Abstract
In this paper, we consider the problem, usual in analog signal processing, to find a continuous linear time-invariant system related to a linear differential equation \(P(D)x = Q(D)f\), i.e. a system \(\mathscr {L}\) such that for every input signal f yields an output \(\mathscr {L}(f)\) which verifies \(P(D)\mathscr {L}(f ) = Q(D)f\). We give a systematic theoretical analysis of the existence and uniqueness of such systems (both causal and non-causal ones) defined on \({L^{p}}\) functions and \({\mathscr {D'}_{L^{p}}}\) distributions (input spaces which include signals with not necessarily left-bounded support), for every p. More precisely, by finding all their possible impulse responses, we characterise all these systems apart two pathologies arising when \(p = \infty \). Finally, we give necessary and sufficient conditions on P, Q for causality and stability of the systems. As an application, we consider the problem of finding the inverse of a causal continuous linear time-invariant system, defined on \({L^{p}}\), related to a simple differential equation. We also show a digital simulation of this inverse system.

Similar content being viewed by others
Notes
For every \(f(t)\in {L^{1}}\), the definition of Fourier Transform we adopt is the following: \({\hat{f}}(\omega ) = \int _{\mathbb {R}} e^{-i\omega t}f(t)\mathrm{d}t\).
References
A.S. Besicovitch, Almost Periodic Functions (Cambridge University Press, Cambridge, 1932)
P. Brémaud, Mathematical Principles of Signal Processing, Fourier and Wavelet Analysis (Springer, New York, 2002)
M. Ciampa, M. Franciosi, M. Poletti, A note on impulse response for continuous, linear, time-invariant, continuous-time systems. IEEE Trans. Circuits Syst. I Regul. Pap. 53(1), 106–113 (2006)
M. Ciampa, M. Franciosi, M. Poletti, Continuous LTI systems defined on \({L^{p}}\) functions and \({{\mathscr {D}^{\prime }}_L^{p}}\) distributions: analysis by impulse response and convolution. IEEE Trans. Circuits Syst. I Regul. Pap. 55(6), 1711–1721 (2008)
S. Fletcher, I. Kirkpatrick, R. Dring, R. Puttock, R. Thring, S. Howroyd, The modelling of carbon-based supercapacitors: distributions of time constants and Pascal equivalent circuits. J. Power Sources 345, 247–253 (2017)
C. Gasquet, P. Witomski, Fourier Analysis and Applications: Filtering, Numerical Computation, Wavelets (Springer, New York, 1999)
T. Kailath, Linear Systems (Prentice-Hall, Englewood Cliffs, 1980)
M. Langenbruch, Continuous linear right inverses for differential operators. Dissertationes Mathematicae 340, 163–181 (1995)
R. Larsen, An Introduction to the Theory of Multipliers (Springer, New York, 1971)
A.V. Oppenheim, A.S. Willsky, Signals and Systems (Prentice-Hall, Upper Saddle River, 1997)
A. Papoulis, The Fourier Integral and its Applications (McGraw-Hill Book Company, New York, 1962)
A. Papoulis, Signal Analysis (McGraw-Hill Book Company, New York, 1977)
A. Papoulis, Limits on bandlimited signals. Proc. IEEE 55(10), 1677–1686 (1967)
I.W. Sandberg, Causality and the impulse response scandal. IEEE Trans. Circuits Syst. I Fundam. Theory Appl. 50(6), 810–811 (2003)
I.W. Sandberg, Notes on the Superposition Scandal, in Proc. of the 9th WSEAS International Conference on Mathematical and Computational Methods in Science and Engineering (2007), pp. 115–118
D. Schlichthärle, Digital Filters—Basics and Design (Springer, Berlin, 2011)
L. Schwartz, Théorie des Distributions (Hermann, Paris, 1966)
L. Schwartz, Mathematics for the Physical Sciences (Hermann, Paris, 1966)
J.S. Vandergraft, Introduction to Numerical Computations (Academic Press, New York, 1983)
M. Vetterli, J. Kovačević, V.K. Goyal, Foundations of Signal Processing (Cambridge University Press, Cambridge, 2014)
G. Zames, On the input-output stability of time-varying nonlinear feedback systems, part I: conditions derived using concepts of loop gain, conicity, and positivity. IEEE Trans. Autom. Control 11(2), 228–238 (1966)
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: Uniformly Almost Periodic Functions
Appendix: Uniformly Almost Periodic Functions
Let \(f(t):\mathbb {R}\rightarrow \mathbb {C}\) be a function, and let \(\varepsilon \not =0\) be a real number. In this Section, we call almost period of f(t) relative to \(\varepsilon \) a real number \(T > 0\) such that for every \(t_{0}\in \mathbb {R}\), there exists \(\tau \in \left[ t_{0},t_{0} + T\right] \) satisfying
The function f(t) is called uniformly almost periodic if \(f(t)\in {C^{0}}\) and for every \(\varepsilon > 0\) there exists an almost period T of f(t) relative to \(\varepsilon \).
