Post Reply 
[VA] SRC #017 - April 1st, 2024 Spring Special
04-01-2024, 06:59 PM
Post: #1
[VA] SRC #017 - April 1st, 2024 Spring Special
      
Hi, all,
Welcome to my new SRC #017 - April 1st, 2024 Spring Special

Once again April 1st is here and I want to celebrate both it and the new season with this Spring Special where I'm proposing a number of mini-challenges for you to tackle with your favorite vintage HP calc, plus interesting facts not widely known (if at all).
    Note: No hard rules, post here whatever you want as long as it's on topic and absolutely NO CODE PANELS (lest I'll consider you a troll only fit for my Ignore list,) but I'd appreciate it if you'd use vintage HP calcs (physical/virtual,) unless you're clueless as far as programming them is concerned.
As April 1st is also referred to as April Fool day, I'll take the last letter from April (L) and the two last ones from Fool (OL) to title each separate Section with a jocular prefix i.e. LOL !


1. LOL the First:  Squares

Let's start nice and easy. Square numbers have been a source of beauty and admiration since Pythagoras did his thing with them millennia ago. Just look at these small (related) beauties !

      375,5012 = 141,001,001,001
      751,0022 = 564,004,004,004

Now, if these are cute just consider the sheer amounts of beauty you'll discover when dealing with their bigger relatives, so this mini-challenge asks you to calculate the following 22 squares:

      1308349044300152392 = ?
      4712877147889716634938992 = ?
      257811083055916284179757382 = ?
      1414220828760672199498050500052 = ?


I know you can simply paste them in Wolfram Alpha or use the multiprecision library in your RPL model or mutiprecision canned software in any device or web site and get done with it, but it would be so lame that you'll risk ridicule. What I'm asking you to do is to program your own multiprecision squaring routine in your vintage HP calc and use it to get the results. It's not that difficult at all, you know ... that is, if you've got what it takes .. Wink

I'll post my own squaring routine that I wrote for the HP-71B from scratch, just a mere 483 bytes long and taking the form of a user-defined function so that it can be called right from the command line. It produces results like this, which you can profitably use for testing your code:

      31479267847267265637800300423742371877512 =

          9909443041999946686063013207932562462851613614976494122324802303779777224438001

Remember: if you decide to solve this mini-challenge, you must post both the (beautiful) results and the code which produces them, do not post results alone. Comments are most welcome, plus you might try getting more results like these.


2. LOL the Second:  GCD

OK, full throttle now. You may remember the GCD function appearing in my recent SRC #016 - Pi Day 2024 Special, where it was used in the Second appearance to count the number of co-prime pairs of random integers in a range and then this count was used to compute an approximation to \(\pi\).

     Sample GCD values:  GCD(35, 13) = 1,  GCD(35, 14) = 7 and GCD(35, 15) = 5

In this mini-challenge you must write code for any HP calc to find out the answer to this simple question:

      As we have that  GCD(15, 4) = 1  and  GCD(15, 5) = 5,  for what value of x is  GCD(15, x) = 2 ?

As always, you must post both results and code. Comments also most welcome. I'll post mine next Sat/Sun.


3. LOL the Third:  Random

You may be aware that advanced 12-digit HP models of old incorporated an excellent RNG (pseudo-Random Number Generator) which could generate a trillion (i.e. 1012) full 12-digit real (pseudo-)random numbers in the interval [0-1) before repeating. This RNG passes the Spectral test so it's extremely reliable for use in simulations and other advanced topics (e.g. multidimensional integration, Monte Carlo algorithms, etc.) without fear of any bias or short period degrading the results.

As far as I know, the implementations for the HP-71B, the RPL models, the HP-42S and Free42 (and perhaps many other HP models) are functionally identical so they produce exactly the same sequence of RNDs from the same seed. Now, this mini-challenge's question is:
    Two consecutive RNDs cannot be exactly equal or the generation of subsequent RNDs would be stuck in a loop. But how close can two consecutive RNDs actually be ?
Use RANDOMIZE 1 or the equivalent instruction to specify the value 1 as the seed and write code to output the list of ever closer consecutive RNDs and their index N in the sequence, like this:
    >RUN
          N       First RND         Next RND          |Difference|
     ----------------------------------------------------------------

          2      .731362440213     .77207218067      .040709740457
         13      5.64471991805E-2  6.30768172146E-2  6.6296180341E-3
        125      .805774019056     .803607575861     .002166443195

        ...
If you intend to go very deep into the sequence you'd better use a fast virtual model such as Emu71/Win or Free42 for faster results and/or greater depths. You might even boldly examine the entire trillion-strong sequence to find out the absolute closest pair !

I'll post my 3-line, ~90 byte solution which produces the above, plus a 4-line, 149-byte variant which stops right before a given pair of consecutive RNDs is generated so that you can generate them manually right from the command line (i.e. >RND;RND) and see for yourself how close they indeed are.

And remember, post code and results. What's your record ? Me, I've found a pair of consecutive RNDs which are only 0.00000000001 apart ! Can you do better ?


4. LOL the Fourth:  Logs

This is not a mini-challenge but a somewhat unexpected fact I've found, so you don't need to do anything but read on and eventually comment.

I've always thought that the two logarithmic functions available in the 10-digit HP-15C (LOG and LN, base 10 and base e respectively) would have essentially the same accuracy overall, but lo and behold, I've found that LOG seems to be substantially more accurate in some cases than LN.

For instance, let's consider computing logb(57) / logb(5), which should return exactly 7 for any base b ≥ 2. But the actual results are:

     LN(57) / LN(5)      ->  7.000000004      { 4 ulp from the exact value }
     LOG(57) / LOG(5)  ->  7                       { exact }

and the same happens with other powers N of 5, e.g. for N = 2 to 9 we have:

     N      LOG          LN        
 ---------------------------------
     2    2.00000001   2.00000001
     3    3            3.00000001
     4    4            4.00000001
     5    5.00000001   5.00000001
     6    6            6.00000002
     7    7            7.00000004
     8    8.00000001   8.00000002
     9    9            9.00000001
-----------------------------------
 Total ulps:    3           13
               { more than 400% larger error overall }

This much larger error puzzled me no end but I thought that perhaps the limited 10-digit accuracy (13-digit internally) might be playing a role and decided to test the same computations using the 12-digit (15-digit internally) HP-71B, which has LN, LOG10 and LOG2 (respectively base e, base 10 and base 2.) These were the results for power 8:

     >STD
     >LN(5^8) / LN(5)              ->  8.00000000004      { 4 ulp from the exact value }
     >LOG10(5^8) / L0G10(5)   ->  8                           { exact }
     >LOG2(5^8) / LOG2(5)      ->  7.99999999999      { 1 ulp from the exact value }

so we see that all three results differ among them, being in error by 4 ulp, 0 ulp and 1 ulp, respectively, and again the same happens with other powers N of 5, e.g. for N = 2 to 9 we have:

     N      LOG10           LN        
 ---------------------------------------
     2    2               2.00000000001
     3    3               3.00000000001
     4    3.99999999999   4.00000000001
     5    5               5.00000000001
     6    6.00000000001   6.00000000001
     7    7               6.99999999999
     8    8               8.00000000004
     9    8.99999999999   9.00000000002
----------------------------------------
  Total ulps:      3           12
               { 400% larger error overall }

and again the LN error is much larger, so having two extra digits didn't help at all. For LOG2 (not shown in the table above) the Total ulps for the range is 6 ulps, half LN's total error but twice LOG10's total error. In the end, LOG10 seems to be the most accurate logarithm available.

If you want to pursue the matter, you can try the above examples in your HP models, both 10-digit and 12-digit, from the first HP-35 up to the latest RPL models, to see if LOG10 and LN differ that much in their respective errors. Testing other range of values (here from 52 = 25 to 59 = 1,953,125) could be revealing as well. Are there arguments with even larger errors ?

At any rate, comments and discussion would be most welcome.


5. LOL the Fifth:  Gamma

Here's something peculiar I noticed while playing around with my HP-71B several decades ago, in Experimental Mathematics fashion.

