Balanced Homodyne Detection Explained

• Balanced homodyne detection is a phase-sensitive measurement that mixes a weak signal with a strong local oscillator and subtracts balanced photodetector outputs to isolate a selected field quadrature.
• Changing the local-oscillator phase measures amplitude, phase, or intermediate quadratures, enabling quantum-state tomography, CV-QKD, QRNG, gravitational-wave readout, and precision optical metrology.
• Practical performance depends on common-mode rejection, detector bandwidth, quantum efficiency, temporal and spatial mode matching, LO power, and calibration, with imperfections causing technical noise, pulse correlations, or measurement loss.

Balanced homodyne detection (BHD) is a phase-sensitive optical measurement in which a weak signal field is interfered with a strong coherent local oscillator (LO) on a balanced beam splitter, and the two output photodetector signals are subtracted. The common-mode optical intensity, dominated by the LO, is ideally canceled, while the antisymmetric signal–LO interference term remains. In the strong-LO limit, the difference current is proportional to a signal-field quadrature selected by the relative optical phase. BHD is used in continuous-variable quantum optics, quantum-state tomography, continuous-variable quantum key distribution (CV-QKD), quantum random-number generation (QRNG), gravitational-wave readout, quantum metrology, optical communications, and cavity characterization.

1. Operating principle and quadrature measurement

For signal and LO annihilation operators $\hat a_{\rm s}$ and $\hat a_{\rm LO}$ entering an ideal $50{:}50$ beam splitter, representative output modes are

$$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$

The output photon-number difference is

$$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$

Treating the LO as a strong coherent field,

$$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$

gives

$$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$

Accordingly, the difference photocurrent has the form

$$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$

where $\eta$ is the detector quantum efficiency and $\hat X_\theta$ is the phase-dependent field quadrature. Depending on normalization, it is written as

$\hat a_{\rm LO}$0

or without the factor $\hat a_{\rm LO}$1. The phase $\hat a_{\rm LO}$2 selects an amplitude-like quadrature, $\hat a_{\rm LO}$3 selects a phase-like quadrature, and intermediate phases select linear combinations of amplitude and phase fluctuations.

The LO therefore performs two related functions. It supplies a phase reference that defines the measured quadrature, and it converts the weak field quadrature into a macroscopic electrical signal. The signal contribution scales linearly with LO amplitude, whereas the shot-noise variance scales linearly with LO power. Increasing the LO does not reduce the normalized quantum uncertainty; it increases the absolute signal and shot-noise levels relative to approximately LO-independent electronic noise.

Repeated measurements at different LO phases yield quadrature distributions,

$\hat a_{\rm LO}$4

which support optical quantum-state tomography. A homodyne detector can also be configured with a shaped LO to project onto a selected spatial or temporal mode. For spatial super-resolution, a local oscillator in the $\hat a_{\rm LO}$5 mode converts a small transverse displacement into a measurable quadrature amplitude (Gosalia et al., 2023). For temporally structured states, the measured photocurrent must be projected onto the relevant temporal mode (Sandner et al., 15 Feb 2026).

2. Balanced subtraction and common-mode rejection

The principal technical advantage of BHD is common-mode noise suppression. LO intensity fluctuations appear approximately identically in both photodiode currents,

$\hat a_{\rm LO}$6

so that

$\hat a_{\rm LO}$7

The signal–LO interference terms have opposite signs at the two beam-splitter outputs and therefore add in the difference. This distinction between common-mode and differential-mode signals is the basis of balanced detection.

Real detectors are never perfectly balanced. Residual common-mode coupling arises from unequal photodiode quantum efficiencies, unequal optical powers, imperfect beam-splitter ratios, different temporal responses, optical-delay mismatch, unequal amplifier gains, frequency-response mismatch, PCB parasitics, and imperfect electronic subtraction. A representative imbalance model is

$\hat a_{\rm LO}$8

where $\hat a_{\rm LO}$9 and $50{:}50$0 are effective channel gains. The balancing condition is correspondingly of the form

$50{:}50$1

The common-mode rejection ratio (CMRR) is generally expressed as

$50{:}50$2

although some experiments define it through spectral-power ratios. Reported values span from approximately $50{:}50$3 for an SNSPD-based BHD to $50{:}50$4 for a silicon-photonics-integrated time-domain detector (Protte et al., 2023, Jia et al., 2023). Other systems reported $50{:}50$5, $50{:}50$6, $50{:}50$7, $50{:}50$8, and up to $50{:}50$9 in particular operating configurations (Chi et al., 2010, Kumar et al., 2011, Zhang et al., 2018, Cooper et al., 2011, Stefszky et al., 2012).

