[VA] SRC#002- Almost integers and other beasties
|
01-23-2019, 10:58 PM
Post: #21
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
Hi again, Dealing with some search theory, I recently came across this short expression: \[ \frac{1}{12}+\frac{\Pi^2}{6Ln^2(2)}+ \frac{2}{Ln(2)} \sum_{k=1}^\infty \frac{(-1)^k}{k(2^k-1)} \] which this straightforward HP-71B program (a command-line expression would work as well) quickly evaluates to the apparently exact integer value 1: 1 DESTROY ALL @ V=0 @ FOR K=1 TO 40 @ V=V+(-1)^K/K/(2^K-1) @ NEXT K 2 V=V*2/LN(2)+1/12+PI^2/6/LN(2)^2 @ DISP V >RUN 1 But it's not ! See if you can get the actual, more accurate almost-integer value with your HP. It certainly qualifies as a most amazing almost-integer one ! V. . All My Articles & other Materials here: Valentin Albillo's HP Collection |
|||
02-09-2019, 09:53 PM
Post: #22
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(01-03-2019 10:33 PM)Valentin Albillo Wrote: For completeness, I forgot to include this remarkable one which I discovered and posted here last March: Beautiful one! Hope you don’t mind if I uglify it a bit: Sin(9*(Sin 1 + Cos 40) + 5/(e*10^8)) Best regards, Gerson. |
|||
02-09-2019, 10:33 PM
Post: #23
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
There are quite a few of these strewn about mathematical sites on the internet. Some have easy explanations. Some have very deep explanations. Some have no seeming reason at all.
Sin(11) = about -1 Exp(pi)-pi = about 20 Cos(pi*Cos(pi*Cos(Ln(pi+20)))) = about 1 Exp(pi*Sqrt(163)) = about 262537412640768744 |
|||
02-09-2019, 10:56 PM
Post: #24
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
.
Hi, Gerson: (02-09-2019 09:53 PM)Gerson W. Barbosa Wrote:(01-03-2019 10:33 PM)Valentin Albillo Wrote: For completeness, I forgot to include this remarkable one which I discovered and posted here last March: I don't understand ... Also, could you accurately evaluate the expression I gave in my post #21 in this thread (the previous one to yours) and if so, what accurate result did you get ? The proximity to the exact integer 1 is not a coincidence, there are deep mathematical reasons for it. Best regards and have a nice weekend. V. . All My Articles & other Materials here: Valentin Albillo's HP Collection |
|||
02-10-2019, 02:41 AM
Post: #25
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(02-09-2019 10:56 PM)Valentin Albillo Wrote: . Hi, Valentin, The factor of e doesn´t appear to fit nicely in the trigonometric expression, but it makes the result get a bit closer to 1: Sin(9*(Sin 1 + Cos 40) + 5/(e*10^8)) = 0.99999999999999999999999997740056822767 (02-09-2019 10:56 PM)Valentin Albillo Wrote: Also, could you accurately evaluate the expression I gave in my post #21 in this thread (the previous one to yours) and if so, what accurate result did you get ? The proximity to the exact integer 1 is not a coincidence, there are deep mathematical reasons for it. %%HP: T(3)A(R)F(.); \<< 126 40 'DIGITS' STO 0 1 ROT FOR k 1 FNEG k FY\|^X k FDIV 2 k FY\|^X 1 FSUB FDIV FADD NEXT DUP FADD 2 FLN FDIV 1 12 FDIV FADD FPI FSQ 6 FDIV 2 FLN FSQ FDIV FADD ZZ\<-\->F DROP \->STR DUP HEAD "." + SWAP TAIL + \>> EVAL -> "1.000000000001237412575736110228719610648" The RPL program above requires the LongFloat library. Number of iterations = Ceil(W(10^n*ln(2))/ln(2)), where n = number of digits and W(x) is the Lambert W function. Best regards, Gerson. |
|||
02-10-2019, 12:48 PM
(This post was last modified: 02-10-2019 12:56 PM by Gerson W. Barbosa.)
Post: #26
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(02-10-2019 02:41 AM)Gerson W. Barbosa Wrote: %%HP: T(3)A(R)F(.); This is a more generic version that takes the desired number of digits, n, as an argument. As I am using an approximation for W(x), the estimation of the required number of iterations might not be exact for small values of n. Both programs are basically a straightforward conversion of Valentin's HP-71B program in post #21. 100 %%HP: T(3)A(R)F(.); \<< RCLF SWAP -105 CF -3 CF DUP 'DIGITS' STO ALOG 2 LN * LN DUP LN - LASTARG SWAP / + 2 LN / CEIL 0 1 ROT FOR k 1 FNEG k FY\|^X k FDIV 2 k FY\|^X 1 FSUB FDIV FADD NEXT DUP FADD 2 FLN FDIV 12 FINV FADD FPI FSQ 6 FDIV 2 FLN FSQ FDIV FADD ZZ\<-\->F NEG SWAP \->STR DUP SIZE ROT \=/ -51 FC? { "." } { "," } IFTE UNROT { DUP TAIL SWAP HEAD } { "0" } IFTE UNROT + + SWAP STOF \>> EVAL -> 1.000000000001237412575736110228719610646672874297732048196548443844171825640530428850913885586193525 (132.59 seconds on the HP 50g) 5 -> 0.99990 6 -> 1.00000 7 -> 1.000000 12 -> 1.00000000001 20 -> 1.0000000000012374126 40 -> 1.000000000001237412575736110228719610648 50 -> 1.0000000000012374125757361102287196106466728742977 |
|||
02-10-2019, 03:26 PM
(This post was last modified: 02-10-2019 04:19 PM by Albert Chan.)
