2. Quantum Entropy Evaluation and Enhancing in Vacuum-Based Quantum RNG
Entropy is defined relative to one’s knowledge of an experiment’s output prior to observation. The larger the amount of the entropy, the greater the uncertainty in predicting the value of an observation. Among types of entropy, min-entropy is a very conservative measure. In cryptography, the unpredictability of secret values is essential. The min-entropy measure the probability that a secret is guessed correctly in the first trial. For mathematically determining min-entropy of a secret, the first thing is to precisely identify the distribution that the secret was generated from [
23].
Quadrature fluctuation of optical quantum vacuum state, the nature initial state of optical field at room temperature, is the noise source for the random bits generation in this scheme. According to Born’ s rule, the measurement outcome of a pure quantum state can be intrinsically random. A single measurement of the quadrature of the vacuum state is completely random and multiple repeated measurements satisfy the Gaussian distribution statistically, so we can extract random bits from the measurement results. Based on homodyne measurement, the microscopic fluctuations of quadrature of the vacuum state are detected, amplified and transferred into an electric signal
is the difference current from the two detectors. Measured quantum quadrature of vacuum state in any local phase is amplified by the factor
, which includes the amplification effects from LO gain and electronics gain in the system [
24]. Without regard to classical noise, the electric signals (voltage or current) obey a Gaussian distribution:
The coefficient
has to be calibrated to rescale histogram of the associated marginal distribution in optical homodyne tomography (OHT) [
25]. In this scheme of quantum random numbers generation,
is associated with the quantum entropy contained in the measured data and it is the critical parameter for digitization of the measured analogue signal.
When classical noise is taken into account, such as electronic noise and local noise resulted from imperfect balancing in balanced homodyne detection (BHD), the observed probability distribution of the electric signal is in the form of a convolution of the scaled vacuum state marginal distribution and the classical noise histogram
without loss of generality, the broadband electric noise and the LO noise distribution can be assumed to be Gaussian:
The vacuum noise and the classical noise as two variables with normal distribution, are independent with each other, thus their sum is also normally distributed with a total variance equal to the sum of the two variances.
According to Equations (2)–(4), the homodyne measurement of the vacuum state yields a signal distribution as follows
with the measurement variance of
where factor 2 is added to renormalize the distribution. Then the quantum and classical noise ratio (
QCNR) in the homodyne measurement system is defined as
The
QCNR related to the signal-to-noise ratio of homodyne detection, is defined as the ratio between the mean square noise of the measured vacuum state and the electronic noise, that is, the quantity
or the clearance between the shot noise power spectrum and electronic noise power in dB units,
, reads on spectrum analyzer. In other words, when the homodyne detection system works in linear region, the
QCNR of the raw data can be indicated from the clearance shown by spectrum analyzer.
In our proposal, as a continuous-variable, the measurement output consisting of scaled quadrature of the vacuum state and the classical noise is discretized by an
n-bit ADC with a dynamical range
. The sampled signals are binned over
bins with width of
and are assigned a corresponding bit combination with length of
n. 3-bit ADC binning is shown in
Figure 1a as an example.
In order to design an entropy source that provides an adequate amount of entropy per output bit string, the developer must be able to accurately estimate the amount of entropy that can be provided by sampling its noise source. The behavior of the other components included in the entropy must also be known clearly since the behavior of the other components may affect the assessment of the entropy. In our system, the randomness or the entropy in the measurements could derive from multiple factors, such as the quantum fluctuation, classical influences on it and even malicious attack from the third part [
5]. Especially and strictly, quantum conditional min-entropy is used to evaluate the maximal amount of randomness extractable from the total entropy of the system [
26]. Firstly, the min-entropy for the Gaussian distribution is defined as
In this scheme, the min-entropy of the probability distribution of quadrature measurements can be accurately predicted from the probability density function of the quantum signal. The maximum probability in (9) can be acquired based on the probability distribution discretized by the bins
Each bin is labelled by an integer , with , the least significant bits (LSB) bin, , the most significant bit (MSB) bin and .
