I am looking for information on (formal) algebraic systems that can be used to transform time-series - in either a practical or academic context.

I hope that there exists (at least one) small, expressive, set of operators - ranging over (finite) time-series. I want to compare and contrast different systems with respect to algebraic completeness, and brevity of representation, of common time-series transformations in various domains.

I realise this question is broad - but hope it is not too vague for datascience.stackexchange. I welcome any pointers to relevant literature for specific scenarios, or the general subject.

Edit... (Attempt to better explain what I meant by an algebraic system...)

I was thinking about "abstract algebras" as discussed in Wikipedia:

http://en.wikipedia.org/wiki/Algebra#Abstract_algebra http://en.wikipedia.org/wiki/Abstract_algebra#Basic_concepts

Boolean Algebras are (very simple) algebras that range over Boolean values. A simple example of such an algebra would consist the values True and False and the operations AND, OR and NOT. One might argue this algebra is 'complete' as, from these two constants (free-variables) and three basic operations, arbitrary boolean functions can be constructed/described.

I am interested to discover algebras where the values are (time-domain) time-series. I'd like it to be possible to construct "arbitrary" functions, that map time-series to time-series, from a few operations which, individually, map time-series to time-series. I am open to liberal interpretations of "arbitrary". I would be especially interested in examples of these algebras where the operations consist 'higher-order functions' - where such operations have been developed for a specific domain.

  • 1
    $\begingroup$ sounds more like something for stats or even maths stack exchange sites... $\endgroup$ – Spacedman Sep 16 '14 at 16:35
  • $\begingroup$ Thanks for those suggestions. I am considering a similar question for the maths site... but hope to discover insights from data-science experts first. The question straddles practical data science and theoretical maths. $\endgroup$ – aSteve Sep 16 '14 at 22:42
  • $\begingroup$ This is a really interesting question and I agree it has practical applications, but I also really think it belongs on Mathematics. It should get a lot more exposure there. $\endgroup$ – shadowtalker Oct 17 '14 at 17:00
  • 1
    $\begingroup$ You can't be successful if you don't have a goal. What kind of insights do you want to get? You have to define that first. Otherwise: Just use element-wise addition and multiplication. You're done now! Or aren't you? If you understand why this doesn't do it - maybe you can make progress. $\endgroup$ – Gerenuk Nov 16 '14 at 16:06
  • 1
    $\begingroup$ @aSteve: I could make up many non-trivial algebras, too. It would be a very tedious process asking you every time why this particular algebra does not suit your needs. It's really much more fruitful if there is an notion of what you want to achieve. $\endgroup$ – Gerenuk Nov 16 '14 at 20:55

It seems like you are overthinking this and getting caught up in your own head a little.

We can define a restricted set of time series as those which are well behaved (non-asymptotic) and single valued. These time series then form a time series algebra which have additive, multiplicative, commutative, and associative properties. There is an identity element (straight line with value 1) and a zero element (straight line with value 0). We even have a well defined inverse member for all time series not crossing zero. I believe this even forms an Abelian group.

Further, previous posters are correct that this group can be perfectly represented by various orthogonal vector spaces such as polynomials and Fourier space. This is true whether we are talking about Lagrange polynomials, Fourier series, or Gauss-Lobatto-Legendre polynomials.

The point then becomes... where does this get you? You got caught up in the suggestion to employ Fourier series in that you didn't want to remain in frequency space, but this is entirely unnecessary. You can, for instance, Fourier transform into frequency space, apply a high frequency filter, and then transform back into the time domain. This is incredibly effective in removing noise (although band pass filters tend to work better).

Stepping back from the question a little, it seems like you might really be asking: Given typical time series and typical time series operations (smoothing, de-seasoning, de-trending, etc) do the time series and their typical operations constitute a group with a well defined algebra? The answer is, yes. All of these operations just involve finding other time series which can be added, subtracted, multiplied, or divided by the original time series in order to yield a more physically tractable (interpretable) representation.

  • $\begingroup$ Thank you for this considered reply - especially the 'Stepping back...' paragraph. I'm aware that various operations on time series constitute groups. The rub is "finding other time series". Rather than rely upon serendipity, I'd like to establish how to find these "other time series" by constructing candidates systematically. I expect the effective strategies to be domain-specific. I am focused on time domain over frequency domain approaches because the time series which interest me most, right now, are finite length and non-cyclical. $\endgroup$ – aSteve Jul 11 '15 at 15:47

The most direct and obvious transformation is from time domain to frequency domain. Possible methods include Fourier transform and wavelet transform. After the transform the signal is represented by a function of frequency-domain elements which can be operated on using ordinary algebra.

It's also possible to model a time series as a trajectory of a dynamical system in a state space (see: http://onlinelibrary.wiley.com/doi/10.1111/j.1467-9892.1980.tb00300.x/abstract, and http://www3.stat.sinica.edu.tw/statistica/oldpdf/A2n16.pdf). Dynamical systems can be modeled symbolically at a course-grain level (see: http://en.wikipedia.org/wiki/Symbolic_dynamics and http://www.math.washington.edu/SymbolicDynamics/) Symbolic dynamics draws on linear algebra.

  • $\begingroup$ Thank you for this interesting answer - it isn't what I anticipated. Time/frequency domain (Fourier/Wavelet) transforms operate on individual time-series, a very restrictive algebra. "Dynamic Systems" - on the other hand - focus on establishing models for time series - as opposed to establishing algebras ranging over time series. I hoped to discover sets of operators ranging over time-series that are 'somehow adequate'. $\endgroup$ – aSteve Sep 17 '14 at 9:59
  • $\begingroup$ Then I'm not sure what you mean by "algebra over time-domain time series". FWIW, you can operate algebraically in frequency domain to express relations over many time-series equivalents. A simple example: formulas for musical scales, incl. harmonics. Also, I don't know why you'd want to restrict yourself to time domain, given that frequency domain contains all the same information. $\endgroup$ – MrMeritology Sep 17 '14 at 12:34
  • $\begingroup$ I didn't intend to 'knock' exploiting the frequency domain... an approach that often yields seemingly 'magical' results. This question is focused exclusively on functions transforming time-domain time-series to time-domain time-series. I've added some details about abstract algebras (and a "Boolean Algebra" analogy) which, I hope, makes my question clearer. I am not excluding the possibility that operations could be defined using the Fourier transform... for example... but I don't want frequency domain representations as "outputs" from the operations which make up the algebra. $\endgroup$ – aSteve Sep 17 '14 at 13:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.