Algorithmic Composition by Autonomous Systems with Multiple Time-Scales

DOI: 10.5920/DivP.2021.01


Dynamic systems have found their use in sound synthesis as well as score synthesis. These levels can be integrated in monolithic autonomous systems in a novel approach to algorithmic composition that shares certain aesthetic motivations with some work with autonomous music systems, such as the search for emergence. We discuss various strategies for achieving variation on multiple time-scales by using slow-fast, hybrid dynamic systems, and statistical feedback. The ideas are illustrated with a case study.


With few exceptions, such as pure drone pieces, most music consists of auditory patterns that vary over time. The variation has often been found to follow a scale-free distribution, or a 1/ƒᵝ power law, which implies variation at all temporal scales and a balance between predictability and unpredictability (Levitin et al., 2012). This power law has been observed across several centuries of Western music, and in different musical dimensions such as pitch and rhythm. The exponent 0 < β < 2 varies between composers and styles, where lower values imply more variation and higher values less variation and more predictability.

Different temporal scales correspond to different zones of perception, from the audio rate variation giving rise to pitch and timbral perception; fast modulation corresponding to grittiness and roughness; slower modulation as in tremolo and vibrato; events sufficiently brief to process in short-term memory forming notes, phrases, gestures or motifs; and longer processes by which we are able to segment the audio stream into formal sections. A similar separation of structural levels can be found in many sound synthesis languages such as Csound. There is the low level of audio rate sound synthesis, an intermediate level of control signals and the high level of discrete time note events.

The audio signal and its lower level attributes such as amplitude, fundamental frequency, and spectral content are often treated as raw material subject to external organisation by compositional procedures or realtime input. It will be useful to distinguish two approaches to the organisation of material, namely those of lattice-based music on the one hand and dynamic morphology on the other (Wishart, 1996). Algorithmic composition often deals with discrete sets of pitches, a temporal grid of onsets and duration values, and a discrete set of symbolic dynamic levels such as $pp$, $p$, $f$, $ff$. Whereas written notes can be ordered on a lattice, dynamic morphology is concerned with the complexity of sound objects that change over time and cannot be easily ordered into scales. We will outline an approach to algorithmic composition that is well adapted to dynamic morphology, although it can also handle the discrete type of events of lattice-based composition.

Historically, various strategies of algorithmic composition have been applied primarily to the note level and used for score generation (Ames, 1987). Some notable exceptions, including the SAWDUST pieces by Herbert Brün and the GENDYN pieces by Xenakis, using what is often referred to as nonstandard synthesis (Döbereiner, 2011), manage to bridge the separation between the low level of sound synthesis and the higher level of large scale form. In this unified approach to the micro and macro-levels, the waveform is composed as much as the entire piece, sometimes applying similar procedures on all levels.

In works inspired by cybernetics and complex systems, typically using networks of feedback systems (Sanfilippo and Valle, 2013), a side-effect may be the upheaval of any meaningful separation of temporal scales. Sound-generating processes that depend explicitly on a few milliseconds of past generated sound can nevertheless result in slower large-scale processes. Feedback systems are ubiquitous where self-organisation and emergence are observed.

Another feedback loop is present in virtually any artistic endeavour, namely the action – perception – evaluation loop. Algorithmic composition is a circular process which may begin with the creation of an algorithm, perhaps with a particular musical expression in mind, then the algorithm generates a piece of music which is evaluated by listening, followed by a cycle of further modifications of the algorithm and evaluations of the output. The process of composition can be highly interactive even if the interaction does not take place in realtime.

The approach outlined in this paper can be situated in the intersection of algorithmic composition and self-organising music systems. Computer simulations of autonomous dynamical systems require an algorithm; hence, when applied to composition it arguably should qualify as algorithmic composition. There are interesting parallels between this type of algorithmic composition and (non-algorithmic) self-organising music systems, not least from an aesthetic perspective.

Martin Supper (2001) distinguishes three categories of algorithmic composition: 1) Modelling traditional compositional procedures, 2) modelling original and novel compositional procedures, and 3) borrowing algorithms from extra- musical disciplines. Style imitation by algorithmic composition will not be our concern. As for Supper’s second and third categories, a strict separation is not necessary. Certain ideas can be borrowed from other sciences and adapted to the needs of musical composition. This is particularly clear when working with dynamic systems and differential equations. When this borrowing has been established over a few years, these previously extra-musical techniques are absorbed into the regular toolbox of musical techniques.

Composition by algorithms or by self-organising processes is detached from the immediate decision making typical of more intuitive approaches. Instead of working directly with the musical material one works through an intermediary, either by running code on a computer or by building some electronic apparatus. The computer follows the instructions of the program code and the electronic machinery follows the laws of physics without the composer’s direct interference. Realtime interaction is optional, but the system might not be fully controllable.

Autonomous systems can be formulated as a set of equations that describe what will happen the next moment as a function of the current state. There is no schedule or plan for future events, nor is there necessarily any memory of past events. That is one source of difficulties often encountered in this kind of work, as Dario Sanfilippo points out.

In my experience, the realisation of an autonomous music system which exhibits a convincing variety and complexity over a relatively long time span has been something difficult to achieve, even when implementing large and articulated networks. (Sanfilippo, 2018, p. 123)

Indeed, the goal is often to achieve some level of complexity and some amount of variety over time, and the solution on offer appears to be to add layers of mechanisms to monitor and automatically adjust system variables.

Agostino Di Scipio’s description of his work on Ecosistemico Udibile is illuminating in this respect (Di Scipio, 2007). The description of the piece begins with a simple electroacoustic feedback loop from which the system grows by a process of accretion. Negative feedback loops are added in order to balance the Larsen effect and positive or nonlinear feedback loops are added to increase the complexity of the system’s dynamics. The system appears to have grown from inside out like an onion, layer upon layer. When designing an autonomous music system one cannot compose the music from beginning to end, one has to design mechanisms that respond to situations that may arise. To be on the safe side, one designs mechanisms also for situations that may not arise, since what will happen in an open and highly complex system is largely unpredictable.

