Deep Frequency Modulation Interferometry

• DFMI is a laser interferometry method that applies strong frequency modulation and unequal-arm delay to encode the interferometric phase and modulation depth into a Bessel-coded harmonic spectrum.
• It enables multi-fringe readout and picometer-level displacement measurements by resolving phase ambiguity and offering sub-wavelength absolute length determination.
• The approach leverages both frequency-domain NLS and time-domain EKF algorithms, balancing high-throughput offline analysis with real-time state tracking for applications in precision sensing.

Searching arXiv for recent and foundational DFMI papers to ground the article. arXiv_search({"query":"\"deep frequency modulation interferometry\" OR DFMI interferometry", "max_results": 10, "sort_by": "submittedDate"}) Deep Frequency Modulation Interferometry (DFMI) is a laser interferometry technique in which a laser is strongly frequency-modulated before being sent into an unequal-arm interferometer. The detected interference signal contains a modulation-induced harmonic structure produced by delayed self-interference, so the path-length difference is encoded simultaneously in the usual interferometric phase and in an effective modulation depth. In the cited literature, this structure underpins multi-fringe readout, picometer-level displacement metrology, and the possibility of absolute length determination with sub-wavelength accuracy (Isleif et al., 2019, Dovale-Álvarez, 31 Jul 2025, Dovale-Álvarez, 15 Aug 2025).

1. Signal formation and harmonic encoding

In DFMI, the detected voltage is written as

$$v(t)=A\left[1+k\cos(\Delta\Phi_{\text{total}}(t))\right],$$

with the exact total phase difference

$$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$

Here $\tau_r(t)=l_r(t)/c$ and $\tau_m(t)=l_m(t)/c$ are the arm delays. Decomposing the phase yields a carrier term proportional to the optical path difference and a modulation term arising from the delayed interference of the modulated laser field with itself. For sinusoidal frequency modulation,

$f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$

with corresponding phase modulation

$$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$

Under the approximation $\omega_m\tau\ll 1$, the signal reduces to

$$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$

where

$$m=\frac{2\pi\Delta f\,\Delta l}{c}, \qquad \Phi=\frac{2\pi f_0\Delta l}{c}.$$

The harmonic amplitudes follow a Bessel-function decomposition,

$$\alpha_n(\mathbf{x})=2C\,J_n(m)\cos\left(\Phi+n\frac{\pi}{2}\right)e^{-in\psi},$$

with $$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$0 and $$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$1 (Dovale-Álvarez, 15 Aug 2025).

This formalism defines the distinctive DFMI measurement mechanism. The phase $$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$2 is precise but ambiguous modulo $$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$3, whereas $$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$4 scales directly with the absolute delay and is therefore unambiguous in the sense used by the absolute-ranging literature. The signal is not interpreted through a single quadrature point or a heterodyne beat note; instead, the harmonic spectrum itself is the observable. A plausible implication is that unequal-arm propagation is not merely an implementation detail but the condition that converts laser frequency modulation into a measurable, delay-dependent self-homodyne encoding (Dovale-Álvarez, 31 Jul 2025).

2. Relation to deep phase modulation interferometry

The mathematical form used in DFMI is closely related to deep phase modulation interferometry, which was developed as a homodyne readout scheme in which one interferometer arm is sinusoidally phase-modulated with a large modulation depth. In that formulation, the photodetector signal is modeled as

$$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$5

and expanded as

$$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$6

with

$$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$7

That work explicitly presented the method as an extension of the “$$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$8” methods, using many harmonics up to an order $$\Delta\Phi_{\text{total}}(t)=\phi_{\text{laser}}(t-\tau_r(t))-\phi_{\text{laser}}(t-\tau_m(t)).$$9 rather than a small fixed subset, and solved for the unknown parameters through nonlinear least squares (Heinzel et al., 2012).

The relation is structural rather than terminological. Deep phase modulation interferometry modulates an arm phase directly, whereas DFMI modulates the laser frequency and relies on unequal-arm delay to convert that modulation into the same Bessel-structured waveform. This suggests that deep phase modulation interferometry functions as a mathematical and algorithmic precursor to DFMI, especially in its use of overdetermined harmonic fitting, continuous tracking over many fringes, and extension to Differential Wavefront Sensing. In the deep phase modulation experiment, useful operating regions were identified roughly in the range $\tau_r(t)=l_r(t)/c$0, with particularly good performance around $\tau_r(t)=l_r(t)/c$1 with $\tau_r(t)=l_r(t)/c$2, or $\tau_r(t)=l_r(t)/c$3 with $\tau_r(t)=l_r(t)/c$4, and the reported implementation chose $\tau_r(t)=l_r(t)/c$5 and $\tau_r(t)=l_r(t)/c$6 (Heinzel et al., 2012).

