Deep Frequency Modulation

• Deep frequency modulation is a regime that utilizes large modulation depths and nested modulation structures to produce enriched harmonic spectra and advanced signal control.
• It spans diverse applications including audio synthesis, high-power RF stabilization, laser interferometry, and learning-based demodulation for improved performance.
• Key insights involve managing non-linear feedback, calibration challenges, and trade-offs between modulation parameters to optimize system design.

Searching arXiv for the cited papers to ground the article in current literature. Deep frequency modulation is not a single universally standardized formalism. In recent literature, the term is used for several technically distinct but structurally related practices: higher-order and feedback FM in synthesis; coherent phase-ramp-based frequency translation for high-power RF stabilization; strongly frequency-modulated laser interferometry for simultaneous estimation of an ambiguous interferometric phase $\Phi$ and an unambiguous effective modulation depth $m$; end-to-end neural FM demodulation from SDR baseband I/Q; and non-linear frequency-domain aggregation schemes such as Log-FSK for over-the-air computation (Lazzarini et al., 2023, Ben-Zvi, 7 Jan 2025, Dovale-Álvarez, 31 Jul 2025, Elbaz et al., 2017, Martinez-Gost et al., 2024). Across these contexts, “deep” denotes the exploitation of large modulation depth, higher-order topology, rich harmonic structure, or algorithmically nontrivial inference beyond the classical one-carrier, one-modulator, local-demodulator view.

1. Signal-theoretic foundations

Classical analog FM is represented in the speech-demodulation literature as

$$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$

where $x_m(t)$ is the message, $f_c$ the carrier frequency, $A_c$ the carrier amplitude, and $f_\Delta$ the peak frequency deviation (Elbaz et al., 2017). In this representation, the modulating signal enters as an integral in the instantaneous phase, and classical receiver design is therefore driven by local instantaneous-frequency estimation.

In synthesis-oriented work, first-order linear FM is written as

$$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$

while sinusoidal PM is written as

$$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$

Differentiation gives the instantaneous-frequency relation

$f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$

so the FM deviation $m$0 and PM index $m$1 satisfy

$m$2

A central consequence is that in PM the modulation index $m$3 directly governs spectrum and bandwidth, whereas in FM the amplitude $m$4 alone is not a modulation index (Lazzarini et al., 2023).

The term “deep” acquires a more specific meaning when the modulation depth is large enough that the signal ceases to be well described by a carrier plus a small number of sidebands. In deep frequency modulation interferometry, the laser is driven strongly enough that the effective phase-modulation depth $m$5 is several radians to tens of radians, and several harmonics carry significant power. In that regime, the useful observable is no longer merely instantaneous frequency, but the full harmonic structure induced by the modulation (Dovale-Álvarez, 31 Jul 2025, Dovale-Álvarez, 15 Aug 2025).

2. Higher-order, stacked, and feedback FM

In musical and signal-synthesis research, deep frequency modulation denotes going beyond the classic one-carrier, one-modulator arrangement into higher-order or stacked FM, multi-operator structures, and feedback FM. The motivating problem is that naïve direct higher-order FM is not spectrally well behaved. If the output of one FM oscillator is used directly as a frequency-deviation signal for another, DC components in the intermediate spectrum become constant frequency offsets at the next stage, so pitch drifts as timbral parameters change (Lazzarini et al., 2023).

The second-order construction developed to avoid this problem makes the next-stage deviation proportional to the modulator’s instantaneous frequency rather than to a fixed nominal frequency. If

$m$6

and

$m$7

then the carrier is defined as

$m$8

This makes the time-varying deviation

$m$9

and yields a second-order FM system equivalent to the issue-free second-order PM form

$$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$0

The same paper generalizes this into an FM operator abstraction. An operator receives amplitude/index, base frequency, and a modulation input, and produces both an audio output and a modulation output suitable for the next stage. In the reference C++ implementation, the modulation output is computed as s*f, i.e. waveform times instantaneous frequency, which is the discrete-time counterpart of the continuous relation $$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$1. This operator formalism supports stacks, arbitrary networks, and self-feedback. Feedback FM is then a practical possibility, with spectra qualitatively similar to feedback PM and approximate $$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$2 spectral decay, although discrete-time integration still introduces small numerical errors and aliasing remains a practical constraint (Lazzarini et al., 2023).

A plausible implication is that, in this domain, “deep” refers less to large deviation alone than to topological depth: multiple nested modulation links whose stability depends on treating instantaneous frequency as a state-dependent quantity rather than a static control amplitude.

3. Reactive-load deep FM in high-power RF systems

In accelerator RF, deep frequency modulation has been proposed as a method for converting the output of a free-running, unstable magnetron into a signal exactly at a stable reference frequency with negligible insertion loss. The architecture uses a ferroelectric fast reactive tuner (FE-FRT), behaving as a tunable, predominantly reactive impedance with only about $$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$3 real part, inserted into the collinear arms of a magic Tee. By changing the tuners’ reactance $$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$4, the system changes the reflection coefficient

$$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$5

and with

$$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$6

one obtains

$$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$7

so the instantaneous reflection phase spans a full $$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$8 (Ben-Zvi, 7 Jan 2025).