In the rest of this paper we focus on deterministic autonomous dynamical systems, beginning with an overview of previous work on ordinary differential equations for sound synthesis. Then we discuss self-organisation and emergence as it relates to composition with autonomous systems. Next, we briefly consider autonomous systems, followed by an introduction to slow-fast systems and hybrid systems that combine discrete time and continuous flow. These concepts are then applied in a review of previous work on feedback systems with feature extractors in the loop. We also discuss how statistical feedback can be used as a means to increase a system’s variability. A case study shows how several of the ideas can be applied to composition. Questions concerning the evaluation of this class of algorithmic composition systems are addressed in the conclusion.

Previous work on sound synthesis by ordinary differential equations

Chaotic maps and ordinary differential equations (ODEs) have long been applied to the generation of note sequences (Bidlack, 1992). The translation of raw data from the orbits of a dynamic system to musical notes requires quantization into discrete values and a choice of mapping from the state variable to musical data. Continuous time systems must be sampled at discrete time steps, for instance by taking a Poincaré section. Therefore, maps are inherently more suitable for the generation of discrete events such as note sequences, whereas the smooth flow of ODEs makes them ideal for sound synthesis. If the oscillations are sufficiently fast the state variables may be used, after proper amplitude scaling, as an audio stream. Slower oscillations are suitable as modulating signals.

ODEs have not quite attained the popularity of other synthesis techniques such as additive, subtractive, granular or physical modelling. Nonlinear oscillators may have a non-trivial relationship between system parameters and the qualitative character of the audio signal they produce. Pitch, loudness and timbre may change simultaneously by the variation of a single system parameter. Analogous codependencies exist in acoustic instruments (e.g., the common correlation between loudness and spectral brightness) and are not necessarily unwanted. Acoustical instruments have been modelled with ordinary and partial differential equations. An analysis method has been proposed that reconstructs an attractor from a recording of an instrument tone and finds a dynamic system capable of producing the attractor (Röbel, 2001). For sound synthesis there is no need to limit oneself to the simulation of acoustic instruments, any dynamic system with a globally stable oscillatory state is potentially interesting.

Chua’s circuit had been explored as a source for sound synthesis early on (Mayer-Kress et al., 1993). It was found to be capable of bassoon-like timbres, as well as percussive sounds by using an initial transient towards a fixed point. Rodet and Vergez (1999) were not so satisfied with Chua’s circuit on its own, but they found that extending the system with a delay line, thereby turning it into a delay differential equation, enriched its sonic register and provided an interesting link to other work on physical modelling of acoustic instruments.

Before digital computers were up to the task, nonlinear differential equations were solved on analog computers. Slater (1998) suggested the use of analog computers in combination with modular synthesizers for chaotic sound synthesis. In a similarly adventurous spirit Collins (2008) introduced a few unit generators for SuperCollider implementing various nonlinear ODEs. There is an ongoing search for new chaotic systems (Sprott, 2010), many of which can be realised as electronic circuits suitable for sound synthesis. In recent years many chaotic oscillators have been introduced as analog modules for modular synthesizers [1] and musicians are exploring sound synthesis using these chaotic systems.

Virtual analog modelling often needs to handle the problem of immediate feedback paths in analog circuits, such as two mutually modulating FM oscillators. Digital implementations usually introduce a one sample delay for such feedback paths, but delay-less, more accurate versions can be constructed from differential equation models of the original system (Medine, 2015). Stefanakis et al. (2015) introduced a few useful techniques by relating complex-valued, time-dependent systems of ODEs with input signals to more familiar concepts of sound synthesis and filtering. Complex variables have the advantage that amplitude and frequency can be modelled in a single variable. Using noise as input, these systems become stochastic differential equations.

Jacobs (2016) describes a system of connected Fitzhugh-Nagumo oscillators on a graph, which are excited by a wave modelled by a partial differential equation. The function of the travelling wave is similar to a higher level control function, and Jacobs describes it as a sequencer or a rudimentary tool for algorithmic composition. This integration of sound synthesis and control level signals is somewhat similar to the approach that we will pursue in this paper.

Although far from complete, this literature survey hopefully shows the diversity of approaches to differential equations in the synthesis of musical signals. Ordinary, as well as delay, stochastic, and partial differential equations have been explored and may be useful as sources of variation on multiple time scales.

Emergence and surprise

An important motivation for making music with autonomous systems is the search for self-organisation and emergence. A useful summary is provided by Wolf and Holvoet (2004), who list a few criteria often considered crucial for emergence:

  1. Global behaviour, properties, patterns, or structures result from interactions at a lower level. These higher level phenomena are called “emergents”.
  2. The global phenomenon is novel and not reducible to the micro-level parts of the system. In the words of Wolf and Holvoet, “radical novelty arises because the collective behaviour is not readily understood from the behaviour of the parts.” We return to this point below.
  3. Emergents maintain a sense of identity or coherence over time.
  4. For emergence to occur, the parts need to interact. Therefore, the system should be highly connected at the low level.
  5. Decentralised control of the system implies that no single part is responsible for the macro-level behaviour; the system as a whole is not controllable.
  6. In turn, the decentralised structure makes the system flexible and robust against small perturbations. Parts of it may be replaced without changing the emergent.

Self-organisation, according to a view that goes back to Ashby, occurs when the degree of organisation or order increases within a system as it evolves by its own dynamics. By this understanding, any dissipative dynamic system that approaches an attractor is a self-organising system.

Some rigorous and quantifiable approaches to emergence and self-organisation have been proposed (Prokopenko et al., 2008). A set of measures introduced by Gershenson and Fernández (2012) relate the amount of information or Shannon entropy at the input to that at the output of a system. Although it is not always clear how the amount of information should be measured, in particular when considering the musical output of a complex system where perceptual criteria should arguably play a decisive role, the idea of comparing input and output information is worth considering. We will return to this point in the Conclusion.

In an overview of some interfaces for self-organising music, Kollias (2018) emphasises electro-acoustic feedback as a primordial element around which many of the works have been structured. Although the openness to the acoustic environment perhaps puts these systems in a special category, feedback can be explored in the digital or analog domain as well, or in any mixture of domains. An emerging category that fits the description of self-organising music interfaces very well is modular synthesizers and what is often referred to as “self generating patches” [2]. These are analog or hybrid analog/digital systems set up in large networks of modules that may run autonomously and produce complex sequences of music.

