Temporal Frequency Modulation (TFM)

• Temporal Frequency Modulation (TFM) is a family of methods that use time‐dependent phase laws and operator frameworks to modulate and analyze frequency behavior in diverse systems.
• It employs mathematical models such as Fourier series, instantaneous frequency functions, and phase-only operations to precisely control spectral characteristics.
• TFM enables advanced applications including finite-mode unitaries in optics, spectrally compact radar waveforms, quantum metrology enhancements, and low-pass filtering in transformer-based recommendation systems.

Searching arXiv for papers on Temporal Frequency Modulation and closely related terminology. Temporal Frequency Modulation (TFM) is a field-dependent term for methods that impose, recover, or exploit explicitly time-dependent frequency content. In the cited literature, it denotes several related constructions rather than a single canonical protocol: phase-sensitive reconstruction of an optical instantaneous-frequency profile, arbitrary unitary transformations on temporal modes through alternating time- and frequency-domain phase operations, direct specification of radar and sonar instantaneous-frequency laws, time-dependent driving of a quantum harmonic oscillator, sequential temporal gating for nonreciprocal frequency conversion, and layerwise temporal-frequency filtering inside LLM recommenders (Jachura et al., 2017, Ashby et al., 2020, Hague, 2020, Kong et al., 13 Jun 2026, Wang et al., 14 Aug 2025). The unifying theme is that a designed temporal law changes phase accumulation, spectral occupancy, modal overlap, or low-frequency energy retention.

1. Core definitions and mathematical structures

Across applications, TFM is most naturally expressed through a time-dependent phase or frequency law. In the MTSFM waveform literature, the transmit signal is written as

$$s(t)=a(t)\,\exp\!\bigl\{j[2\pi f_c t+\varphi(t)]\bigr\},$$

with instantaneous frequency offset

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$

and the modulation function expanded as a finite Fourier series in $\{a_k,b_k\}$ (Hague, 2020). In the optical phase-retrieval setting, the directly reconstructed quantity is the temporal phase $\phi_{\rm mod}(t)$, from which the temporal frequency modulation profile follows as

$\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$

(Jachura et al., 2017). In quantum metrology, the modulation is transferred from phase to the oscillator itself through

$$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$

so that the integrated temporal profile $F(t)$ governs the scaling of the quantum Fisher information (Kong et al., 13 Jun 2026).

A second recurring structure is operator-theoretic. Ashby et al. formulate temporal modes in a travelling-pulse picture by expanding the positive-frequency field envelope as

$E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$

with orthonormal temporal modes $\{\psi_n(t)\}$, and then implement transformations through phase-only operators in time and frequency, $\hat U_t[\phi]=\exp[i\phi(t)]$ and $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$0 (Ashby et al., 2020). In machine learning, TFM becomes a frequency-domain operator on sequence hidden states,

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$1

where $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$2 is a Butterworth low-pass mask applied along the temporal axis (Wang et al., 14 Aug 2025).

This suggests that TFM is best understood as a family of temporal-spectral operators whose physical meaning depends on the state variable being modulated: optical field envelope, radar waveform phase, oscillator frequency, intermode coupling, or sequence representation.

2. Optical temporal modes, phase retrieval, and quantum-light reshaping

In quantum photonics, TFM is used both to transform temporal modes and to measure the RF phase profiles required for those transformations. Ashby et al. show theoretically that any finite-mode unitary acting on temporal modes can be realized by a sequence of time- and frequency-domain phase operations. In discretized form, a well-known unitary-diagonal-Fourier factorization implies

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$3

and in the continuum limit this becomes a sequence of alternating temporal and spectral phase operations,

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$4

The construction is phase-only and therefore unitary in principle; Ashby et al. explicitly note that this implies no quantum noise added, although real components have loss (Ashby et al., 2020).

Representative simulated transformations are summarized below.

Transformation	Conditions	Reported result
$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$5th HG mode $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$6th HG mode	$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$7 ps; $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$8 GHz; $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$9 steps	$\{a_k,b_k\}$0
2-mode Hadamard between HG$\{a_k,b_k\}$1 and HG$\{a_k,b_k\}$2	$\{a_k,b_k\}$3 GHz; $\{a_k,b_k\}$4	$\{a_k,b_k\}$5, $\{a_k,b_k\}$6
4-mode Hadamard on $\{a_k,b_k\}$7	$\{a_k,b_k\}$8 GHz; $\{a_k,b_k\}$9	$\phi_{\rm mod}(t)$0, $\phi_{\rm mod}(t)$1
3-mode temporal demultiplexer to Gaussian bins 200 ps apart	$\phi_{\rm mod}(t)$2 ps; $\phi_{\rm mod}(t)$3	average $\phi_{\rm mod}(t)$4, time-bin SNRs $\phi_{\rm mod}(t)$5–$\phi_{\rm mod}(t)$6