Back then, I executed this loop directly from the command line to list the values of \(\Gamma\)(10-1), \(\Gamma\)(10-2), ..., \(\Gamma\)(10-10), and got the following unexpectedly peculiar results:

   >DESTROY ALL @ FOR N=1 TO 10 @ N;GAMMA(10^(-N))@ NEXT N

      N     \(\Gamma\)(10-N)
    -------------------
      1  9.51350769867
      2  99.4325851191
      3  999.423772485
      4  9999.42288323
      5  99999.4227942
      6  999999.422785
      7  9999999.42278
      8  99999999.4228
      9  999999999.423
     10  9999999999.42


where the evergrowing integer part of each value is clearly 10N - 1, while the fractional part seems to be quickly converging to some limit around ~0.42 but the 12-digit HP-71B lacks the accuracy needed to refine it further, so this mini-challenge is:

You must write code to find this limit to much greater accuracy (say 10-12 digits or more,) perhaps by simply using Free42 Decimal to compute the values. Once done, answer these two questions:
  • Can you estimate the most accurate value you got ? Because, as the integer part is exponentially growing, even the 34-digit Free42 Decimal will soon begin to lose digits in the fractional part, as it happened with the HP-71B results above. In that case, an RPL model with arbitrary multiprecision capability would surely help.
          
  • Can you identify the symbolic, closed form of that numeric value ? My IDENTIFY program v2.0 (unreleased) sure can.
As always, post here your results and code (comments also welcome,) and please try to steer clear of references such as Wolfram, MathWorld, Wikipedia, ... as you'd just ruin the pleasure of doing a little Experimental Mathematics all on your own.


6. LOL the Sixth:  Miscellanea

  1. [HP-71B specific] Try and deduce what result will be output by this expression without actually executing it: some regular numeric value ? Perhaps Inf or NaN ? An error message ? A long-running or indefinite internal loop ?

          NEIGHBOR(INTEGRAL(FNROOT(EPS,EPS,EPS),EPS,EPS,EPS),GAMMA(EPS))

    Once you've thoughtfully made your deduction, execute the expression and see what comes out. Was your deduction correct ? Can you explain the result obtained ?

    Please post your deduction & comments on the result.
          
  2. [HP-71B specific] I executed this command-line expression on my HP-71B but it just resulted in System Error. What did I do wrong ? Bad addresses ? Forbidden range for PEEK$ ? Can you find out what's wrong ?

          FOR I=64204 TO 64215 @ DISP CHR$(HTD(REV$(PEEK$(DTH$(2*I),2)))); @ NEXT I
          
  3. Some nice results I got (you can check the s with Free42 Decimal):

          95888      = 309.657875727390470000000975517...
          22008840 = 4691.3580123456789961013...      

          (HP-71B specific, RPL/12-digit models too, maybe other 12-digit models) In radians:

        >EXP(ACOS(430/433))  ->  1.12500000006     { ~  9/8 }
        >EXP(ACOS(538/541))  ->  1.11111111113     { ~ 10/9 }

    Notice that 433 = 430 + 3 and 541 = 538 + 3, i.e. both fractions are of the form N / (N+3) and produce results of the form (N+1) / N so perhaps there's some hidden pattern at large here. Or not !
          
  4. [HP-71B specific] What does this HP-71B user-defined function compute ?

          10  DEF FNC(M,N)=M!/N!/(M-N)!

          
  5. [HP-71B specific] The HP-71B does not allow for variable names beginning with a 2-letter or more prefix (e.g. AB, CD7, MYTAXES, etc.) as many other pocket computers do, but executing this assignment in a program or from the keyboard ...

          FORM=STOP

    ... doesn't result in a Syntax Error so what gives ?
          
  6. Last but not least, you're probably acquainted with the number 42, not only because it's the model number of the renowned HP-42S, later simulated by the Free42 virtual version, but also because according to The Hitchhiker's Guide to the Galaxy 42 is "The Answer to the Ultimate Question of Life, the Universe, and Everything", which is saying a whole lot.

    However, this doesn't end 42's importance to this Universe and to prove the point you must write code to compute to full accuracy this simple function S(x) given the argument x:
    \[ S(x) = \frac{1}{\sum_{n=0}^{\infty} \binom {2 n}{n}^3 \frac{{nx \space + \space 5}}{2^{12 n \space + \space 4}}} \]Use your code to compute S(42). Astounding, isn't it ?  Smile

    Do not use any built-in summation capability but write and use instead your own code and post here both code and result. I'll post my simple 3-line HP-71B user-defined function in a few days.



That's all. Waiting for your clever solutions and/or comments galore, I'll post mine next Saturday/Sunday so you've got plenty of time to concoct and polish your code. And no CODE pannels, please.

V.

  
All My Articles & other Materials here:  Valentin Albillo's HP Collection
 
Visit this user's website Find all posts by this user
Quote this message in a reply
04-02-2024, 08:29 AM (This post was last modified: 04-02-2024 08:36 AM by J-F Garnier.)
Post: #2
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-01-2024 06:59 PM)Valentin Albillo Wrote:        As we have that  GCD(15, 4) = 1  and  GCD(15, 5) = 5,  for what value of x is  GCD(15, x) = 2 ?

OK, let's try to solve GCD(15,X)=2 for X on the 71B w/ Math ROM.
Since GCD only accepts integer values, I had to cast the X variable to the INT type.
Choosing 0 and 10 as the initial guesses:

>FNROOT(0,10,GCD(15,INT(FVAR))-2)
12.9999999999

We can safely round the result to X=13.
An interesting result, isn't it?


Quote:[*] [HP-71B specific] Try and deduce what result will be output by this expression without actually executing it: some regular numeric value ? Perhaps Inf or NaN ? An error message ? A long-running or indefinite internal loop ?

      NEIGHBOR(INTEGRAL(FNROOT(EPS,EPS,EPS),EPS,EPS,EPS),GAMMA(EPS))

I miserably failed to predict the result !
But I learnt something about the 71B solver...


Quote:[*] [HP-71B specific] I executed this command-line expression on my HP-71B but it just resulted in System Error. What did I do wrong ? Bad addresses ? Forbidden range for PEEK$ ? Can you find out what's wrong ?

      FOR I=64204 TO 64215 @ DISP CHR$(HTD(REV$(PEEK$(DTH$(2*I),2)))); @ NEXT I

Ah ! Nothing wrong, I even didn't need to try ...
      
Quote:[*] [HP-71B specific] What does this HP-71B user-defined function compute ?

      10  DEF FNC(M,N)=M!/N!/(M-N)!

It doesn't do much, I'm afraid.

Quote:[*] [HP-71B specific] The HP-71B does not allow for variable names beginning with a 2-letter or more prefix (e.g. AB, CD7, MYTAXES, etc.) as many other pocket computers do, but executing this assignment in a program or from the keyboard ...

      FORM=STOP

... doesn't result in a Syntax Error so what gives ?

Ah! another new "hidden feature" of the 71B ?

I well remember the undocumented MEMORY function, that accepts up to 2 parameters, but isn't that useful (do first DESTROY ALL to get consistent results):
>MEMORY(1,2)
1


More results later...

J-F
Visit this user's website Find all posts by this user
Quote this message in a reply
04-02-2024, 07:38 PM
Post: #3
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
Interesting challenges as always, Valentin!

I couldn't resist the last one, S(42). Here is a quick and dirty RPL program using a FOR loop as a fake DO loop with a counter.

\<< 0. 0. 99.
FOR k OVER k * 5. +
k DUP 2. * SWAP COMB 3. ^ *
2. k 12. * 4. + ^ /
OVER + DUP ROT \=/
1. 99. IFTE
STEP INV SWAP DROP
\>>


I also noticed that using the bottom part of the expression (remove the INV from the last line), a value of about 8.45 will return a number close to 1/10 of S(42).
Find all posts by this user
Quote this message in a reply
04-02-2024, 07:41 PM (This post was last modified: 04-02-2024 07:51 PM by J-F Garnier.)
Post: #4
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-01-2024 06:59 PM)Valentin Albillo Wrote:  I've always thought that the two logarithmic functions available in the 10-digit HP-15C (LOG and LN, base 10 and base e respectively) would have essentially the same accuracy overall, but lo and behold, I've found that LOG seems to be substantially more accurate in some cases than LN.

I remember this has been reported a few times in the past, maybe in connection with some solutions of your challenges/SRCs, but I can't find any reference again (can you?).

I've never been convinced by this effect (decimal LOG better than natural LN), because internally the decimal log is computed by:
LOG(x) = LN(x) / LN(10) , using 3 extra guard digits.
So there is no reason for LOG to be more accurate, on the contrary it may be marginally less accurate.