Immediate consequences of the definition are
-
(a)
Every continuous periodic function is uniformly almost periodic
-
(b)
If f(t) is uniformly almost periodic then for every \(\lambda \in \mathbb {R}\), \(f(\lambda t)\) and \(f(t + \lambda )\) are uniformly almost periodic
Moreover, the following results are well know:
-
(c)
The set of uniformly almost periodic functions is a vector subspace of \({C^{0}}\) (see [1, Theorem 6 and Theorem 12])
-
(d)
Every linear combination of continuous periodic functions is uniformly almost periodic (follows by items (a) and (c))
-
(e)
A uniformly almost periodic function is bounded and uniformly continuous (see [1, Theorems 4 and 5])
-
(f)
If f(t) is uniformly almost periodic then \(\mathrm{Re}f(t)\) and \(\mathrm{Im}f(t)\) are uniformly almost periodic (follows by [1, Theorem 6] and item (c))
The following theorem proves that every non null uniformly almost periodic function cannot be null at infinity.
Theorem 7
Let f(t) be a uniformly almost periodic function; the following statements are equivalent:
-
a)
\({\displaystyle \lim _{t \rightarrow +\infty } }{f(t)} = 0\)
-
b)
\({\displaystyle \lim _{t \rightarrow -\infty } }{f(t)} = 0\)
-
c)
\(f(t) = 0\) for every \(t\in \mathbb {R}\)
Proof
a)\(\Rightarrow \)c). Let \(n > 1\). By the assumption there exists a real number \(t_{n} > 0\) such that: \(|\,f(t_{n} + t)\,| < 1/n\) for every \(t > 0\).
Let \(T_{n}\) be an almost period of f(t) relative to 1 / n. Then, there exists \(\tau _{n}\in [-n-t_{n}-T_{n},-n-t_{n} ]\) such that: \(|\,f(t + \tau _{n}) - f(t)\,| < 1/n\) for every \(t\in \mathbb {R}\). Writing \(\tau _{n} = -n-t_{n}-\sigma _{n}\), with \(0\leqslant \sigma _{n}\leqslant T_{n}\), we obtain \(|\,f(t-n-t_{n}-\sigma _{n}) - f(t)\,| < 1/n\) for every \(t\in \mathbb {R}\). Substituting t with \(t_{n} + t\) we obtain: \(|\,f(-n-\sigma _{n}+t) - f(t_{n} + t)\,| < 1/n\) for every \(t\in \mathbb {R}\), and hence \(|\,f(-n-\sigma _{n}+t)\,| < |\,f(t_{n}+t)\,| + 1/n\) for every \(t\in \mathbb {R}\).
As a consequence it is \(|\,f(-n-\sigma _{n}+t)\,| < 2/n\) for every \(t > 0\), and hence \(|\,f(t)\,| < 2/n\) for every \(t > -n\).
b)\(\Rightarrow \)c). Let \(g(t) = f(-t)\) and apply the previous result to g(t).
Obviously c)\(\Rightarrow \)a), c)\(\Rightarrow \)b). \(\square \)
The following theorem proves that every regularization of a uniformly almost periodic function is uniformly almost periodic.
