See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/298636581

Quantum random flip-flop and its applications in random frequency

synthesis and true random number generation

Article  in  Review of Scientific Instruments · March 2016

DOI: 10.1063/1.4943668

CITATIONS READS
5 665

1 author:

Mario Stipčević
Ruđer Bošković Institute
209 PUBLICATIONS   4,310 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

HOLOGRAPHY AND INTERFEROMETRY UNDER WEAK ILLUMINATION View project

All content following this page was uploaded by Mario Stipčević on 03 December 2017.

The user has requested enhancement of the downloaded file.

 REVIEW OF SCIENTIFIC INSTRUMENTS 87, 035113 (2016)

Quantum random flip-flop and its applications in random frequency

synthesis and true random number generation
Mario Stipčevića)
Photonics and Quantum Optics Research Unit, Center of Excellence for Advanced Materials and Sensing
Devices, Rud̄er Bošković Institute, Bijenička 54, 10000 Zagreb, Croatia
(Received 22 January 2016; accepted 27 February 2016; published online 16 March 2016)
In this work, a new type of elementary logic circuit, named random flip-flop (RFF), is proposed,
experimentally realized, and studied. Unlike conventional Boolean logic circuits whose action is
deterministic and highly reproducible, the action of a RFF is intentionally made maximally unpre-
dictable and, in the proposed realization, derived from a fundamentally random process of emission
and detection of light quanta. We demonstrate novel applications of RFF in randomness preserving
frequency division, random frequency synthesis, and random number generation. Possible usages of
these applications in the information and communication technology, cryptographic hardware, and
testing equipment are discussed. C 2016 AIP Publishing LLC. [http://dx.doi.org/10.1063/1.4943668]

I. INTRODUCTION assume that a flip-flop performs its action when clock pulse
(CP) input changes state from LOW (logic 0) to HIGH (logic
Today life is unimaginable without microprocessors that
1) that is upon the “rising edge” of the input pulse. In this
are in the heart of computers, cell phones, and other devices
study, we will predominantly use D-type flip flop (DFF).
which depend on the information processing. Boolean logic
Conventional edge-triggered DFF, whose symbol is shown in
circuits are the basis technology for all contemporary micro-
Fig. 1(a), is an important building element of logic circuits
processors, computers, mobile phones, communication equip-
usually used as a one bit data storage, frequency divider-by-
ment, and many other information and communication tech-
two, or a digital counter. With each occurrence of the rising
nology (ICT) devices. Logic circuits are divided into non-
edge at the CP input, logic level present at the data input D
sequential (e.g., , , and ) whose output state only
is copied to the output Q and stays memorized there until
depends on the current state(s) of input(s), and the sequential
the new clock pulse. This operation is described exactly by
(e.g., flip-flop) whose output state depends on both current
truth Table I (3rd column), and illustrated in Fig. 1(b). DFF
and past states of the input(s). Their purpose in ICT applica-
may optionally feature the inverted output Q̄. It can also have
tions and equipment is information processing and, depending
either or both Set (S) and Reset (R) inputs which, upon being
on how data are interpreted by humans, they appear to be
brought to HIGH unconditionally (asynchronously and irre-
performing a variety of tasks like mathematical calculations,
spective of the CP input) set (to HIGH) or reset (to LOW) the
playing music and video, and word processing. Logic circuits
output Q. DFFs and other logic circuits are specifically de-
can also be used for the purpose of analog and digital signal
signed to behave deterministically, with a very small proba-
processing: frequency division, multiplication or synthesis,
bility of error in order to ensure stable and predictable opera-
clock restoration, signal delaying, measuring of relative phase
tion of ICT equipment.
shift, detecting coincident events, etc. Boolean logic circuits
There are a few technical details of marginal importance
are explicitly made to behave deterministically, that is, to give
for this study which, for simplicity, we do not take into consid-
the same output if fed by the same sequence of input signals.
eration: (1) setting time (t SET): D input has to hold a vale for
This is their great advantage but also a limitation.
some time before the clock; (2) propagation time (t CPQ): it is
Here we propose a new elementary logic building block
a finite time delay between the clock input CP and setting of
that behaves non-deterministically, the random flip-flop (RFF),
the correct value at the output Q; (3) (re)set time (t RSQ): it is
and discuss a few of its applications. We show that RFF is
the time delay between (reset)set input and the output Q. In the
a convenient way of “packaging” of randomness in a circuit
experimental technique that we will use, all these time delays
that can be mixed with conventional logic circuits in order to
are on the order of 1 ns and of little or no importance for the
arrive to a rich set of properties and applications that cannot
effects under study as long as they are constant or small with
be achieved by Boolean logic alone.
respect to the average clock rate.
By definition, the D-type random flip-flop (DRFF) (sym-
II. DEFINITION OF THE RANDOM FLIP-FLOP bol shown in Fig. 1(c)) operates similar to DFF with only
difference that, upon the clock pulse, state of the data input
In this study, we only consider the edge-triggered master- D is transferred randomly to the output Q with probability of
slave flip-flops.1 Henceforth, without loss of generality, we ½, which ensures maximal unpredictability. This is described
by truth Table I (4th column) and illustrated in Fig. 1(d).
a)Author to whom correspondence should be addressed. Electronic mail: Similarly, we define other types of RFF such as T-type
mario.stipcevic@irb.hr random flip-flop (TRFF). While Boolean T-type flip-flop (TFF)

0034-6748/2016/87(3)/035113/9/$30.00 87, 035113-1 © 2016 AIP Publishing LLC

Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
 035113-2 Mario Stipčević Rev. Sci. Instrum. 87, 035113 (2016)

FIG. 2. The RFFs can emulate each other in exactly the same way as their
deterministic counterparts do.