Let's see your 15c example (a 10-digit machine) - with minor corrections:

    N    LOG(5N)/LOG(5)  LN(5N)/LN(5)
 ---------------------------------
     2    2.000000001   2.000000001
     3    3             3.000000001
     4    4             4.000000001
     5    5.000000001   5.000000001
     6    6             6.000000002
     7    7             7.000000004
     8    8.000000001   8.000000002
     9    9             9.000000001
-----------------------------------
 Total ulps:    3           13
               { more than 400% larger error overall }

but let's try another number: 11N instead of 5N:

    N    LOG(11N)/LOG(11)  LN(11N)/LN(11)
 ---------------------------------
     2    2               2
     3    3               3
     4    4.000000001     4
     5    5.000000001     4.999999998
     6    6.000000001     6.000000001
     7    7.000000001     7
     8    8.000000001     7.999999998
     9    9.000000001     9.000000001
-----------------------------------
 Total ulps:    6           6
               { now similar error overall }

Doing now the same test on the 71B (a 12-digit machine):

    N    LGT(11N)/LGT(11)  LN(11N)/LN(11)
 ---------------------------------------
     2    2               2
     3    2.99999999999   3
     4    3.99999999999   4
     5    4.99999999999   5
     6    5.99999999999   6
     7    6.99999999999   7
     8    7.99999999999   8
     9    8.99999999999   9
----------------------------------------
  Total ulps:      7           0
               { here, much larger error overall for the decimal LOG case }

So it's clear to me that we can't say that the decimal LOG provides more accurate results than the natural LN, overall.

The question we may ask is: why does it seem that LOG is better for expressions using a certain number such as 5N, and why is LN better for other numbers such as 11N ?

J-F
Visit this user's website Find all posts by this user
Quote this message in a reply
04-03-2024, 06:13 PM (This post was last modified: 04-04-2024 07:04 PM by C.Ret.)
Post: #5
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-02-2024 08:29 AM)J-F Garnier Wrote:  
(04-01-2024 06:59 PM)Valentin Albillo Wrote:        As we have that  GCD(15, 4) = 1  and  GCD(15, 5) = 5,  for what value of x is  GCD(15, x) = 2 ?

OK, let's try to solve GCD(15,X)=2 for X on the 71B w/ Math ROM.
Since GCD only accepts integer values, I had to cast the X variable to the INT type.
Choosing 0 and 10 as the initial guesses:

>FNROOT(0,10,GCD(15,INT(FVAR))-2)
12.9999999999

We can safely round the result to X=13.
An interesting result, isn't it?

This is a surprising result!

I believe that looking for x such that GCD(15,x)= 2 is like looking for a hairy fish (a very common fish at the very beginning of the April River).

Here is my code for any HP-71B to find x: 10 DISP “GCD( 15 , NaN ) = 2” and what it display:                  GCD( 15 , NaN ) = 2 
[Image: attachment.php?aid=13424]
   


But, I may have start by the first apriL foOL :

Here is my program to compute a large square on the HP-71B:

10 DESTROY ALL @ DIM A$[39] @ INPUT A$ @ L=LEN(A$) @ DIM R$[2*L] @ R$=SPACE$(48,2*L)
20 FOR K=L TO 1 STEP -1 @ C=0 @ FOR J=L TO 1 STEP -1 @ I=K+J
30 X=C+VAL(R$[I,I])+VAL(A$[J,J])*VAL(A$[K,K]) @ R$[I,I]=CHR$(48+MOD(X,10)) @ C=X DIV 10
40 NEXT J @ R$[K,K]=CHR$(48+C) @ DISP R$[K+L] @ NEXT K @ DISP R$ @ BEEP @ PAUSE


The calculation takes longer as the number is larger.
One can follow the progress of the computation as the figures are displayed as soon as they are determined. That is to say starting from the last towards the first.

[RUN]
? _
? 141422082876067219949805050005_ [END LINE]
                              (prgm)
> 5                            (prgm)
> 25                           (prgm)
> 025                           (prgm)
> 0025                          (prgm)
> 00025                          (prgm)
> 500025                         (prgm)
. . .
> 2520222202205205000550500025           (prgm)
> 22520222202205205000550500025          (prgm)
~ biiiip ~
> 020000205525005225202000505202222520222202205205000550500025     (susp)


You will note the presence of a leading zero which is important but which is not systematic. Try to calculate 99999999².

A small figure which shows how the calculation is done:[Image: attachment.php?aid=13426]
   

several edit to correct broken English, code syntax, insert illustrations and correct typos
Find all posts by this user
Quote this message in a reply
04-03-2024, 08:55 PM (This post was last modified: 04-03-2024 09:28 PM by Gerson W. Barbosa.)
Post: #6
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
LOL the First

HP-75C program:

10 OPTION BASE 0
15 INTEGER I,J,N
20 N=5
25 B=1000000
30 DIM A(9),B(5),C(12)
35 REM DIM A(N+4), B(N), C(2*N+2); N>1
40 FOR I=0 TO 2*N+2 @ C(I)=0 @ NEXT I
45 FOR I=0 TO N+4 @ A(I)=0 @ NEXT I
50 FOR I=1 TO N
55 READ A(I)
60 B(I)=A(I)
65 NEXT I
70 FOR I=N TO 1 STEP -1
75 T1=0 @ A1=0
80 FOR J=N-I+4 TO -1 STEP -1
85 C2=(T1+B(I)*A(J+1))/B
90 A1=FP(C2)*B
95 C(I+J)=C(I+J)+A1
100 T1=IP(C2)
105 NEXT J
110 NEXT I
115 FOR I=2*N TO 2 STEP -1
120 T=C(I)/B
125 C(I)=FP(T)*B
130 C(I-1)=C(I-1)+IP(T)
135 NEXT I
140 FOR I=0 TO 2*N-1
145 DISP C(I);
150 NEXT I
155 END
160 DATA 141422,82876,67219,949805,50005

>RUN
20000 205525 5225 202000 505202 222520 222202 205205 550 500025


That is,

141422082876067219949805050005^2 =

20000205525005225202000505202222520222202205205000550500025



-----

LOL the Fifth:

I have digressed on this one and haven't done what has been asked (No cigar, I guess :-). I'll just say the fractional part tends to 1 minus a known mathematical constant to which the following is a pandigital appoximation in RPL algebraic expression format, good to twelve digits:

'INV(√(3+7/((29/10)^8-INV(SQ((4^5)^INV(6))))))'

On the HP 50g in approximate mode, '1 - INV(√(3+7/((29/10)^8-INV(SQ((4^5)^INV(6))))))' should return the same numeric result as '1+Psi(1)'.

Edited to fix a typo
Find all posts by this user
Quote this message in a reply
04-04-2024, 09:03 PM (This post was last modified: 04-04-2024 09:09 PM by J-F Garnier.)
Post: #7
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-01-2024 06:59 PM)Valentin Albillo Wrote:  Two consecutive RNDs cannot be exactly equal or the generation of subsequent RNDs would be stuck in a loop.

It's not correct. This would assume that the next RND values are fully determined by the last RND value, and this is wrong.
The RND value is determined by the internal seed, which is stored with 15 digits.
Or, in other words,we can't predict the next RND value from the last one (well, the possibilities for the next RND are limited).
To illustrate it, we can get the same RND value in two different sequences:

>RANDOMIZE .636248123586
>RND, RND, RND
.14159265359     .494478890124     .825547412541

>RANDOMIZE .90365957958
>RND, RND, RND, RND
2.71250347884E-2     .14159265359     .416751399163     .130703385847

(done on a 71B, but as Valentin noted, the results are the same for many/all Saturn-based machines)

Quote:But how close can two consecutive RNDs actually be ?

