Remarks on The Logistic Lattice in Random Number Generation. Neal R. Wagner

Transcription

1 Remarks on The Logistic Lattice in Random Number Generation Nea R. Wagner 1. Introduction Pease refer to the quoted artice before reading these remarks. I have aways been fond of this particuar random number generator (RNG) because of its unusua character, as expained in the main artice. These remarks just present a few additiona experiments that I ve carried out. 2. The Logistic Equation The ogistic equation is given by. I experimented on cyce engths with various hardware, but not with the IEEE standard for doube cacuations. Tabe 1 shows cyce engths obtained from random starting vaues. Notice that of the time it headed into the cyce of ength. (The software checked that this was not the other cyce of ength :!" #.) This cyce structure is simiar in form to that given in the earier artice for other hardware Cyce Structure Number Hard- Preci- Cyce Percent Average of starting ware sion ength occurrence initia run vaues IEEE doube % standard % (Pent II) % % % Tabe 1. Cyces of the ogistic equation. $( I aso carried out experiments on the interva of numbers around %$'& that is mapped into and then into the cyce ) by. The foowing is a sma piece of C code that checks for this interva. Beow getdoube() fetches the next doube vaue, and putdoube(x) prints the vaue of x in exponent and hex (interna) form. x = getdoube(); y = 4.*x*(1.0 - x); putdoube(y); z = 4.*y*(1.0 - y); putdoube(z); Here are resuts of a run. Notice that 3ff is an exact one, and a zero bits is an exact zero. The number 3fefffff ffffffff is as cose to as one can get from beow. User input is in bod red, whie the bue comments inside square brackets are not part of the output.

2 y: Dec: e-01, Hex: 3fefffff ffffffff [amost one] z: Dec: e-16, Hex: 3cbfffff ffffffff y: Dec: e+00, Hex: 3ff [exacty one] z: Dec: e+00, Hex: y: Dec: e-01, Hex: 3fefffff ffffffff [amost one] z: Dec: e-16, Hex: 3cbfffff ffffffff y: Dec: e+00, Hex: 3ff [exacty one] z: Dec: e+00, Hex: %$(, The above shows that the entire interva %$* ++++,+++ -,. /0%$'&,,,!.-, of ength. #& - is mapped to $(, and then on into the cyce. The hex vaues show that no arger interva wi map this way. This might not sound ike so much, but two smaer intervas map into the above interva, and so on with doube the number of intervas at each stage, each smaer than those at the previous stage. To get the two intervas that map into the interva around %$'&, one can sove for 2, 3 in the formua 1, to get :9 ; <" - 1. Now pug in the endpoints of the first interva above to get the endpoints of the two intervas that map into it. The foowing resuts, using the same C code as before, gives vaues for the two intervas that map into the interva around %$'&. Again the hex vaues show that no arger interva wi map this way. (Beow, amost one means as cose to as you can get with a doube vaue.) y: Dec: e-01, Hex: 3fe e y: Dec: e-01, Hex: 3fe y: Dec: e-01, Hex: 3fdfffff fbff812c y: Dec: e-01, Hex: 3fdfffff fc y: Dec: e-01, Hex: 3fe e y: Dec: e-01, Hex: 3fe ffaede y: Dec: e-01, Hex: 3fdfffff fbff812c y: Dec: e-01, Hex: 3fdfffff fc Tabe 2 beow shows the first interva and the two intervas mapping into it.

3 E Interva = Length What maps where >? ( , ) , for < >? = ( , ) , for < >? = ( , ) Tabe 2. Intervas mapped to 1., for in = 3. The Re-mapped Logistic Equation Whie I was studying the ogistic equation, I had an intuitive feeing that the equation coud be re-structured in a way that woud improve its performance. I didn t know exacty what I was ooking for. I spent an entire afternoon one day scratching around, trying a sorts of ideas, drawing endess pictures. The resut was the equation in this section. It very much out-performs the ordinary ogistic equation of the previous section in terms of the ength of its cyces and the average initia run up to a cyce. The re-structured or re-mapped ogistic equation is defined by: BDC F G -IH H - H H KJ LM H H ON - 6; H H 6PKJ LMKNRQ H H?, (1) where NS T;U V" 9 -. The earier artice gives a graph of the origina (expanded by a factor of two) and re-mapped versions (Figure 1 of the earier artice). Tabe 3 beow gives the cyces obtained from & random starting vaues of the remapped equation using the IEEE standard. (In the earier artice, I ony used. & random starting vaues.) Cyce Structure Number Hard- Preci- Cyce Percent Average of starting ware sion ength occurrence initia run vaues IEEE doube % standard % (Pent II) % % % Tabe 3. Cyces of the re-mapped ogistic equation. In the earier artice, I stated: In infinite precision, this re-mapped equation behaves exacty ike the origina, but with foating point numbers there is no onger any convergence to the cyce ) of ength. This was my intuition a aong, but I sti don t have a proof of it. However, numerica experiments simiar to those above show that the conjecture must be true. So here again I carried out experiments in the re-mapped setting. The numbers must be )X$Y&Z\[]%$Y&_^[0 near but ess than and near but greater than W. So what corresponds a]ax^b[ to an interva and 6_\[]]c. is repaced in this re-mapped setting by the intervas ` I used the foowing C code to evauate the behavior under using different vaues of [. #define BETA x = getdoube(); putdoube(x); x = fabs(x);