III. PRACTICAL REALIZATION OF AN OPTICAL

QUANTUM RANDOM FLIP-FLOP
In this study, all experiments were performed using logic
gates and flip-flops available in the Altera MAX7000 family
reconfigurable Programmable Logic Chip (PLD). The flip-
flops therein conform with the above description and truth
tables. The setting time t SET is about 1 ns, both propagation
delays t CPQ and t RSQ are about 2 ns, and maximum counting
frequency exceeds 250 MHz. Output rise and fall times are
below 1 ns.
Before addressing practical realization of RFF, let us
FIG. 1. Examples of input/output waveforms for (a) and (b) the conventional
D-type flip-flop (DFF); (c) and (d) D-type random flip-flop (DRFF). introduce notion of random pulse train (RPT), an important
concept in this presentation. By definition, RPT is a sequence
of electrical logic pulses of constant amplitude, typically but
TABLE I. The truth tables for deterministic D-type flip-flop (DFF) and
D-type random flip-flop (DRFF). RND means that upon the clock, the cor-
not necessarily of constant duration, wherein time intervals t i
responding output is set to randomly chosen value 0 or 1 regardless of the between subsequent pulses (also known as “waiting times”)
previous states. follow the exponential probability density function. It is best
to picture a RPT as being a result of an underlying stationary
DFF DRFF and memoryless physical process generating random events
D Qn Qn+1 Qn+1 and that each event generates one logic electrical pulse whose
0 0 0 0 leading (i.e., positive-going) edge corresponds to the time of
0 1 0 RND the event. A RPT, illustrated in Fig. 3, is characterized by a
1 0 1 RND single parameter: the average pulse rate henceforth referred to
1 1 1 1 as “random frequency” or just “frequency” and denoted by f .
Quite generally, we can define the average frequency f of
a digital signal as
TABLE II. The truth tables for deterministic T-type flip-flop (TFF) and
1
T-type random flip-flop (TRFF). RND means that upon the clock, the corre- f = lim n (1)
n→ ∞ 1
i=1 t i
sponding output is set to randomly chosen value 0 or 1 regardless of previous
n
states.
which is as an average number of rising edges appearing in a
TFF TRFF very long time interval divided by the length of the interval.
This definition holds quite generally: for periodic and non-
T Qn Qn+1 Qn+1
0 0 0 0
periodic signals, including RPTs. Of course, for periodic sig-
0 1 1 1 nals, it is enough to take n = 1 because all intervals t i are equal
1 0 1 RND and the averaging over n periods is unnecessary.
1 1 0 RND A RPT that precisely obeys the exponential probability
distribution function (p.d.f.) is important for realization of the
RFF and in some other applications. A particularly precise
RPT can be realized by use of quantum randomness. Of all
toggles output state with each clock if the input T is held HIGH, possible sources of randomness, those relying on quantum
TRFF toggles with probability of ½. In both TFF and TRFF, processes of photon emission and detection are probably the
clock input is disabled while T = 0, as exactly described in
Table II. Any of those RFFs (deterministic or random) may
optionally feature the inverted output Q̄ and Set (S) and Reset
(R) inputs which unconditionally set (to HIGH) or reset (to
LOW) the output Q.
Interestingly, the random flip-flops can emulate each
other in exactly the same way as their deterministic counter-
FIG. 3. A random pulse train (RPT) is a sequence logic pulse usually
parts, as shown in Fig. 2, but as we shall see, the analogy of constant duration. By definition, pulse waiting times t i (i = 1, 2, 3, . . .)
between random and deterministic FFs generally does not between rising edges of subsequent pulses follow exponential probability
exist. density function.
Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
 035113-3 Mario Stipčević Rev. Sci. Instrum. 87, 035113 (2016)

FIG. 4. An optical random pulse train generator (ORPTG) consists of

InGaAs red light emitting diode (LED) operated in constant-power mode
that shines light upon a single-photon detector (SPD) through a dispersive
attenuating filter (DF). The light power, and thus the mean pulse rate, can be
adjusted via potentiometer P1.