High CMRR does not eliminate differential technical noise, electronic noise, detector saturation, beam-pointing noise, or single-arm scattering. At low Fourier frequencies, beam motion across spatially nonuniform photodiodes can generate noise that is not common to both detectors. Scattering after the beam splitter can produce single-arm fluctuations, while scattering before the beam splitter can be rejected more effectively. In audio-band experiments, low-flicker-noise resistors, mode-cleaning cavities, mechanical enclosure, beam dumps, optical isolation, and environmental stabilization were required to obtain a flat shot-noise spectrum below $$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$0 (Stefszky et al., 2012).

3. Detector architectures and performance parameters

A practical BHD comprises a signal and LO input, a balanced optical mixer, two matched photodetectors, a subtraction stage, amplification, and a calibrated acquisition system. Implementations differ substantially in bandwidth, detector technology, temporal regime, and integration level.

Architecture or application	Representative performance
Broadband GMCS-QKD BHD	$$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$1 bandwidth, $$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$2 electronic-noise clearance, $$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$3 CMRR
Wideband tomography BHD	$$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$4 pulsed bandwidth, $$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$5 CW bandwidth, $$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$6 clearance, $$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$7 CMRR
RF-integrated BHD	$$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$8 bandwidth, approximately $$\hat b_1=\frac{\hat a_{\rm s}+\hat a_{\rm LO}}{\sqrt{2}}, \qquad \hat b_2=\frac{\hat a_{\rm s}-\hat a_{\rm LO}}{\sqrt{2}}.$$9 QCNR, $$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$0 CMRR
Die-level vacuum-noise BHD	$$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$1 bandwidth, $$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$2 integrated QCNR over $$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$3
MIR BHD	$$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$4 bandwidth, detector quantum efficiency up to $$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$5, clearance up to $$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$6
SNSPD BHD	$$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$7 bandwidth, $$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$8 shot-noise clearance
Silicon-photonics TBHD	$$\hat n_1-\hat n_2 = \hat a_{\rm s}^{\dagger}\hat a_{\rm LO} + \hat a_{\rm LO}^{\dagger}\hat a_{\rm s}.$$9 optical circuit, $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$0 CMRR
Feedback-free pulsed BHD	$$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$1 operation, approximately $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$2 maximum SNR, $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$3 CMRR

Conventional high-speed detectors often use low-noise operational amplifiers in transimpedance configurations. The photodiode capacitance, amplifier feedback, PCB parasitics, and power-supply inductance jointly determine the response. The detector in “Versatile Wideband Balanced Detector for Quantum Optical Homodyne Tomography” used Hamamatsu S5972 photodiodes and an OPA847 transimpedance amplifier, achieving $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$4 pulsed and $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$5 CW bandwidths (Kumar et al., 2011). A broadband RF design using an Agilent ABA-52563 integrated circuit and low-capacitance InGaAs photodiodes extended the measured bandwidth to approximately $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$6 while retaining a QCNR of approximately $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$7 (Zhang et al., 2018).

A different architecture connects matched photodiodes at a common electrical node and amplifies the resulting voltage with a broadband voltage amplifier rather than using a feedback TIA. The feedback-free current-to-voltage conversion avoids some nonlinearities, dynamic instabilities, overshoot, ringing, and pulse broadening associated with ultrashort-pulse operation. A $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$8 implementation at $$\hat a_{\rm LO}\rightarrow\alpha_{\rm LO} = |\alpha_{\rm LO}|e^{i\theta},$$9 achieved linear shot-noise scaling, approximately $$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$0 SNR, and negligible intrinsic inter-pulse correlations (Altilia et al., 8 Apr 2026).