4 ` if ( x <= BETA ) y = (2.0*x*(2.0-x)); ese y = (-2.0*(1.0-x)*(1.0-x)); putdoube(y); y = fabs(y); if ( y <= BETA ) z = (2.0*y*(2.0-y)); ese z = (-2.0*(1.0-y)*(1.0-y)); putdoube(z); Here are resuts of the second run. Notice that 3fefffff ffffffff is the argest number ess than, whie bfefffff ffffffff is the smaest number greater than a. The number is usuay referred to as minus zero, that is, 8, but in fact the hardware treats it just ike the usua vaue of zero. (Again, amost one means as cose to as you can get with a doube vaue.) x: Dec: e-01, Hex: 3fefffff ffffffff [amost one] y: Dec: e-32, Hex: b z: Dec: e-32, Hex: x: Dec: e+00, Hex: 3ff [exacty one] y: Dec: e+00, Hex: [exacty -zero] z: Dec: e+00, Hex: [exacty zero] x: Dec: e-01, Hex: bfefffff ffffffff [amost -one] y: Dec: e-32, Hex: b z: Dec: e-32, Hex: x: Dec: e+00, Hex: bff [exacty one] y: Dec: e+00, Hex: [exacty -zero] z: Dec: e+00, Hex: [exacty zero] The above resuts show that no proper intervas with eft end at a or with right end at wi map into, but ony the singe vaues W and. 4. Runs Into the Cyces (0.75) and (-0.5) The ordinary ogistic equation has the obvious cyce ) of ength and one other cyce of ) ength, namey %$. &. In the re-mapped ogistic equation, the corresponding cyces are and 8%$Y&,. With the ordinary ogistic equation in singe precision (foat type), the origina artice showed that a arge number of initia vaues ( +! ) end up in, whie there are no runs at a into %$. &. Using the re-mapped ogistic equation, there were very few runs into ), and again none at a into 6d%$'&. In the artice, I stated without proof: The cyce search in singe precision is exhaustive because every cyce%$ wi eventuay enter the range from!" to. I meant to say incusive. &Xec at the end, that is, `. I aso meant to incude, except for the cyce. Here is a proof, with )X$ a tighter statement of the assertion. (Beow, f means a cyce containing just, and. &g]c means those for which X$. &hqi.) %]c Here are cases based on the vaue of kj If m, then it is in the cyce ), the specia case. If m %$. & then it is in the cyce )%$. & %$. &Xec, so it is certainy in the range `. If jn )X$. &g]c then oqx$. &, so no cyce can be entirey inside %$. &X]c. :