Surprise, as well as emergence, are frequently mentioned as desired qualities in work involving self-organising music systems. Emergence is often described in terms of the expectancies of an observer, defining “the quality of unexpectedness of the results” (Sanfilippo and Valle, 2013, p. 18); see also the already quoted view that novelty arises because collective behaviour is not readily understood from the behaviour of the parts (Wolf and Holvoet, 2004). This would seem to imply that emergence is a mere side-effect of not knowing exactly what to expect, of lacking a full understanding of the system’s dynamics. Yet, one can argue, having an inkling of what the system is capable of is necessary for building up an expectation – that can then be thwarted when the system behaves in an unexpected way. In any case, as Kivelson and Kivelson (2016) point out, the definition of emergence according to which “something is qualitatively new if it cannot be straightforwardly understood in terms of known properties of the constituents” suffers from many shortcomings; “perhaps the most glaring is that it implies that as soon as something is understood it ceases to be emergent.”

Novelty effects also wear out with repetition, which shows the troubling ephemerality of emergence as defined in terms of the observer’s reactions. Nevertheless, anticipation of what will happen next in a piece of music is a crucial part of the listening experience, as Huron (2007) discusses at length. From an evolutionary perspective, surprise indicates a failure to predict an event in our environment, which ultimately can be bad for our prospects of survival. As Huron points out, we actually enjoy being right in our predictions of what is going to happen next in a piece of music, including correctly predicting the regular recurrence of a downbeat or the chord sequence of a cadence. This seemingly contradicts the common wisdom that we enjoy surprises in music. Violated expectations, after all, produce reactions like frisson, awe, or laughter.

The surprise of a practitioner of self-organising music as the system, for some poorly understood reason, generates an output that is more complex than expected is very different from the reaction of a listener who is not aware of what is going on inside the system. It is by no means illegitimate to seek out these surprising situations as a practitioner, but we should be aware that for the uninformed listener, the surprise is a function of previous listening experiences and whatever expectations the piece itself sets up, in contrast to the expectations of the composer, the one who built the system and knows a few things about its inner workings.

As for emergence, “musical form emerges from interactions composed at the signal level” as Di Scipio (2007) puts it concerning his own work. Indeed, this would be a good example of emergence independent of the observer and their reactions.

Autonomous systems

We now turn to a more technical description of some aspects of dynamic systems that are of importance in multi-scale algorithmic composition. The term ’autonomous system’ has a rather precise meaning in the context of dynamic systems, and a less stringently defined meaning in the context of algorithmic or generative music.

Let us recall the definition of a dynamic system in continuous time, t, with a state variable 𝑥 ∈ ℝ. A general ODE is described by an equation evolving from an initial condition 𝑥(0) = 𝑥₀ with a constant or time-variable parameters p(t) ∈ ℝ. An autonomous system has no explicit time dependence, so it has the form ẋ = ƒ (𝑥; p) where p is constant.

From a musician’s point of view, autonomy rules out realtime interaction and external control of the system. Thus, without abuse of terminology, autonomous systems have no place in interactive live-electronics where system parameters or the state variable itself may be put under the performer’s influence, nor can the system receive an input signal.

The opposite of an autonomous system is a forced or driven system. To complicate things, we note that the distinction is also a matter of perspective.

Consider the equation ẋ = −𝑥 + sin t which has the non-autonomous forcing term sin t. This system can be reformulated as an autonomous system in two ways. First, one could introduce a new time variable, τ, and write the system as

Alternatively, sin t can be expressed as the orbit of a harmonic oscillator which takes two new state space variables and initial conditions.

Similar distinctions can be discussed in the context of self-generating patches on modular synthesizers. The goal is to build a patch that generates interesting musical sequences of its own accord without manual interference. Is the patch truly autonomous (“self-generating” or autopoietic) if there is a sequencer driving it, or an LFO or noise source? One could argue that the patch is more autonomous—a matter of degrees—the fewer sources of modulation or discrete events there are that are not themselves modulated by other parts of the patch.

Interconnectedness may therefore be a more useful criterion than autonomy (see Fig. 1). In fully connected systems all parts receive input from all other parts. In the limit everything is bidirectionally coupled to everything else. Then there can be no external sources of modulation or control; hence, the system must be autonomous. Notice also that interconnectedness was listed as one of the criteria for emergence (see point 4 in the list in the previous section).

Slow-fast systems

Relaxation oscillators such as the van der Pol oscillator or the stick-slip mechanism in bowed string instruments are well-known examples of slow-fast systems.

Figure 1: Left: driven system, right, autonomous system.

A number of special techniques have been introduced to simplify the treatment of such multi-scale systems (see Strogatz (1994) for some of them). Our motivation for discussing slow-fast systems is that they provide a convenient conceptual framework for describing compositional models that integrate audio synthesis and larger scale levels.

A general slow-fast system in two time-scales may be written:

– Equation (1) –

where ε is a small positive time scaling factor, 𝑥 ∈ ℝ is the slow subspace and 𝑦 is the fast subspace. For example, an audio rate oscillator and an LFO modulating each other could be modeled as a slow-fast system. With mutual coupling between the slow and fast subspaces their dynamics are intertwined, although with a sufficient separation of time-scales or a loose enough coupling some simplifying assumptions can be made for a qualitative understanding of the dynamics of each subspace.

In particular, in the fast subspace   = g(𝑥, 𝑦) the variable 𝑥 may be regarded as a set of slowly drifting parameters. As the parameters drift by some small amount, the fast subsystem may change smoothly. The drifting parameters may also cause a bifurcation producing a qualitatively different dynamics in the fast subsystem. As long as the fast system does not bifurcate one could think of it as an attractor continuously varying in shape and the fast orbit as permanently being in a transient state chasing the current attractor (Ruelle, 1987).

Conversely, with the slow subsystem modulated by the fast subsystem the rapid oscillations may have an effect similar to noise, where the average position ⟨𝑦⟩ in the fast system’s phase space will act like a constant parameter value with an added ”noise” term ξ. Then we may write the slow system as

The usefulness of these simplifications depends on the strength of the coupling between the two subsystems, the separation of their characteristic time-scales and the exact form of the equations.