Secondly, some restrictions must be taken into account in analog-digital conversion process. Those samples go off-scale, that is, points in saturation will be recorded as extrema values as depicted in
Figure 1b. So, underestimating the range will induce too many blocks of zeros and ones. Conversely, overestimating the signal range will lead to undue unused bins (
Figure 1c). In either situation, some bit combinations are too frequent to be considered random. It is necessary to adjust the amplitude of the analogue signal and the ADC dynamical range in order to employ the full
n-bit sampling properly whenever possible.
Further, considering the influence of classical noise on the measurement outcome, ADC dynamical range should be optimized over the classical noise shifted quantum signal probability distribution. In application scenario, inevitable classical noise excursion in the measurement system will result in nonzero mean in the measured signal probability distribution. On the other hand, eavesdropper may induce a deliberate offset over the sampling period. In a word, a noticeable classical noise excursion, , need to be considered in the optimization of the sampling dynamical range.
Taking into account all these factors offered above, we rewrite the discretized probability distribution as,
where,
is the probability density distribution of the quantum signal given full knowledge of the classical noise
, where
with an excursion of
. Finally, the quantum conditional min-entropy is expressed as
In the best-case scenario of ADC sampling range, the measurement outcome probability in the center bin is equal to the higher one of the first and the last bins. In this way, the quantum conditional min-entropy is information theoretically provably estimated and the amount of quantum-based randomness in the total noise signal is rigorously evaluates. In applications with the requirement of information security, a random sequence is demanded to be truly unpredictable and have maximum entropy [
27].
At the same time, the conditional min-entropy sets the lower bound of extractable randomness from the raw measurements and quantifies the least amount of randomness possessed by each sample or bit per raw bit. Quantum randomness can be distilled from raw data by applying information theoretically provable Toeplitz-hash extractor. As discussed above, the key point is to find out the QCNR and derive the probability distribution of the quantum signal. The higher the QCNR, the more true randomness can be extracted from the raw measurement. Only when QCNR is high enough, both the quality and the security of the random number generator are guaranteed. Fulfilling the condition of optimal dynamical sampling range R, minimum-entropy of the quantum signal for growing clearance is theoretically analyzed. Proceeding from the directly measurable quantity, homodyne clearance, corresponding QCNR is derived from Equation (8). Then quantum noise variances are expressed as multiples of the . For different clearance, probabilities of middle bin and the LSB/MSB are compared and the optimal sampling range R is decided based on Equation (11). Finally, based on Equation (13), the quantum conditional min-entropy in optimal sampling range scenario as a function of different classical noise excursion is analyzed.
The classical noise excursions in our raw data have been collected from multiple measurements, which range from almost 3 to 29 times of classical noise standard deviation
. In application scenario, much larger DC offset may be induced deliberately by the eavesdropper. In
Figure 2, we show the quantum conditional min-entropy,
, as a function of homodyne detection clearance for three different classical noise excursions under the precondition of optimal sampling range.
is the smallest classical noise excursion among our multiple measurements,
, a larger classical noise excursion for comparison and
is the excursion in the raw data from which we extract true random numbers. As shown in
Figure 2, the extractable random bits are robust against the decline of
QCNR while the classical excursion is subtle. Whereas if classical noise excursion is evidence, one can achieve high secure randomness only when clearance is high enough.
The clearance relies on the total gain in homodyne detection system (also
in Equation (1)), including the LO amplification and the electrical gain. In quantum state measurements and reconstructions, the clearance needed between shot noise and classical noise is dependent on the amount of squeezing and entanglement one wishes to measure. Empirically, the homodyne system should satisfy the condition that the measured shot noise is 10 dB higher than the classical noise among the analysis frequency range [
28,
29]. High TIA gain and moderate dynamical range are required so that shot noise is the dominant spectral feature among the detection frequency range. In this scheme of quantum RNG, however, high
QCNR, but also large detection bandwidths, are pursued, since the cut-off frequency of the homodyne detector upper bound the sampling frequency in random numbers generation process [
30].