The same study reports that the $\phi_{\rm mod}(t)$7th HG $\phi_{\rm mod}(t)$8th HG conversion reaches $\phi_{\rm mod}(t)$9 fidelity with only $\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$0 steps, and that experimentally feasible specifications include EOPM bandwidths $\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$1–$\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$2 GHz, $\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$3 V, CFBG group-delay dispersion up to $\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$4–$\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$5 ps$\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$6, pulses of duration $\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$7–$\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$8 ps, and up to $\omega_{\rm inst}(t)=\frac{d\phi_{\rm mod}(t)}{dt}$9 phase steps (Ashby et al., 2020). The principal caveats are equally explicit: EOPM speed sets the minimum feature size in the phase profiles, maximum GDD is limited by grating length, synchronization of multiple EOPMs is nontrivial, and higher bandwidth reduces the required number of steps but increases electronic complexity.

Measurement of the required temporal phase patterns is addressed by spectral interferometry with chirped pulses. The interferometric method of (Jachura et al., 2017) uses a highly chirped optical pulse in the far field of dispersion, where

$$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$0

so that retrieved spectral phase maps directly to temporal phase through $$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$1. The demonstrated temporal resolution is $$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$2 ps for $$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$3 m SMF and $$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$4 ps for $$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$5 m SMF, with RF bandwidth demonstrated up to the $$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$6 GHz limit of the EOPM and sensitivity to RF pulses with peak voltages as low as $$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$7 V (Jachura et al., 2017). The same work reports arbitrary RF pulses up to $$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$8 ps duration with sub-$$\Omega(t)=\omega_0 f(t), \qquad F(t)=\int_0^t f(s)\,ds,$$9 ps resolution and identifies direct temporal-phase measurement as critical for deterministic spectral-temporal reshaping of quantum light pulses.

3. Radar, sonar, and direct-antenna realizations

In radar and sonar, TFM is instantiated most explicitly by Multi-Tone Sinusoidal Frequency Modulation (MTSFM). The core model expands the modulation function as

$F(t)$0

with $F(t)$1, so that the coefficients $F(t)$2 act as discrete design parameters for spectrum, ACF, CCF, and ambiguity-function shaping (Hague, 2020). The optimization is formulated as a weighted multi-objective cost balancing Integrated Sidelobe Ratio and total CCF sidelobe energy under an RMS-bandwidth constraint. The optimized waveform families possess thumbtack-like ambiguity functions, ideally low Peak-to-Average Power Ratios, and high Spectral Efficiency; the detailed report gives $F(t)$3 $F(t)$4 and $F(t)$5–$F(t)$6 (Hague, 2020).

A complementary treatment derives closed-form expressions for the spectrum, autocorrelation, and ambiguity function via K-dimensional Generalized Bessel Functions, and then formulates explicit optimization problems for ACF-sidelobe energy and AF volume under nearly fixed time-bandwidth product (Hague, 2020). In the reported thumbtack-like example, $F(t)$7 with ISR optimization reduces ACF-sidelobe-ISR from $F(t)$8 dB to $F(t)$9 dB while retaining $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$0; increasing $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$1 to $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$2, $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$3, and $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$4 yields ISR $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$5 dB, $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$6 dB, and $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$7 dB with $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$8, $E(t)=\sum_n \alpha_n \psi_n(t)e^{-i\omega_0 t},$9, and $\{\psi_n(t)\}$0, respectively (Hague, 2020). The same analysis emphasizes constant-envelope transmission and high spectral compactness as intrinsic consequences of smooth frequency modulation rather than chipwise phase coding.

A distinct RF realization uses time-modulated arrays to generate frequency-hopped M-ary FSK directly at the antenna. There the element switching functions are periodic and binary, with duty cycle $\{\psi_n(t)\}$1, and their Fourier expansion produces discrete sidebands at $\{\psi_n(t)\}$2 for $\{\psi_n(t)\}$3 (Maneiro-Catoira et al., 2024). The first positive sideband carries almost all the power,

$\{\psi_n(t)\}$4

and the worst unwanted sideband is at $\{\psi_n(t)\}$5 dB. In the reported comparison, a conventional FH-MFSK + VPS array has insertion loss $\{\psi_n(t)\}$6 dB, whereas the TMA approach yields $\{\psi_n(t)\}$7 dB, for a net gain $\{\psi_n(t)\}$8 dB; the same $\{\psi_n(t)\}$9 dB SNR advantage appears in the BER curves (Maneiro-Catoira et al., 2024). Beam steering is embedded through continuous delays $\hat U_t[\phi]=\exp[i\phi(t)]$0, and for the $\hat U_t[\phi]=\exp[i\phi(t)]$1, $\hat U_t[\phi]=\exp[i\phi(t)]$2 case study the half-power beamwidth is $\hat U_t[\phi]=\exp[i\phi(t)]$3, while a $\hat U_t[\phi]=\exp[i\phi(t)]$4 ns delay step at $\hat U_t[\phi]=\exp[i\phi(t)]$5 kHz corresponds to phase error $\hat U_t[\phi]=\exp[i\phi(t)]$6.