As a consequence of the above analysis, it could be possible to get two consecutive RND values that are equal.
Matter of fact, I have one example. Can you find it? :-)

J-F
Visit this user's website Find all posts by this user
Quote this message in a reply
04-06-2024, 01:16 AM
Post: #8
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
For the function S(X); notice that:
1). 2^(12N+4) = 2^(12N)*16, so 16 can go outside the summation.
2). 2^(12(N+1))/2^(12N) = 2^12 = 4096.
3). COMB(2N, N) = (2*N)!/(N!^2).
COMB(2(N+1), N+1)/COMB(2N, N) = ((2*(N+1))!/((N+1)!^2))/((2*N)!/(N!^2))
= (4*N+2)/(N+1) = 2*(2-1/(N+1)).
Having (COMB(2N, N)^3)/(2^(12N)) stored in register 03, you can get to (COMB(2(N+1), N+1)^3)/(2^(12(N+1)) by multiplying register 03 by ((2*(2-1/(N+1)))^3)/4096
= ((2-1/(N+1))^3)/512.
With X in register 01, N+1 in register 02 and summation in register 04, we have the next program for the free42:
PHP Code:
01 LBL “S”
02 STO 01
03 0
04 STO 02
05 1
06 STO 03
07 5
08 STO 04
09 LBL 00
10 2
11 1
12 RCL
02
13 STO 02
14 1
/X
15 
-
16 3
17 Y↑X
18 512
19 ÷
20 RCL× 03
21 STO 03
22 RCL 01
23 RCL× 02
24 5
25 
+
26 ×
27 STO
04
28 1E-34
29 X
<Y?
30 GTO 00
31 RCL 04
32 1
/X
33 16
34 ×
35 END 
S(42) = 3.141592653589793238462643383279504
wow, pi again!
Find all posts by this user
Quote this message in a reply
04-06-2024, 10:30 AM (This post was last modified: 04-06-2024 05:09 PM by J-F Garnier.)
Post: #9
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-01-2024 06:59 PM)Valentin Albillo Wrote:  how close can two consecutive RNDs actually be ?
Use RANDOMIZE 1 or the equivalent instruction to specify the value 1 as the seed and write code to output the list of ever closer consecutive RNDs and their index N in the sequence

OK, let's try the brute force on a super-fast HP-71B emulator:

10 RANDOMIZE 1 @ X=RND @ I=1 @ M=1
20 Y=RND @ I=I+1 @ IF ABS(X-Y)>=M THEN X=Y @ GOTO 20
30 M=ABS(X-Y) @ PRINT I,X,Y,M @ GOTO 20

2 .731362440213 .77207218067 .040709740457
13 5.64471991805E-2 6.30768172146E-2 6.6296180341E-3
125 .805774019056 .803607575861 .002166443195
316 .128424219936 .126476247276 .00194797266
378 .128629571043 .127765838222 .000863732821
1746 .657235932954 .658084547243 .000848614289
...
later, much later ...
38181163 .151576837566 .151576827517 1.0049E-8
... and even much later
157373808 .477539626899 .477539622968 3.931E-9
322950317 .325324468276 .325324470156 1.88E-9
... getting closer to the record:
431380423 .275319512003 .275319511289 7.14E-10
...[edit:]got it:
1838286534 .564079556829 .564079556839 1.E-11

[edit2:]
Sopped with index exceeding 5 x 109 with no better result.
No identical consecutive RNDs found, but only a small fraction of the RND period has been explored.

J-F
Visit this user's website Find all posts by this user
Quote this message in a reply
04-07-2024, 03:01 AM (This post was last modified: 04-07-2024 10:25 PM by Valentin Albillo.)
Post: #10
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
  
Hi, all,

I've got already fully formatted and ready-to-post my looong Solutions post (I might actually split it in three parts ... or not) so this is the last chance for those of you who still want to post something (code & results, comments) before I post my original solutions & comments revealing it all. The status of things done or still pending is as follows:
  • LOL the First: Several people have posted code but so far only one of the 4 squares has been computed and posted, all other three are still pending, and no comments whatsoever on the nice pattern they all exhibit or on the code posted.
      
  • LOL the Second: Only one entry so far. More detail or comments on it ? Any additional takers ?
      
  • LOL the Third: This was well covered by J-F, interesting and revealing comments. 
      
  • LOL the Fourth: Again, this one has been well covered by J-F but no one else commented anything or offered any results for the questions asked.
      
  • LOL the Fifth: Only Gerson offered some hints about this one but he posted neither code nor results, and no one else posted a thing.
      
  • LOL the Sixth:

    1. Very little detail or comment and the result has not been posted.
        
    2. Ditto, only minimal detail and no explanation or comments posted.
        
    3. No one posted any comments or similar nice results, it's been totally ignored.
        
    4. Only the briefest of comments (like Earth's entry in Hitchhiker: "Mostly harmless") and no explanations.
        
    5. An additional mention of a similar case but no explanations for any of them.
        
    6. This has been more successful, with working code posted but only the one result.

That's all. If you need a few more days to finish something, just say it and I'll oblige. Frankly, I'm surprised at the lack of entries and comments, I deemed all sections interesting and perfectly within the reach of most calc programmers with minimal effort. Live and learn !

Waiting for your 11th-hour productions,
V.
Edit: typo

  
All My Articles & other Materials here:  Valentin Albillo's HP Collection
 
Visit this user's website Find all posts by this user
Quote this message in a reply
04-07-2024, 11:10 PM (This post was last modified: 04-08-2024 03:12 AM by Gerson W. Barbosa.)
Post: #11
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-07-2024 03:01 AM)Valentin Albillo Wrote:    
[*] LOL the Fifth: Only Gerson offered some hints about this one but he posted neither code nor results, and no one else posted a thing.
  

Hello, Valentín,

Unusually busy week here. I'll try to fix that, even if only by a litte.

As I have suggested, that has to do with the Euler-Mascheroni constant, denoted by γ (lower-case gamma) .

The wp34s has γ built-in, but I guess Free42 should be more appropriate here. The following table, produced with help of one small RPN program, illustrates our attempt to get as many digits as possible on Free42, using that approach:

     γ = 0.5772156649015328606065120900824024 
          
 1 - γ = 0.4227843350984671393934879099175976 

      N     FRAC(Γ(10⁻ᴺ))         ABS(1 - γ - FRAC(Γ(10⁻ᴺ)))
    ---------------------------------------------------------
      1  0.51350769866873183629        9.07233635703E-02
      2  0.43258511915060371353        9.80078405214E-03
      3  0.42377248459546611498        9.88149496999E-04   
      4  0.42288323162419080574        9.89965257237E-05
      5  0.42279422556767349323        9.89046920635E-06
      6  0.42278532415355498927        9.89055087500E-07
      7  0.42278443400405759740        9.89055904580E-08
      8  0.42278434498902700193        9.89055986253E-09
      9  0.42278433608752313381        9.89055994420E-10
     10  0.42278433519737273892        9.89055995237E-11
     11  0.42278433510835769935        9.89055995288E-12
     12  0.42278433509945619539        9.89055997912E-13
     13  0.42278433509856604495        9.89055555121E-14
     14  0.42278433509847702999        9.89059651209E-15
     15  0.4227843350984681235         9.84106512090E-16
     16  0.422784335098467242          1.02606512090E-16
     17  0.42278433509846684           2.9939348791OE-16 
   ---------------------------------------------------------


From the table, we notice that the constant approaches 1 - γ as N increases. The maximum accuracy is reached when N = 16, when the first 15 digits are correct. From this point on, the accuracy degrades, as the Free42 precision is limited to 34 digits.

Here we also notice that the mantissas of the errors, the second column in the table, appear to tend to another constant. Regardless of any attempt to identify it, we can try to use it to get a few more correct digits, starting with N = 16 and down to N = 13, when the maximum accuracy is reached:
  
     16  0.422784335098467242 - 9.89055997912E-17 = 0.4227843350984671430944002088
     15  0.4227843350984681235 - 9.89055997912E-16 = 0.422784335098467134444002088
     14  0.42278433509847702999 - 9.89055997912E-15 = 0.42278433509846713943002088
     13  0.422784335098566044949 - 9.89055997912E-14 = 0.4227843350984671393492088                             
     12  0.4227843350994561953914 - 9.89055997912E-13 = 0.422784335098467139393488

Now we have about 50% more correct digits, 23. This scheme should give up to eight or nine correct digits on the HP-42S, but I have it to check it out.

Best regards,

Gerson.

------------------------------------------------------------

00 { 58-Byte Prgm }
01▸LBL "L5th"
02 1
03 0.5772156649015328606065120900824024
04 -
05 X<>Y
06 +/-
07 10↑X
08 GAMMA
09 FP
10 -
11 LASTX
12 X<>Y
13 ABS
14 END

------------------------------------------------------------

P.S.:

By using the constant 9.89055995288 we can get up to 7 correct digits of γ on the HP-42S and up to 23 on Free42:


00 { 38-Byte Prgm }
01▸LBL "E_M"
02 RCL ST X
03 +/-
04 10↑X
05 GAMMA
06 FP
07 X<>Y
08 NOT
09 10↑X
10 9.89055995288
11 ×
12 -
13 +/-
14 1
15 +
16 END

4 XEQ "E_M" ->

0.5772156756

11 XEQ "E_M" ->

0.57721566490153286060651

 ------------------------------------------------------------
Find all posts by this user
Quote this message in a reply
04-08-2024, 08:34 AM
Post: #12
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
LOL the Sixth:

I've no HP-71B, but I had a Commodore 64 so I can throw a couple of guesses:

LOL 6.2:

A famous challenge in those naive days was to write a program resulting in the largest number of errors. Of course the easiest way to accomplish the task is to dump the ROM area where the messages are actually stored and embellish the output to make it look like a real error message...