best understood from the first principles and offer theoretically

perfect randomness.
In this study, we use a stationary optical random pulse
train generator (ORPTG), similar to those described in Refs. 2
and 3, shown in Fig. 4. It consists of InGaAs red light emit- FIG. 5. Histogram of time intervals between subsequent pulses (waiting
times) generated by the optical random pulse train generator (ORPTG) at
ting diode (LED) which shines light attenuated by a disper- a frequency of 1 Mcps.
sive neutral density filter (NDF) upon a single-photon detec-
tor (SPD). Upon each successful detection of a photon, SPD
generates a logic pulse and in that way a train of pulses which 2. the action probability of RFF should be as close as possible
appear randomly in time is generated. to ½, for applications in this work within ±10−4; and
The LED (Hamamatsu L7868, λ = 670 nm) is operated in 3. the propagation delay between clock pulse and the
a constant-current mode. Such a source is well-known to yield output(s) should either be negligible with respect to the
photons emitted at random times due to quantum effects.2,4 average clock rate or be constant, in order to allow for
Filter NDF is conveniently chosen such that the resulting rate synchronization of operation of several RFFs. This is
of photon detections falls in the desired range. Light power important in some applications like the random frequency
can be fine-tuned by potentiometer P1. The detector SPD is synthesis (Sec. VI) and the Bell test.7
home-made photon counting module based on a Geiger-mode It is not known to us whether all three conditions can be
avalanche photodiode (Laser Components, SAP500) and the met exactly, even in theory. However, the three conditions can
active quenching circuit described in Ref. 5. Upon each detec- be met asymptotically by the use of the RPT.
tion, SPD generates a transistor-transistor logic (TTL) pulse D-type RFF can now be constructed as shown in Fig. 6.
of duration τp = 12 ns. No two output pulses can come closer Under the control of an ORPTG, the output Q of the T-type
than the dead time of the detector, τdead = 24 ns. Maximum flip-flop FF1 toggles upon each received pulse. The ORPGT
photon detection rate of random light is over 38 Mcps. Given is made such that its average frequency (in the sense of
the spectral width of the emitted light, ∆λ ≈ 20 nm, one can Eq. (1)) stays unchanged over time. Because of that, FF1
estimate the coherence time τcohr = λ 2/(c∆λ) ≈ 75 fs. As long spends exactly equal time in HIGH and LOW states while
as τcohr is much smaller than the smallest period between changes happen at random (unpredictable) times. If a logic
successive photon detections, waiting times between succes- pulse appears at the global CP input, FF2 captures momentary
sive photon detections obey exponential probability density
function. Main correlations among events come from imper-
fections of the SPD notably the afterpulsing and the dead
time. Afterpulsing of SAP500 is relatively high (3%-5%) at
“usual” operating conditions that are optimized for low noise
and high detection efficiency (temperature in the range −25
to −10 ◦C, overvoltage 15 V).6 In order to improve on after-
pulsing, we operated SAP500 at an unusually high junction
temperature of +35 ◦C and overvoltage of 9 V which resulted
in the afterpulsing probability of only 0.28% and an estimated
afterpulsing lifetime of about 10.5 ns. The near-exponential
empirical distribution of photon waiting times of the ORPTG
built with this detector is shown in Fig. 5. The exponential
p.d.f. hypothesis probed by the chi-square test, using statistics
of 106 recorded intervals, is accepted with probability greater
than 0.999 999 confirming that the measured distribution con-
forms with what one would expect from a random source.
There are three important considerations when conceiving
a practical RFF, which are as follows: FIG. 6. Realization of the D-type random flip-flop by using an optical quan-
tum random pulse train generator (ORPTG) made of (1) an LED light source
1. the action of a RFF should be truly random and indepen- (LED); (2) neutral density filter (NDF); and (3) a single-photon detector
dent of all other RFFs used in the same circuit; (SPD).
Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
 035113-4 Mario Stipčević Rev. Sci. Instrum. 87, 035113 (2016)