Integrated photonics reduces optical footprint and can provide common-wafer fabrication of matched photodiodes. A silicon-photonics TBHD used a $$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$1 multimode-interference coupler, PIN-based variable optical attenuators, integrated germanium photodiodes, and automatic balancing. It achieved $$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$2 coherent-state tomography fidelity and demonstrated CV-QKD over a $$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$3 multiplexed channel, although approximately $$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$4 edge-coupling loss, lower detection efficiency, balance drift, and external delay lines remained limitations (Jia et al., 2023).

4. Noise, bandwidth, and temporal-mode effects

The detector output can be modeled as the convolution of the optical quadrature with an impulse response plus electronic noise:

$$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$5

The finite impulse response determines temporal resolution, frequency response, and mode selectivity. Electronic noise adds a random quadrature contribution and is often modeled as an equivalent optical loss,

$$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$6

This effective efficiency depends on LO power, detector bandwidth, temporal-mode weighting, and signal regime rather than being solely a fixed circuit property.

For a CW LO, the measured quadrature variance is determined by the detector spectrum and the weighting function. If $$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$7 is the electronic integration function and $$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$8 is the impulse response, the effective optical weighting is

$$\hat n_1-\hat n_2 = |\alpha_{\rm LO}| \left( \hat a_{\rm s}e^{-i\theta} + \hat a_{\rm s}^{\dagger}e^{i\theta} \right).$$9

The mode-matching efficiency is

$$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$0

Consequently, a nominal bandwidth or a single shot-noise clearance does not fully specify tomography performance. The relevant quantity is the electronic-noise spectrum integrated with the Fourier spectrum of the selected temporal mode.

For a pulsed train of period $$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$1, the integrated measurement associated with pulse $$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$2 has the form

$$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$3

where

$$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$4

The temporal-mode efficiency is

$$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$5

If the response extends into neighboring pulse slots, successive quadrature samples become correlated. Finite-bandwidth overlap was explicitly identified as an excess-noise contribution in GMCS-QKD. For a detector bandwidth $$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$6 and repetition rate $$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$7, the modeled input-referred overlap noise is

$$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$8

where $$\hat i_-(t)\propto \eta e|\alpha_{\rm LO}|\hat X_\theta(t),$$9 is Alice’s modulation variance. An adjacent-sample correlation coefficient of $\eta$0 yielded $\eta$1 for $\eta$2 (Chi et al., 2010).

Digital deconvolution can reduce pulse overlap, but it requires accurate calibration of the impulse response and can amplify electronic noise. Digital notch filters can also introduce ringing and long response tails. In the $\eta$3 feedback-free detector, residual correlations over roughly $\eta$4 pulses were attributed primarily to ringing from narrow notch filters rather than to intrinsic detector memory (Altilia et al., 8 Apr 2026).

5. Calibration, noise metrics, and measurement models

Shot-noise calibration is generally performed by blocking the signal and varying LO power. The measured variance is modeled as

$\eta$5

or, in a more detailed pulsed model, as the sum of shot noise, residual LO noise, and electronic noise. In a practical imbalanced detector, the variance may be written as

$\eta$6

where the quadratic term is residual LO-intensity noise, the linear term is shot noise, and the constant term is electronic noise (Chi et al., 2010).

For broadband detectors, the quantum-to-classical noise ratio is commonly defined by

$\eta$7

where $\eta$8 and $\eta$9 denote quantum and electronic noise powers. In the $\hat X_\theta$0 RF-integrated BHD, the frequency-domain QCNR was approximately $\hat X_\theta$1 at $\hat X_\theta$2 LO power (Zhang et al., 2018). A die-level detector achieved $\hat X_\theta$3 integrated QCNR over $\hat X_\theta$4–$\hat X_\theta$5 at $\hat X_\theta$6 LO power (Honz et al., 2021). The SNSPD implementation achieved $\hat X_\theta$7 clearance, but at an effective bandwidth of $\hat X_\theta$8 (Protte et al., 2023).

LO power has an optimum rather than an unlimited benefit. Electronic noise decreases in relative shot-noise units as LO power increases, whereas residual LO-fluctuation noise increases when imbalance converts LO intensity fluctuations into differential noise. A representative model is

$\hat X_\theta$9

The resulting optimum LO level in one GMCS-QKD detector was approximately $\hat a_{\rm LO}$00 photons per pulse (Chi et al., 2010). Excessive LO power can also saturate photodiodes, amplifiers, digitizers, or SNSPDs. In SNSPD BHD, the linear regime ended near $\hat a_{\rm LO}$01 photons per second because detector dead time caused multiphoton arrivals to be registered as single clicks (Protte et al., 2023).