The magnetron output is written as

$$y(t)=A_c \cos\Big(2\pi f_c t+2\pi f_{\Delta }\int_{0}^{t}x_m(\tau)d\tau\Big),$$9

A sample of the magnetron output is mixed with the master oscillator at $x_m(t)$0, producing the baseband difference signal

$x_m(t)$1

which encodes both amplitude and frequency error. A DSP then drives the FE-FRTs so that the reflection phases satisfy

$x_m(t)$2

These are explicit phase-ramp laws whose time derivative corresponds to a constant frequency shift $x_m(t)$3. With the magic Tee recombination, the result is

$x_m(t)$4

meaning that all magnetron power appears at the reference frequency at the load port, while no power is reflected back to the magnetron apart from small insertion losses (Ben-Zvi, 7 Jan 2025).

The method is conceptually distinct from injection locking. The magnetron is not directly tuned; instead, its emission at $x_m(t)$5 is processed by a high-power, reflection-type phase-modulator network. The same structure can also apply amplitude and phase feedback, and the paper states that the magic Tee modulation can be used to achieve a highly accurate drive to demanding applications like high loaded-$x_m(t)$6 superconducting cavities, limited just by the DSP capabilities. The treatment is primarily theoretical: residual frequency error, phase-noise spectral density, amplitude ripple, loop bandwidth, and efficiency are not numerically evaluated, and a CW high-power prototype has yet to be demonstrated (Ben-Zvi, 7 Jan 2025).

4. Deep frequency modulation interferometry

Deep Frequency Modulation Interferometry (DFMI) uses a strongly frequency-modulated laser in an unequal-arm interferometer so that time-of-flight delay converts laser frequency modulation into deep phase modulation of the interferometric signal. With

$x_m(t)$7

integration gives

$x_m(t)$8

For arm delay $x_m(t)$9, the small-$f_c$0 approximation yields an effective phase modulation of amplitude

$f_c$1

superposed on the static interferometric phase

$f_c$2

The detector output is then

$f_c$3

or, in AC-coupled form,

$f_c$4

(Dovale-Álvarez, 31 Jul 2025, Dovale-Álvarez, 15 Aug 2025).

This signal contains harmonics at integer multiples of the modulation frequency. The complex harmonic amplitudes are

$f_c$5

with parameter vector $f_c$6. The parameter $f_c$7 governs the Bessel-function envelope of harmonic magnitudes, $f_c$8 gives a harmonic-dependent rotation $f_c$9, and $A_c$0 modulates even and odd harmonics differently through $A_c$1 (Dovale-Álvarez, 31 Jul 2025).

The central DFMI measurement principle is that $A_c$2 is precise but ambiguous modulo $A_c$3, whereas $A_c$4 is coarser but unambiguous. A coarse absolute length estimate follows from

$A_c$5

and a coarse unwrapped phase estimate from

$A_c$6

Comparing $A_c$7 with the wrapped phase $A_c$8 yields the fringe order

$A_c$9

provided the total coarse-plus-wrapped error remains below $f_\Delta$0 (Dovale-Álvarez, 31 Jul 2025).

Two estimator classes are emphasized. A frequency-domain Non-Linear Least Squares (NLS) fit minimizes

$f_\Delta$1

using Levenberg–Marquardt, while a time-domain Extended Kalman Filter (EKF) tracks parameters from the nonlinear measurement model

$f_\Delta$2

The paper derives the Cramér–Rao Lower Bound for $f_\Delta$3 and $f_\Delta$4, identifies intrinsic dead zones where precision degrades, and gives asymptotic limits as a function of signal quality and integration time achievable by both NLS and EKF. It also develops analytical models for dominant systematic biases, including modulation non-linearity and residual amplitude modulation, and identifies valleys of robustness where those biases are strongly suppressed (Dovale-Álvarez, 31 Jul 2025).

DeepFMKit extends this framework into software. It is an open-source Python library for end-to-end simulation and analysis of DFMI systems, with a high-fidelity physics engine modeling time-of-flight delays in dynamic interferometers, arbitrary laser modulation waveforms, and colored noise from user-defined $f_\Delta$5 spectral densities. It includes a highly-optimized, parallelized frequency-domain NLS, multiple time-domain EKF implementations with random walk and integrated random walk process models, and a high-throughput experimentation framework for parameter sweeps and Monte Carlo analyses (Dovale-Álvarez, 15 Aug 2025).