Theorem 8
Let f(t) be a uniformly almost periodic function, and let \(\varphi (t)\in {\mathscr {D}}\). Then \((f*\varphi )(t)\) is uniformly almost periodic.
Proof
If \(\varphi (t) = 0\), the statement is obvious. Let \(\varphi (t) \not = 0\), and let \(\mu = \int {}{}{|\varphi |} > 0\).
Let \(\varepsilon > 0\), let \(T > 0\) be an almost period of f(t) relative to \(\varepsilon /(2\mu )\), and let \(t_{0}\in \mathbb {R}\).
There exists \(\tau \in [t_{0}, t_{0} + T]\) such that: \(|\,f(t + \tau ) - f(t)\,| < \varepsilon /(2\mu )\) for every \(t\in \mathbb {R}\). Write \(f_{\tau }(t) = f(t + \tau )\). For every \(t\in \mathbb {R}\) it is
As a consequence T is an almost period of \((f *\varphi )(t)\) relative to \(\varepsilon \). \(\square \)
As corollaries of Theorems 7 and 8, we can now state the results used in the proofs of the theorems of Sects. 2 and 3.
Corollary 1
Let \(\omega _{1},\dots , \omega _{\nu }\in \mathbb {R}\) be pairwise distinct, and \(c_{1},\dots , c_{\nu }\in \mathbb {C}\) be not all equal to 0. Then, there exists \(\mu > 0\) and two sequences \(\tau _{k}', \tau _{k}''\in \mathbb {R}\) such that
and
Proof
Simply observe that \(\sum _{l = 1}^{\nu }c_{l}e^{i\omega _{l}t}\) is uniformly almost periodic (as a linear combination of continuous periodic functions), and nonzero (since the family \(e^{i\omega _{l}t}, l = 1,\dots ,\nu \) is linearly independent); then apply Theorem 7. \(\square \)
Corollary 2
Let \(\omega _{1},\dots ,\omega _{\nu }\in \mathbb {R}\) be pairwise distinct, and let \(c_{1},\dots ,c_{\nu }\in \mathbb {C}\). The following statements are equivalent
-
a)
\({\displaystyle \sum _{l = 1}^{\nu }c_{l}e^{i\omega _{l}t}}\) is a distribution null at \(\infty \)
-
b)
\({\displaystyle \sum _{l = 1}^{\nu }c_{l}e^{i\omega _{l}t}}\) is a distribution null at \(-\infty \)
-
c)
\(c_{1} = \dots = c_{\nu } = 0\)
Proof
Obviously a)\(\Rightarrow \)b), and c)\(\Rightarrow \)a).
b)\(\Rightarrow \)c). Let \(B\in \mathbb {R}\) be such that \(0< B < \min \{\frac{\pi }{4\,|\,\omega _{l}\,|}: l \text {such that } \omega _{l} \not = 0\}\) and let \(\varphi (t)\in {\mathscr {D}}\) be such that \({\mathrm{supp~}}\varphi \subset (-B,B)\), \(\varphi (t) \geqslant 0\) for every \(t\in \mathbb {R}\), \(\int _{}^{}\varphi \not = 0\). A straightforward argument proves that \({\hat{\varphi }}(\omega _{l}) \not = 0\) for every l (here \({\hat{\varphi }}(\omega )\) denotes the Fourier Transform of \(\varphi (t)\)).