Fujimoto and Kaneko (2003) have shown that in a chain of coupled chaotic systems, each slower than the next one by a constant factor, under certain conditions the fastest system can influence the slowest. For this to happen the separation of time-scales must be in a certain range; there must be a bifurcation in the fastest system and the bifurcation must cascade through to the slower systems. In their particular model, despite mutual coupling, the influence went only from faster to slower systems. It is probably more common to have slow subsystems influence the faster subsystems.

Slow-fast systems commonly describe spiking or bursting dynamics, such as firing neurons, where the state variable moves slowly for most of the time and then jumps or oscillates rapidly. In contrast, the slow-fast systems we are interested in should have a more or less permanent fast time-scale for audio signals and slower time-scales for synthesis parameters to ensure variation over time.

Hybrid systems

Other useful ideas are differential equations with discontinuous right-hand side and hybrid systems that combine the continuous flow with the discrete time of maps. Many theoretical results about flows assume smooth vector fields, but non-smooth systems such as piecewise smooth functions are useful models of many mechanical and electronic phenomena. In particular, what makes this class of systems interesting is that they allow for sudden changes.

Hybrid systems can have their discontinuities imposed at certain points in time, such as

where Tᵢ are disjoint sets of time intervals whose union covers all time for which the system is defined, such that the flow follows different equations at different time intervals. Alternatively, the system can switch between different sets of continuous flow equations when the state variable passes from one region of the state space to another. Hysteretic switching is also possible, where the switching happens only if the state variable approaches the switching point from a certain direction (Saito, 2020).

Figure 2: Dynamics near the switching manifold. a) The vector fields on both sides point in roughly the same direction and the flow crosses the switching manifold. b) If the vector fields point away from the switching manifold the solution might not be unique. c) When the vector fields point inwards to Σ the trajectory may stick or slide along it.

Since we are interested in autonomous systems we will consider discontinuities of the form

induced by switchings depending on position in state space. For simplicity we consider a system separated into two distinct regions, Ω₁,₂ but the idea easily extends to any number of regions.

Suppose each region Ωᵢ of continuous flow is governed by the equation = ƒᵢ (𝑥). The border between regions is called the switching manifold, Σ. What happens when 𝑥 ∈ Σ is not obvious, the system might not even have a unique solution (Danca, 2010). Three different situations near a switching manifold are illustrated in Figure 2. When the vector fields on one side point into the other side and the vector field on the other side points further away (a) the trajectory will pass through. In the case of vector fields pointing away from the switching manifold there may be no unique solution for an orbit starting on Σ itself (b), and finally (c), if the vector fields on both sides point inwards to Σ the trajectory may approach it and start sliding along it.

In addition to the familiar bifurcations that occur in smooth systems some new bifurcation scenarios are only observed in discontinuous or non-smooth systems (Makarenkov and Lamb, 2012). So called grazing happens when a periodic orbit touches the switching manifold, which models situations such as a swing- ing clapper just touching a church bell. Another example is the friction causing the squealing noise of brakes.

Modular synthesizers are essentially hybrid systems. Oscillators and filters generate and process continuous time signals, whereas triggers and clock signals are discrete time events. Similarly to the switching functions discussed above, the continuous signals can be segmented with sample & hold or analog shift registers.

In autonomous ODEs used for algorithmic composition, hybrid systems allow for smooth flows suitable for audio signals to have points of instant change. Applied to frequency, one can articulate a discrete set of pitches instead of having a constant glissando. Combining a slow-fast system with switching is particularly interesting when the discontinuities are defined on the slow subspace, as we will demonstrate in the case study below.

Feedback from feature extractors

In the search for mechanisms that equilibrate a feedback system it is quite natural to turn to feature extractors, such as in Di Scipio’s Feedback Study (Di Scipio, 2007). My previous research centred on a class of autonomous systems comprised of an oscillator or signal generator, a feature extractor analysing the oscillator’s output, and a mapping unit that transforms the feature extractor’s output to synthesis parameters for the signal generator (Holopainen, 2012); I have proposed to call these systems Feature Extractor Feedback Systems, or FEFS for short.

FEFS network of five coupled oscillators using adaptive zero crossing rate as feature extractor.

For a feature extractor to be useful inside a feedback system, it should process short segments of the most recent audio output and be able to provide output without too much latency. Time domain feature extractors that generate audio rate output are particularly suitable for this purpose. Block-based processing using DFT or other transforms typically deliver output values at a much slower rate. Unless the output is interpolated, block-based processing imposes its own regular pace of updates which tends to become a dominant and easily audible effect.

Figure 3: Left: generic FEFS model, Right: filtered map.

By adapting to their own output FEFS have useful applications such as automatic pitch correction of nonlinear oscillators with unknown functional relations between parameters and pitch. Another interesting scenario would be to connect several FEFS in networks where each unit analyses and responds to the other units.

A few detailed studies of various FEFS models led me to conclude that the feature extractor’s most salient contribution was a smoothing effect, which we will explain shortly. A broad class of FEFS can be described by the equations

where 𝑥$_{n}$ is the audio output, θ$_{n}$ is an internal state or phase variable of the signal generator, $\pi_{n}$ are the synthesis parameters, and ϕ$_{n}$ is the output of a feature extractor which operates on the last L output samples (see Fig. 3). All variables may be vector valued.

Let us consider feature extraction using zero crossing rate (zcr) as an example. The zero crossing rate can serve as a crude pitch estimator or a descriptor of spectral balance. There is a well-known trade-off between temporal acuity and precision; a longer feature extractor window provides more accurate frequency estimates but also smears out sudden changes in the analysed signal, whereas shorter windows respond faster to changes but with less accuracy.

There are a few different implementations of zcr, a popular choice being to count the number of zero crossings during the past L samples and divide by L. Another, probably less common way is to tally up to a certain fixed number of N zero crossings and then divide by the number of samples elapsed since the first counted crossing. This method adapts to the signal’s content and uses a longer effective window length for lower frequencies than for high frequencies.

