Kramers-Kronig Receiver: Principles & Applications

• The Kramers-Kronig receiver is a direct-detection scheme that reconstructs complex optical fields from intensity measurements by enforcing the minimum-phase condition with a strong carrier or DC bias.
• It utilizes digital signal processing operations—such as square-root, logarithm, and Hilbert transforms—to accurately recover amplitude and phase without traditional coherent detection hardware.
• This receiver finds applications in optical, THz, quantum, and orbital-angular-momentum systems, offering simplified optical front-ends while shifting complexity to DSP.

A Kramers–Kronig receiver is a direct-detection architecture that reconstructs a complex-valued field from an intensity measurement when the received waveform satisfies the minimum-phase condition and, in several formulations, a single-sideband constraint. In its canonical form, a strong continuous-wave carrier or local-oscillator-like tone is transmitted together with the information-bearing signal, the photocurrent is sampled with a single-ended photodiode, and DSP recovers the field amplitude and phase by combining square-root, logarithm, Hilbert transform, and complex reconstruction operations. Across optical fiber transmission, THz wireless links, orbital-angular-momentum metrology, and quantum-network protocols, the method is used to recover quadratures without a conventional coherent front end built around an optical hybrid and balanced detection (1711.02550, Hoang et al., 2017, Harter et al., 2019, Pousset et al., 2024).

1. Mathematical principle and minimum-phase reconstruction

The theoretical core of the KK receiver is the analyticity of the logarithm of a minimum-phase signal. If $s(t)$ is a minimum-phase analytic signal, then

$$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$

and the real and imaginary parts of $\ln s(t)$ form a Hilbert-transform pair. A standard reconstruction formula is

$$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$

with

$$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$

Equivalent formulations appear in frequency-domain and amplitude-phase notation, for example

$E(\omega)=A(\omega)e^{j\phi(\omega)},$

with phase recovered from $\ln A$ by a Cauchy principal value integral, and in time-domain notation

$$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$

after measuring $I(t)=|E(t)|^2$ (1711.02550, Hout et al., 2021, Heide et al., 2021).

The minimum-phase condition is not a secondary implementation detail. In several formulations it is stated as the condition that the complex field has no zeros in the upper half of the complex plane, or that the complex trajectory never encircles the origin. In practical communication systems this condition is enforced by adding a strong carrier tone, pilot tone, or DC bias so that the composite field remains strictly nonzero over time. For single-sideband constructions, the signal is also constrained spectrally so that analyticity is preserved on one side of the spectrum (Hoang et al., 2017, Pousset et al., 2024, Liu et al., 8 Sep 2025).

The same structure reappears outside one-dimensional temporal communications. In single-shot orbital-angular-momentum retrieval, the periodic azimuthal variable $\phi$ leads to the KK relation

$$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$0

after which the analytic field is reconstructed from $$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$1 and its KK phase. The periodic $$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$2-kernel is the azimuthal analogue of the Hilbert-transform kernel used in standard KK receivers (Lin et al., 2022).

2. Canonical architectures and DSP workflow

In the simplest optical communication architecture, the transmitter superposes a digitally generated carrier with the modulated waveform, launches the combined field through fiber, and a single photodiode measures the intensity. The DSP chain then performs square-root, logarithm, Hilbert transform, complex-field reconstruction, digital frequency downshift, equalization, and symbol decisions. A representative textual chain is

$$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$3

with variations that add DC correction, static frequency-domain equalization, or adaptive LMS stages (Hout et al., 2021, Heide et al., 2021).

Antonelli et al. described two early architectural branches. The PAM-based KK transceiver forms a nonnegative waveform

$$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$4

then generates a single-sideband field

$$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$5

After direct detection and KK reconstruction, digital chromatic-dispersion compensation is applied and the real PAM waveform is extracted. The second branch, the two-sided polarization-multiplexed KK transceiver, interleaves two real signals around a guard band, performs de-interleaving and local-oscillator mixing at the receiver, and applies KK reconstruction independently to each sideband and polarization (1711.02550).

More elaborate receiver structures preserve the same nonlinear front end while changing the optical front-end geometry. The Stokes-Vector Kramers-Kronig transceiver uses transmitted digital carriers, a Stokes-vector receiver, Stokes-parameter measurements $$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$6–$$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$7, estimation of a real $$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$8 polarization-rotation matrix, and a real-valued $$I(t)=|s(t)|^2,\qquad \ln s(t)=\ln|s(t)|+i\,\arg s(t),$$9 MIMO filter with 61 taps after KK recovery on each polarization branch. The generalized THz KK receiver replaces the square-law photodiode with a Schottky-barrier diode envelope detector and inserts an inverse nonlinearity

$\ln s(t)$0

before the logarithm and Hilbert transform so that phase retrieval is performed on $\ln s(t)$1 rather than on the response of an ideal quadratic detector (Hoang et al., 2017, Harter et al., 2019).