4. Quantum metrology with temporally modulated oscillators

In continuous-variable quantum metrology, TFM denotes direct modulation of the oscillator frequency itself. The model Hamiltonian is

$\hat U_t[\phi]=\exp[i\phi(t)]$7

with $\hat U_t[\phi]=\exp[i\phi(t)]$8 the unknown base frequency to be estimated (Kong et al., 13 Jun 2026). The dynamics are expressed through a classical complex solution $\hat U_t[\phi]=\exp[i\phi(t)]$9 of

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$00

whose modulus, phase, and logarithm determine shearing, squeezing, and rotation parameters in the Gaussian unitary. Under the adiabatic condition $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$01, a WKB analysis shows that the relevant accumulated phase is

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$02

and that the dominant contribution to the QFI scales as

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$03

This is the central metrological statement of the framework. For static encoding, the coherent-state benchmark is $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$04, whereas the driven scheme yields an enhancement ratio

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$05

under a fair final-time energy constraint $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$06 (Kong et al., 13 Jun 2026). The reported examples make the scaling explicit. For polynomial modulation $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$07, one has $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$08, $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$09, and $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$10, giving $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$11 growth at long times. For exponential modulation $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$12, the QFI scales as $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$13, and $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$14 grows as $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$15 in the long-time limit (Kong et al., 13 Jun 2026).

The output state remains Gaussian, so optimal measurement is operationally simple. Time-dependent homodyne detection of a quadrature

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$16

achieves $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$17 for both polynomial and exponential drivings in the reported numerical and analytic studies, indicating that the quantum Cramér–Rao bound is asymptotically saturable with real-time homodyne readout (Kong et al., 13 Jun 2026).

5. Sequential temporal modulation, nonreciprocity, and frequency conversion

A major photonic use of TFM is the creation of nonreciprocal or selective frequency conversion without magnetic bias. In the three-mode scheme of (Pour et al., 15 Apr 2026), mode amplitudes evolve under two sequential, non-overlapping rectangular gates,

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$18

with duty cycles $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$19 and an idle interval satisfying $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$20. Harmonic-balance expansion in Floquet sidebands and a Dyson–Born treatment yield a closed-form isolation ratio in which forward and reverse pathways differ because they spend unequal dwell times in the lossy intermediate mode 2. The practical consequences are explicit: optimal isolation is found for $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$21–$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$22, zero isolation occurs when $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$23, and one should choose $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$24 so that $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$25–$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$26 (Pour et al., 15 Apr 2026). Time-domain simulations show peaks of $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$27 dB isolation near $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$28, with analytic and full numerical results agreeing to better than $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$29 dB across the practical parameter regime.

The Bragg Frequency Convertor realizes a different TFM mechanism by combining spatial and temporal periodicities in a quarter-wave stack. Its refractive index is modulated as

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$30

with the modulation applied selectively to either high-index or low-index layers (Taravati, 7 Mar 2026). The coupled-mode analysis predicts layer-dependent phase matching: modulation of the high-index layers selectively yields down-conversion, whereas modulation of the low-index layers leads to up-conversion. Because the static Bragg stopband suppresses the carrier and the non-phase-matched sideband, the output is dominated by a single converted tone $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$31, with residual signals and higher harmonics suppressed by $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$32–$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$33 dB (Taravati, 7 Mar 2026). The same report states that $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$34–$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$35 is sufficient for $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$36 conversion in a few-mm device, and that conversion can be tuned continuously through the modulation phase $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$37.

These magnet-free proposals coexist with a stringent system-level critique. Khurgin’s analysis of optical isolation by temporal modulation shows that, independent of whether modulation is achieved by carrier injection, Pockels and acousto-optic effects, or any other conceivable method, full isolation without excessive insertion loss is constrained by footprint, modulation frequency, and power consumption (Khurgin, 2022). The central scale is

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$38

so $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$39 GHz implies $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$40 mm and $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$41 GHz implies $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$42 mm. For practical mm-scale devices targeting $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$43 dB isolation with $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$44 dB insertion loss, the required drive power is estimated to be on the order of at least $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$45 W (Khurgin, 2022). A plausible implication is that temporal modulation can provide chip-compatible nonreciprocity, but only under nontrivial frequency–length–power trade-offs.