A complete quantum measurement model must specify the observable directly measured by the photodetectors. Ideal multimode descriptions consider either a Glauber photon-number observable,

$\hat a_{\rm LO}$02

or an optical-power observable,

$\hat a_{\rm LO}$03

Both produce the same mean quadrature in the strong-LO, balanced limit. Their finite-LO quantum-noise spectra differ because the power operator contains field-ordering terms and explicit LO-vacuum contributions. The conventional two-photon spectrum includes vacuum fluctuations entering through the main interferometer but normally treats the LO as a classical phase reference. The detector-model-dependent corrections are negligible under realistic strong-LO gravitational-wave conditions (Nakamura, 2021, Nakamura, 2021).

6. Applications and extensions

Quantum-state tomography

BHD directly samples quadrature distributions and supports maximum-likelihood reconstruction of density matrices and Wigner functions. A high-stability ultrafast detector operating at $\hat a_{\rm LO}$04 achieved $\hat a_{\rm LO}$05 SNR, approximately $\hat a_{\rm LO}$06 overall quantum efficiency, $\hat a_{\rm LO}$07 CMRR, and a stability interval of approximately $\hat a_{\rm LO}$08. Its time-bandwidth product was

$\hat a_{\rm LO}$09

The detector reconstructed a coherent state with fidelity $\hat a_{\rm LO}$10 and a heralded single-photon state whose reconstructed Wigner function had $\hat a_{\rm LO}$11 (Cooper et al., 2011).

Continuous-variable quantum key distribution

In GMCS-QKD, Alice modulates amplitude and phase quadratures of weak coherent pulses, while Bob randomly measures one quadrature per pulse. The BHD must therefore resolve individual pulses, suppress LO noise, and maintain a calibrated shot-noise reference. A $\hat a_{\rm LO}$12 detector was designed for $\hat a_{\rm LO}$13 GMCS-QKD and predicted secure key rates of a few megabits per second over a few kilometers, with no secure key beyond approximately $\hat a_{\rm LO}$14 under the selected assumptions (Chi et al., 2010).

The $\hat a_{\rm LO}$15 RF-integrated detector projected secret-key rates of $\hat a_{\rm LO}$16 at $\hat a_{\rm LO}$17 and approximately $\hat a_{\rm LO}$18 at $\hat a_{\rm LO}$19 under finite-size simulation parameters (Zhang et al., 2018). The silicon-photonics TBHD demonstrated positive-key-rate CV-QKD experiments at $\hat a_{\rm LO}$20-dB, $\hat a_{\rm LO}$21-dB, and $\hat a_{\rm LO}$22-km channel conditions, with measured excess noises of approximately $\hat a_{\rm LO}$23, $\hat a_{\rm LO}$24, and $\hat a_{\rm LO}$25, respectively (Jia et al., 2023).

Quantum random-number generation

When one input port contains vacuum, the BHD converts vacuum quadrature fluctuations into an electrical noise signal. A die-level detector produced an estimated extracted random-bit rate of approximately $\hat a_{\rm LO}$26 using a lower-bound min-entropy of $\hat a_{\rm LO}$27 bit per acquired bit, and the generated data passed all $\hat a_{\rm LO}$28 tests of NIST SP800-22-rev1a (Honz et al., 2021). A $\hat a_{\rm LO}$29 detector projected a QRNG rate of $\hat a_{\rm LO}$30 from vacuum-fluctuation measurements, although the result was presented as a potential rate based on an entropy model rather than as a complete cryptographic validation (Zhang et al., 2018).