On the other hand, the classical effects, which blur the distribution and cause classical entropy in the raw bit sequence, include imperfect balancing of LO, non-unit quantum efficiency and electronic noise of the detectors [
31,
32,
33,
34]. The non-unit detector efficiency can almost completely overcome by using special fabricated diodes and the quantum efficiencies of more than 99% have been reported [
35]. The detrimental electronic noise depends on numerous components in the circuit part as expressed by
One term is from the photodiode (PD) and comprise of thermal noise and dark current noise of PD, both of which are usually negligible thanks to its big shunt resistance
and low dark current
[
36]. The other term is from the TIA circuit including thermal noise
, input noise current
and input noise voltage
of the operational amplifier. The electrical gain of TIA amplifies quantum fluctuations as well as the electronic fluctuations, so the electronic noise included in the homodyne raw measurements comes mainly from the amplified TIA circuit noise. LO effectively acts as a noise-less amplifier for the quantum fluctuations of the vacuum state and the electrical noise is independent of the LO. In fact, the optical fluctuations seen by the detector can be made much larger than the electronic fluctuations by increasing the laser intensity of LO beam to enhance the
QCNR signally [
37].
At the same time, the gain of a typical op-amp is inversely proportional to frequency and characterized by its gain–bandwidth product (GBWP). As a trade-off, lower electrical gain put up with higher op-amp bandwidth. In fact, theoretically, vacuum quadrature fluctuates with unlimited bandwidth in the frequency domain. The random number generation rate in this scheme is ultimately limited by the bandwidth of the homodyne detector. Increased bandwidth of op-amp allows higher sampling rate.
3. Experiment and Results
Experimentally, we dedicate to enhance quantum entropy in quantum RNG by enhancing the laser power of LO beam to noise-independently amplify quadrature fluctuation of vacuum state on the premise of optimizing ADC sampling range. An extraction ratio of true randomness of 85.3% is achieved by finite enhancement of the LO power when classical noise excursions of the raw data is obvious and the extracted random sequences passed the NIST (National Institute of Standards and Technology), Diehard and the TestU01 tests.
The experimental setup is depicted in
Figure 3. A 1550 nm laser diode (LD) is driven by constant current with thermoelectric temperature control with a maximal out power of 15 mW. A half-wave plate and a polarizing beamsplitter (PBS2) were combined to serve as accurate 50/50 beamsplitting. Single-mode continuous-wave laser beam from the laser incident into one port of the beamsplitter and acts as the LO, while the other port was blocked to ensure that only the vacuum state could enter in. The vacuum field and the LO interfere on the symmetric beamsplitter to form two output beams with balanced power. The outputs are simultaneously detected by balanced homodyne detector (PDB480C, Thorlabs Inc., Newton, MA, USA) to cancel the excess noise in LO while amplify the quadrature amplitude of the vacuum state, which fluctuates randomly and is independent of any external physical quantities.
Classical noise in the photocurrents is rejected effectively over the whole detection band while the clearance has dependence on frequency as shown in
Figure 4. We filtered out a part of the vacuum spectrum, where the clearance is almost consistent, to extract true randomness based on a certain quantum conditional min-entropy and analyze the effect of LO intensity on the conditional min-entropy. The shot noise limited signal from the homodyne detector is mixed down with a 200 MHz carrier (HP8648A) and then passes through a low-pass-filter (LPF) with 50 MHz cut-off frequency (BLP50+, Mini-Circuits Corp., Brooklyn, NY, USA), that is, we actually use 100 MHz vacuum sideband frequency spectrum centered at 200 MHz to act as the random noise resource.
In OHT, BHD system is established and locked to every relative phase to measure the marginal distributions of electromagnetic field quadrature for completely reconstruction of quantum states [
25]. While the random numbers generation scheme discussed here focus on a marginal distribution of vacuum state in any one phase thanks to the space rotational invariance of its distributions in the phase space, that is no active modulation or phase (or polarization) stabilization is required.