Post: #27
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(02-10-2019 12:48 PM)Gerson W. Barbosa Wrote: Number of iterations = Ceil(W(10^n*ln(2))/ln(2)), where n = number of digits and W(x) is the Lambert W function. Above 50 digits numbers are confirmed corrrect. Can you explain how the iteration count formula is derived ? (01-23-2019 10:58 PM)Valentin Albillo Wrote: \[ \frac{1}{12}+\frac{\Pi^2}{6Ln^2(2)}+ \frac{2}{Ln(2)} \sum_{k=1}^\infty \frac{(-1)^k}{k(2^k-1)} \] I see that the sum gained about 1 bit precision with each iteration, so I use n / log10(2) ~ 3.322n It is slightly over-estimated, but not by much. Edit: the difference is the effect of 1/k, iterations ~ 3.322n - log2(3.322n) |
|||
02-10-2019, 04:06 PM
(This post was last modified: 02-10-2019 04:14 PM by Gerson W. Barbosa.)
Post: #28
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(02-10-2019 03:26 PM)Albert Chan Wrote:(02-10-2019 12:48 PM)Gerson W. Barbosa Wrote: Number of iterations = Ceil(W(10^n*ln(2))/ln(2)), where n = number of digits and W(x) is the Lambert W function. Starting with log₁₀(k.2ᵏ) = n, I solved k.2ᵏ = 10ⁿ for k. Well, actually W|A did :-) k = W(10ⁿ.ln(2))/ln(2) BTW, it's better to replace ALOG 2 LN * LN with 10 LN * 2 LN LN + just in case you want to evaluate it to one thousand digits or more: 1000 \<< RCLF SWAP -105 CF -3 CF DUP 'DIGITS' STO 10 LN * 2 LN LN + DUP LN - LASTARG SWAP / + 2 LN / CEIL 0 1 ROT FOR k 1 FNEG k FY\|^X k FDIV 2 k FY\|^X 1 FSUB FDIV FADD NEXT DUP FADD 2 FLN FDIV 12 FINV FADD FPI FSQ 6 FDIV 2 FLN FSQ FDIV FADD ZZ\<-\->F NEG SWAP \->STR DUP SIZE ROT \=/ -51 FC? { "." } { "," } IFTE UNROT { DUP TAIL SWAP HEAD } { "0" } IFTE UNROT + + SWAP STOF \>> EVAL -> 1.000000000001237412575736110228719610646672874297732048196548443844171825640530428850913885586193524976268453340086191658374509030019046729786005370140207590865397221066886209167246612158255597136947833662811711180501522046958297318386956749813586119403326983996836799698362386464361717810944715248515847063950123049027855289479337807074973722174863007602234598952082713436126867407223085711221417206013336683950248036912034243322848607544096465559742710057944068020597818546946376873631661338090760132715563114425400886965240835824220034845681146540332945848091156055661073808986770237768671181359710868112079802546002171398844199048674600407150411381977070159608769770037395721001869135492839448159377839257477067778776337799415286212226231921875049198549974749265675547171167195366657491492695699893916926664962342406045357897998136027548661020448361327035579555228205809418530092189232789163297481121766653027554098532310918458342580878445369891507372744436069036208883146409368525831685839774710 (363.97 seconds on the emulator) # B529h, 363 bytes |
|||
02-10-2019, 08:42 PM
Post: #29
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(02-10-2019 03:26 PM)Albert Chan Wrote: Edit: the difference is the effect of 1/k, iterations ~ 3.322n - log2(3.322n) Quite good estimation! It allows me to save 10 bytes by replacing LN + DUP LN - LASTARG SWAP / + 2 LN with SWAP DUP2 / LN + SWAP SQ Now # 62Fh, 353 bytes. |
|||
02-13-2019, 01:49 PM
(This post was last modified: 02-13-2019 01:51 PM by Albert Chan.)