The mapping $\pi_{n+1}$ = m(ϕ$_{n}$ ) from feature extractors to synthesis parameters also plays an important role. Making this mapping highly nonlinear increases the susceptibility of the FEFS to exhibit wild behaviour, whereas smoother mappings are likely to make ϕ$_{n}$ and $\pi_{n}$ settle on some fixed values, thus causing the generator to output a signal with constant synthesis parameters. As a rule of thumb, increasing the feature extractor’s analysis window length has the opposite effect of increasing the nonlinearity of the mapping.

In order to emphasise an essential point of FEFS we may grossly simplify and reformulate the system as a filtered map (Fig. 3, right part). The feature extractor operates on a running block of samples, and its effect can be modelled as a time average combined with some nonlinearity. The simplified system

lumps together the oscillator and mapping in ƒ(𝑥, 𝑦), and 𝑦 represents the output of the feature extractor; g(𝑥) represents some nonlinear function of the audio signal, whose time average ⟨𝑥⟩ is taken over the last L samples, corresponding to the effective window length of the feature extractor.

Clearly, the averaging smooths out any rapid changes, so 𝑦 is a slow variable. Notice that even if ƒ(𝑥, 𝑦) were a chaotic map, a long enough smoothing of its orbit in the feature extractor’s averaging part makes it likely that the parameter 𝑦 will approach a constant value with perhaps some small fluctuations. In other words, this situation corresponds exactly to the slow-fast system (Eq. 1) where we approximated the slow variable as a constant plus noise term.

It has proven difficult to design FEFS that exhibit non-trivial behaviour over extended time. Often there is a more or less prolonged initial transient phase after which the slow variables (ϕ$_{n}$ and $\pi_{n}$) approach some fixed state. The analysis in terms of a slow-fast system explains why this often happens. Chaotic solutions are also attainable with short effective window lengths and strongly nonlinear mapping functions.

Statistical feedback and mapping with memory

Deterministic FEFS, as formulated above, are limited by the fact that the mapping from feature extractors to synthesis parameters involves no memory of its past beyond the effective length of the feature extractor. No monitoring mechanism allows the system to discover if it has got stuck on some repeating cycle that extends beyond the length of the feature extractor. It might resemble an improviser suffering from a loss of short-term memory, who would not be able to guide the performance in any particular direction other than wherever the haphazard steps of a Markov chain brings it.

Algorithmic compositions using autonomous systems may stray far from the syntax and material of the common-practice music that has been extensively studied by music scholars. Nevertheless, certain insights from music theory and music psychology may be relevant, even though its focus has been lattice-based structures rather than dynamic morphology, to borrow Wishart’s terminology again. Research in music psychology has provided ample evidence that statistical learning of musical patterns plays an important part in forming listener expectations (Huron, 2007). First order probabilities, such as the distribution of pitches or scale degrees, allow us to identify scales and their tonal centres. Higher order probabilities, including transition probabilities between pitches or durations, contribute to the recognition of different styles.

As noted in the introduction, it is a common experience that the goal of variety over time may be elusive in autonomous music systems. Variety can always be increased by adding another regulatory mechanism, a new layer of the onion that makes up the algorithm. A strikingly simple and efficient technique called dissonant counterpoint was pioneered by James Tenney. Originally it was used to enforce a certain amount of variation on randomly generated pitch sequences (or dynamics, durations, whatever parameter one wishes to control). Random sequences may have occurrences of short subsequences that appear more orderly than one might naively expect, such as immediate repetitions or alternations between elements. Tenney’s method guarantees that such close repetitions are ruled out. Simply put, Tenney’s scheme goes as follows (Polansky et al., 2011):

  1. Initialise an array of N entries all with the same positive number (say, 1). Each element corresponds to a unique pitch.
  2. Interpret the values stored in the array as relative probabilities. Pick an element randomly using its relative probability.
  3. Reset the value of the chosen entry to zero and increase the values of all other entries.
  4. Repeat from step 2.

This simple scheme will make a direct repetition of a pitch impossible and a close repetition unlikely. Tenney’s method cannot be used as such in a deterministic autonomous system since it relies on random choices, but the idea of keeping track of the statistics of past states can be generalised to suit our needs.

In a FEFS, one can keep track of the relative frequency of past synthesis parameter or feature extractor values. In practice, one would use a number of discrete bins for a histogram and introduce a mechanism tipping the system in a direction that favours the production of less often occurring values. This, in effect, expands the dynamic system’s memory of its past behaviour without necessitating storage of the entire time series. There is no guarantee that the system will not settle on cyclical patterns, but at least the cycles are likely to be longer than they would otherwise be (see Holopainen (2012, pp. 275-280) for further details).

The goal distribution of synthesis parameter values does not have to be uniform. Uneven distributions lead to a differentiation of common and rare events. If these events are perceptually distinguishable, the rare ones are those that are prone to carry some significance for the listener by being outliers or exceptions. Statistical learning is at play also in the course of listening to a piece for the first time (this is what Huron (2007) calls dynamic expectations, in contrast to those expectations that pertain to a whole genre which he refers to as schematic). Expectations are shaped by the probability distribution of events as they occur in a piece.

The information content and information density of pieces of music (amount of information over time) has been studied in music theory for decades. The less probable an event is, the higher its information content. As Temperley (2019) points out, it would be a formidable task to quantify the amount of information in any musical piece, since everything (melodic pitches, harmony, rhythm, timbre, and other aspects) contributes to information. And this is still only considering common-practice music with large corpora available for study and the information-carrying units relatively easy to identify.

Algorithmic composition beginning from sound synthesis, where higher levels emerge from processes at lower levels, poses the additional difficulty of relating synthesis parameters to resulting audio signals, and the audio signals to their perceptual correlates, and of defining the units that carry information. There is a gap between theories of information applicable to the symbolic note level and the dynamic morphology articulated by sound synthesis with freely flowing synthesis parameters. Some possible approaches to information density or complexity will be discussed in the conclusion.

Case study: The auto-detuning system

We turn now to a simple three-dimensional ODE which serves as a building block in two algorithmic compositions. It is an example of a slow-fast system with discontinuous derivative on its right-hand side. The system consists of two oscillators that are detuned by an amount that depends on the amplitude of the sum of the oscillators.

