Networks, Not Chains: What Happens When Delays Feed Each Other

The Architecture of Interconnection

In 1982, Miller Puckette and John Stautner published a paper in Computer Music Journal describing a structure for artificial reverberation that departed from the dominant approach of the time. Where earlier reverbs used comb filters in parallel or series, each delay line isolated and feeding back only into itself, Puckette and Stautner proposed something different: a network where multiple delay lines cross-fed through a mixing matrix. Every delay influenced every other delay.

They weren't the first. Michael Gerzon had described the same architecture a decade earlier in Studio Sound, a British journal with limited distribution outside the UK. Jean-Marc Jot formalized the mathematics in 1991, adding frequency-dependent damping that made the structure practical for realistic room simulation. The term that stuck was Puckette and Stautner's: feedback delay network, or FDN.

The research goal was always reverberation: creating the dense, diffuse decay of acoustic spaces without the metallic ringing that plagued simpler designs. But the architecture itself is more general than any single application. A feedback delay network is a system where multiple delays are interconnected, where the output of each delay feeds into the inputs of others according to some mixing rule. What emerges from this interconnection is behavior that no single delay could produce alone.

For readers interested in the technical foundations, Julius O. Smith's Physical Audio Signal Processing provides rigorous treatment of FDN theory and stability conditions. The heritage runs deep, with over fifty years of research into how interconnected delays behave.

Chains vs. Networks

Most delay effects are chains. Signal enters, passes through a delay line, exits. The delay might feed back into itself, creating repetition, but it's still a closed loop: one delay, talking to itself.

                FEEDBACK DELAY (CHAIN)

     Input ───▶ [  Delay  ] ───▶ Output
                    │
                    │ feedback
                    └────┘

     One delay. Feeds back into itself.
     Produces: repetition, echo, simple decay.

Even "complex" delays are often just parallel chains (multi-tap delay, where several delays run independently) or serial chains (ping-pong, where signal bounces between two points). The delays don't talk to each other. They share an input or pass signal in sequence, but each one operates in isolation.

A network is structurally different. Every delay feeds into every other delay. The output of delay 1 doesn't just feed back into delay 1; it feeds into delays 2, 3, and 4 as well. And their outputs feed back into delay 1. The signal doesn't just recirculate. It cross-pollinates.

               FEEDBACK DELAY NETWORK

                  ┌──────────┐
     Input ──────▶│  Delay 1 │◀─────┐
                  └────┬─────┘      │
                       │      ┌─────┴─────┐
                       ▼      │           │
                  ┌──────────┐│     ┌──────────┐
                  │  Delay 2 │◀────▶│  Delay 4 │
                  └────┬─────┘      └────┬─────┘
                       │                 │
                       ▼                 │
                  ┌──────────┐           │
                  │  Delay 3 │◀──────────┘
                  └────┬─────┘
                       │
                       ▼
                    Output
                   (layered)

     Four delays. Each feeds all others.
     Produces: density, diffusion, emergent texture.

The difference isn't just architectural. It's behavioral. A single feedback delay produces echoes that repeat at a fixed interval. Two cross-coupled delays produce interference patterns as their outputs combine and recombine. Four or more produce behavior that becomes difficult to predict from the individual components: dense textures, evolving timbres, resonances that emerge from the interaction rather than from any single delay line.

Why Interconnection Creates Complexity

Consider what happens when a single impulse enters a network of four delays.

In a simple delay, that impulse would echo at the delay time, then again, then again, each repetition quieter than the last. The pattern is predictable: one rhythm, one decay curve.

In a network, the impulse enters all four delays. Each delay outputs its copy at a different time (assuming the delays are set to different lengths). Those four outputs then feed into all four inputs. Now each delay contains not just the original impulse but fragments of what the other delays produced. On the next cycle, each delay outputs a more complex signal: the original impulse, plus the cross-fed contributions, plus the contributions of contributions.

        IMPULSE PROPAGATION IN A 4-DELAY NETWORK

     Time 0:    Impulse enters all delays
                [·] [·] [·] [·]

     Time 1:    Delay 1 outputs (shortest)
                [→] [·] [·] [·]
                Output feeds into all delays

     Time 2:    Delay 2 outputs
                [·→] [→] [·] [·]
                Now contains: original + delay 1's output

     Time 3:    Delay 3 outputs
                [··→] [·→] [→] [·]
                Contains: original + d1 + d2 outputs

     Time 4:    Delay 4 outputs (longest)
                [···→] [··→] [·→] [→]
                All delays now contain cross-fed material

     Time 5+:   Density increases exponentially
                Each cycle adds more cross-contributions
                Pattern becomes increasingly complex

The rate at which this complexity builds depends on the feedback amount (how much of each output feeds forward) and the relationship between the delay times. If all four delays are the same length, the outputs align and reinforce, producing a louder, more resonant version of a simple delay. If the delays are different lengths, especially if those lengths share no common factors, the outputs scatter across time, filling in gaps, creating density.

This is why FDN reverbs sound more natural than simple comb filter reverbs. Real rooms don't produce evenly-spaced echoes; they produce a wash of reflections arriving from every surface at different times. The feedback network approximates this by generating a dense, irregular pattern of echoes from the interaction of a few delay lines.

