AMPAdvances in Mathematical Physics1687-91391687-9120Hindawi10.1155/2021/66968956696895Research ArticleThe New Scramble for Faure Sequence Based on Irrational Numbershttps://orcid.org/0000-0001-9378-2344O.Ali Mogharrabi1https://orcid.org/0000-0002-2403-9271V.Behrooz Fathi2https://orcid.org/0000-0002-1849-5588BehzadiM. H.1https://orcid.org/0000-0002-9058-2491FarnooshR.3GrecoLeopoldo1Department of StatisticsScience and Research BranchIslamic Azad UniversityTehranIranazad.ac.ir2Department of StatisticsUniversity of GuilanRashtIranguilan.ac.ir3Department of MathematicsIran University of Science and TechnologyTehranIraniust.ac.ir202128420212021241220201112021103202128420212021Copyright © 2021 S. A. Mogharrabi et al.This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

This article intends to review quasirandom sequences, especially the Faure sequence to introduce a new version of scrambled of this sequence based on irrational numbers, as follows to prove the success of this version of the random number sequence generator and use it in future calculations. We introduce this scramble of the Faure sequence and show the performance of this sequence in employed numerical codes to obtain successful test integrals. Here, we define a scrambling matrix so that its elements are irrational numbers. In addition, a new form of radical inverse function has been defined, which by combining it with our new matrix, we will have a sequence that not only has a better close uniform distribution than the previous sequences but also is a more accurate and efficient tool in estimating test integrals.

1. Introduction

It is well known that Monte Carlo calculations are based on the generation of random numbers on interval (0,1). Therefore, the generation of random numbers that have more uniformity on (0,1) guarantees better approximations in these calculations. In recent years, some researchers have employed quasirandom sequences instead of random numbers to aim producing extra uniformity of the randomly generated numbers on (0,1). Due to the breadth and complexity of some problems that are mostly unsolvable by classical mathematical methods or solving them with classical methods is associated with more time and computational cost, the stochastic solving of such cases with numerical methods and using the Monte Carlo method plays a key role. The quasirandom sequences are common in Monte Carlo calculations such as Faure, Halton, Niederreiter, and Sobol sequences, but due to the lack of complete success of these sequences in Monte Carlo computation, we use scrambled versions of them, all of which are designed to increase the uniformity of randomly (quasirandom) generated numbers on (0,1), so that we can estimate the obtained solution to the desired unknown solution of the problem.

To resolve this problem, researchers are competing on the use of scrambled quasirandom generators based on their version of random number generation to provide more accurate results in Monte Carlo calculations.

Today, Monte Carlo and quasi-Monte Carlo methods are widely used to solve the computations of physical and mathematical problems. Quasi-Monte Carlo (QMC) methods play an alternative role for Monte Carlo methods. The advantage of these methods is that they use numbers to provide extera uniformity on unit hypercube. This feature has led to the use of these methods to estimate high-dimensional integrals (Niederreiter, 1992; Spanier and Maize, 1994) .

So far, several quasirandom sequences (or low discrepancy sequences) have been introduced for the QMC method. Such as the Faure sequence, the Halton sequence, and the Sobol sequence. Despite the fact that among these three sequences, the Faure sequence has better features in terms of discrepancy bound, but in practice, it is less used. Because, the convergence rate of this class of sequences is not so good compared with the other sequences . In addition, due to the correlation between the different dimensions of the Faure sequence, the distribution of the sequence points is not very favorable, and we see poor two-dimensional projections (Figure 1). To overcome this problem, many scrambling methods have been proposed for the Faure sequence (see ). In almost all of these scrambles, there are attempts to define a new matrix by shifting numbers in the generating matrix or the placement of the elements, and there has been less talk about the properties of irrational numbers. The matrix that we introduce in this paper has been selected from several proposed matrices. Because it has good two-dimensional projections and it is also at a very high level in terms of integral estimation.

512 points from the original Faure sequence in several bases.

In the next section, the structure of the original Faure sequence is given. We then briefly list the scramblers that have already been introduced in Section 3. In Section 4, we have brought our proposed matrix. Sections 5 and 6 give the evaluation criteria for the quality of the sequence generated by our proposed matrix and compare it with previous sequences, and in Section 7, the conclusion is stated.

2. The Faure Sequence

Suppose p2 is a prime number, and suppose N is the number of points we want to generate, and n=a0,a1,,am1T is a vector of integers whose components are from the expansion of the number n on the base p, where 0ajp and m=logpN. We define the radical inverse function, ϕpn, as (1)ϕpn=a0p+a1p2++am1pm.

