Load a guitar cabinet impulse response and it sounds convincing. The cabinet resonances, the frequency shaping, the high-end rolloff, all of it is there. Then push the gain and something shifts. The compression, the harmonic bloom, the way the upper mids stack differently at high levels, none of that changes. The IR cannot represent it, not because it was captured carelessly, but because a linear impulse response is mathematically incapable of encoding what a nonlinear system does.
The framework that completes the picture is the Volterra series, a hierarchy of higher-order impulse responses derived by Italian mathematician Vito Volterra (1860-1940) as part of his theory of functionals. Audio engineers rarely encounter it by name, but every modern neural amp model is, in a precise sense, an implicit approximation of a Volterra description.
What a Linear Impulse Response Assumes
The convolution operation works because of superposition. If you know how a system responds to an impulse, you can predict how it responds to any signal by summing scaled and delayed copies of that impulse response. This assumption has a specific mathematical name: linearity. A system is linear if doubling the input exactly doubles the output, and if feeding two signals simultaneously gives exactly the sum of the two individual outputs.
The critical consequence is that a linear system cannot create new frequencies. Feed a 1 kHz sine into a linear filter. The output contains only 1 kHz, possibly at a different amplitude and phase. The filter cannot produce 2 kHz or 3 kHz or 500 Hz components that were not in the input.
This is the exact point where a tube amplifier diverges from the model. Feed a 1 kHz sine into a triode at moderate to high amplitude. The plate current follows the transconductance curve, which bends. The output is not a pure 1 kHz sine; it contains 2 kHz (second harmonic), 3 kHz (third harmonic), and higher orders. The tube created frequencies the input did not have. No linear impulse response can encode this, because the math of convolution does not allow it.
The Volterra Series
The Volterra series extends convolution to handle systems where superposition breaks down. It describes the output of any time-invariant, fading-memory nonlinear system as an infinite sum of terms:
y(t) = h₀
+ ∫ h₁(τ)·x(t-τ) dτ
+ ∫∫ h₂(τ₁,τ₂)·x(t-τ₁)·x(t-τ₂) dτ₁dτ₂
+ ∫∫∫ h₃(τ₁,τ₂,τ₃)·x(t-τ₁)·x(t-τ₂)·x(t-τ₃) dτ₁dτ₂dτ₃
+ ...Each term is a multi-dimensional convolution. The first-order kernel h₁(τ) is the standard linear impulse response. The second-order kernel h₂(τ₁, τ₂) captures all interactions between two time-delayed copies of the input. The third-order kernel h₃ captures three-way interactions.
The physical interpretation maps onto tube behavior directly. The second-order term generates sum and difference frequencies: feed in two tones at f₁ and f₂ and h₂ produces components at 2f₁, 2f₂, f₁+f₂, and f₁-f₂. These are second-order intermodulation products and second harmonics. Single-ended triode stages, which clip asymmetrically, have a large h₂ contribution. Third-order terms produce 3f₁, 3f₂, 2f₁±f₂, 2f₂±f₁. A symmetric (but nonlinear) push-pull stage cancels most even-order terms, leaving h₃ as the dominant nonlinear contribution.
Linear IR measurement captures only h₁. Everything above it is invisible to any technique that assumes superposition.
Measuring the Kernels: The Swept-Sine Trick
Measuring h₁ is a standard IR measurement. Measuring h₂ and h₃ requires observing how the system responds to combinations of its own delayed input, which demands a different approach.
The practical breakthrough came from Angelo Farina at AES Convention 108 in Paris in 2000. His paper "Simultaneous Measurement of Impulse Response and Distortion with a Swept-Sine Technique" showed that an exponentially swept sine has a specific property when deconvolved: harmonic distortion products land at predictable time offsets, separated from the linear IR.
The reason is geometric. With a linear sweep, the nth harmonic of the instantaneous frequency is a random distance from that frequency in the time record. With an exponential sweep, frequency doubles at a constant rate, so the second harmonic of any frequency was played by the sweep exactly one octave-worth of sweep-time earlier. After deconvolution with the inverse filter, each harmonic order collapses into a narrow time-domain feature at a known offset before time zero.