Beyond Reverb: The Network as Instrument

The researchers who developed FDNs were solving an engineering problem: how to make artificial reverb that doesn't sound artificial. The solution was architectural; interconnection produces complexity that isolated elements cannot.

But reverb is just one application of the architecture. The same structure that creates diffuse spatial decay can create other things entirely, depending on how you set the parameters.

Short delay times create pitch. A delay line with feedback is also a resonator. Set the delay to 10 milliseconds and you get a resonance around 100 Hz. Set it to 2 milliseconds and you get 500 Hz. A network of short delays becomes a network of coupled resonators, a resonant body with multiple modes like the cavity of an instrument or the body of a drum.

        DELAY TIME AND PITCH

        Delay Time          Resonant Frequency
        ──────────          ──────────────────
        20 ms               50 Hz
        10 ms               100 Hz
        5 ms                200 Hz
        2 ms                500 Hz
        1 ms                1000 Hz

        f = 1000 / delay_ms  (approximate)

        Network of short delays = network of resonances

Long delay times create rhythm. Set the delays to musical intervals (quarter notes, eighth notes, dotted values) and the network becomes a polyrhythmic engine. The cross-feeding means that a single input event triggers a cascade of echoes at different rhythmic positions, each feeding back into the others, creating patterns more complex than any single delay line could produce.

The feedback amount determines character. Low feedback produces discrete echoes that decay quickly; the individual delay lines remain audible as separate events. High feedback produces sustained textures where the echoes blur into continuous sound. Push feedback above unity and the network becomes generative, producing sound indefinitely, potentially growing rather than decaying, until the energy is tamed or the limiter catches it.

Filters and saturation shape the evolution. If each delay line passes through a lowpass filter before feeding forward, high frequencies decay faster than low frequencies and the texture darkens over time, like sound in a room with absorptive walls. If each delay line passes through saturation, the signal accumulates harmonics with each cycle and the texture gets denser, grittier, more complex. These per-node processors determine the material of the network, how it colors and transforms what passes through it.

Excitation: What Feeds the Network

A feedback network needs something to process. In reverb applications, that something is the audio you're trying to add space to: a vocal, a snare hit, a full mix. The network transforms the input, smearing it across time, adding density and decay.

But a network can also be played. If short delay times create pitched resonances and the network has enough feedback to sustain, then the network is an instrument waiting for an excitation source to set the resonances ringing.

The character of that excitation shapes what emerges. A short impulse excites all frequencies equally; the network's resonances emerge based purely on the delay times and feedback structure. A filtered impulse emphasizes certain frequencies, steering which resonances dominate. A noise burst produces a more diffuse, less pitched result and the network becomes a texture generator rather than a pitched resonator. A sustained tone feeds energy continuously, building up resonances over time, potentially overloading if the feedback is too high.

This is physical modeling in miniature. A struck drum is an excitation (the stick) feeding a resonant system (the membrane and shell). A plucked string is an excitation (the finger or pick) feeding a different resonant system (the tensioned string and body). A feedback network with an impulse exciter and short delay times operates on the same principle: excitation provides energy, the network provides resonant structure, and what you're hearing is the interaction.

Freezing: Capturing System State

Most sounds decay. We accept this as natural. Play a note, it fades. Feedback networks decay too, unless the feedback is set to unity or above. But there's another option: freezing.

To freeze a network is to capture its current state and hold it indefinitely. The delays continue to recirculate, the signal continues to evolve within the network, but the overall energy level stops decaying. It's not a sample. You're not capturing audio and looping it. You're capturing a system in motion and sustaining that motion.

The distinction matters. A frozen sample is static; it repeats exactly. A frozen network is dynamic; the internal circulation continues, the signal evolves, the texture shifts. Depending on the network topology and delay times, a frozen state might be nearly static (if the delays align into a short loop) or continuously evolving (if the delays are incommensurate and the signal never quite repeats).

Freeze also interacts with new input. In full freeze, new audio can't enter and the network sustains what it had at the moment of freezing. In freeze-with-input, the network sustains its current state while allowing new audio to layer on top, creating accumulating textures that grow denser over time. In partial freeze, feedback is set high but below unity, so the texture sustains but slowly evolves as old material fades and new material enters.

This is a compositional tool as much as a sound design tool. Freeze a network, let it sustain, add new material, freeze again. The network becomes a canvas for layering, a way to build complex textures from simple inputs.

The Network as Instrument

The researchers who built feedback delay networks were trying to disappear, to create artificial spaces so natural that listeners wouldn't notice the artifice. The goal was transparency, invisibility, the sense of being in a room rather than hearing an effect.

But the architecture they created has applications beyond simulation. A feedback network is a system with inputs, internal state, and emergent behavior. Feed it an impulse and it rings. Feed it continuously and it builds. Freeze it and it sustains. Adjust its topology and its character changes.

This is what it means to play a network: not to trigger samples or sequence notes, but to excite a system and shape its evolution. The network has its own tendencies, its own resonances, its own ways of transforming what enters it. The player's role is to understand those tendencies and work with them, or against them when the goal is surprise rather than predictability.

Static sounds are easy to build. Dynamic sounds require systems that evolve. A feedback delay network is a small, controllable system with rich emergent behavior: complex enough to be interesting, simple enough to be playable. The delays provide the skeleton; the interconnections provide the life.