logic state of the FF1 by its input D. If D is LOW, nothing

happens. However, if D is HIGH, then output Q of the FF2
goes HIGH and then back to LOW after a short period of
time due to the feedback between its output Q and reset input
R, generating a short pulse of duration equal to the delay
through buffer DLY. We assume that duration of this pulse
is shorter than the time to the next global clock pulse. Since
probability of finding FF1 in HIGH state is exactly equal to ½,
FF2 generates the pulse with probability of ½. Thus FF3, and
in fact the whole DRFF, will act as an ordinary DFF but with
clock action probability of exactly ½ which, by definition, is
the action of a DRFF. As explained in Ref. 3, any short-lived
correlations among pulses of the ORPTG as well as theoretical
imperfections of this method will have a vanishingly small
effect on randomness of the sampled state of FF1 as long FIG. 7. Basic dividers with RFFs: (a) random divider by 2; (b) random
divider by 4; (c) divider which performs random division by 2 followed by
as the sampling period is sufficiently long with respect to
deterministic division by 2 which has the same functionality as (d).
lifetimes of imperfections (dead time, afterpulsing, etc.) and
the average toggling period of the FF1. In order to ensure good output Q. The subscript “R” denotes “random division” action.
randomness at sampling rates of up to 3 MHz, the frequency Random division preserves exponential distribution of pulse
of the ORPTG (i.e., the photon detection frequency) was set waiting times. Stacking two DRFFs as shown in Fig. 7(b), one
to about 24 Mcps. We emphasize that the action probability achieves division by 4R, etc. Note that these particular dividers
of the RFF does not depend on the frequency of the QRPTG always produce a sequence of short pulses that is an RPT.
which is one of the strengths of this design. Consequently its A surprising possibility of dividing frequency by 4 with a
choice is not critical. An alternative approach would be to use single DRFF is shown in Fig. 7(c). This division is composed
a fast-response random bit generator described in Ref. 7. of first dividing by 2 randomly and then by 2 deterministically,
The construction laid out in Fig. 6 presents the basic which we denote as 1/2R/2D, where subscript “D” stands for
DRFF used in this work. All logic circuits required for the “deterministic.” Result of such a division is a signal with
RFF as well as data acquisition are made by programming variable pulse width and bell-shaped waiting time p.d.f. shown
a single Altera MAX7000 family reconfigurable PLD chip in Fig. 8. Note that random and deterministic divisions do
complemented with a Cypress CY7C68013 communication not commute: 1/2R/2D , 1/2D/2R. This circuit is functionally
chip for transfer of data to the PC computer via USB2 link. equivalent to the circuit in Fig. 7(d).
In comparison to the random division, deterministic divi-
IV. RANDOMNESS PRESERVING sion of a RPT by n results in a pulse train with frequency of f /n
FREQUENCY DIVIDERS too; however its pulse waiting times get modified and follow
Erlang distribution with the shape parameter n,
In this section, we are concerned with frequency division
of RPTs. f nt n−1e− f t
Erlang (t, f , n) = (2)
Dividing frequency of periodic signals, by an integer or Γ(n)
fractional number, can be performed by various well known which for large n tends to the normal distribution due to
techniques.8–10 For example, a single TFF will divide by two the central limit theorem. Note that as n goes to infinity, the
frequency of its clock signal while division by n can be made
by a counter-to-n. If a RPT is divided by such a divider, the re-
sulting signal will have a correct frequency but the distribution
of waiting times will not be exponential anymore and will be
less random. In fact, the direct division of frequency of RPTs
in such a way that the resulting signal is again a RPT was not
feasible so far. Now, this can be made by “random frequency
dividers” made with random flip-flops. A few simple random
frequency dividers are shown in Fig. 7.
A random frequency divider by 2 is shown in Fig. 7(a).
For each clock pulse, the DRFF decides randomly whether to
set output Q to HIGH. If it does, the bootstrap between Q and
R resets Q to LOW essentially after a propagation time of the
non-inverting delay buffer DLY, thus generating a short pulse.
Since the random decisions are not correlated to input pulses,
the output pulses are memoryless (exponentially distributed)
but only a half of them will yield an output pulse. In another FIG. 8. Experimentally obtained waiting time distributions of a RPT (expo-
words, if a RPT of random frequency f is brought to the nential) and of a RPT divided by factors: 2D (Erlang p.d.f.), 2D2R, 2D4R, and
input CP, a RPT of random frequency f /2R will appear at the 2D16R.
Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
 035113-5 Mario Stipčević Rev. Sci. Instrum. 87, 035113 (2016)

FIG. 9. Deterministic divider by 2 (top) and deterministic multiplier by

two (bottom). The input and output waveforms are shown in the middle.
The output pulse width of the multiplier equals to propagation delay of the
non-inverting buffer.