Specifically:
- The linear impulse response h₁ appears at time zero
- Second-harmonic distortion appears approximately one octave (in sweep duration) before zero
- Third-harmonic distortion appears log(3)/log(2), roughly 1.585 octaves before zero, and so on
Each of these offset responses is a diagonal slice through the corresponding higher-order Volterra kernel, the function h_n(τ, τ, ..., τ) evaluated along the main diagonal where all delay arguments are equal. It is not the full multi-dimensional kernel, but it is the most accessible part, and it is sufficient for identifying many common distortion types.
Room acoustics engineers who use this measurement daily typically discard the distortion peaks and keep only the linear IR. For nonlinear system identification, the discarded peaks are the point.
The Coefficient Explosion
Measuring and using the full Volterra kernels runs into a counting problem. For a kernel of memory length M (M delay taps), the number of independent coefficients grows rapidly by order:
- First order: M coefficients
- Second order (symmetric kernel): M(M+1)/2 coefficients
- Third order (symmetric kernel): M(M+1)(M+2)/6 coefficients
At M = 100 (roughly 2 ms at 48 kHz), first order costs 100 coefficients. Second order costs 5,050. Third order costs 171,700. A fourth-order kernel would require over 4 million coefficients for the same 2 ms window. Each higher order also requires increasingly complex measurement procedures to isolate from lower-order contributions.
This is why full Volterra models are rare in real-time audio. Researchers addressing tube amp emulation have used truncated models (first through third order), reduced-coefficient models where only a structured subset of the full kernel is estimated, or diagonal-only approximations.
The diagonal approximation deserves attention. If you keep only h₂(τ, τ) and h₃(τ, τ, τ), those diagonal kernels are each one-dimensional functions of length M, the same size as the linear IR. You have tripled the model complexity but approximated the second and third-order contributions. The Farina swept-sine technique measures approximately this diagonal directly, which is one reason it remains the standard tool for audio nonlinear measurement.
What Neural Amp Models Are Actually Doing
The mathematical structure of the Volterra series explains why block-oriented models have limits. A Hammerstein model, a static nonlinearity followed by a linear filter, or a Wiener model with the order reversed, can approximate low-order Volterra contributions only when the nonlinear and linear parts are separable. Real tube circuits have reactive elements (coupling capacitors, output transformers, Miller capacitance) whose behavior interacts with the nonlinearity in ways that no fixed separation captures.
Recurrent neural networks bypass this constraint without explicitly computing Volterra kernels. The hidden state of an LSTM or GRU cell maintains a learned summary of recent signal history. The nonlinear activations applied to combinations of that history and the current input produce outputs that, expanded as polynomial functions, have exactly the structure of a truncated Volterra series. The network learns a compressed representation of higher-order kernels through training, using measured input/output pairs from the real hardware.
This is why measuring a neural amp model requires nothing more than recording what the real hardware does with a designed test signal. The training process discovers Volterra-like structure implicitly, without ever measuring a kernel directly. The coefficient explosion is sidestepped because the neural network's parameter count scales with hidden state size, not with the polynomial degree of the nonlinearity.
The Measurement Gap
What the Volterra framework makes explicit is also what current tools still cannot capture. Full multi-dimensional kernels beyond second or third order are practically unmeasurable for real-time audio systems. Time-varying behavior, the way a tube's operating point drifts as coupling capacitors charge and the output transformer core shifts under sustained load, falls outside the time-invariant assumption that makes the series definition clean.
The series describes a class of systems precisely. Real tube circuits sit near the boundary of that class. The mathematical framework is not a perfect fit for a decades-old amplifier running hot, but it is the closest analytic structure audio engineering has for describing why a linear IR captures only part of the picture, and what the rest of the picture would require.
For engineers building or evaluating nonlinear audio models, the Volterra framework provides vocabulary and measurement tools. The swept-sine method is practical and standardized. The kernel structure predicts which distortion types each order captures. And the coefficient problem explains, without handwaving, why neural approaches are not just empirical shortcuts but structurally necessary for capturing systems that exceed what tractable direct kernel measurement can reach.