For the Faure sequence, we define a different generator matrix for each dimension. If P be the Pascal matrix, then for a s-dimensional Faure sequence the generator matrix of the jth dimension is Cj=Pj1, 1js, where the member on the row c and the column r is defined as follows: (2)Pj1=c1r1j1crmodp,c1,r1.

Thus, let xnjbe represents the number n in the dimension j in the Faure sequence, then (3)xnj=ϕpPj1n,and so the s-dimensional Faure sequence is ϕpP0n,ϕpP1n,,ϕpPs1n.

3. Scrambling the Faure Sequence

Since the introduction of the Faure sequence, several methods were proposed to scramble it. In this section, we give an overview of some of such scrambles.

3.1. The Generalized Faure Sequence

Tezuka  proposed the generalized Faure sequence, GFaure, with the jth dimension generator matrix Cj=AjPj1 and the Aj for j=1,2,,s are arbitrary nonsingular lower triangular matrices over Fp. A special case for Aj is that all members are one for all dimensions .

3.2. Random Linear (Digit) Scrambling

After reviewing different versions of the Owen’s method, Matoušek introduced a scramble matrix and a transfer vector for various dimensions . The sequences obtained by Matoušek have the following general form: (4)xn=ϕpA1P0n+g1,ϕpA2P1n+g2,,ϕpAsPs1n+gs.

For the random linear scrambling, the matrices Aj and the vectors gj for j=1,,s are of the form (5)Aj=h1,1000h2,1h2,200h3,1h3,2h3,30,gj=g1g2g3,where the gj’s and the hi,j with ij are chosen randomly and independently from 0,1,,b1, the hj,j’s are chosen randomly and independently from 1,2,,p1.

Random linear digit method is the basis of other scrambles that will follow. Even the GFaure method is a subset of this method in which the members of the shift vectors are all zero.

3.3. I-Binomial Scrambling

A subset of the family of random linear scrambling methods is called left I-binomial scrambling . Here, the Aj is defined as (6)Aj=h10000h2h1000h3h2h100h4h3h2h10,where h1 is chosen randomly and independently from 1,2,,p1 and also hi’s i>1 are chosen randomly and independently from 0,1,,p1.

3.4. Striped Matrix Scrambling

The scrambling matrix Aj for Striped Matrix Scrambling method, has the following form: (7)Aj=h10000h1h2000h1h2h300h1h2h3h40,where the hi’s are chosen randomly and independently from 1,2,,p1. The different types of this scramble are examined in , only for problems in the first dimension. Of course, it does not say what changes should be made to the matrix for higher dimensions.

3.5. Chi’s Optimal Scramble

When we use the I-binomial method to scramble the Faure sequence, the value of (number) h1 causes all the expansion digits of each number to be replaced (permuted). Now, if we leave out the first digit, the value of (number) h2 causes all the remaining digits to be replaced (permuted).

So by cleverly selecting these two members, we can achieve better Faure sequences.

In , Chi has shown that the best choice for these two values can be obtained based on the primitive roots of the p. Finally, the Chi’s optimal scramble matrix is as follows: (8)Aj=h1j1000h2h1j1000h2h1j1000h2h1j1,

3.6. Inverse Scrambling

Fathi and Eskandari  adapted Chi’s optimal matrix and introduced two kind of matrix Aj as follows: (9)A1j=dj10000dj10000dj10,A2j=dj10000dj0000dj+10,where d is chosen from 1,2,,p1 for j=1,,s.

Based on nonlinear congruential method, they proposed another scrambling method for the Faure sequence for which the jth coordinate of the nth point has the general form (10)xnj=ϕpΦ1AjΨPj1n+gj,where Φx and Ψx are bijections that map a digit vector x to another digit vector.

4. Scrambling Matrix with Irrational Members

In this section, corresponding to the method of random linear digits, we introduce a scrambling matrix that its members are a function of square root of base p (that is, they are irrational numbers).

After testing many functions, we found the following function that has the most performance: (11)Aj=p1000p1p100p1p1p10p1p1p1p1.

In (1), we introduced the common form of radical inverse function ϕ for base p. Now, we define a new form of this function as follows: (12)ϕpn=a0pm+a1pm1++am1p.

We call this as reverse radical inverse function, and we denote the sequences that are made in this way, with the suffix “rev.” Therefore, by combining the matrix Aj and function ϕ, the general form of number n in dimension j will be as follows: (13)xnj=ϕpAjPj1n.