Observe that \((\sum _{l = 1}^{\nu } c_{l}e^{i\omega _{l}t}) *\varphi (t) = \sum _{l = 1}^{\nu } c_{l}\,{\hat{\varphi }}(\omega _{l})\,e^{i\omega _{l}t}\) is a function null at \(-\infty \) in the usual sense. As a consequence, by Corollary 1, for every l we have \(c_{l}\,{\hat{\varphi }}(\omega _{l}) = 0\), and hence \(c_{l} = 0\). \(\square \)
Corollary 3
Let \(\omega _{1},\dots ,\omega _{\nu }\in \mathbb {R}\) be pairwise distinct, and \(M_{1}(t),\dots ,M_{\nu }(t)\) be polynomial functions not all constant. Then, there exist two sequences \(\tau _{k}', \tau _{k}''\in \mathbb {R}\) such that
and
Proof
Write the polynomial functions \(M_{l}(t)\) in the form \(M_{l}(t) = \sum _{h = 0}^{p} m_{lh}t^{h}\) with \(p \geqslant 1\) and at least one of the coefficients \(m_{1p},\dots ,m_{\nu p}\) non null. By Corollary 1, there exists \(\mu > 0\) and two sequences \(\tau _{k}', \tau _{k}''\in \mathbb {R}\) such that
and
Then observe that for every k it is
and
where \({\displaystyle \lim _{k \rightarrow \infty } }{\alpha _{k}^{\prime }} = 0\) and \({\displaystyle \lim _{k \rightarrow \infty } }{\alpha _{k}^{\prime \prime }} = 0\). \(\square \)
Corollary 4
Let \(s_{1} = {\sigma _{1}+i\omega _{1}},\dots ,s_{\nu } = {\sigma _{\nu }+i\omega _{\nu }}\) be pairwise distinct complex numbers, \(M_{1}(t),\dots ,\)\(M_{\nu }(t)\) be polynomial functions not all null, and \(P_{0}(t)\) be a polynomial function. If \(\sigma _{l} \not = 0\) for every \(l= 1,..., \nu \) then
-
a)
\({\displaystyle \sum _{l = 1}^{\nu }M_{l}(t)e^{s_{l}t}} + P_{0}(t)\) is not a slowly increasing function in the usual sense
-
b)
\({\displaystyle \sum _{l = 1}^{\nu }M_{l}(t)e^{s_{l}t}}\not \in {\mathscr {S'}}\)
Proof
Changing, if necessary, the family of indices, we may assume \(M_{l}(t)\) not null for every l.
a) Assume first that there exists l such that \(\sigma _{l} > 0\), and let \(\sigma = \max \{\,\sigma _{1},\dots ,\sigma _{\nu }\,\}\)\(> 0.\) Then it is
where \(\lim _{t \rightarrow +\infty }\alpha (t) = 0\). By Corollaries 1 and 3, there exist \(\mu > 0\) and a sequence \(\tau _{k}\in \mathbb {R}\) such that
As a consequence \(|\sum _{l = 1}^{\nu } M_{l}(t)e^{s_{l}t} + P_{0}(t)|\) cannot be bounded by any \(|\,F(t)\,|\) where F(t) is a polynomial function.
Assume now that \(\sigma _{l} < 0\) for every l, and let \(\varPhi (t) = \sum _{l = 1}^{\nu } M_{l}(-t)e^{-s_{l}t} + P_{0}(-t)\). By the previous result \(\varPhi (t)\) is not slowly increasing; as a consequence, also \(\varPhi (-t) = \sum _{l = 1}^{\nu } M_{l}(t)e^{s_{l}t} + P_{0}(t)\) it is not.
b) Let \(f(t) = \sum _{l = 1}^{\nu } M_{l}(t)e^{s_{l}t}\). By [17, Chapter VII, Theorem VI] \(f(t)\in {\mathscr {S'}}\) if and only if there exist \(n\in \mathbb {N}, g(t)\in {C^{0}}\) such that g(t) is slowly increasing and \(f(t) = g^{(n)}(t)\). It is well known that every \(g(t)\in {\mathscr {D'}}\) such that \(f(t) = g^{(n)}(t)\) may be written in the form \(g(t) = \sum _{l = 1}^{\nu } N_{l}(t)e^{s_{l}t} + Q_{0}(t)\) where \(N_{l}(t)\) and \(Q_{0}(t)\) are polynomial functions. Obviously there exists \(l \in \{\,1,\ldots ,\nu \,\}\) such that \(N_{l}(t) \not = 0\). By a), g(t) cannot be slowly increasing. \(\square \)
Corollary 5
Let
where all the \(f_{l}(t)\) are uniformly almost periodic functions, \(p \geqslant 1, f_{p}(t) \not = 0\). Then \(F(t), F(t)H(t), F(t)H(-t)\not \in {\mathscr {D'}_{L^{\infty }}}\).