LOL 6.5:

The Commodore 64's screen editor doesn't accept more than 80 characters when editing a program line: there are a lot of dirty tricks to circumvent this limitation, but of course one always starts with the cleanest ones.

For example you could omit useless blanks between instructions and gain a little space to squeeze that extra statement... but sometimes you end up with an ambiguous statement that leaves the parser (and you!) scratching its head in puzzlement:

FORM=STOP is a variable assignment or the beginning of a loop: FOR M = S TO P ?

Cheers!
Find all posts by this user
Quote this message in a reply
04-09-2024, 05:19 PM (This post was last modified: 04-10-2024 01:50 AM by Gerson W. Barbosa.)
Post: #13
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-07-2024 11:10 PM)Gerson W. Barbosa Wrote:  By using the constant 9.89055995288 we can get up to 7 correct digits of γ on the HP-42S and up to 23 on Free42:

00 { 38-Byte Prgm }
01▸LBL "E_M"
02 RCL ST X
03 +/-
04 10↑X
05 GAMMA
06 FP
07 X<>Y
08 NOT
09 10↑X
10 9.89055995288
11 ×
12 -
13 +/-
14 1
15 +
16 END

4 XEQ "E_M" ->

0.5772156756

11 XEQ "E_M" ->

0.57721566490153286060651

The constant actually goes like

K = 9.890559953279725553953956515

But there’s a better way to get those extra digits. We only have to use the GAMMA function twice:

00 { 29-Byte Prgm }
01▸LBL "E_M"
02 1
03 -11
04 10↑X
05 GAMMA
06 LASTX
07 R↓
08 FP
09 -
10 R↑
11 +/-
12 GAMMA
13 FP
14 -
15 2
16 ÷
17 END


This will return

0.577215664901532860606565,

good to 22 digits, but nine bytes shorter.

Or we can add a couple of steps and get all 34 digits right, at the cost of another nineteen bites:

17 5.29099175976ᴇ-23
18 -


But then again the following should be more simple and two bytes shorter:

00 { 46-Byte Prgm }
01▸LBL "E_M"
02 5.772156649015328606065120900824024ᴇ-1
03 END


------------------------------------------------------------

Update

If all we want to do is to obtain γ from Γ then this 42-byte Free42 program is better:

00 { 42-Byte Prgm }
01▸LBL "E_M"
02 -11
03 10↑X
04 GAMMA
05 LASTX
06 +/-
07 GAMMA
08 +
09 -2
10 ÷
11 5.29099175976ᴇ-23
12 -
13 END

XEQ "E_M" ->

5.772156649015328606065120900824024E-1
Find all posts by this user
Quote this message in a reply
04-10-2024, 08:51 PM
Post: #14
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
      
Hi, all,

First of all, thanks to all of you who were interested in this April 1st thread and in particular to those who posted code, results and/or comments, namely J-F Garnier, Juan14, C.Ret, John Keith, Gerson W. Barbosa and JoJo1973. Thank you very much for your interest and continued appreciation.

However, I feel somewhat let down by the fact that only 6 people 6 took the trouble to post anything at all (out of the ~9,400 members currently registered,) but that's life and now it's time for my original solutions and comments, which for sheer message-length reasons I'll post two LOL at a time, so let's get the party started with the first two, i.e. LOL the First: Squares and LOL the Second: GCD ...


1. LOL the First:  Squares

This mini-challenge asks you to calculate the following 22 squares. You must program your own multiprecision squaring routine in your vintage HP calc and use it to get the results.

      1308349044300152392 = ?
      4712877147889716634938992 = ?
      257811083055916284179757382 = ?
      1414220828760672199498050500052 = ?


My original solution is this 10-line, 415-byte HP-71B user-defined function, which accepts the number to be squared as a string and returns its square as another string:

  1  DEF FNS$[200](P$) @ STD @ DIM M,L,P,I,J,U,A$[206]
  2  OPTION BASE 1 @ A$=FNL$(P$) @ U=LEN(A$) DIV 6 @ K=10^6
  3  DIM A(U),C(2*U),C$[12*U] @ MAT A=ZER @ MAT C=ZER
  4  FOR I=1 TO U @ A(U+1-I)=VAL(A$[I*6-5,I*6]) @ NEXT I @ FOR I=1 TO U
  5  M=A(I) @ L=I @ FOR J=1 TO U @ P=M*A(J) @ C(L)=C(L)+P @ P=RES
  6  IF P>=K THEN C(L)=RMD(P,K) @ C(L+1)=C(L+1)+P DIV K
  7  L=L+1 @ NEXT J @ NEXT I @ I=2*U+1 @ REPEAT @ I=I-1 @ UNTIL C(I)
  8  C$=STR$(C(I)) @ FOR I=I-1 TO 1 STEP -1
  9  C$=C$&FNL$(STR$(C(I))) @ NEXT I @ FNS$=C$ @ END DEF
 10  DEF FNL$[206](A$) @ P=RMD(LEN(A$),6) @ FNL$=RPT$("0",6*(P#0)-P)&A$

    Note: Branching isn't needed thanks to the keywords provided by the JPC ROM (REPEAT..UNTIL). It also uses MAT..ZER from the Math ROM and RPT$ from the STRNGLEX LEX file for speed and convenience, all of them replaceable by slower BASIC code if necessary.
    • Lines 1-3 do the initialization.
    • Line 4 splits the input string into an array of 6-digit numeric values.
    • Lines 5-7 perform the multiprecision squaring and locate the leftmost nonzero element.
    • Lines 8-9 conjoin the result array elements into a result string and return it.
    • Line 10 is an auxiliary user-defined function which adds needed leftmost 0s, if any.
My program chops the input into 6-digit values (multiplying two 6-digit numbers never exceeds the 12-digit range of the HP-71B) instead of doing the squaring digit by digit, which would be about 36x more processing and so ~30x slower.

Also, using strings for input/ouput is extremely convenient for multiprecision values, as the user simply enters the value between quotes and the result can be directly stored in a string variable for further processing e.g. combined with other similar multiprecision UDFs. For instance, raising a multiprecision value to the fourth power is as simple as FNS$(FNS$(value))

Finally, there's no limit to the size of the result squares other than maximum string length (65,500+ characters i.e. digits). As listed above, it allows for up to 200-digit squares but that limit can be changed by simply replacing the constants 200 and 206 by the desired maximum size.

Now let's use FNS$ from the command line to quickly compute the requested squares:
    >DESTROY ALL
    >FNS$("130834904430015239")

          17117772217211221211117217772227121

    >FNS$("471287714788971663493899")

          222112110111011100020110111110102200012010222201

    >FNS$("25781108305591628417975738")

          664665545464645645665646644665564654546645556644644

    >FNS$("141422082876067219949805050005")

          20000205525005225202000505202222520222202205205000550500025
As you can see, these beautiful large squares contain solely the digits (1,2,7), (0,1,2), (4,5,6) and (0,2,5), respectively. Alas, they aren't the only such squares, see if you can find some more !  Smile

Additional comments:  No one posted all four squares so three of them were left missing (though once you've created your multiprecision squaring program it's all too easy to use it to compute the squares and post them) and no one posted any more examples of this pattern so here you are, a few more:

      774700591300020347197007492 =

            6001610061606011616611060006010661000616100111161001

      9425754295779433269877982 =

            888448440444044400080440444440408800048040888804


2. LOL the Second:  GCD

You may remember the GCD function. In this mini-challenge you must write code for any HP calc to find out the answer to this simple question:

      As we have that  GCD(15, 4) = 1  and  GCD(15, 5) = 5,  for what value of x is  GCD(15, x) = 2 ?



In the question above there's no mention of the GCD keyword from the JPC ROM, which is only available for the HP-71B and is limited to integer arguments, while this mini-challenge (which can be solved using almost any programmable scientific HP calc, from the HP-19C/29C upwards) deals with the GCD (Greatest Common Divisor) math function.