Sound example: the auto-detuned system with $\omega=1$ and $c = 0.8, 0.9 1.0$

In its simplest form, the system is given by

– Equation (2) –

with detuning δ ≥ 0, phase variables θ$_{1,2}$ and frequency ω. The parameter c ≥ 0 indirectly affects the amount of detuning. Time series of sin($\theta_{i}$) and δ are shown in Fig. 4.

For now, let ω = 1 and let Ω $_{i}$ = 1±δ be the frequencies of the two oscillators. If the frequency were stable the two oscillators would be tuned to a constant ratio R = Ω$_{1}$/Ω$_{2}$ which is known as the rotation number, and may be calculated as

For 0 < c < 1/4 the two oscillators are perfectly locked in sync, so R = 1. As c increases from c = 1/4 the rotation number decreases.

Figure 4: Time series of the detuning system with ω = 1 and c = 1, apparently chaotic at these parameter values.
Top to bottom: sin($\theta_1)$, sin($\theta_2)$, and $\delta$

Assuming that δ approaches some constant value, the phases will increase at a constant rate. Under such conditions we can calculate an average value of the expression in the third equation of Eq (2), for which we introduce the variable

If we knew the time average ⟨ϑ⟩ we could solve the equation

separately, which is easy; it will simply approach c⟨ϑ⟩ asymptotically. Under certain assumptions (δ being constant and R irrational), the average can be found by evaluating

over the region S = [0, 2$\pi$] × [0, 2$\pi$], which turns out to be ⟨ϑ⟩ = 8/$\pi^2$ ≈ 0.81.

Although δ is not constant but fluctuates over time, it is revealing to plot its average value as a function of c (Fig. 5). Doing so, we find the familiar fractal graph of the devil’s staircase, implying that for certain intervals of c the average ⟨δ⟩ locks to a constant value.

Let us see how this 3D ODE resembles a FEFS. The two oscillators make up the signal generator; there is a rudimentary feature extractor (taking the absolute value of the oscillators’ sum and lowpass filtering works as an envelope follower); and the estimated amplitude envelope is mapped to the detuning synthesis parameter. In terms of slow-fast systems, the oscillators are the fast variables (the first one faster than the second), and the detuning is the slow variable.

The auto-detuning system is deeply embedded within a program that generates a piece called Auto-detune.

Early sketch for the piece Auto-detune, dating from 21 December, 2017.
This is the raw output from the program. The final version differs in many ways and has been substantially processed.

As used there, it has been bestowed with a few additional time-varying parameters which in turn are parts of other regulating mechanisms; there are also several instances of the system as well as other systems all connected in a complicated web. The system then takes the form

with certain functions doing the updating of all time-varying parameters. Notice that division by Kᵢ ≫ 1 is a way to transform the phases originally used for audio signals into much slower variables.

Figure 5: Top: A devil’s staircase appears when plotting the time average of δ against c. Bottom: Poincaré section of δ over the same parameter range.

Another variant of the auto-detuning system is used as part of a larger autonomous system in the piece Megaphone.

A variant of the megaphone system, here using 25 oscillators.

The piece unfolds as a series of beating sinusoids fading in and out, sometimes making jumps in pitch. Several trials while tuning the system parameters resulted in processes that produced variation for a few minutes and then approached an equilibrium state.

Megaphone combines ideas from the previous piece Bourgillator [3] and the above described detuning system. Bourgillator consists of a network of oscillators whose frequencies are updated by a function of the output amplitudes of the oscillators. Specifically, the ordering of their relative amplitudes determines the frequency of each oscillator. The oscillators are also phase coupled as in the Kuramoto system (e.g. Strogatz, 2000), which allows for synchronisation of all oscillators or clusters of oscillators.

Megaphone is built on a larger system of coupled subsystems, and is roughly described by the following set of equations. First there is a modified auto- detuning part,

where i = 1,2,…,N, j = i + 3 (modN), with N = 13 oscillators used in the realisation of the piece. The auxilliary variables u, v are defined by

– Equation (3) –

The main difference from the bare-bones detuning system (Eq. 2) is the phase coupling and the amplitude variables $\alpha_i$. Next, we introduce a slow variable in the form of an envelope follower applied to the oscillators’ outputs. A simple envelope follower tracing the amplitude of a faster variable u(t) can be realized with the equation

using a time scaling constant 0 < τ ≪ 1. The envelope followers are applied also to the variables v$_i$. Then the amplitude envelopes are used as inputs to a function g($A$),

– Equation (4) –

where U is the Heaviside step function and β$_{j}$ > 0 is a decreasing sequence of coefficients, the output of which sets the oscillator frequencies to

The step function of course is discontinuous, and so is the sum of step functions in Eq. (4). Since the envelopes are slow variables the output of g($A$) can be expected to remain constant for certain intervals of time and then jump to another value. Thus, the function g($A$) produces a stepped sequence of pitches. As soon as one of the envelopes overtakes another envelope in amplitude the function will produce a new pitch for at least one of the oscillators.

Now, the goal is to keep these pitch changes happening. The system should be designed such that the amplitude envelopes do not immediately settle on a fixed order of relative loudness, because when that happens the piece no longer evolves, it has frozen into a final fixed state. Obviously an initial condition must be chosen so as to avoid landing directly on such a steady state. A trick to introduce some more variability is to inject a small amount of neighbouring envelopes into each envelope follower,

and also modifying the function g($A$) to compare the sizes of amplitude differences between pairs of envelopes instead of the envelopes by themselves.

Next we introduce a set of even slower variables

and use them to update the oscillator amplitudes

as well as other parameters such as Mᵤ and Mᵥ as defined in Eq. (3) that are supposed to change at a slow pace.

This description still omits a few details but provides the gist of this particular autonomous system. To summarise, there are the fast variables (the audio output and oscillator phases), the slower amplitude envelopes A, and the lethargic second order envelope followers ψ. The function g() was designed to produce stepwise changes at a rate that can be controlled to some extent by setting appropriate parameter values. Nevertheless, this system goes through an extended initial transient over a period of a few minutes (depending on sampling rate and many other parameters) before ending up on a steady state (see Figure 6). It is not impossible that there are stable periodic oscillatory or chaotic states in parts of the parameter space, however, it is not very practical to search the space for different dynamics given the long duration of transients.