So, for example, we denote the 40-dimensional Faure sequence generated on the base 41 by the scrambled matrix Aj and the function ϕ with Aj.41rev.

In the following sections, we have examined (studied) the quality of this sequence along with its performance compared to other sequences.

5. Investigation of the Uniformity of Generated Sequences5.1. Two-Dimensional Projections

The first step in evaluating the performance of a sequence is to see how the points in the 2D projections are distributed. From Figure 2, in these designs, the Faure sequence points are located within parallel lines, which shows that there is a linear correlation between points in successive dimensions. Also in these designs, we see a lot of empty spaces. Therefore, the distribution of good points has not been done. In Figure 3, we draw the same designs using the by Aj.41rev.

1000 points from the original Faure sequence.

1000 points from our scrambled Faure sequence.

These figures show that two-dimensional projections of Aj.41rev sequence are better than two-dimensional projections of the original Faure sequence.

5.2. Discrepancy

One way to measure the quality of a sequence is to calculate its discrepancy . Warnock shows that (14)TN2=1N2k=1Nm=1Ni=1s1maxxki,xmi21sNk=1Ni=1s1xki2+3s.where xki is the ith component of the point xk.

Figure 4 compares TN between the original Faure sequence and the some scrambled Faure sequences that introduced in Section 3. From Figure 4, we see that the discrepancy p diagram for our new sequence is at all points below the other sequences. This is a good indication of the high quality of our sequence points.

L2-discrepancy for various randomized scramblings of a 40-dimensional Faure sequence.

6. Numerical Integration

Another way to compare the quality of sequences is to use them to solve high-dimensional integration problems with numerical methods. Consider the following test integrals: (15)I1f=0101i=1sπ2sinπxidx1dxs=1,(16)I2f=0101i=1s4xi2+ai1+aidx1dxs=1,where the ai are parameters. There are four choices of parameters as follows:

ai=0 for 1is

ai=1 for 1is

ai=i for 1is

ai=i2 for 1is

Note that the most difficult case is when a=0. Because in this case, the importance of all variables is the same and the superposition dimension is approximately the same as the truncation dimension. It is important to know that the larger ai, the less important the variables are, and therefore, the effective dimension becomes smaller. The last three choices of the parameters will be considered here .

In numerical solution of problems with qMC methods, an accepted procedure is to omit the starting points of the sequence. For example, Fox  has suggested that we consider the starting point of the sequence as n=QS41. Although, this may lead to better results, note that with this selection, a large amount of points must be omitted. For example, for s=40, we have to skip the initial 2825760 points with this formula, which is practically impossible. We found in our research that if we select the p as the starting point, it will significantly improve the results. We will probably get the best result when we start from the p2+1th point.

Therefore, we skip the first 41 points and start n=42 in our calculations. For comparison purposes, we present numerical results for original Faure (Faure), our sequence (Aj-41rev), and two types of scrambled Faure sequences, I-binomial (IB) and random linear digits (RLD).

Now, we compare the numerical results of different scrambled Faure sequences presented in this paper. The estimated values for the test functions are given in Figures 58. These figures show that our proposed scramble has a very acceptable convergence compared to other scrambles.

Estimates of the integral I1f by using various Faure sequences.

Estimates of the integral I2f with ai=1 by using various Faure sequences.

Estimates of the integral I2f with ai=i by using various Faure sequences.

Estimates of the integral I2f with ai=i2 by using various Faure sequences.

An observation is that estimated values by the matrix Aj-41rev very close to the actual value. This can be seen in Tables 14. The estimation error obtained with this scramble in dimension 40, for the function (16) with parameters ai=1, ai=i, and ai=i2 are at most 11.68%, 0.32%, and 0.64%, respectively. However, there are some exceptions. For example, in the first function, when the number of dimensions increases, the accuracy of the estimation decreases. So that, for the dimension 40, the maximum estimated relative error value is 43.7%.

Estimates of I1f by using Faure sequences.