divided signal becomes increasingly periodic and eventually FIG. 10. Synchronous 4-bit random counter. With each clock pulse, the
its randomness vanishes. Thus, the deterministic division counter randomly advances up similar to random walker on a line.
drives a random input signal away from randomness. Interest-
ingly, by virtue of the Poisson limit theorem (see, e.g., Ref. 11),
random division brings a pulse train obeying any initial time Figure 10 shows the well-known synchronous counter to-
interval distribution asymptotically towards exponential distri- pology, however, realized with TRFFs instead of TFFs. In our
bution that is towards maximal randomness. experiment, it has been built using four independent TRFFs,
This is illustrated in Fig. 8 by experimentally measured each realized by an independent DRFF, built according to
data, where an initial Erlang distributed pulse train is repeat- Fig. 6, via the emulation circuit shown in Fig 2. Both the deter-
edly divided by 2R until it almost resembles exponential p.d.f. ministic counter (made with TFFs) and the random counter
In that respect, random frequency division and synthesis of (made with TRFFs) feature 16 possible (output) states which
RPTs (Section VI) are stable, self-healing processes which can be represented by a binary number {Q3,Q2,Q1,Q0} whose
tend to eliminate effects of imperfections in hardware. value is in the range from 0 to 15. The deterministic version
Note that in the special case of dividing a RPT by 2D (for of this counter advances by one upon each (rising edge of the)
example, by means of a deterministic flip-flop, as shown in clock pulse, that is, it counts from 0 to 15, then falls back to
Fig. 9), no information is lost: every edge (rising or falling) 0, etc. Thus the counter can only advance in steps of 1 or −15.
corresponds to one pulse from the original RPT, and therefore, Starting from 0, the counter advances by 1 in 15 consecutive
it is possible to restore the original RPT form the divided one. steps and then advances by −15 in a single step, thus exhibiting
The restoration can be done by a deterministic multiplier an average advance of 0 in the long run.
by 2D, such as the one shown in Fig. 9, which upon every edge Regarding counting, unlike its deterministic counterpart,
(rising or falling) generates one short pulse of width equal to upon a clock pulse random counter can go up, down, or stay
the propagation delay of the non-inverting buffer (DLY). On unchanged, that is, the output numbers appear “random.” On
the contrary, when a RPT is randomly divided by 2R, then half a long run, probabilities of all 16 possible output states (num-
of the information (every second pulse, on average) is lost and bers) are equal for both deterministic and random counters.
there is no way to recover the original RPT. The difference is in the variety of possible step sizes by
Random frequency dividers are the basis for the random which the counter advances and their probabilities. Possible
counters and random frequency synthesis that are described in advancement steps are in the range from −15 to +8 since
Secs. V and VI. advances greater than 8 (the most significant bit) are impos-
sible. Namely in order to achieve step of 8 or greater, it is
necessary that Q3 goes from LOW to HIGH. But to enable this,
V. RANDOM COUNTER
considering the three  gates, we see that all other outputs
Counters are important building blocks, made of flip- must be HIGH. Therefore, the step of 8 is the greatest possible.
flops, that can be used for counting events or for frequency A histogram of probabilities of advancement step upon a clock
division. To illustrate the complexity of circuits built with pulse of the random counter, obtained by computer simulation,
RFFs in this section, we analyze the “random synchronous is shown in Fig. 11. In order to determine these probabilities
counter,” defined as a synchronous up-counter in which all flip- theoretically, one would need to determine probabilities of
flops are replaced by their random counterparts. 24 possible steps for each of the 16 possible states, in total
A few of its characteristics are investigated: counting, M = 384 values. The distribution is skewed and grouping of
random number generation, and frequency division. values in clusters of 1, 2, 4, and 8 steps is apparent.
Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
 035113-6 Mario Stipčević Rev. Sci. Instrum. 87, 035113 (2016)

FIG. 11. Probabilities of step sizes of the 4-bit synchronous random counter. FIG. 12. Autocorrelation coefficients with lag 1-36 of 4-bit numbers gener-
ated by 3 devices: random counter (blue triangles), periodic up counter (red
round dots), and random number generator (black squares).
By far the most probable step sizes are 0 (no change)
and 1, each with roughly equal probabilities. This could be ministically by 2 (in our notation: f (Q i ) = f in/(2i+1) R/2 D ).
expected from a forward counter topology constructed such Unfortunately, RPTs appearing at outputs Q0 − Q3 are strongly
that the counter mostly advances by 1, where deterministic correlated: any RPT contains all pulses from any lower fre-
flip-flops are replaced by random ones which have the prob- quency RPT; therefore, they cannot be combined to obtain
ability of action equal to ½; thus, the most probable steps are RPTs of frequencies other than those. Thus only divisions by
indeed 0 and 1. However, interestingly, the average step size 2i are possible.
is 0, as is for the deterministic counter. The number of prob- In this example of a random counter, we see that behavior
abilities (M) that need to be determined in order to calculate of circuits containing only a few RFFs may exhibit a very
probabilities of steps rises faster than exponentially with the rich behavior. In particular, the random synchronous counter
number of flip-flops (N): M = (2 N )2 + 22N −1, indicating high counts erratically up and down can be used for the genera-
complexity of this seemingly simple circuit. tion of correlated random numbers and for the randomness-
Regarding random number generation, state of the random preserving frequency division by a power of two. However,
counter changes unpredictably in such a manner that the 16 frequency division by an integer number n is yet unknown.
possible outcomes are distributed uniformly, but numbers so Instead, random frequency synthesis can be used.
generated are strongly correlated and therefore not completely
random. Namely, when considered individually, each of the
outputs Q i (taken in synchronization with the Strobe output) VI. RANDOM FREQUENCY SYNTHESIS
yields a random bit sequence with zero bias but with a non- Frequency synthesis of periodic signals can be accom-
zero serial autocorrelation coefficients of lag k, defined, for plished by well known techniques, for example, phase-locked
example, in Ref. 12 loop (PLL) or direct digital synthesis (DDS). These enable
 N −k generation of a set of frequencies separated by an arbitrary
(x i − x̄) (x i+k − x̄)
ak = i=1 N −k 2
. (3) small, constant increment in a desired frequency interval.
i=1 (x i − x̄) Using RFF, one is, for the first time, able to build a full-
Experimentally determined, autocorrelation coefficients featured random frequency synthesizer (RFS) with digitally
obey geometric law with respect to k: ak (Q i ) = (1 − 2i )k controlled rate. In this example, the output signal is a RPT
indicating Markov process. Indeed, one expects Markovian even though this is not the only possibility. A RSF that makes
behavior since sequentiality of RFF dictates that the next state possible synthesis of RPTs having frequency in an arbitrary
of the counter depends solely (albeit non-deterministically) interval and with arbitrary small steps is shown in Fig. 13. A
on the previous state. Of course, autocorrelation among indi- random pulse splitter (RPS) conveys a pulse from the input IN
vidual bits causes autocorrelation of the sequence of 4-bit randomly to one and only one of its two outputs, O1 or O2.
binary numbers {Q3,Q2,Q1,Q0} generated by the random Pulses of frequency f 1 from the input are randomly picked up
counter. Triangular dots (blue) in Fig. 12 represent first 36 and directed towards switches Si with mutually independent
autocorrelation coefficients (1 ≤ k ≤ 36) experimentally ob- probabilities of 1/2i . Depending on the setting of switches,
tained by the random counter (statistics 3 · 108 numbers, error none, some, or all pulses from the input reach the input of the
are bars smaller than the dot size) whereas round dots (red) 1 gate which sums their frequencies by simply interleaving
correspond to the deterministic counter and square dots (black) the incoming pulses.
correspond to completely random numbers. Of course, stream Frequency of the resulting RPT is equal to ( 2kN ) f 1, where
of random values does not show any autocorrelation. k can take on any integer value in the range 0 ≤ k ≤ 2 N .
Regarding usability of the random counter for frequency Parameter k depends on setting of the switches,
division, we note that if the input is fed by a RPT, each output N
Q i gives another RPT. The frequency of ith output (Q i ) is equal