Notice also that the discrete output range of the g function means that when it enters a stable attracting state (that is, the A’s end up in a certain order of magnitudes), no small fluctuations or a gradual approach to the stable state takes place, it ends up there quite abruptly. Instead of trying to scaffold further layers of control to add complexity and longevity to the system’s errant dynamics one might consider having a sensor analysing the output and turning it off as soon as the equilibrium state is reached.

Figure 6: Dynamics of the Megaphone system. (a) One of the frequency variables ω over time in seconds. (b) All thirteen slowest variables. (c) Detuning over time, and (d) ψ against A showing the nature of the long transient towards a fix point.

Concluding remarks

Autonomous dynamic systems can be used for generative music or algorithmic composition with systems that integrate all levels, from sound synthesis to phrase level and formal sections. Monolithic systems of this kind have channels between their fast and slow subsystems through which the different levels can influence each other.

Music created with an uncompromising insistence on using the autonomous system’s output as is, without editing or mixing with other material, may offer a certain conceptual clarity while also bearing the marks of dynamic systems. One frequently observed phenomenon is a prolonged initial transient as the system approaches an attractor.

Long-lived chaotic transients have been observed in various settings, e.g. in networks of pulse-coupled oscillators (Zumdieck et al., 2004). In this type of network the transient length depends on the connectivity between oscillators. If either a small number of oscillators or most of the oscillators are connected the transients are short, but at intermediate degrees of connectivity there can be very long transients. At intermediate connectivity the average transient length grows exponentially with the total number of oscillators. When long transients are observed in algorithmic composition with networks of oscillators and other signal processing units, these transients could conceivably follow a similar law of scaling with connectivity and network size.

Even if the system does not reach an equilibrium state after a prolonged transient phase, another common observation is that it seems to enter recognisable patterns after a while. As a composer one is tempted to compensate for any lack of variety by adding layers of control to ensure development also on longer time-scales. Since the system’s behaviour may differ dramatically between different positions in its parameter space as well as initial conditions, one may need to search for a “sweet spot”. The introduction of multiple temporal scales by explicit design of slow-fast systems, optionally with statistical feedback, is another convenient way to ensure variation on multiple levels. And, it should be added, although variation over multiple time-scales perhaps characterises most music, the deliberate avoidance of variation on some time-scale might be an interesting avenue to explore.

A musical motivation for using autonomous algorithmic composition systems is to create as much complexity as possible using as simple means as possible. As with fusion reactors, one hopes, so to speak, to get more energy out of them than one puts into them to ignite the process, and perhaps the old saying that fusion energy is always 30 years away also holds for this flavour of algorithmic composition. Working with closed, deterministic, autonomous systems imposes strict limits on what is possible, the contours of which are not as easily seen in open, interactive, and stochastic systems.

As mentioned above (in the section Emergence and surprise), there have been efforts to quantify notions such as complexity, emergence and self-organisation by comparing the information content at the system’s input to that at its output. A related concept is the Kolmogorov complexity, developed independently by Kolmogorov, Solomonoff and Chaitin, and also known as algorithmic complexity (Prokopenko et al., 2008), which is defined in terms of universal Turing machines. The Kolmogorov complexity of an object’s description as a text string is defined as the length of the shortest program that produces that output string. Using the same computer and programming language, several output strings can be compared to find out which one is more complex than the other. Still, there may be no practical way of finding the shortest possible program when the output is a soundfile consisting of a piece of complex music. This is equivalent to finding an optimal compression scheme for the soundfile.

Although Kolmogorov complexity is of greater value from a theoretical than from a practical point of view, it suggests an interesting challenge for algorithmic composition – namely, to generate as much complexity as possible with as little code as possible. Of course it may take more effort to formulate a concise program than to write a longer piece of equivalent code. The program’s brevity also resembles the concept of elegance that Sprott (2010) has applied in his search for algebraically simple chaotic flows. Elegance, according to Sprott, is defined as the simplicity of a system of equations where the number of linear and nonlinear terms are counted; the fewer and simpler the more elegant the system is. This concept of elegance is obviously applicable to algorithmic composition using dynamic systems.

These ideas of simplicity or elegance of program code, while producing complex output, have been a guiding principle for a collection of works including those mentioned in the Case study [4]. We have also noted the opposite temptation of adding more layers of mechanisms to increase the complexity of the output. Musical complexity and conciseness of code are two goals usually at odds with one another.

Preference for musical complexity has been thought to follow an inverted U-curve; very simple or extremely complex pieces are less liked than pieces of moderate complexity. Recent research indicates that it is more revealing to consider two groups of subjects: those who prefer simplicity and those who prefer complex stimuli (Güçlütürk and van Lier, 2019). The inverted U-curve in fact appears to be an artefact of pooling these two groups together. It was found that preference for complex stimuli was more common among men, young subjects and those with high scores of a systemising quotient, although other factors may contribute. Inasmuch as algorithmic composition requires a systemising mentality, one should not be surprised to find that composers who engage with this type of systems have a predilection for complex results.

Given that the output of these autonomous systems is a musical composition, we would like to evaluate its complexity according to perceptual criteria. There is no single agreed upon definition of musical complexity, although it may be best thought of as a multi-dimensional concept. Quantifiable approaches to measuring musical complexity from audio recordings have been proposed in music information retrieval, including one that takes structural change over multiple temporal scales into consideration (Mauch and Levy, 2011). In the end it is the composer’s judgement of the algorithm’s output that matters.

Apart from any sensory appeal music generated by autonomous systems might have, the medium also has a certain scientific appeal as well as a conceptual flavour. As for science, one might come across phenomena at the forefronts of dynamic systems research. And the terseness of a few equations even surpasses that of the equivalent computer code. The conceptual aspect is well illustrated by the ease of communicating the generating formula – jot down a few equations and there you have your composition.


This project has been realised with funding from The Audio and Visual Fund, Norway.


Ames, C. (1987). Automated composition in retrospect: 1956-1986. Leonardo, 20(2):169–185.

Bidlack, R. (1992). Chaotic systems as simple (but complex) compositional algorithms. Computer Music Journal, 16(3):33–47.