5 ` If %$ Oj? )X$ - &g0%$. & then \j? )%$. &X]]c so no cyce can be entirey inside - &X0%$. &, and in fact in this case, f jn )%$. &X]]c. If pj %0%$ - &qc then rf s T hit, so Irf vu -. Using the Mean Vaue Theorem, A <"x ku there is an r Qw such 8 <"/ f R r r, so - or ou -. Ifz'}{ we f ~ take iterates %$ A f < y Dz'P6{, etc., then eventuay either - & z'} ){ f ~ X$. & z(}v{ T %$ - & z'} ){< f ƒ %$ so that, or so that. &, so that iterates go into )X$. &g]c. Thus no cyce can be entirey inside )%0X$ - &xc. This %$ competes the proof. This shows that every cyce except must eventuay be inside. &X]c (incuding the cyce %$. &.) It aso shows that no cyce can be entirey inside %0%$ - &qc, or entirey inside )X$ - &g0%$. &, or entirey inside %$. &X]c. However, much to my surprise, there is a run of ength &g -, ending in the cyce )%$. &, that is entirey inside )%0X$. &xc. The same situation occurs with the re-mapped ogistic equation. So now I ook into possibe runs into %$. & (ordinary ogistic equation) and 6dX$Y& (remapped ogistic equation) in doube precision. First, focus on the ordinary ogistic equation, and work backwards from %$. &. In infinite precision, there are in genera two vaues that map by into any singe vaue. We can use the equation ƒ o7 9 " - 1 to get approximate vaues, athough to quaify, the vaue must be exact. Cacuating, both X$ - & and %$. & map into %$( %$. &. There are two vaues that map approximatey into X$ - & : %$(+!!,% -. XVt+ -- V+! and +,t.- +txv.. t,. -. However, as the cacuation beow shows, ony the second of these vaues actuay works. (Beow, the input consists of exact hex vaues, rather than approximate decima vaues.) 3feddb3d 742c2655 x : Dec: e-01, Hex: 3feddb3d 742c f1(x): Dec: e-01, Hex: 3fd succes f2(x): Dec: e-01, Hex: 3fe sive vaues 3feddb3d 742c2656 x : Dec: e-01, Hex: 3feddb3d 742c f1(x): Dec: e-01, Hex: 3fcfffff fffffff5 f2(x): Dec: e-01, Hex: 3fe7ffff fffffffa 3fb e9ecd56 x : Dec: e-02, Hex: 3fb e9ecd56 f1(x): Dec: e-01, Hex: 3fd f2(x): Dec: e-01, Hex: 3fe For the number that starts with %$'+,!!% -,., you can see two successive hex vaues, one too %$( sma and one too arge, so that no vaue can work exacty. The number that starts with +,t.- +t %$ has a cose hex vaue that works exacty. So ony one precursor vaue from - & works. Continuing to move backwards, at each stage the vaue near works, whie the vaue near does not. One gets a singe run into %$. & of ength &X - : 00118bc4 418cafe1 x : Dec: e-308, Hex: 00118bc4 418cafe1 f 1(x): Dec: e-308, Hex: 00318bc4 418cafe1 f 2(x): Dec: e-307, Hex: 00518bc4 418cafe1 f 3(x): Dec: e-306, Hex: 00718bc4 418cafe1... [480 ines omitted, a ending with 18bc4 418cafe1 ] f484(x): Dec: e-17, Hex: 3c918bc4 418cafe1 f485(x): Dec: e-16, Hex: 3cb18bc4 418cafe1

6 f486(x): Dec: e-16, Hex: 3cd18bc4 418cafe0 f487(x): Dec: e-15, Hex: 3cf18bc4 418cafdb f488(x): Dec: e-14, Hex: 3d118bc4 418cafc8 f489(x): Dec: e-14, Hex: 3d318bc4 418caf7b f490(x): Dec: e-13, Hex: 3d518bc4 418cae47 f491(x): Dec: e-13, Hex: 3d718bc4 418ca978 f492(x): Dec: e-12, Hex: 3d918bc4 418c963a f493(x): Dec: e-11, Hex: 3db18bc4 418c4943 f494(x): Dec: e-11, Hex: 3dd18bc4 418b1567 f495(x): Dec: e-10, Hex: 3df18bc f6 f496(x): Dec: e-09, Hex: 3e118bc f497(x): Dec: e-09, Hex: 3e318bc f498(x): Dec: e-08, Hex: 3e518bc4 3ff234b9 f499(x): Dec: e-08, Hex: 3e718bc4 3b22c33c f500(x): Dec: e-07, Hex: 3e918bc4 27e4fd53 f501(x): Dec: e-06, Hex: 3eb18bc3 daede659 f502(x): Dec: e-06, Hex: 3ed18bc2 a f503(x): Dec: e-05, Hex: 3ef18bbd d7a0f869 f504(x): Dec: e-05, Hex: 3f118baa 99e912e4 f505(x): Dec: e-04, Hex: 3f318b5d a3b24972 f506(x): Dec: e-03, Hex: 3f518a29 d563b9c0 f507(x): Dec: e-03, Hex: 3f71855b 44e5d92f f508(x): Dec: e-02, Hex: 3f91722b 8b740eb3 f509(x): Dec: e-02, Hex: 3fb e9ecd56 f510(x): Dec: e-01, Hex: 3fd f511(x): Dec: e-01, Hex: 3fe The notation f 3(x) stands for f(f(f(x))), and so on for vaues other than!. The argest possibe IEEE doube ess than is 3fefffff ffffffff. Appy to this vaue to get /$(, t+ - +t&,, - &.. This is the smaest vaue, other than, that can occur in a run using the ogistic equation, athough one coud choose a smaer initia vaue. Thus everything in the ist above preceding f486(x) coud ony occur as an initia vaue. A numbers before f485(x) end with 18bc4 418cafe1, and in this range each successive hex vaue is obtained by adding to the previous vaue, that is, by mutipying by, since for numbers very cose to, f is exacty mutipication by. The smaest possibe IEEE doube is - $ -- &,.!t&,t&.- %]!t, so that the run cannot be made one onger. Of course a subsequent vaues in the run wi be the cyce vaue X$. &. 8%$'& With the re-mapped there is an exact corresponding run of ength &X - into the cyce : 00218bc4 418cafe1 x : Dec: e-308, Hex: 00218bc4 418cafe1 f 1(x): Dec: e-307, Hex: 00418bc4 418cafe1 f 2(x): Dec: e-307, Hex: 00618bc4 418cafe1 f 3(x): Dec: e-306, Hex: 00818bc4 418cafe1... [504 ines omitted] f508(x): Dec: e-02, Hex: 3fa1722b 8b740eb3 f509(x): Dec: e-01, Hex: 3fc e9ecd56 f510(x): Dec: e-01, Hex: 3fe f511(x): Dec: e-01, Hex: bfe There is one type of addition to this run. In the case of the re-mapped ogistic equation, at any stage of the run, you can start with the negative of an entry. There is no precursor to this vaue, but the$'! subsequent vaues are the same positive ones as above. Thus, for exampe, the. /.!. V,+!%.,+ numbers, X$ˆV!,!+. #&+ - x&,& V!&, %$'&, dx$y& form an initia run up to 8%$'&. Otherwise the same remarks appy to these cacuations as to the ones for the ordinary ogistic equation.