5. Learning-based demodulation and computation-oriented FM

In software-defined radio, deep learning has been used to replace a classical FM demodulator with a stacked bidirectional LSTM that maps baseband I/Q sequences directly to a speech waveform. The receiver assumes conventional heterodyning to baseband and takes, for each audio sample, five in-phase and five quadrature samples, so the per-step network input is a 10-dimensional real vector and the output is one audio sample. Training uses truncated backpropagation through time over 100 time steps, batch size 512, and RmsProp optimization, with network state preserved between batches to maintain long-term dependencies (Elbaz et al., 2017).

The significance of this design is not that the FM law itself changes, but that demodulation is recast as inference with a learned speech prior. The paper reports a noise-free output SNR of $f_\Delta$6 dB and a PESQ score of $f_\Delta$7. It further reports that a memory-limited network can reconstruct audio reasonably well in the noise-free case with SNR $f_\Delta$8 dB, but at $f_\Delta$9 dB amplitude noise reconstruction is not possible without memory and demodulation fails. Under both amplitude noise and combined phase-noise-plus-amplitude-noise conditions, the LSTM demodulator outperforms a conventional FM demodulator based on the MATLAB communication toolbox, particularly at low SNR (Elbaz et al., 2017).

A different reworking of frequency modulation appears in Log-FSK for over-the-air computation. Here the transmitted discrete-time waveform for user $$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$0 is

$$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$1

with $$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$2 defined from a DCT basis function. Under ideal synchronous superposition and exponential post-processing, the received nonlinearly transformed signal contains a component whose DCT frequency index equals

$$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$3

Demodulation applies a DCT and declares

$$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$4

For $$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$5, the amplitude of the sum-frequency component is

$$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$6

and the approximate symbol error probability is

$$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$7

The paper uses $$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$8, corresponding to a destination SNR threshold of approximately $$c(t)=\cos\!\left(2\pi\int_0^t (f_c+m(x))\,dx\right),$$9 dB. Above this threshold, Log-FSK outperforms linear AirComp implemented with double sideband, whereas below threshold DSB has lower MSE. The method is explicitly described as suited to small $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$0, such as $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$1–$$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$2 users, and not to massive access (Martinez-Gost et al., 2024).

Taken together, these works suggest that one contemporary meaning of deep frequency modulation is algorithmic: FM becomes a substrate for learned reconstruction or function-specific non-linear computation rather than a fixed analog modulation format.

6. Mechanism selectivity, limitations, and recurrent design constraints

In room-temperature $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$3C DNP of diamond powder, frequency-modulated microwave irradiation has been shown to affect four DNP mechanisms simultaneously present in the same material system: the solid effect (SE), cross effect (CE), truncated cross effect (tCE), and Overhauser effect (OE). The two key FM parameters are the modulation frequency $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$4 and modulation amplitude $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$5, and the paper states explicitly that frequency modulation during DNP not only allows improvement of DNP enhancement but also provides a way to control which DNP mechanism is most active. By choosing appropriate modulation parameters, some mechanisms can be selectively enhanced while others are simultaneously suppressed (2207.14731).

The experimental and theoretical trends are strongly mechanism dependent. At the SE peak, maximum MCW enhancement is about $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$6, while FM yields at best about $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$7, i.e. only a $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$8 gain. At the CE peak, MCW enhancement is about $$c(t)=\cos(\phi(t)), \qquad \phi(t)=2\pi f_c t+ z\sin(2\pi f_m t).$$9 and optimal FM yields about $f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$0, a $f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$1 gain. At the OE and tCE frequencies, FM yields more modest improvements of about $f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$2. Across mechanisms, an important operating condition is

$f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$3

so that the electron does not fully repolarize between sweep passes. Large $f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$4 and low $f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$5 are particularly detrimental for SE and can strongly reduce OE and tCE as well, whereas CE often shows an optimum near $f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$6 (2207.14731).

Across the broader literature, several recurrent limitations appear. Naïve higher-order FM exhibits DC-induced carrier drift and pitch–timbre coupling, and high indices produce aliasing-prone wideband spectra (Lazzarini et al., 2023). The magnetron-stabilization scheme depends on tuner speed, linearity, loss, hybrid imperfections, and DSP latency, and a CW high-power prototype has yet to be demonstrated (Ben-Zvi, 7 Jan 2025). DFMI is bounded by dead zones, calibration error in $f_{\text{inst}}(t)=f_c+z f_m \cos(2\pi f_m t),$7, laser drift, and systematic bias from modulation non-linearity and residual amplitude modulation; the paper shows that long baselines severely shrink the usable design space (Dovale-Álvarez, 31 Jul 2025). Log-FSK requires tight synchronization, channel equalization, and operation above a threshold SNR, while the effective noise after exponentiation is multiplicative and log-normal rather than additive Gaussian (Martinez-Gost et al., 2024).

This suggests that deep frequency modulation is best understood not as a monotonic increase in modulation depth, but as a design regime in which spectral richness, estimator structure, and control bandwidth become decisive. In every domain covered by the recent literature, performance depends on operating at the right modulation depth, sweep range, harmonic set, or algorithmic operating point rather than on “deeper” modulation in the abstract.