Real-time software-defined implementations preserve the same mathematics while exposing the receiver as a sequence of GPU kernels and batched FFTs. Reported pipelines use $\ln s(t)$2 ADC-sample buffers, overlap-and-save FFT Hilbert transforms with 1,024-point blocks, static frequency-domain equalization, and short decision-directed LMS equalizers. One field-trial implementation ran all DSP on the GPU in five asynchronous CUDA streams; another used four GPU streams in a pipeline; another used three streams in parallel to overlap DMA and compute (Heide et al., 2020, Heide et al., 2021, Heide et al., 2021).

3. Carrier engineering, guard bands, sampling, and quantization control

KK reception is strongly shaped by carrier engineering. The carrier-to-signal power ratio is defined as

$\ln s(t)$3

or, in other contexts, $\ln s(t)$4. The carrier must be sufficiently strong to guarantee the minimum-phase condition and to suppress signal-signal beat interference, but excessive carrier power reduces the usable dynamic range for the data signal. Reported thresholds and operating regions are application-dependent: a digitally added CW tone for a 64-QAM, 25 Gbaud optical KK receiver is described as requiring roughly $\ln s(t)$5–$\ln s(t)$6; Stokes-vector KK experiments reported typical optimum values in the $\ln s(t)$7–$\ln s(t)$8 range; generalized THz KK reported an optimum $\ln s(t)$9–$$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$0; and single-shot OAM retrieval identified a worst-case threshold of about $$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$1 (Hout et al., 2021, Hoang et al., 2017, Harter et al., 2019, Lin et al., 2022).

Guard-band placement is equally important. In digital-carrier optical implementations, the carrier is placed beyond the upper edge of the signal bandwidth by a guard band $$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$2, for example

$$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$3

so that the up-converted carrier tone and the double-sideband signal do not overlap. Reported values include $$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$4–$$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$5 in the Stokes-vector experiment, a CW offset of $$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$6 in the 64-QAM transmitter with digitally added CW tone, and $$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$7 in the THz experiment (Hoang et al., 2017, Hout et al., 2021, Harter et al., 2019).

Sampling density is a recurring non-ideal constraint because the logarithm broadens the effective spectral range. The Stokes-vector experiment recommended $$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$8 samples/symbol and reported a large SNR loss of $$\phi(t)=\arg s(t)=\tfrac12\,\mathcal H\{\ln I(t)\}, \qquad \hat s(t)=\sqrt{I(t)}\,e^{\,i\phi(t)},$$9 at $$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$0 samples/symbol due to discretization errors in log/Hilbert processing. In the OAM experiment, 71 physical angles were acquired and the result was 11× digitally up-sampled before the discrete Hilbert transform; fewer points or no up-sampling led to clear degradation (Hoang et al., 2017, Lin et al., 2022).

When the carrier is inserted digitally at the transmitter, DAC quantization becomes a central issue. Adding a strong CW tone digitally increases the peak-to-average ratio seen by the DAC and aggravates quantization noise in the signal band. The digital resolution enhancer reshapes the DAC quantization noise so that most of it falls outside the signal + CW band. Its workflow is

$$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$1

followed at the receiver by the inverse filter $$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$2. Without DRE the quantization-noise PSD is flat,

$$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$3

whereas with DRE

$$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$4

The reported implementation used a filter length $$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$5 taps, and a “soft” quantizer with 3 intermediate levels to linearize quantization around CW peaks (Hout et al., 2021).

4. Communication-system realizations

Across communication systems, KK reception is primarily valued for replacing conventional coherent optical hardware with direct detection plus DSP while retaining access to complex modulation, chromatic-dispersion compensation, and equalization. The specific realizations differ in how the minimum-phase condition is enforced, how polarization is handled, and whether the detector is square-law or generalized through a calibrated nonlinearity.

Implementation	Distinguishing feature	Reported result
Stokes-Vector Kramers-Kronig transceiver	Stokes-vector receiver, digital carrier, real $$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$6 MIMO	480 Gb/s over 80 km SSMF with BER below $$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$7
Generalized KK receiver for coherent THz communications	Schottky-barrier diode with polynomial inverse nonlinearity	net 115 Gbit/s at 0.3 THz over 110 m
KK receiver combined with digital resolution enhancer	digitally added CW tone and DRE shaping filter	OSNR gain $$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$8 back-to-back and $$\mathcal H\{f(t)\}=\frac{1}{\pi}\,\mathrm{p.v.}\!\int_{-\infty}^{\infty}\frac{f(\tau)}{t-\tau}\,d\tau.$$9 after 50 km SSMF
Real-time GPU-based minimum-phase KK receiver	software-defined DSP on GPU	10,000 km for 4-QAM and 5 Gbps net data rate using 64-QAM on a 91 km field link