Post: #30
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
Hi, Gerson W. Barbosa
Your code might be shortened and faster by removing (-1)^k factor. Since it pre-calculated iterations required, summing backwards may be more accurate. Example: With 4 digits precision, sum 4 terms, t = |sum| = 1 - 1/(2*3) + 1/(3*7) - 1/(4*15) Steps: t = 0 t = 1/60 - t = 0.01667 t = 1/21 - t = 0.04762 - t = 0.03095 t = 1/6 - t = 0.1667 - t = 0.1358 t = 1 - t = 0.8642 |
|||
02-15-2019, 04:29 PM
Post: #31
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(02-13-2019 01:49 PM)Albert Chan Wrote: Your code might be shortened and faster by removing (-1)^k factor. Sure I am aware of these, but initially I was not interested in optimizing the code for speed, only for size. Anyway, since the summation can be evaluated without resorting to exponentiation, as you've pointed out, there's no reason to use it, at least inside the loop: enter k; sign := -2*mod(k, 2) + 1; sum := 0; a := 2^k - 1; for i = k to 1 step -1; sum := sum + sign/(a*i); sign := -sign; a := a div 2 next i; display sum The RPL code, however, is now slightly slower, because either it's not properly optimized yet or the LongFloat FY^X function is very efficient for integer arguments. %%HP: T(3)A(R)F(.); \<< RCLF SWAP -105 CF -3 CF DUP 1 + 'DIGITS' STO 10 LN * 2 LN SWAP DUP2 / LN + SWAP / CEIL 2 OVER ^ 1 - OVER 2 MOD -2 * 1 + SWAP 0 4 ROLL 1 FOR i OVER i FMULT 4 PICK FMULT FINV FADD ROT NEG ROT 2 FDIV FIP ROT -1 STEP UNROT DROP2 DUP FADD 2 FLN FDIV 12 FINV FADD FPI FSQ 6 FDIV 2 FLN FSQ FDIV FADD ZZ\<-\->F NEG SWAP \->STR DUP SIZE ROT \=/ -51 FC? { "." } { "," } IFTE UNROT { DUP TAIL SWAP HEAD } { "0" } IFTE UNROT + + DUP SIZE " " REPL " " "" SREPL DROP SWAP STOF \>> # 1A8h, 428.5 bytes 141.34 seconds on my HP-50g, for 100 (previously, 132.59 seconds) 1 -> 0.9 2 -> 1.0 3 -> 0.999 4 -> 0.9999 5 -> 0.99998 6 -> 1.00000 7 -> 0.9999997 12 -> 1.00000000000 20 -> 1.0000000000012374125 40 -> 1.000000000001237412575736110228719610646 50 -> 1.0000000000012374125757361102287196106466728742977 100 -> 1.000000000001237412575736110228719610646672874297732048196548443844171825640530428850913885586193525 Regards, Gerson. |
|||
02-15-2019, 06:50 PM
Post: #32
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(02-15-2019 04:29 PM)Gerson W. Barbosa Wrote: a := 2^k - 1; Despite RPL does not see speedup, this is a great optimization ! Can the sign flipping code be removed, and possibly gain some speed ? Code: >>> sum, k = 0, 35 # for 12 digits accuracy |
|||
02-15-2019, 10:48 PM
(This post was last modified: 02-16-2019 01:15 AM by Gerson W. Barbosa.)
Post: #33
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
(02-15-2019 06:50 PM)Albert Chan Wrote: Can the sign flipping code be removed, and possibly gain some speed ? Oh, I see! %%HP: T(3)A(R)F(.); \<< RCLF SWAP -105 CF -3 CF DUP 1 + 'DIGITS' STO 10 LN * 2 LN SWAP DUP2 / LN + SWAP / CEIL 2 OVER ^ 1 - 0 ROT 1 FOR i OVER i FMULT FINV SWAP FSUB SWAP 2 FDIV FIP SWAP -1 STEP FSUB DUP FADD 2 FLN FDIV 12 FINV FADD FPI FSQ 6 FDIV 2 FLN FSQ FDIV FADD ZZ\<-\->F NEG SWAP \->STR DUP SIZE ROT \=/ -51 FC? { "." } { "," } IFTE UNROT { DUP TAIL SWAP HEAD } { "0" } IFTE UNROT + + DUP SIZE " " REPL " " "" SREPL DROP SWAP STOF \>> # 10B6h, 393.5 bytes Now 129.59 seconds for 100 digits. P.S.: Replaced NIP FNEG with FSUB at loop exit. I’d imagined the former were faster, but it turns our it’s actually 0.05 slower (although this is not conclusive after one measurement only). |
|||
02-23-2019, 04:58 AM
Post: #34
|
|||
|
|||
RE: [VA] SRC#002- Almost integers and other beasties
For things like (-1)^k in a list with k running over a large range, one can process by unrolling the loop twice. There needs to be either cleanup at the end or pre compute the first step for odd range.
|
|||
« Next Oldest | Next Newest »
|
User(s) browsing this thread: 4 Guest(s)