Proof
Changing if necessary F(t) with iF(t), we may assume \(\mathrm{Re}\,f_{p}(t) \not = 0\). Changing if necessary F(t) with \(-F(t)\), by Theorem 7 we may assume that there exist a real number \(\mu > 0\) and a sequence \(\tau _{k}\in \mathbb {R}\) such that: \(\lim _{k\rightarrow \infty }{\tau _{k}} = +\infty \) and \(\mathrm{Re}\,f_{p}(\tau _{k}) > \mu \) for every k.
Let T be an almost period of \(\mathrm{Re}\,f_{p}(t)\) relative to \(\frac{1}{8}\mu \).
For every \(\tau _{k}\), there exists a real number \(-\tau _{k} + \sigma _{k}\in [-\tau _{k}, -\tau _{k} + T]\) (i.e. a real number \(\sigma _{k}\in [0,T]\)) such that: \(|\,\mathrm{Re}\,f_{p}(t - \tau _{k} + \sigma _{k}) - \mathrm{Re}\,f_{p}(t)\,| < \textstyle \frac{1}{8}\mu \) for every \(t\in \mathbb {R}\) and hence, changing t with \(t + \tau _{k}\), such that
Since every \(\sigma _{k}\in [0,T]\), changing if necessary the sequence \(\tau _{k}\) with a subsequence, we may assume that there exists \(\sigma \in [0,T]\) such that \(\lim _{k\rightarrow \infty }\sigma _{k} = \sigma \).
Since \(\mathrm{Re}\,f_{p}(t)\) is uniformly continuous, there exists a real number \(\zeta > 0\) such that for every pair \(( t', t'')\) such that \(|\,t'' - t'\,| < \zeta \) it is \(|\,\mathrm{Re}\,f_{p}(t'') - \mathrm{Re}\,f_{p}(t')\,| < \frac{1}{8}\mu \). Changing if necessary the sequence \(\tau _{k}\) with a subsequence, we may assume that \(|\,\sigma _{k} - \sigma \,| < \zeta \) for every \(\tau _{k}\), and hence
As a consequence of Eqs. (A.1) and (A.2) it is
Taking \(t = 0\) in Eq. (A.3) we obtain \(|\mathrm{Re}\,f_{p}(\sigma ) - \mathrm{Re}\,f_{p}(\tau _{k})| < \frac{1}{4}\mu \). Since we have \(\mathrm{Re}\,f_{p}(\tau _{k}) > \mu \), it is \(\mathrm{Re}\,f_{p}(\sigma ) > \mu - \frac{1}{4}\mu = \frac{3}{4}\mu \). Then there exists a real number \(B > 0\) such that: \(\mathrm{Re}\,f_{p}(t + \sigma ) > \textstyle \frac{3}{4}\mu \) for every \(t\in (-B,B)\). As a consequence of Eq. (A.3), it is also
Last, observe that, for every \(t \not = 0\), F(t) may be uniquely written in the form \(F(t) = t^{p}(\,\psi (t) + f_{p}(t)\,)\) with \(\psi (t)\) continuous on \(\mathbb {R}\setminus \{\,0\,\}\), and such that there exists a real number \(C > 0\) such that: \(|\,\psi (t)\,| < \textstyle \frac{1}{4}\mu \) for every \(|\,t\,| > C\).
Let \(\varPhi (t) = F(t)H(t).\)
Let \(\varphi (t)\in {\mathscr {D}}\) be such that \({\mathrm{supp~}}\varphi \subset (-B,B)\), \(\varphi (t) \geqslant 0\) for every \(t\in \mathbb {R}\), \(\int _{}^{}\varphi = 1\). For every \(\tau _{k} > C + B\) it is
Since for every \(t\in (-B,B)\) it is: \(\varphi (-t) \geqslant 0\), \((\tau _{k} + t)^{p} \geqslant (\tau _{k} - B)^{p} > 0\) and (since \(\tau _{k} + t > C\)) \(\mathrm{Re}\,\psi (\tau _{k} + t) + \mathrm{Re}\,f_{p}(\tau _{k} + t)> -\frac{1}{4}\mu + \frac{1}{2}\mu = \frac{1}{4}\mu > 0\), then
Then \((\varPhi *\varphi )(t)\) is not bounded in \(\mathbb {R}\), and hence \(\varPhi (t) = F(t)H(t)\not \in {\mathscr {D'}_{L^{\infty }}}\).