In the Stokes-vector implementation, a 60.1 Gbaud PDM-16QAM single-carrier experiment over 80 km SSMF achieved a raw 480 Gb/s with BER $E(\omega)=A(\omega)e^{j\phi(\omega)},$0, while 32.5 Gbaud PDM-16QAM over 80 km achieved BER $E(\omega)=A(\omega)e^{j\phi(\omega)},$1. The study also reported that variation in SNR was $E(\omega)=A(\omega)e^{j\phi(\omega)},$2 across arbitrary SOPs, optimum CSPR increased with guard band, and small roll-off $E(\omega)=A(\omega)e^{j\phi(\omega)},$3 incurred a $E(\omega)=A(\omega)e^{j\phi(\omega)},$4 penalty in KK due to PAPR-related violation of minimum phase (Hoang et al., 2017).

In THz wireless communications, the generalized KK receiver transferred the optical KK scheme to $E(\omega)=A(\omega)e^{j\phi(\omega)},$5 carrier frequencies. Using 16QAM at up to 33 GBd, it achieved a line rate of 132 Gbit/s and a net 115 Gbit/s after 7% FEC over 110 m free space, with BER $E(\omega)=A(\omega)e^{j\phi(\omega)},$6 at $E(\omega)=A(\omega)e^{j\phi(\omega)},$7 at the SBD input. The same study reported that generalized KK processing yields up to an order-of-magnitude BER reduction over conventional square-law KK or direct heterodyne reception without a guard band, especially at high THz power where SBD saturation sets in (Harter et al., 2019).

In transmitter-limited optical links, the digital-resolution-enhanced KK receiver targeted the interaction between digitally inserted CW carriers and finite DAC resolution. For 64-QAM at 25 Gbaud with a 100 GSa/s DAC and a CW offset of 13.9 GHz, the study reported back-to-back NGMI improvement from 0.972 to 0.980 for a 6 bit DAC, and from 0.862 to 0.972 for a 4 bit DAC; after 50 km SSMF, NGMI improved from 0.936 to 0.949 for a 6 bit DAC and from 0.803 to 0.935 for a 4 bit DAC. The nominal DAC resolution was reduced from 6 bits to 5 bits, and even 4 bits, with only a small NGMI penalty when DRE was enabled (Hout et al., 2021).

Real-time implementations establish that KK detection is not confined to offline DSP. A metropolitan field trial over a 91 km metro ring processed 1 GBaud KK-QPSK and KK-16-QAM in real time at 4 Gs/s with five asynchronous CUDA streams and reported stable operation over a continuous 6 s run with $E(\omega)=A(\omega)e^{j\phi(\omega)},$8 Q-factor fluctuation, no packet loss, and no buffer underruns. Another real-time minimum-phase KK receiver evaluated 4-, 8-, 16-, 32-, and 64-QAM over a 100-span straight-line optical link and reported maximum reaches of 10,000 km, 7,600 km, 5,600 km, 3,600 km, and 1,600 km, respectively, at the 20% OH HD-FEC criterion. A related 91 km field-deployed experiment processed MP-QAM and geometrically shaped constellations in real time and reported a net data rate of 5 Gbps using 64-QAM (Heide et al., 2020, Heide et al., 2021, Heide et al., 2021).

5. Extensions to quantum networking, OAM metrology, and spectral tomography

KK reception has been extended beyond classical fiber and wireless data transmission into quantum communications and wavefront metrology. In continuous-variable quantum key distribution, the field is converted into a minimum-phase single-sideband signal through frequency shifting and addition of a DC bias,

$E(\omega)=A(\omega)e^{j\phi(\omega)},$9

with $\ln A$0 to guarantee that the field never winds around the origin. Bob measures

$\ln A$1

with a single low-noise photodiode, then reconstructs $\ln A$2, computes $\ln A$3, applies the Hilbert transform, and recovers $\ln A$4 after subtracting the DC bias and down-shifting by $\ln A$5. The detection stage contains no local oscillator, balanced homodyne, or interferometric arms at Bob’s end; the entire detection is a single-ended PD plus DSP (Liu et al., 8 Sep 2025).

The same work extended the protocol to a $\ln A$6 downstream quantum access network and reported $\ln A$7 per user with an aggregate throughput of $\ln A$8, over up to 20 km fiber and 4 dB channel loss, using $\ln A$9, $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$0, $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$1, $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$2, $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$3, and $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$4. The loss tolerance $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$5 matched that of interferometric heterodyne systems within 10%. A cost model expressed in units of $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$6 gave

$$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$7

so that the direct-detection KK-CV-QKD architecture rises only $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$8 per user (Liu et al., 8 Sep 2025).