Collins, N. (2008). Errant sound synthesis. In Proc. of the ICMC 2008, Belfast, Ireland.

Danca, M. (2010). On the uniqueness of solutions to a class of discontinuous dynamical systems. Nonlinear Analysis: Real World Applications, 11:1402– 1412.

Di Scipio, A. (2007). Emergence du son, son d’emergence. essai d’epistemologie expérimental par un compositeur. Intellectica, 48(49):221–249.

Döbereiner, L. (2011). Models of constructed sound: Nonstandard synthesis as an aesthetic perspective. Computer Music Journal, 35(3):28–39.

Fujimoto, K. and Kaneko, K. (2003). How fast elements can affect slow dynamics. Physica D, 180:1–16.

Gershenson, C. and Fernández, N. (2012). Complexity and information: Measuring emergence, self-organization, and homeostasis at multiple scales. Complexity, 18(2):29–44.

Güçlütürk, Y. and van Lier, R. (2019). Decomposing complexity preferences for music. Frontiers in Psychology, 10.

Holopainen, R. (2012). Self-organised Sound with Autonomous Instruments: Aesthetics and experiments. PhD thesis, University of Oslo.

Huron, D. (2007). Sweet Anticipation. Music and the psychology of expectation. MIT Press.

Jacobs, B. A. (2016). A differential equation based approach to sound synthesis and sequencing. In Proc. of the ICMC, pages 557–561.

Kivelson, S. and Kivelson, S. A. (2016). Defining emergence in physics. npj Quantum Materials, 1(1):16024.

Kollias, P. (2018). Overviewing a field of self-organising music interfaces: Autonomous, distributed, environmentally aware, feedback systems. Tokyo, Japan. ACM Conference on Intelligent User Interfaces, Intelligent Music Interfaces for Listening and Creation.

Levitin, D., Chordia, P., and Menon, V. (2012). Musical rhythm spectra from Bach to Joplin obey a 1/f power law. PNAS, 109(10):3716–3720.

Makarenkov, O. and Lamb, J. (2012). Dynamics and bifurcations of nonsmooth systems: A survey. Physica D, 241:1826–1844.

Mauch, M. and Levy, M. (2011). Structural change on multiple time scales as a correlate of musical complexity. In Proc. of the 12th International Society for Music Information Retrieval (ISMIR 2011), pages 489–494, Miami, USA.

Mayer-Kress, G., Choi, I., Weber, N., Bargar, R., and Hübler, A. (1993). Musical signals from Chua’s circuit. IEEE Transactions on Circuits and Systems–II: Analog and Digital Signal Processing, 40(10):688–695.

Medine, D. (2015). Unsampled digital synthesis: Computing the output of implicit and non-linear systems. In Proc. of the ICMC, pages 90–93, University of North Texas.

Polansky, L., Barnett, A., and Winter, M. (2011). A few more words about James Tenney: dissonant counterpoint and statistical feedback. Journal of Mathematics and Music, 5(2):63–82.

Prokopenko, M., Boschetti, F., and Ryan, A. (2008). An information- theoretic primer on complexity, self-organization, and emergence. Complexity, 15(1):11–28.

Röbel, A. (2001). Synthesizing natural sounds using dynamic models of sound attractors. Computer Music Journal, 25(2):46–61.

Rodet, X. and Vergez, C. (1999). Nonlinear dynamics in physical models: Simple feedback-loop systems and properties. Computer Music Journal, 23(3):18–34.

Ruelle, D. (1987). Diagnosis of dynamical systems with fluctuating parameters [and discussion]. Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences, 413(1844):5–8.

Saito, T. (2020). Piecewise linear switched dynamical systems: A review. Non- linear Theory and Its Applications, IEICE, 11(4):373–390.

Sanfilippo, D. (2018). Time-variant infrastructures and dynamical adaptivity for higher degrees of complexity in autonomous music feedback systems: the order from noise (2017) project. Musica/Tecnologia, 12(1):119–129.

Sanfilippo, D. and Valle, A. (2013). Feedback systems: An analytical framework. Computer Music Journal, 37(2):12–27.

Slater, D. (1998). Chaotic sound synthesis. Computer Music Journal, 22(2):12– 19.

Sprott, J. C. (2010). Elegant Chaos. Algebraically Simple Chaotic Flows. World Scientific, Singapore.

Stefanakis, N., Abel, M., and Bergner, A. (2015). Sound synthesis based on ordinary differential equations. Computer Music Journal, 39(3):46–58.

Strogatz, S. (1994). Nonlinear Dynamics and Chaos. With Applications to Physics, Biology, Chemistry, and Engineering. Westview Press.

Strogatz, S. (2000). From Kuramoto to Crawford: exploring the onset of synchronization in populations of coupled oscillators. Physica D, 143:1–20.

Supper, M. (2001). A few remarks on algorithmic composition. Computer Music Journal, 25(1):48–53.

Temperley, D. (2019). Uniform information density in music. Music Theory Online, 25(2).

Wishart, T. (1996). On Sonic Art. Harwood Academic Publishers, Amsterdam, new and revised edition.

Wolf, T. and Holvoet, T. (2004). Emergence versus self-organisation: Different concepts but promising when combined. volume 3464, pages 1–15.

Zumdieck, A., Timme, M., Geisel, T., and Wolf, F. (2004). Long chaotic transients in complex networks. Physical Review Letters, 93(244103).

About the Author

Risto Holopainen is a Swedish composer based in Oslo, Norway. He studied composition at the Norwegian State Academy, and in 2012 he completed a PhD on self-organised sound with autonomous instruments. His compositions include instrumental and electroacoustic works, collaborations with dancers, radiophonic plays, and videos. He is also a part time visual artist currently focusing on print making, and has published a novel and several essays.


[1] Among the notable builders of chaotic modules are Ian Fritz and Andrew Fitch. A list of existing chaotic modules in the Eurorack format is maintained on the Modwiggler forum:

[2] For sound examples, patch ideas and general discussion, see the thread “Self generating patches….tips and ideas ?” started at the Modwiggler forum on March 24, 2011: https: //

[3] See

[4] Titled Kolmogorov Variations or Eleven Hard Pieces. Source code for generating the pieces is available at