7 - %$. & The same behavior occurs with IEEE foat cacuations, except that the run into or 8%$'& is ony 64 ong. Here the smaest possibe IEEE foat is $. & +,!&!t. The ist of vaues beow shows both the ordinary and re-mapped ogistic equations side-by-side. Here before the f52(x) terms, each term is obtained from the previous one by adding , that is, mutipying by. Any run for the re-mapped ogistic equation can start with a negated entry, as above. In the origina artice, the exhaustive search of the ordinary equation for foat X$ precision. &g]c, so it found the cyce )X$. &,, but missed the run ony covered the range in the interva ` beow atogether, and simiary for the re-mapped equation. Ordinary Logistic Equation Re-mapped Logistic Equation 008c5e22 010c5e22 x : e-38, Hex: 008c5e e-38, Hex: 010c5e22 f 1(x): e-38, Hex: 018c5e e-37, Hex: 020c5e22 f 2(x): e-37, Hex: 028c5e e-37, Hex: 030c5e22 f 3(x): e-37, Hex: 038c5e e-36, Hex: 040c5e22... [46 ines omitted, a ending with c5e22 ] f50(x): e-08, Hex: 328c5e e-08, Hex: 330c5e22 f51(x): e-08, Hex: 338c5e e-07, Hex: 340c5e22 f52(x): e-07, Hex: 348c5e e-07, Hex: 350c5e21 f53(x): e-06, Hex: 358c5e1f e-06, Hex: 360c5e1f f54(x): e-06, Hex: 368c5e e-06, Hex: 370c5e15 f55(x): e-05, Hex: 378c5def e-05, Hex: 380c5def f56(x): e-05, Hex: 388c5d e-04, Hex: 390c5d55 f57(x): e-04, Hex: 398c5aed e-04, Hex: 3a0c5aed f58(x): e-03, Hex: 3a8c514f e-03, Hex: 3b0c514f f59(x): e-03, Hex: 3b8c2ada e-03, Hex: 3c0c2ada f60(x): e-02, Hex: 3c8b915c e-02, Hex: 3d0b915c f61(x): e-02, Hex: 3d8930a e-01, Hex: 3e0930a3 f62(x): e-01, Hex: 3e e-01, Hex: 3f f63(x): e-01, Hex: 3f e-01, Hex: bf Discussion Suppose one tried to use the ordinary ogistic equation directy as a RNG. Start with a seed (some initia doube vaue). Then ++X$'+! of the time, you get some miion vaues before cycing ( ) or ending in ( ). This is compared to a singe cyce of about V biion numbers with an od-fashioned! - -bit generator. (Of course you have to iterate maybe times between samping for a random number.) Anyway, this isn t so terribe, perhaps, except for the remaining X$'. of the time. The main artice provides three ways to improve the basic ogistic equation based RNG: 1. Iterate the generator to fi with random noise. 2. Use the re-mapped ogistic equation that has much onger cyces and initia runs, and has no convergence to. 3. Use a couped attice of ogistic equation generators that wi perturb one another. Of the above, item 1 is essentia, but fairy easy to think of. Item 2 is the one I m most proud of, because it s unexpected and origina. But item 2 is not essentia as ong as one uses item 3, which reay is essentia. The basic idea of item 3 was suggested to me by Kay Robbins. (Couped attices of ogistic equations are a standard simpe mode of chaos, but I didn t know that initiay.)