3. Readout algorithms and digital estimation

Two estimator families recur in the recent DFMI literature: frequency-domain Non-Linear Least Squares (NLS) and time-domain Extended Kalman Filters (EKF). In the NLS formulation, the first $\tau_r(t)=l_r(t)/c$7 harmonic quadratures are extracted from a full modulation cycle and fit to the analytic harmonic model through a cost function of the form

$\tau_r(t)=l_r(t)/c$8

The fit is usually performed with Levenberg–Marquardt,

$\tau_r(t)=l_r(t)/c$9

and is described as accurate for static or slowly varying signals, asymptotically efficient under the stated assumptions, and natural for offline or block-based processing. The EKF instead processes raw samples sequentially using the nonlinear measurement equation

$\tau_m(t)=l_m(t)/c$0

together with a predict-update recursion and an augmented state that includes rates of change. In the ideal constant-parameter limit, the two approaches access the same Fisher information and can achieve the same asymptotic statistical limit for $\tau_m(t)=l_m(t)/c$1 (Dovale-Álvarez, 31 Jul 2025).

The software ecosystem described in the cited work formalizes this split between batch precision and streaming estimation. DeepFMKit provides a high-fidelity physics engine with time-of-flight delays in dynamic interferometers, arbitrary laser modulation waveforms, and colored noise from user-defined $\tau_m(t)=l_m(t)/c$2 spectral densities, together with a highly-optimized, parallelized frequency-domain NLS and multiple time-domain EKF implementations using both random walk and integrated random walk process models. It also includes a high-throughput experimentation framework for parameter sweeps and Monte Carlo analyses, with an illustrative example comprising a 2D sweep over modulation depth and second-harmonic distortion amplitude with 500 Monte Carlo trials per grid point, totaling 50 million simulate-and-fit runs (Dovale-Álvarez, 15 Aug 2025).

For DFMI practice, the estimator choice is therefore architectural rather than doctrinal. NLS is identified with high-throughput offline analysis of buffered data, while EKF is identified with real-time state tracking sample-by-sample. The underlying measurement physics is unchanged; what changes is latency, handling of dynamics, and the way model imperfections are incorporated.

4. Implementations and demonstrated performance

A compact DFMI implementation reported a single-component interferometer “smaller than a cubic inch” for each sensor element. The system used a fiber-coupled external cavity diode laser at $\tau_m(t)=l_m(t)/c$3, frequency-modulated by about

$\tau_m(t)=l_m(t)/c$4

at

$\tau_m(t)=l_m(t)/c$5

The modulated light was distributed to a frequency-reference interferometer and to a “Test Mass in the Middle” sensor based on two prism interferometers. The readout algorithm performed a non-linear fit using a Bessel-function decomposition of the complex amplitudes at the modulation frequency and roughly the first ten harmonics, extracting the interferometric phase, the signal amplitude, the modulation index $\tau_m(t)=l_m(t)/c$6, and the modulation phase in real time. Reported effective modulation indices were

$\tau_m(t)=l_m(t)/c$7

The phasemeter was software-based and ran on a PC, using an 8-channel DAQ at $\tau_m(t)=l_m(t)/c$8, with software feedback for test-mass tilt control, test-mass pathlength control, laser frequency control, and modulation-deviation control; the control bandwidth was limited to about $\tau_m(t)=l_m(t)/c$9 by the phasemeter (Isleif et al., 2019).

The same work reported displacement sensing at the $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$0 level between roughly $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$1 and $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$2 in the residual phase noise combination $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$3, sub-picometer precision summarized in the abstract as better than $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$4 at frequencies below $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$5, and tilt sensing around $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$6 above about $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$7. The paper also emphasized a dynamic range greater than two orders of magnitude between the motion signal and the residual noise at some frequencies (Isleif et al., 2019).

The deep phase modulation predecessor demonstrated the same general multi-harmonic logic in a homodyne setting on a stable optical bench, using a Mach–Zehnder-type setup with phase modulation applied by a ring piezoelectric transducer wrapped with single-mode fiber. The modulation frequency was $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$8, the data were sampled at $f_{\text{mod}}(t)=\Delta f\cos(\omega_m t+\psi),$9, the FFT block length was $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$0 samples, and the phase readout update rate was $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$1. After mitigation of laser frequency noise and DAQ transfer-function distortions, it achieved optical pathlength sensitivity of about $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$2 above $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$3 and angular resolution better than $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$4 above about $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$5, with more than a factor of 35 improvement relative to the initial uncorrected result (Heinzel et al., 2012).

Taken together, these implementations exemplify the recurring DFMI trade: optical minimalism is exchanged for model-based digital reconstruction. In the cited compact prism realization, that trade is explicit: one modulated laser source, one reference interferometer per laser, and many compact sensor heads can share the same stabilized laser and phasemeter infrastructure (Isleif et al., 2019).