k = SN +1 + 2 N −i Si . (4)
to the input frequency divided randomly by 2i+1 and then deter- i=1
Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
 035113-7 Mario Stipčević Rev. Sci. Instrum. 87, 035113 (2016)

frequency of an adjustable RPT generator.15,16 The minimum

time required to stabilize loop to a given relative precision rises
quadratically with the mean period of the RPT. For example,
to stabilize a variable RPTG to 1 Hz within precision 1%, it
would take at least 10 000 s because it takes that long time to
measure the random frequency of 1 Hz to 1% precision. To
stabilize random frequency of 0.1 Hz, it would take 100 times
longer, etc. On the other hand, governed by the same law, high
random frequency can be stabilized much faster. For example,
10 MHz RPTG can be stabilized to 1% in only ∼0.001 s. Then,
dividing such signal with a randomness preserving divider,
one can quickly obtain lower frequency with the same relative
precision regardless of the division factor.
FIG. 13. A full-featured random frequency synthesizer in the range [ f 0, f 0 + Frequency synthesis is useful in a wide variety of applica-
f 1] in steps of f 1/2 N . Output frequency f out can be set via switches S1 − S5. tions from telecommunications to laboratory equipment. RTPs
of precisely known frequency could find application in spread
Switch setting Si refers to ith switch and can take on either spectrum communication, test equipment, signal generators,
value of 0 (switch open) or 1 (switch closed) thus enabling the research of random number generators, simulation and cali-
frequency of RPT coming out of 1 to take on a value between bration of nuclear and radiation detectors, in secure authenti-
0 and f 1 in 2 N equally sized steps. The same summing property cation methods, spurless fractional frequency division,17 and
of  gate can be used once again to add an arbitrary frequency in computing with random pulse trains.18
f 0. With addition of base frequency f 0 by means of the gate
2, it is possible to generate a RPT of frequency in the range
[ f 0, f 0 + f 1] in steps of f 1/2 N . Both the starting frequency VII. RANDOM NUMBER GENERATOR (RNG)
( f 0) and range ( f 1) are arbitrary and step size ( f 1/2 N ) can be Probably the simplest application of the random flip-flop
made as small as desired. Note that when switches are re-set, is the RNG, a physical device that produces one or more
the random frequency jumps virtually instantaneously to the random bits upon a request. Two equivalent circuits, shown
new value (after combined propagation delay of 1 and 2) in Fig. 15, would generate one random bit upon each rising
which is in contrast to other methods that would have finite edge of the Request signal in synchronization with the Strobe
settling time.13,14 signal. If one bit per request is not enough, a random number
Using four DRFFs and a single ORPTG of frequency generator that produces multiple random bits per single request
f 1 = 1 Mcps (and f 0 = 0), the circuit in Fig. 13 was built and can be obtained by parallelizing a desired number of mutually
tested. It allows generation of 16 frequencies in the range independent bit generators that share the same Request signal.
0-1 Mcps in steps of 0.0625 Mcps. Waiting time distribu- For example, parallelizing 8 RFFs, one could generate one
tions, shown in logarithmic scale in Fig. 14, indicate expo- random byte (random integer number between 0 and 255) per
nential p.d.f. as expected. The largest measured deviation of single request pulse.
any output frequency from the expected theoretical value was To test the quality of such a RNG, we constructed four
230 cps. independent DRFFs (each according to Fig. 6). Internal
One of the advantages of this approach is the possibility ORPTGs (each according to Fig. 4) were running at a fre-
of generating RPTs with a very low yet precisely known fre- quency of ≈24 Mcps each. Since all four RNGs share the
quency. The current art uses negative feedback loop to set a same Request line, each Request signal results in simultaneous
generation of 4 bits. By applying a periodic request signal of
3 MHz, bits were generated at a total rate of 1.2 · 107 bits/s
and sent to a PC computer via a USB2 interface. Several
random sequences, each 109 bits long, have been generated
this way, stored in a PC computer, and their randomness tested
with well-known NIST’s Statistical Test Suite, version 2.1.1,19
which they passed without any post-processing. The NIST
STS suite is a collection of stringent statistical tests that are