Mathematically, the GCD of two integers is the largest positive integer that divides each of them. But as it happens with the factorial, which can be extended to non-integer arguments via the \(\Gamma\) function, or with the Harmonic series, for which the digamma function does likewise, the GCD function can also be extended to non-integer arguments, and even to complex ones. This is done using this simple formula (which you can find in Wikipedia, no need to hunt obscure scholar papers for it):

      [Image: GCD%2002.jpg]

where n has to be odd and ≥ 1 but m can be complex.

This 2-line, 89-byte HP-71B user-defined function accepts a complex argument K and a positive odd integer argument N and returns GCD(K,N) as per the above formula:
    5  DEF FNG(K,N) @ COMPLEX P,R @ P=1 @ R=-2*PI*K/N @ FOR M=0 TO N-1
    6  P=P*(1+RECT((1,R*M))) @ NEXT M @ FNG=LOG(P)/LOG(2) @ END DEF
We can check that it works fine by using this 4-line, 106-byte driver program to test it for 100 pairs of random integer arguments, excluding co-prime pairs from the listing. The first integer is passed as a complex value and the second integer is forced to be odd:
    1  DESTROY ALL @ RANDOMIZE 1 @ RADIANS
    2  FOR I=1 TO 100 @ A=INT(RND*50)+1 @ B=2*INT(RND*25)+1
    3  G=IROUND(REPT(FNG((A,0),B))) @ IF G#1 THEN DISP USING "4X,4D,4D";A,B,G
    4  NEXT I @ END


    >RUN
           6   3      3
          38  19     19
          22  33     11
           9  21      3
          30  45     15
            ...
so it seems to be working Ok. Of course we can also call the UDF from the keyboard, e.g.:
    >FNG((30,0),45)  ->  (15.000000002, 3.81114434334E-9), i.e. 15
    >FNG((1/2,0),3)  ->  (1.79248125035, -2.26618007092)
Now, for real arguments we can use  cos x = ½·(eix + e-ix)  to reduce the formula to the real product:

      [Image: GCD%2003.jpg]

The user-defined function now becomes simpler (71 bytes, no comples variables or functions) and faster:
    5  DEF FNG(N,M) @ P=1 @ R=PI*M/N @ FOR K=1 TO (N-1)/2
    6  P=P*COS(R*K) @ NEXT K @ FNG=N+LOG2(P*P) @ END DEF


          Note: do not "optimize" LOG2(P*P) to 2*LOG2(P)
Let's apply it to this mini-challenge's question. First, let's evaluate GCD(15, x) for x in [4..5] at steps of 0.2:
    >DESTROY ALL
    >FOR I=4 TO 5 STEP .2 @ DISP USING "3X,D.D,2X,S2D.7D";I,FNG(15,I) @ NEXT I

      x    GCD(15, x) 
    ------------------
     4.0   +1.0000000  <- GCD(15,4) = 1
     4.2   +3.0660268
     4.4   +1.6288312
     4.6   +2.2305063
     4.8   +5.0211098
     5.0   +5.0000000  <- GCD(15,5) = 5
and a cursory inspection reveals that there are at least three solutions to the original question "For what value of x is  GCD(15, x) = 2 ?". Let's find one of them using this 1-line main program together with the UDF:
    1  DESTROY ALL @ RADIANS @ STD @ DISP FNROOT(4,5,FNG(15,FVAR)-2)
    5  DEF FNG(N,M) @ P=1 @ R=PI*M/N @ FOR K=1 TO (N-1)/2
    6  P=P*COS(R*K) @ NEXT K @ FNG=N+LOG2(P*P) @ END DEF

    >RUN
         x3 =  4.59247568122
Now we can check that this value of x really makes  GCD(15, x) = 2 :
    >FNG(15,RES)  ->  2
In the table above we can see that there are likely two other solutions for x, one in [4.0, 4.2] and another in [4.2, 4.4], namely:
    FNROOT(4.0, 4.2 ...) ->  x1 = 4.06191388928, FNG(15,RES) = 2.0000000002
    FNROOT(4.2, 4.4 ...) ->  x2 = 4.38207921034, FNG(15,RES) = 1.9999999996
and that completes all three solutions to this mini-challenge in the interval [4, 5], which you can see listed with 20 decimal digits in the comments below.

Additionally, they are duly located and labeled in my graph of  y = GCD(15, x)  right below, where you can also see the perhaps unexpected fact that  GCD(15, 4.5)  tends to -\( \infty \) because in this case the cosine product includes one cosine which is 0, thus so is the product itself and its log2 tends to -\( \infty \). Notice theres's a fourth solution just outside the [4..5] interval:

      [Image: GCD%20graph.jpg]

Additional comments:  These are the three solutions computed using the HP-42S Solver via the 34-digit Free42 Decimal simulator:
    01  LBL "GCD15"  13   x         01  LBL "LOL2"       
    02  PI           14  DSE 00     02  MVAR "X"
    03   x           15  GTO 00 ►   03  RCL "X"
    04  15           16  X^2        04  XEQ "GCD15"
    05   ÷           17  LOG        05   2
    06   7           18   2         06   -
    07  STO 00       19  LOG        07  END
    08  SIGN         20   ÷
    09 ►LBL 00       21  15
    10  RCL 00       22   +
    11  RCLx ST Z    23  END
    12  COS

    [SOLVER]  ->  Select Solve Program
    Touch [LOL2], 4.0, [X], 5.0, [X], [X]  ->  4.59247568122 (...121 559541826...)
                  4.0, [X], 4.2, [X], [X]  ->  4.06191388926 (...926 216056306...)
                  4.2, [X], 4.4, [X], [X]  ->  4.38207921034 (...034 278186774...)
You can press [SHOW] after each solution to see up to 34 digits (at least 30 will be correct.) The fourth "solution" just outside the [4..5] interval is at x4 = 5.22459905977 (222552325...).



That's all for now. I'll post the next two LOLs in a couple' days or so. Meanwhile, let's see your comments.

Regards.
V.

  
All My Articles & other Materials here:  Valentin Albillo's HP Collection
 
Visit this user's website Find all posts by this user
Quote this message in a reply
04-15-2024, 08:17 PM (This post was last modified: 04-15-2024 08:48 PM by Gerson W. Barbosa.)
Post: #15
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-10-2024 08:51 PM)Valentin Albillo Wrote:  Additional comments:  No one posted all four squares so three of them were left missing (though once you've created your multiprecision squaring program it's all too easy to use it to compute the squares and post them) and no one posted any more examples of this pattern so here you are, a few more:

      774700591300020347197007492 =

            6001610061606011616611060006010661000616100111161001

      9425754295779433269877982 =

            888448440444044400080440444440408800048040888804

Here's a less sloppy HP-75 program with all four squares plus one found at OEIS. When I was editing it, I noticed there was a typo in line 115 of my previous program, current line 125: the second 2 should be 1. The numbers to be squared are placed in DATA lines, divided in chunks of six digits with no leading zeros, preceded by the number of chunks.

10 OPTION BASE 0
15 INTEGER I,J,K,L,N
25 B=1000000
30 DIM A(9),B(5),C(12)
35 REM DIM A(N+4), B(N), C(2*N+2); N>1
40 FOR K=1 TO 5
45 READ N
50 FOR I=0 TO 2*N+2 @ C(I)=0 @ NEXT I
55 FOR I=0 TO N+4 @ A(I)=0 @ NEXT I
60 FOR I=1 TO N
65 READ A(I)
70 B(I)=A(I)
75 NEXT I
80 FOR I=N TO 1 STEP -1
85 T1=0 @ A1=0
90 FOR J=N-I+4 TO -1 STEP -1
95 C2=(T1+B(I)*A(J+1))/B
100 A1=FP(C2)*B
105 C(I+J)=C(I+J)+A1
110 T1=IP(C2)
115 NEXT J
120 NEXT I
125 FOR I=2*N TO 1 STEP -1
130 T=C(I)/B
135 C(I)=FP(T)*B
140 C(I-1)=C(I-1)+IP(T)
145 NEXT I
150 L=1
155 IF C(0)<>0 THEN DISP STR$(C(0));ELSE DISP STR$(C(1)); @ L=2
160 FOR I=L TO 2*N-1
165 C$=STR$(C(I)) @ C$=RIGHT$(C$,LEN(C$))
170 IF 6-LEN(C$)=0 THEN GOTO 185
175 C$="0"&C$
180 GOTO 170
185 DISP C$;
190 NEXT I
195 DISP
200 NEXT K
205 END
210 DATA 3,130834,904430,15239
215 DATA 4,471287,714788,971663,493899
220 DATA 5,25,781108,305591,628417,975738
225 DATA 5,141422,82876,67219,949805,50005
230 DATA 5,100,990098,979999,970099,500001

>RUN
17117772217211221211117217772227121
222112110111011100020110111110102200012010222201
664665545464645645665646644665564654546645556644644
20000205525005225202000505202222520222202205205000550500025
10199000091990191001091091099001091999900190199000001


Edited to remove an unnecessary BASIC line
Find all posts by this user
Quote this message in a reply
04-16-2024, 06:07 AM
Post: #16
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
Oh, a nice implied pair of challenges at a page linked from that OEIS entry - no squares known consisting only of digits 0,1,3 or 6,7,8.
Find all posts by this user
Quote this message in a reply
04-17-2024, 01:47 AM
Post: #17
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
      
Hi, all,

VA (i.e. me) Wrote:I'll post the next two LOLs in a couple' days or so. Meanwhile, let's see your comments.