We present the
QCNR as a function of the LO power arriving at the PD. The electrical noise variance is relatively consistent for certain TIA gain. The clearance depends only on the LO power. The noise power is given by
where
is the electron charge,
is the quantum efficiency of the photodiode (Hamamatsu G8376),
the resolution bandwidth,
the transimpedance gain of the photo detector and
the load impedance [
38]. For each power value the distribution of the random data was analyzed in time domain in the form of histogram to calculate the
QCNR.
QCNR as a function of the LO power figured out from the measured clearance levels is plotted with open circles in
Figure 5. The LO power received by each PD is gradually increased from 300 μW to 6 mW by rotating the HWP before PBS1. Here we interpolate between the experimental points to obtain the dependence of
QCNR on LO power. It is shown as the black dashed line in
Figure 5. The experimental results are given by red open circles and can be fitted well by the theoretical curve with a transimpedance gain of
. The experimental results are about 2 dB lower than the theoretically excepted
QCNR, which is due to uncertainties in determining the transimpedance of the detector and the transmission losses in the LPF.
We increase the LO power up to 6 mW to achieve the largest
QCNR of 17.8 dB in our system, limited by the maximal output power of the laser. The signal is sampled with a rate of 100 MHz, upper limit of twice the LPF band for the sampling rate to avoid temporal correlation between samples. The resolution is 12 bits and the dynamical range is optimized according to the histogram of the time series acquired with reasonably larger sampling range. The amplitude acquisition scale of oscilloscope (SDA806Zi-A, LeCroy, New York, NY, USA) is continuously adjustable. By choosing the analog-digital conversion range appropriately and tuning the LO intensity finely, the amount of off-scale points can be controlled within allowed statistical deviation. The number of saturated points is easy to restrain on-line from the oscilloscope. The distributions of the random data in time domain and in histogram are shown as insets of
Figure 5. The measured total variance of the raw data and electrical noise variance are
and
, respectively. The classical noise excursions of the raw data are about 17.2 times of the classical noise standard deviation
. Then the probability distribution of the quantum signal is derived and the conditional min-entropy in the quantum signal is worked out to be 10.13 bit per sample, as circled in red in
Figure 2.
Finally, information-theoretically provable post-processing scheme, Toeplitz-hashing extractor, is constructed on an FPGA to extract true randomness from the raw data and uniform the Gaussian biased binary stream [
39]. A binary Toeplitz matrix of
is constructed with a seed of
random bits (the seed can be reused since the Toeplitz-hashing extractor is a strong extractor).
m final random bits are extracted by multiplying the matrix and
n raw bits, where
and
. We employ
Toeplitz Hash extractor to distil 10.13 bits/sample. The extraction ratio of 85.3% is the highest as ever reported. We recorded the data with the size of 1 G bits to undergo random test. 1000 sequences with each one 1 M bits are applied to the NIST test and significant level is set as
. The NIST test is successful if final P-values of all sequences are larger than
with a proportion within the range of
for 15 test suits [
40]. P-value shown in the
Figure 6 are the worst cases of our test outcomes.
Results of the Diehard statistical test suite for the same data file is shown in
Figure 7. Kolmogorov-Smirnov (KS) test is used to obtain a final
p-value to measure the uniformity of the multiple
p-values. The test is considered successful if all the final
p-values lies in the range from 0.01 to 0.99 [
41].
Constrained by the computational power of crush of TestU01, small crush test is performed with a data size of 8 G bits [
42]. The random numbers can pass all the statistical tests successfully. The
p-value from a failing test converges to 0 or 1. Where the test has multiple
p-values, the worst case is tabled in
Figure 8. All the test items are passed successfully.
On the other hand, we reduce the LO power in the homodyne system to 400 μW and correspondingly, the clearance declines to 4.06 dB. The time series of the system outcomes are collected and statistically analyzed. Classical noise excursion in the Gaussian distribution is about 19.3 times of the classical noise standard deviation. Based on theoretical calculation, the min-entropy is worked out to be 7.73 bits/sample. The hash extraction results with maximum extraction ratio of 0.63 can pass the NIST, Diehard and TestU01 tests finally.