FIG. 15. Two equivalent random bit generators: by each rising edge at the
FIG. 14. Pulse waiting time distributions for 16 possible output frequencies. Request input circuit generates a fresh new random bit synchronously with
For k = 0, output frequency is 0. the Strobe pulse.
Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
 035113-8 Mario Stipčević Rev. Sci. Instrum. 87, 035113 (2016)

used as a standard randomness assessing tool. Passing of the M

1 
tests is important because it indicates good randomness of the H (M) = − pi log (pi ) , (5)
log (M) i=1
four DRFFs that have been used in experimental demonstra-
tions in this study. where pi are empirical probabilities of each outcome. The
Physically generated random numbers find use in cryp- entropy H (M) is equal to 1 for an infinite sequence of perfectly
tography (securing mobile communications, online payments, fair dice. However for a sequence of finite length (in our case
e-banking, etc.), Monte Carlo numerical simulations, scientific N = 107), the entropy is smaller than unity and there is a
research, etc. statistical error associated with it. In our analysis, we follow
the theory developed in Ref. 21.
VIII. AN EXAMPLE OF APPLICATION TO HAZARD The first test (M = 6) runs over probabilities of individual
GAMES: THE ELECTRONIC DICE throws and is sensitive to evenness of the probabilities of
the 6 possible outcomes (generalized bias), while the second
Global gambling market is in steady rise since 2003 and test (M = 36) runs over probabilities of all combinations of
has reached $32 billion revenue for online games and over two successive throws and is sensitive to correlations be-
$310 billion in total hazard games.20 While not all games tween successive throws (serial autocorrelation), if any. The
require random numbers, most lucrative ones do. Fig. 16 obtained results are H(6) = 0.999 999 77(15) for singles and
shows an elegant construction of electronic dice made of three H(36) = 0.999 999 16(30) for pairs. They agree with expected
random flip-flops that could be perfectly suited as a substitute theoretical values of 0.999 999 86 and 0.999 999 02 within
for 6-sided dice in online games. statistical errors. One sigma statistical error and expected
It functions in the following way. A signal at the Request values for the measured set were estimated according to
input triggers three independent TRFFs to generate a binary Ref. 21. The first result shows that the 6 outcomes appear with
random number in the range 0-7 at its Output. If 0 or 7 gets equal probabilities on average while the second shows that
generated, then the periodic oscillator keeps sending addi- knowing outcome of (any) one throw will not help narrowing
tional requests until a random number in the range 1-6 is gener- down the probability for the outcome of the next. Since this
ated by a chance. When this happens, further requests from the system has no intentional memory, it is reasonable to assume
periodic oscillator are suppressed and logic “1” appears at the that any higher-order correlations are even smaller than those
output Ready signifying completion of the “throw.” We per- tested and that consequently this electronic dice passes test of
formed an experiment where the electronic dice was connected randomness. In order to cross-check this assumption, we eval-
to a PC computer via USB2 interface and “thrown” at a rate of uate autocorrelation coefficients with lags from 1 to 36 on an
1.5 · 106 throws/s to generate a sample of N = 107 numbers. experimentally obtained set of 108 throws. The average value
One cannot use NIST STS or similar tests that operate of the 36 coefficients is +0.08 · 10−4 with an expected error of
on long strings of random bits to test this sample. Namely, 1.00 · 10−4 and RMS of 0.99 · 10−4 confirming that the auto-
because 6 is not an integer power of 2, convert string of random correlation coefficients up to lag 36 are consistent with zero.
throws into a string of random binary bits is quite a demanding The sustained throughput of 1.5 · 106 throws/s should be
mathematical task. But even if such conversion would be more than enough to feed a whole server serving web-based
made, testing the resulting binary string would miss the point. hazard games. In a similar fashion, one could envisage a sys-
What we need to assess is any un-equalness of probabilities tem for mixing of play cards, drawing lottery numbers, etc.
of outcomes and any correlations among successive outcomes
that a player could use to his advantage to win the game
more frequently that if the outcomes were perfectly random. IX. CONCLUSION
Therefore in order to test randomness of the dice, since there
are no readily available standard randomness tests for this case, A new sequential logic element, the RFF, is defined, built
two ad hoc tests were devised. Randomness of the sequence exploiting quantum randomness present in light emission and
was tested by evaluating normalized Shannon entropy for detection, and tested. Some entirely new applications have
M = 6 throws and for M = 36 combinations of two consec- been identified and experimentally realized such as random-
utive throws, that is, by evaluating ness preserving frequency division, random frequency syn-
thesis, and simultaneous multi-bit random number generation.
Further novel applications are under investigation. We believe
that RFF will find many more useful applications in the future
and if put on chip could become a potent enabler of new
technologies.