Well, it seems that I was being overoptimistic, as usual, because after 5 days 5 elapsed only the illustrious Gerson W. Barbosa bothered to post an interesting new HP-75C program to solve LOL 1 and (drum roll) also explicitly listed all four beautiful squares I wanted everyone to behold, plus a bonus fifth square in the same fashion. Thanks a lot, Gerson, and here you are, another bonus square featuring \(\pi\) no less than three times i.e. at the very beginning, in the middle and near the end, a worthy apropos appearance:
    1773248751146694430800869081882 =

          31444111334433114334141133143444444313434111431113141443344
Now, these are my original solutions and comments for the next two sections, i.e. LOL the Third: Random and LOL the Fourth: Logs. The show must go on ...


3. LOL the Third:  Random
 
This mini-challenge's question is:
    How close can two consecutive RNDs actually be ? Use RANDOMIZE 1 as the seed and write code to output the list of ever closer consecutive RNDs and their index N in the sequence.

My original solution is this 3-line, 89-byte never-ending program which will produce the goods:
    1  DESTROY ALL @ RANDOMIZE 1 @ STD @ N=0 @ X=99 @ L=1
    2  N=N+1 @ Y=RND @ D=ABS(X-Y) @ IF D<L THEN DISP N;X;Y;D @ L=D
    3  X=Y @ GOTO 2


    >RUN
           N     First RND         Next RND          |Difference|
     -------------------------------------------------------------

             2  .731362440213     .77207218067      .040709740457
            13   5.64471991805E-2  6.30768172146E-2  6.6296180341E-3
           125  .805774019056     .803607575861     .002166443195
           316  .128424219936     .126476247276     .001947972660
           378  .128629571043     .127765838222     .000863732821
          1746  .657235932954     .658084547243     .000848614289
          1864  .724574750035     .724711386925     .000136636890
          3091  .804652037305     .804530031878     .000122005427
          4983  .900183907166     .900119502104     .000064405062
          5002  .185041013513     .184988311770     .000052701743
          5964  .350363678669     .350333411003     .000030267666
         33971  .800461443203     .800483558857     .000022115654
         56943  .322507676917     .322505089514     .000002587403
        144113  .468047416778     .468049379480     .000001962702
        192237  .771445619886     .771443980639     .000001639247
       1606781  .176400853304     .176399416567     .000001436737
       1732702   3.20935575951E-2  3.20944940649E-2  9.364698E-7
       1840905  .862173808769     .862174478752     .000000669983
       1969683  .573449101200     .573448667982     .000000433218
       2143212  .948380454319     .948380365276     .000000089043
      10684317  .555880267196     .555880325796     .000000058600
      38181163  .151576837566     .151576827517     .000000010049
     157373808  .477539626899     .477539622968     .000000003931
     322950317  .325324468276     .325324470156     .000000001880
     431380423  .275319512003     .275319511289     .000000000714
    1838286534  .564079556829     .564079556839     .000000000010
at which time I interrupted the search, with my record being the last line shown in the output, i.e. two consecutive RNDs, namely 0.564079556829  and  0.564079556839 , which differ by just 0.00000000001.

Of course this finding took waaaay long to run even on a very fast emulator, so eventually I had to stop the search without looking at the entire one-trillion-long sequence. Thus, I can't confirm whether closer consecutive RNDs are possible or not in this specific sequence generated by the seed 1, in particular consecutive ones identical to 12-digit accuracy (i.e. difference = 0).

Also, I feel that there's something eerie (IMHO) in seeing two consecutive RNDs come out as the almost identical value (or even identical just on screen if you use FIX 4 or FIX 6, say), so if you want to experience this feeling yourself try this 4-line, 149-byte ad-hoc variant of the above program:
    Note: Yes, I know that you can go directly to the vicinity of the two close consecutive RNDs using PEEK$ and POKE so saving much running time but I like it better this way.
   1  DESTROY ALL @ RANDOMIZE 1 @ STD @ N=0 @ X=99 @ L=1 @ INPUT "N=";M
   2  IF N=M-2 THEN DISP @ DISP "Execute RND ; RND ..." @ PAUSE
   3  N=N+1 @ Y=RND @ D=ABS(X-Y) @ IF D<L THEN DISP N;X;Y;D @ L=D
   4  X=Y @ GOTO 2


Now, in the list above we have this line:

      1864  .724574750035     .724711386925     .000136636890

so to see by yourself that those two very close RNDs are indeed produced consecutively at that point in the sequence, do the following:
    >RUN ->  N=  { key in 1864 and [END LINE] }
              ->  { normal output as above, until ... }
              ->  Execute RND ; RND ...    { the program stops; execute the following: }

                          >FIX 4 @ RND;RND  ->  0.7246    0.7247
or for a much closer pair of consecutive RNDs:
    >RUN ->  N=  { key in 56943 and [END LINE] }
              ->  { normal output as above, until ... }
              ->  Execute RND ; RND ...    { the program stops; execute the following: }

                          >FIX 6 @ RND;RND  ->  0.322508    0.322505
and last, still closer so even more impressive:
    >RUN ->  N=  { key in 10684317 and [END LINE] }
              ->  { normal output as above, until after a really, really long while ... }
              ->  Execute RND ; RND ...    { the program stops; execute the following: }

                          >FIX 7
                          >RND  ->  0.5558803    
                          >RND  ->  0.5558803
and le voilà!. Reminds me of Groundhog Day, it looks like RND is broken ...
  

Additional comments:  J-F Garnier did his best to try and solve this mini-challenge. He used a "super-fast" HP-71B emulator with his own 3-line BASIC program which was very similar to my own original solution above (although he forgot to include the nearly-mandatory DESTROY ALL statement at the very beginning Smile ) and let it run for presumably large amounts of time until at last he exactly matched my own record, namely:

         1838286534  .564079556829     .564079556839     1.E-11

but although he let it run for 5 billion-deep values in the sequence before stopping it for good, he was unable to get two identical (to 12-digit) consecutive random numbers in this particular sequence created by the original seed 1, as he theorized to be possible. He adds:
    "but only a small fraction of the RND period has been explored"
and I agree, as 5 billion is only 0.5% of the whole 1-trillion-strong sequence so 99.5% of it remains unexplored.

Last but absolutely not least, J-F left some intriguing observations which he never fully developed. For example, he said:
    "The RND value is determined by the internal seed, which is stored with 15 digits. Or, in other words,we can't predict the next RND value from the last one (well, the possibilities for the next RND are limited)."
but he never explained the underlying algorithm used and why and how the possibilities for the very next RND value are limited. Knowing that, perhaps it'd be much easier to see if two consecutive values can indeed be identical to 12-digit accuracy or not, without having to generate and check tens or hundreds of billions of RNDs.

He also posted two sequences generated by different seeds which include the same 11-digt (not 12-digit) value .14159265359 but he didn't tell how he found those particular seeds, nor did he explain why the very next RND values after the .14159265359 do differ from their second decimal digit on (namely .494478890124 and .41675139916,) while it would seem that, as the possibilities are "limited", they should be much closer and not differ so markedly.

Adding to the mysteries, J-F also said:
    "[...] It could be possible to get two consecutive RND values that are equal. Matter of fact, I have one example. Can you find it? :-)"
and again he never bothered to post that "one example" he found nor to explain how did he find it.

In short: for sure J-F is under no obligation whatsoever to share or post anything at all but I've always thought that one of the goals of these mini-challenges is to provide entertainment while also introducing useful math concepts & programming techniques to learn from. So, having at hand the solutions by me and by others serves as an effective way of sharing knowledge, but if you won't share what you know or what you found, then what's the point ? I don't give prizes, you know.