5. Absolute ranging, precision limits, and systematic effects

The absolute-ranging formulation of DFMI uses the unambiguous estimate of $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$6 to resolve the fringe order of the wrapped phase $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$7. With

$$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$8

the coarse length and coarse phase are formed from $$\phi_{\text{mod}}(t)=\frac{\Delta f}{f_m}\sin(\omega_m t+\psi).$$9, and the fringe order is estimated by

$\omega_m\tau\ll 1$0

The ambiguity-resolution condition is stated as

$\omega_m\tau\ll 1$1

or conservatively,

$\omega_m\tau\ll 1$2

A central observation is that $\omega_m\tau\ll 1$3 is usually much larger than $\omega_m\tau\ll 1$4 because the coarse phase amplifies modulation-depth error by the factor $\omega_m\tau\ll 1$5 (Dovale-Álvarez, 31 Jul 2025).

The same analysis derives Cramér–Rao bounds for both $\omega_m\tau\ll 1$6 and $\omega_m\tau\ll 1$7. For $\omega_m\tau\ll 1$8,

$\omega_m\tau\ll 1$9

while for the interferometric phase,

$$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$0

The literature also identifies intrinsic “dead zones” where the uncertainty in $$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$1 sharply worsens because the Bessel-function structure renders the Jacobian ill-conditioned. These are signal-structure limitations rather than failures of a particular estimator. Using more harmonics reduces sensitivity to any single dead zone, but does not eliminate the underlying structure (Dovale-Álvarez, 31 Jul 2025).

Systematic effects dominate once statistical noise is sufficiently reduced. The principal analyzed sources are modulation non-linearity, residual amplitude modulation (RAM), ghost beams, scale-factor calibration errors, and laser drift. For second-harmonic distortion, the modulation is modeled as

$$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$2

while RAM is represented through a time-varying amplitude

$$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$3

A major result is the existence of “valleys of robustness,” operating points where bias is strongly suppressed. For modulation non-linearity, bias in $$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$4 is suppressed when $$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$5 and bias in $$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$6 is suppressed when $$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$7; for RAM, the corresponding conditions are $$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$8 and $$v(t)=A\left[1+k\cos\left(\Phi+m\cos(\omega_m t+\psi)\right)\right],$$9. The cited analysis stresses that there is no single universally optimal $$m=\frac{2\pi\Delta f\,\Delta l}{c}, \qquad \Phi=\frac{2\pi f_0\Delta l}{c}.$$0, because robustness valleys for $$m=\frac{2\pi\Delta f\,\Delta l}{c}, \qquad \Phi=\frac{2\pi f_0\Delta l}{c}.$$1 and $$m=\frac{2\pi\Delta f\,\Delta l}{c}, \qquad \Phi=\frac{2\pi f_0\Delta l}{c}.$$2, and for RAM and modulation non-linearity, do not generally coincide (Dovale-Álvarez, 31 Jul 2025).

6. Related methods, applications, and bibliographic issues

DFMI is closely related to other frequency-modulation-based metrology, but the relation is not identity. A conceptually adjacent example is the calibration of LIGO displacement actuators via laser frequency modulation, where sinusoidal laser frequency modulation creates a calibrated effective length variation in a resonant cavity according to

$$m=\frac{2\pi\Delta f\,\Delta l}{c}, \qquad \Phi=\frac{2\pi f_0\Delta l}{c}.$$3

That work uses the same physical equivalence between frequency modulation and effective optical path variation, but as a displacement fiducial for actuator calibration in a single-arm locked interferometer rather than as a DFMI readout architecture. Its role in the DFMI context is therefore foundational and metrological rather than definitional (Goetz et al., 2010).

The application domain emphasized in the cited DFMI literature includes optical gradiometers for satellite geodesy, LISA-like gravitational reference systems, ground-based gravity experiments, and precision accelerometers and inertial sensors. The compact multi-fringe prism implementation is presented in exactly that context: a route to low-frequency displacement and tilt sensing that is compatible with compact sensor heads and distributed readout architectures (Isleif et al., 2019).

A separate bibliographic issue concerns misattribution. The supplied record for “Towards a FPGA-controlled deep phase modulation interferometer” (Terán et al., 2014) states that the document is in fact a JPCS manuscript-preparation guide and contains no interferometry content, no optical setup, no modulation equations, no FPGA or LEON3 implementation, and no DFMI-relevant results. As a consequence, it should not be used as a technical source for Deep Frequency Modulation Interferometry or for deep phase modulation interferometry (Terán et al., 2014).

DFMI, as represented in these sources, is therefore best understood as a structured family of unequal-arm, strongly frequency-modulated self-homodyne readouts in which absolute and wrapped length observables are jointly inferred from a Bessel-coded harmonic spectrum. Its central technical questions are no longer confined to raw displacement sensitivity; they include estimator architecture, calibration fidelity, robustness against systematic distortion, and operating-point selection for ambiguity-resolved absolute ranging.