6. Temporal-frequency filtering in LLM-based recommendation

In recommender systems, TFM has acquired a computational meaning that is explicitly spectral. The FreLLM4Rec framework first applies a Global Graph Low-Pass Filter to pretrained item embeddings,

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$46

where $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$47 is the symmetric normalized Laplacian of the global item–item co-occurrence graph (Wang et al., 14 Aug 2025). TFM then acts inside each Transformer layer by taking a 1D DFT along the temporal axis of the hidden states, multiplying by a Butterworth mask,

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$48

and applying the inverse transform. The intended effect is to preserve low temporal frequencies, which the paper connects theoretically to low graph frequencies under a ring-graph approximation.

The theoretical justification is given as an informal spatio-temporal equivalence theorem: under a Spatio-Temporal Locality assumption, temporal low-pass filtering reduces the graph Laplacian quadratic form,

$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$49, and therefore shifts energy toward smaller graph eigenvalues (Wang et al., 14 Aug 2025). This is important because an exact local graph filter for each user sequence would require an $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$50 eigendecomposition per layer, whereas TFM adds only FFT and mask costs: $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$51

compared with standard Transformer complexity $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$52.

The implementation described in the paper uses Qwen2.5-7B-Instruct or another frozen pretrained LLM, a two-layer fusion MLP with GELU activations, $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$53, and Butterworth order $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$54, with best results at $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$55, $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$56 (Wang et al., 14 Aug 2025). Reported NDCG@10 values are:

Backbone	Dataset	NDCG@10 Base $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$57 +Fre
Qwen2.5-7B	All Beauty	$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$58
Qwen2.5-7B	Movies & TV	$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$59
Qwen2.5-7B	LastFM	$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$60
Llama3.1-8B	All Beauty	$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$61
Llama3.1-8B	Movies & TV	$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$62
Llama3.1-8B	LastFM	$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$63

The same study states that vanilla LLMs lose $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$64 of collaborative energy by layer 6, whereas FreLLM4Rec maintains it nearly intact, and reports improvements of up to $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$65 in NDCG@10 over the best baseline (Wang et al., 14 Aug 2025). Here TFM no longer means modulation of a physical carrier; it denotes an intra-layer frequency-domain correction that preserves collaborative low-frequency structure.

7. Scope, misconceptions, and recurring constraints

A common misconception is that TFM always means direct frequency sweeping of a carrier. The cited literature does not support such a narrow definition. It includes derivative-based recovery of instantaneous frequency from measured temporal phase, phase-only time/frequency operators for temporal-mode unitaries, finite-Fourier-series waveform design, oscillator-frequency driving, sequential time-gated couplings that produce unequal dwell times in a lossy mode, and FFT-based low-pass filtering of hidden-state sequences (Jachura et al., 2017, Ashby et al., 2020, Hague, 2020, Kong et al., 13 Jun 2026, Pour et al., 15 Apr 2026, Wang et al., 14 Aug 2025). The term is therefore polysemous but not arbitrary: in every case, a temporal law is used to reshape spectral or modal behavior.

Another misconception is that temporal modulation is automatically inexpensive or lossless. The quantum temporal-mode framework is phase-only and unitary in principle, but Ashby et al. note real component losses of EOPM insertion loss $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$66–$m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$67 dB and CFBG $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$68 dB, together with synchronization and RF-waveform-generation challenges (Ashby et al., 2020). The spectral-interferometric TFM measurement technique is phase-sensitive and low-jitter, but requires a stable interferometer and high-resolution OSA, and its acquisition time of a single OSA sweep is $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$69 s, much slower than electronic scopes in current implementations (Jachura et al., 2017). In nonreciprocal photonics, small duty cycles enhance asymmetry yet reduce conversion efficiency if they approach zero, and the general isolation problem remains bounded by the $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$70 and power-consumption constraints identified by Khurgin (Pour et al., 15 Apr 2026, Khurgin, 2022).

Taken together, these works present TFM as a methodological family for engineering time dependence in frequency-domain behavior. In optics and quantum information it enables arbitrary finite-mode unitary maps and direct phase-sensitive diagnostics; in radar and sonar it supplies constant-envelope, spectrally compact waveform families with controllable ACF/CCF and AF structure; in quantum metrology it changes the scaling law of phase accumulation through $m(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},$71; in photonics it supports magnet-free nonreciprocal conversion and Bragg-enabled frequency translation; and in recommender systems it serves as a computationally efficient low-pass correction that preserves collaborative signal components layer by layer (Ashby et al., 2020, Hague, 2020, Kong et al., 13 Jun 2026, Taravati, 7 Mar 2026, Wang et al., 14 Aug 2025).