Original worksheet page 2: question and worked solution for 4-9-009

Original worksheet layout. Use Enlarge or open the PDF for a closer view.