GeneratorNs=5s=10s=20s=30s=40
Faure5001.21752.230113.4022.64640.0001
IB5000.00170.00160.00140.00060.0002
RLD5000.00130.00130.00110.00060.0001
Aj-41rev5001.03921.04391.52860.80951.1542
Faure50000.98031.01283.15850.78480.0314
IB50000.03400.03970.04430.04020.0306
RLD50000.04690.04750.05480.07370.0341
Aj-41rev50001.00151.01541.02191.14010.6935
Faure100000.95840.94451.99190.65830.0464
IB100000.05690.05640.05930.07740.0496
RLD100000.05390.05410.05660.07580.0418
Aj-41rev100000.99920.99591.01820.87800.5630
Faure200000.99431.05221.49090.54400.0575
IB200000.06100.06070.06280.07910.0561
RLD200000.05830.05850.05920.06700.0637
Aj-41rev200000.99991.00211.06140.93010.7249
Faure500000.99651.01270.99720.51590.0674
IB500000.06170.06180.06210.06830.0628
RLD500000.06040.06050.05820.06570.0587
Aj-41rev500000.99840.99831.02410.99010.9173
Faure700000.99641.00580.96800.50290.0669
IB700000.08090.08060.08770.20002.3467
RLD700000.07110.07130.06710.07020.0583
Aj-41rev700000.99850.99541.02601.08410.9623
Faure1000001.00081.02481.04590.54660.1665
IB1000000.46100.46160.46670.46941.7654
RLD1000000.27420.27460.27720.25290.1366
Aj-41rev1000000.99791.00071.02151.12511.3969

Estimates of I2f with ai=1 by using Faure sequences.

GeneratorNs=5s=10s=20s=30s=40
Faure5001.07591.23701.79940.69130.1727
IB5001.50151.50001.55721.34451.1510
RLD5001.51231.48201.54561.46611.2191
Aj-41rev5000.98790.90140.93470.89300.9629
Faure50000.99630.99081.03420.74060.4475
IB50001.48361.47851.47871.42121.2841
RLD50001.48071.47701.48401.44551.2824
Aj-41rev50000.99760.98400.99640.98981.0902
Faure100000.99350.98641.00620.78600.5359
IB100001.47691.47501.46971.43701.3296
RLD100001.47791.47591.47841.44381.3261
Aj-41rev100000.99900.98650.99250.99681.0920
Faure200000.99450.99540.98610.81400.5897
IB200001.47531.47551.47061.42441.3306
RLD200001.47631.47521.46401.42881.3271
Aj-41rev200001.00001.00021.01511.03361.0965
Faure500000.99920.99690.96090.83970.6360
IB500001.47501.47441.46971.43971.3342
RLD500001.47551.47511.47251.43981.3372
Aj-41rev500001.00151.00391.01921.02221.0509
Faure700000.99830.99350.96130.84280.6458
IB700001.46541.46501.45821.42491.3238
RLD700001.47101.47051.46531.43621.3491
Aj-41rev700001.00201.00401.02371.04181.1168
Faure1000000.99971.00030.97980.88560.7312
IB1000001.27851.27831.27271.24761.1649
RLD1000001.38481.38441.38211.35081.2698
Aj-41rev1000001.00271.00501.02331.03711.1048

Estimates of I2f with ai=i by using Faure sequences.

GeneratorNs=5s=10s=20s=30s=40
Faure5001.02431.02001.00770.99450.9817
IB5001.50101.50221.50311.50301.5023
RLD5001.50551.50291.50431.50511.5045
Aj-41rev5000.99510.97460.97680.97610.9770
Faure50000.99720.99440.99080.98770.9851
IB50001.48371.48351.48331.48331.4832
RLD50001.48101.48091.48071.48061.4804
Aj-41rev50000.99900.99570.99630.99680.9974
Faure100000.99770.99640.99460.99290.9915
IB100001.47701.47691.47681.47681.4766
RLD100001.47811.47801.47791.47781.4777
Aj-41rev100000.99980.99710.99730.99770.9984
Faure200000.99790.99770.99650.99560.9949
IB200001.47521.47531.47531.47521.4752
RLD200001.47641.47631.47631.47621.4762
Aj-41rev200001.00021.00021.00071.00121.0017
Faure500000.99990.99960.99910.99870.9983
IB500001.47501.47501.47501.47501.4749
RLD500001.47551.47551.47551.47551.4755
Aj-41rev500001.00091.00121.00191.00241.0027
Faure700000.99930.99890.99860.99830.9980
IB700001.46541.46541.46541.46541.4653
RLD700001.47101.47101.47091.47091.4709
Aj-41rev700001.00111.00131.00211.00261.0029
Faure1000000.99970.99960.99930.99900.9989
IB1000001.27851.27851.27851.27851.2784
RLD1000001.38481.38481.38481.38481.3848
Aj-41rev1000001.00131.00161.00241.00281.0032

Estimates of I2f with ai=i2 by using Faure sequences.