In orbital-angular-momentum metrology, KK reception was adapted to a single-shot interferometric analyzer. A strong Gaussian reference $$\hat E(t)=A(t)\exp\{j\phi(t)\}, \qquad \phi(t)=\mathcal H\{\ln A(t)\},$$9 is added to the OAM field $I(t)=|E(t)|^2$0, the interferogram

$I(t)=|E(t)|^2$1

is recorded, and the periodic KK relation reconstructs the complex field from a single intensity trace. Reported performance on random spectra was $I(t)=|E(t)|^2$2 average accuracy for orders $I(t)=|E(t)|^2$3, and $I(t)=|E(t)|^2$4 for orders $I(t)=|E(t)|^2$5, with clear degradation when the sampling density or CSPR was insufficient (Lin et al., 2022).

A different quantum extension proposed a spectral-tomography protocol for single-photon states inspired by KK detection. After spectral engineering with an electro-optic modulator and narrowband filtering, a synthetic LO-like component is combined with a single-sideband single-photon envelope so that the reconstructed probability histogram can be processed by a discrete KK transform. Numerical simulations with EOM modulation at 80 MHz, photon bandwidth $I(t)=|E(t)|^2$6, detector jitter $I(t)=|E(t)|^2$7, dark counts $I(t)=|E(t)|^2$8, and efficiency 40% showed fidelities $I(t)=|E(t)|^2$9 with $\phi$0–$\phi$1 events (Pousset et al., 2024).

6. Noise, complexity, misconceptions, and open interpretations

A common misconception is that KK reception is simply “direct detection with a Hilbert transform.” The literature instead treats it as a constrained reconstruction problem whose validity depends on minimum-phase and, in several formulations, single-sideband signal design. The strong carrier, guard band, DC bias, or spectral asymmetry is therefore part of the receiver concept rather than a peripheral engineering choice. The corresponding penalties are explicit: bias in KK-PAM slightly reduces spectral efficiency, guard bands reduce usable bandwidth in TS-KK and carrier-aided systems, and higher CSPR consumes dynamic range and can aggravate DAC quantization noise unless compensated by techniques such as DRE (1711.02550, Hout et al., 2021).

Another misconception is that KK reception universally eliminates coherent-receiver DSP complexity. The optical hardware is simplified, but the digital workload increases. Reported real-time implementations rely on batched FFT/IFFT operations, overlap-and-save Hilbert transforms, static frequency-domain equalizers, down-sampling, and adaptive equalizers on GPUs processing $\phi$2-sample buffers at 4 Gs/s. One implementation reported GPU occupancy $\phi$3, peak memory bandwidth use $\phi$4, and sustained PCIe DMA $\phi$5; another emphasized that each $\phi$6-sample block had to be fully processed in $\phi$7 to preserve real-time operation (Heide et al., 2020, Heide et al., 2021, Heide et al., 2021).

The quantum-noise interpretation of KK detection is an active point of divergence across the cited literature. One analysis of coherent-state detection through a minimum-phase signal concluded that the KK receiver keeps the radial quantum fluctuation the same as balanced heterodyne detection, reduces the tangential fluctuation to $\phi$8 times the radial one at the decision time, and therefore achieves $\phi$9 times the signal-to-noise ratio of balanced heterodyne detection, with an asymmetric, time-varying noise ellipse (Zhang et al., 2020). A later quantum treatment showed that, up to first order in the local-oscillator amplitude, KK detection acts as a coherent detection able to measure both quadratures, making it a Gaussian measurement similar to double homodyne detection, with the same 3 dB extra vacuum-noise picture as heterodyne detection on each spectral mode (Pousset et al., 2024). This suggests that the quantum interpretation remains sensitive to the signal model, the reconstruction formalism, and the approximation regime adopted in the analysis.

Taken together, the literature presents the Kramers–Kronig receiver as a family of carrier-aided, minimum-phase direct-detection schemes that relocate complexity from the optical front end to DSP. Its established uses include chromatic-dispersion-tolerant PAM links, polarization-multiplexed direct detection, Stokes-vector and THz coherent reception, real-time software-defined receivers, orbital-angular-momentum spectrum retrieval, and continuous-variable quantum networking. Its continuing research themes are equally consistent across domains: tighter minimum-phase enforcement, lower sampling and quantization penalties, more efficient real-time implementations, and a more unified account of quantum-noise behavior (1711.02550, Hoang et al., 2017, Hout et al., 2021, Liu et al., 8 Sep 2025).