The proof for F(t) and \(F(t)H(-t)\) are straightforward consequences. \(\square \)
Corollary 6
Let
where all the \(f_{l}(t)\) and \(g_{l}(t)\) are uniformly almost periodic functions. The following statements are equivalent:
-
a)
\(F(t)H(t) + G(t)\in {\mathscr {D'}_{L^{\infty }}}\)
-
b)
For every \(l \not = 0\) it is \(f_{l}(t) = 0\) and \(g_{l}(t) = 0\)
Proof
a)\(\Rightarrow \)b). Assume for a contradiction that \(g_{l}(t) \not = 0\) for some \(l \not = 0\). The argument adopted in the Proof of Corollary 5 may be used to prove that \(F(t)H(t) + G(t)\not \in {\mathscr {D'}_{L^{\infty }}}\) (simply use a sequence \(\tau _{k}\) such that \(\lim _{k\rightarrow \infty }\tau _{k} = -\infty \)) which is absurd. Hence, it is \(g_{l}(t) = 0\) for every \(l \not = 0\). In particular, \(G(t) = g_{0}(t)\) is a uniformly almost periodic function, and hence \(G(t)\in {\mathscr {D'}_{L^{\infty }}}\).
As a consequence \(F(t)H(t)\in {\mathscr {D'}_{L^{\infty }}}\), and hence by Corollary 5 it is \(f_{l}(t) = 0\) for every \(l \not = 0\).
b)\(\Rightarrow \)a). The function \(F(t)H(t) + G(t)\) is locally integrable and bounded; hence \(F(t)H(t) + G(t)\in {\mathscr {D'}_{L^{\infty }}}\). \(\square \)
Corollary 7
Let F(t) and G(t) be uniformly almost periodic functions. If \(F(t)H(t) + G(t)\in {\mathscr {\dot{D}'}_{L^{\infty }}}\), then \(F(t) = G(t) = 0\).
Proof
Let \(\varPhi (t) = F(t)H(t) + G(t)\), \(\varphi (t)\in {\mathscr {D}}\), and \(\widetilde{\varphi }(t) = \varphi (-t)\). By the argument pointed out in the last paragraph of [17, Chapter VI, section 8] it is \(\lim _{|\,t\,|\rightarrow \infty }(\varPhi *\widetilde{\varphi })(t) = 0\). Obviously, there exists \(t_{0} < 0\) such that for every \(t < t_{0}\) it is \((\varPhi *\widetilde{\varphi })(t) = (G *\widetilde{\varphi })(t)\), hence \(\lim _{t\rightarrow -\infty }(G *\widetilde{\varphi })(t) = 0\).
By Theorem 8, \((G *\widetilde{\varphi })(t)\) is uniformly almost periodic, hence by Theorem 7, \((G *\widetilde{\varphi })(t) = 0\) for every \(t\in \mathbb {R}\); in particular
Hence, for every \(\varphi \in {\mathscr {D}}\) it is \({\left\langle G(t),\varphi (t) \right\rangle } = 0\). As a consequence \(G(t) = 0\) and \(\varPhi (t) = F(t)H(t)\).
Let again \(\varphi (t)\in {\mathscr {D}}\), and \(\widetilde{\varphi }(t)=\varphi (-t)\). Observe that there exists \(t_{0} > 0\) such that for every \(t > t_{0}\) it is \((\varPhi *\widetilde{\varphi })(t) = (F *\widetilde{\varphi })(t)\), hence \(\lim _{t\rightarrow +\infty }(F *\widetilde{\varphi })(t) = 0\). As a consequence, the argument previously applied to G(t) proves that \(F(t) = 0\). \(\square \)
Rights and permissions
About this article
Cite this article
Ciampa, M., Franciosi, M. & Poletti, M. Linear Differential Equations and Related Continuous LTI Systems. Circuits Syst Signal Process 38, 4465–4503 (2019). https://doi.org/10.1007/s00034-019-01080-7
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00034-019-01080-7