GeneratorNs=5s=10s=20s=30s=40
Faure5001.01251.01151.01101.01081.0106
IB5001.50031.50051.50061.50061.5006
RLD5001.50171.50151.50151.50151.5015
Aj-41rev5000.99660.99360.99360.99360.9936
Faure50000.99880.99860.99850.99840.9984
IB50001.48381.48381.48381.48381.4838
RLD50001.48101.48101.48101.48101.4810
Aj-41rev50000.99930.99890.99890.99890.9989
Faure100000.99910.99900.99890.99890.9988
IB100001.47701.47701.47701.47701.4770
RLD100001.47811.47811.47811.47811.4781
Aj-41rev100001.00000.99960.99960.99960.9996
Faure200000.99920.99910.99910.99910.9991
IB200001.47531.47531.47531.47531.4753
RLD200001.47641.47641.47641.47641.4764
Aj-41rev200001.00031.00021.00031.00031.0003
Faure500001.00001.00001.00001.00001.0000
IB500001.47501.47501.47501.47501.4750
RLD500001.47551.47551.47551.47551.4755
Aj-41rev500001.00061.00061.00071.00071.0007
Faure700000.99980.99970.99970.99970.9997
IB700001.46551.46551.46551.46551.4655
RLD700001.47101.47101.47101.47101.4710
Aj-41rev700001.00061.00061.00071.00071.0007
Faure1000000.99980.99980.99980.99980.9998
IB1000001.27851.27851.27851.27851.2785
RLD1000001.38481.38481.38481.38481.3848
Aj-41rev1000001.00071.00071.00071.00081.0008
7. Conclusion

We studied the original Faure sequence and some of its recent years introduced scrambles. Then, we introduced a new scrambling matrix based on irrational numbers that its elements are function of square root of base p. In Figures 58, we have shown that this modified scrambled Faure sequence provides better results than the previous versions of its scrambles. Also, we presented that this modified scrambled Faure sequence has greatly improved the distribution of points. The 2D designs confirm this claim, good two-dimensional projections in successive dimensions. As mentioned in the previous section, using this scramble leads to very small estimated relative errors that can often be ignored. In our next research, we will use deterministic scrambling matrices based on another irrational numbers and primitive roots. The results proved the improvements of accuracy using our new scramble.

Data Availability

The data is contained in the article itself.

Conflicts of Interest

The authors declares that there is no conflict of interest regarding the publication of this paper.

OwenA. B.NiederreiterH.SpanierJ.Monte Carlo, quasi-Monte Carlo, and randomized quasi-Monte CarloMonte Carlo and Quasi-Monte Carlo Methods1998Berlin HeidelbergSpringer-Verlag8697TezukaS.A generalization of Faure sequences and its efficient implementation1994Technical Report RT0105, IBM. Tokyo Research LaboratoryVandewoestyneB.ChiH.CoolsR.Computational investigations of scrambled Faure sequencesMathematics and Computers in Simulation200381522535FaureH.Variations on (0, s)-sequencesJournal of Complexity2015174741753Matouˇsek’sJ.On the L2-discrepancy for anchored boxesJournal of Complexity199814527556TezukaS.FaureH.I-binomial scrambling of digital nets and sequencesJournal of Complexity200319674475710.1016/S0885-064X(03)00035-92-s2.0-0348233983BratleyP.FoxB. L.NiederreiterH.Implementation and tests of low-discrepancy sequencesACM Transactions on Modeling and Computer Simulation19922319521310.1145/146382.1463852-s2.0-84976815495ChiH.Scrambled Quasirandom Sequences and their Applications, [Ph.D. thesis]2004Tallahassee, FL, USAFlorida State UniversityFathi VajargahB.Eskandari ChechaglouA.New modified scrambled faure sequencesCommunications in Statistics - Simulation and Computation201244366668210.1080/03610918.2011.6502552-s2.0-84870878814VandewoestyneB.Quasi-Monte Carlo Techniques for the Approximation of High-Dimensional Integrals, [Ph.D. theses]2008Advisor Ronald CoolsWangX.FangK. T.The effective dimension and quasi-Monte Carlo integrationJournal of Complexity200319210112410.1016/S0885-064X(03)00003-72-s2.0-0037389675FoxB. L.Algorithm 647: implementation and relative efficiency of quasirandom sequence generatorsACM Transactions on Mathematical Software198612436237610.1145/22721.3561872-s2.0-0022954116