ACKNOWLEDGMENTS
This work was in part financed by MoSES No. 533-19-
14-0002.

FIG. 16. Electronic dice generates random numbers in the range 1-6 upon a 1See http://pdf.datasheetcatalog.com/datasheet/philips/74f74.pdf for 74F74
request. Completion of a throw is signified by appearance of logic 1 at the Dual D-type flip-flop, Philips Product specification (last accessed June 27,
output Ready. 2015).
Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
 035113-9 Mario Stipčević Rev. Sci. Instrum. 87, 035113 (2016)

2T. F. d. M. Wahl, M. Leifgen, M. Berlin, T. Roehlicke, H. J. Rahn, and 12D. E. Knuth, The Art of Computer Programming, 3rd ed. (Addison-Wesley,

O. Benson, “An ultrafast quantum random number generator with provably Reading, 1997), Vol. 2.
bounded output bias based on photon arrival time measurements,” Appl. 13A. S. Robinson, “Pulse rate function generator,” U.S. patent US3217148
Phys. Lett. 98, 171105 (2011). (priority date 28 June 1960).
3M. Stipčević, “Fast nondeterministic random bit generator based on weakly 14J. A. Hanby and S. J. Redman, “Random pulse train generator with

correlated physical events,” Rev. Sci. Instrum. 75, 4442–4449 (2004). linear voltage control of average rate,” Rev. Sci. Instrum. 42, 657–662
4M. Stipčević and B. M. Rogina, “Quantum random number generator based (1971).
on photonic emission in semiconductors,” Rev. Sci. Instrum. 78, 045104 15G. White, “The generation of random-time pulses at an accurately known

(2007). mean rate and having a nearly perfect Poisson distribution,” J. Sci. Instrum.
5M. Stipčević, “A novel active quenching circuit for single photon detec- 41, 361–364 (1964).
tion with Geiger mode avalanche photodiodes,” Appl. Opt. 48, 1705–1714 16T. Noriyoshi and Y. Hiroyasu, “Random pulse generator and random number

(2009). generation device and probability generation device utilizing the random
6M. Stipčević, D. Wang, and R. Ursin, “Characterization of a commer- number generator,” patent EP1094603A1 (20 October 1999).
cially available large area, high detection efficiency single-photon avalanche 17V. S. Reinhardt, R. P. Verdes, I. Shahriari, inventors, and Hughes Aircraft

diode,” J. Lightwave Technol. 31, 3591–3596 (2013). Company, asignee, “Spurless fractional divider direct digital frequency
7M. Stipčević and R. Ursin, “An on-demand optical quantum random number synthesizer and method,” U.S. patent US4815018 (21 March 1989).
generator with in-future action and ultra-fast response,” Sci. Rep. 5, 10214 18R. C. Lawlor, “Computer ulitizing random pulse trains,” U.S. patent

(2015). US3612845 (12 October 1971).

8V. B. Minuhin, “3/2 frequency divider,” U.S. patent US4866741 (priority 19A. Rukhin, J. Soto, J. Nechvatal, M. Smid, E. Barker, S. Leigh, M. Levenson,

date 12 September 1989). M. Vangel, D. Banks, A. Heckert, J. Dray, and S. Vo, “A statistical test suite
9N. Regimbal, F. Badets, D. Bordeaux, J. B. Begueret, inventors, and CNRS, for the validation of random number generators and pseudo random num-
STMicroelectronics S.A., applicants, “Fractional frequency divider,” U.S. ber generators for cryptographic applications,” NIST Special Publication,
patent US2012/0001665 (29 June 2011). 800–22rev1a, April 2010, URL: http://csrc.nist.gov/groups/ST/toolkit/rng/
10R. E. Best, Phase-locked Loops: Design Simulation and Applications documents/sts-2.1.1.zip.
(McGraw-Hill, 2003), ISBN: 0-07-141201-8. 20See http://www.statista.com/markets/409/topic/438/gambling/ for statistics
11A. Papoulis and S. U. Pillai, Probability, Random Variables, and and market data on gambling (last accessed March 8, 2016).
Stochastic Processes, 4th ed. (McGraw-Hill Europe, 2002), ISBN-13: 21M. S. Roulston, “Estimating the errors on measured entropy and mutual

978–0071226615. information,” Physica D 125, 285–294 (1999).

Reuse of AIP Publishing content is subject to the terms at: https://publishing.aip.org/authors/rights-and-permissions. Download to IP: 75.143.18.145 On: Thu, 17 Mar
2016 01:13:13
View publication stats