Gravitational-wave detection

BHD permits arbitrary quadrature readout of an interferometer output. This flexibility is relevant to variational readout, speed meters, and frequency-dependent squeezing. For a LO amplitude $\hat a_{\rm LO}$31 and residual signal-port carrier amplitude $\hat a_{\rm LO}$32, the linearized difference current contains

$\hat a_{\rm LO}$33

The desired signal quadrature is amplified by the LO amplitude, but LO technical noise couples through residual signal-port carrier power. Consequently, LO requirements depend strongly on the unwanted carrier in the signal port rather than solely on LO power. The analysis gives an amplitude-noise condition

$\hat a_{\rm LO}$34

and a differential path-stability condition scaling as

$\hat a_{\rm LO}$35

BHD combined with squeezed-vacuum injection has also been proposed for reducing phase-quadrature noise in Michelson interferometers. In the idealized treatment, the squeezed quadrature variance reaches $\hat a_{\rm LO}$36, while the conjugate quadrature variance increases to $\hat a_{\rm LO}$37 (Ben-Aryeh, 2010). Parametric-amplifier-assisted BHD has been proposed to pre-amplify correlations before detection, reducing the impact of detector inefficiency and temporal-mode mismatch for continuous-variable entanglement measurements (Li et al., 2018).

Other optical measurements

Quadrature-averaged homodyne detection slowly sweeps the LO phase through $\hat a_{\rm LO}$38 and averages the measured fluctuation power over all quadratures. For cavity parameter estimation, this removes the dependence of the transduction function on unknown non-resonant channels and produces a Lorentzian-like detuning dependence, at a sensitivity penalty of at most $\hat a_{\rm LO}$39 relative to the optimal fixed quadrature (Gala et al., 2022).

BHD also extends to nonstandard detector technologies. Click-based BHD with multiplexed on–off detectors defines a bounded nonlinear quadrature,

$\hat a_{\rm LO}$40

which approaches conventional BHD as the number of detector elements $\hat a_{\rm LO}$41 becomes large. Negative normally ordered moment determinants provide experimentally accessible nonclassicality criteria without full state reconstruction (Sperling et al., 2014). SNSPD-based BHD similarly uses time-multiplexed click statistics to measure weak coherent-state quadratures while achieving exceptionally high shot-noise clearance (Protte et al., 2023).

7. Limitations and design trade-offs

BHD performance is governed by coupled optical, electronic, statistical, and security constraints.

Bandwidth versus temporal overlap: Increasing bandwidth permits higher repetition rates, but a detector response that extends into neighboring slots produces correlated quadrature samples and, in CV-QKD, additional input-referred excess noise.

LO power versus technical noise: Higher LO power suppresses relative electronic noise, but residual imbalance converts LO fluctuations into differential noise that grows with LO intensity. Saturation establishes an upper operating limit.

CMRR versus bandwidth: Optical and electrical matching become more difficult at high frequency. A detector can have a large nominal bandwidth but insufficient common-mode rejection across the full measurement band.

Quantum efficiency versus mode matching: Photodiode inefficiency, propagation loss, spatial mismatch, polarization mismatch, and temporal mismatch act as effective loss and inject vacuum. These effects reduce nonclassical signatures and increase inferred input-referred noise.

Stability versus coherent-amplitude preservation: Baseline drift shifts the inferred quadrature origin and is particularly harmful for displaced coherent states. Repetition-rate notch filtering can suppress technical peaks but can also remove information about coherent displacement.

Detector technology versus operating regime: Analog photodiodes provide high bandwidth and linear intensity response; SNSPDs provide low dark noise and high clearance but require cryogenic operation, have dead-time saturation, and operate at lower demonstrated BHD bandwidth.

Security modeling: In CV-QKD, electronic noise may be treated as internal to Bob’s receiver, whereas LO-fluctuation noise and finite-bandwidth pulse overlap may be conservatively assigned to Eve-controllable excess noise. Omitting these contributions can overestimate the secure key rate (Chi et al., 2010).

Quantum measurement specification: The optical diagram alone does not completely define the quantum measurement. The directly measured detector observable, temporal and spatial mode, LO quantum treatment, response function, normalization, and calibration procedure determine the measured signal and noise spectrum (Nakamura, 2021, Nakamura, 2021).

Balanced homodyne detection is therefore not a single fixed detector design but a family of phase-sensitive measurement architectures. Its invariant principle is balanced interference followed by differential detection; its practical behavior depends on the detector response, mode basis, LO stability, optical balance, electronic implementation, and application-specific calibration.