Frankly, I was expecting J-F to eventually provide answers to the above matters so that me and other interested people would learn something new and be enlightened in the process, but to my big surprise he never did. At least so far, several weeks (as of 2024-04-17) since he posted his message. As Chloe B. would say: "C'est décevant, totalement décevant". Smile

    
4. LOL the Fourth:  Logs

This wasn't a mini-challenge per se but simply me reporting an unusual finding, namely that LOG10 seemed to be the most accurate logarithm available in some HP vintage calcs (namely the 10-digit HP-15C and the 12-digit HP-71B). I then suggested the following for interested people to explore and post their findings:
  • Try the above examples in your HP models, both 10-digit and 12-digit, from the first HP-35 up to the latest RPL models, to see if LOG10 and LN differ that much in their respective errors.
      
  • Test other range of values (here from 52 = 25 to 59 = 1,953,125).
      
  • Find out if there are arguments with even larger errors.

Regrettably, only J-F Garnier investigated the matter in this excellent post of his, coming to the conclusion that "we can't say that the decimal LOG provides more accurate results than the natural LN, overall", conclusion with which I mostly agree.

He also asks a related question which is pending further investigation but this LOL the Fourth seems to have been largely ignored so far (J-F excepted, of course,) though I think it's an interesting topic related to the innards of HP algorithms and even perhaps to their architecture. And answering J-F's question, no, I couldn't find any references in older threads either nor do I remember this topic having been discussed here ever. I took my observations from ancient (20-25 year-old or more) notes which I wrote at the time but never published before.



Enough for now. I'll post my original solutions and comments to the two final LOL 5 and LOL 6 sections either in an unspecified number of days (a few, a week, a month ...) or right after the next comment(s), whichever comes first.

In the meantime, you might want to have a look at some of my previous April 1st challenges, they feature a true plethora of interesting topics, solutions and comments which are sure to keep you entertained while you wait:
Regards.
V.

  
All My Articles & other Materials here:  Valentin Albillo's HP Collection
 
Visit this user's website Find all posts by this user
Quote this message in a reply
04-19-2024, 09:51 PM
Post: #18
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
      
Hi, all,

Well, at long last these are my original solutions and comments for the last two sections, i.e. LOL the Fifth: Gamma and LOL the sixth: Miscellanea. The party's almost over ...


5. LOL the Fifth:  Gamma

Executing this loop to list the values of \(\Gamma\)(10-1), \(\Gamma\)(10-2), ..., \(\Gamma\)(10-10), gives:

   >DESTROY ALL @ FOR N=1 TO 10 @ N;GAMMA(10^(-N))@ NEXT N

      N     \(\Gamma\)(10-N)
    -------------------
      1  9.51350769867
      2  99.4325851191
            ...
      9  999999999.423
     10  9999999999.42


where the fractional part seems to be quickly converging to some limit around ~0.42, so you must write code to find this limit to much greater accuracy (say 10-12 digits or more). Once done, answer these questions:
  • Can you estimate the most accurate value you got ?
  • Can you identify the symbolic, closed form of that numeric value ?

As the table above demonstrates, the 12-digit HP-71B lacks the precision to resolve the fractional part to at least 12 digits so we use the 34-digit Free42 Decimal with this small 15-step, 27-byte RPN program which displays just the fractional part of \(\Gamma\)(10-N) for N= -1, -2, ...

   01  LBL "GAM10"   09  FP
   02  ALL           10  STOP
   03   1            11  R
   04 ►LBL 00        12  ISG ST X
   05  ENTER         13  LBL 00
   06  +/-           14  GTO 00 ►
   07  10^X          15  END
   08  GAMMA


   XEQ "GAM10" -> 

    { I in ST Y, fractional part in ST X[R/S] to continue }

      1   .513507698669     { using SHOW from now on }
      2   .432585119151
      3   .423772484595     12   .4227843350994561953914
      4   .422883231624     13   .422784335098566044949
      5   .422794225568     14   .42278433509847702999 
      6   .422785324154     15   .4227843350984681235
      7   .422784434004     16   .422784335098467242   { best estimate }
      8   .422784344989     17   .42278433509846684    
      9   .422784336088     18   .4227843350984742     { error grows, insufficient accuracy }
     10   .422784335197
     11   .422784335108


and we see that the fractional part seems to be converging to  0.4227843350984672  for N = 16, truncating at 16 digits by ignoring two "guard" digits, and the values for N = 17 and N = 18, as the accuracy is clearly worsening (the error begins to grow). Thus, my estimate for the most accurate value obtained is:

          0.4227843350984672

Can this value be identified ? Well, yes, we can try any of the programs out there online or using your own or your calc's identification software. I did use my HP-71B's IDENTIFY program which identified the value as  1-EulerGamma.

If proceeding manually, one can notice that  1 - 0.4227843350984672 = 0.5772156649015328 , which is immediately recognizable but if not, just searching this value online immediately reports it as the Euler–Mascheroni constant, aka \(\gamma\) constant. So, the fractional part's limit is symbolically identified as:

          1 - \(\gamma\)

which is as can be expected because the Laurent series expansion of the \(\Gamma\) function near zero is:

         [Image: Gamma%20Laurent.jpg]

and you can see the  - \(\gamma\)  term there.

Additional comments:  Solver extraordinaire Gerson W. Barbosa tackled this mini-challenge extensively, correctly produced the limit to great accuracy and identified it as the Euler-Mascheroni \(\gamma\) constant, but he also tried to significantly improve the accuracy and said (my bold):
    "Here we also notice that the mantissas of the errors, the second column in the table, appear to tend to another constant. Regardless of any attempt to identify it, we can try to use it to get a few more correct digits, [...] Now we have about 50% more correct digits, 23"
but I think that his reasoning is circular because he's using the actual, externally-obtained 34-digit \(\gamma\) value (0.5772156649015328606065120900824024, as seen in his "L5th" program,) to compute the second column of the table in order to obtain the second constant (0.98905599...,) and use it to improve the result. In other words, he's using the real \(\gamma\) to improve the accuracy of his computed \(\gamma\), i.e. a circular way of proceeding.

As for the second constant's (0.98905599...) identification, it's mainly the O(z) term in the Laurent expansion above, namely:
\[ \frac{1}{12} (6 \gamma^2 + \pi^2) z\]
6. LOL the Sixth:  Miscellanea

  1. Try and deduce what result will be output by this HP-71B expression without actually executing it:

          NEIGHBOR(INTEGRAL(FNROOT(EPS,EPS,EPS),EPS,EPS,EPS),GAMMA(EPS))

    Was your deduction correct ? Can you explain the result obtained ?


    The expression returns the value 10, exactly. EPS is the constant 1E-499 and GAMMA(EPS) is thus 1E499, which is just used to indicate NEIGHBOR the direction used to return the nearest machine-representable number to its first argument (the value of INTEGRAL,) so its value isn't relevant as long as it's ≥ 10. The breakdown is:

          >FNROOT(EPS,EPS,EPS)            ->  -9.99999999999E499 = -MAXREAL
          >INTEGRAL(-MAXREAL,EPS,EPS,EPS) ->   9.99999999999
          >GAMMA(EPS)                     ->   1.E499
          >NEIGHBOR(9.99999999999,1E499)  ->  10

    so:
          >NEIGHBOR(INTEGRAL(FNROOT(EPS,EPS,EPS),EPS,EPS,EPS),GAMMA(EPS) -> 10

          
  2. I executed this command-line expression on my HP-71B but it just resulted in System Error. Can you find out what's wrong ?

          FOR I=64204 TO 64215 @ DISP CHR$(HTD(REV$(PEEK$(DTH$(2*I),2)))); @ NEXT I

          
    A trick question. The expression above just peeks characters from the specified addresses in the System ROMs, which happen to be the ones which form the error message "System Error" so that's what you get. The expression itself is perfectly correct and returns the pertinent text.

    Additional comments:  JoJo1973 had the right intuition here. He wrote:

      "[...] to write a program resulting in the largest number of errors. Of course the easiest way to accomplish the task is to dump the ROM area where the messages are actually stored [...]"

    which goes to the gist of the matter.

          
  3. Some nice results I got:

          95888      = 309.657875727390470000000975517...
          22008840 = 4691.3580123456789961013...      

    In radians:

        >EXP(ACOS(430/433))  ->  1.12500000006     { ~  9/8 }
        >EXP(ACOS(538/541))  ->  1.11111111113     { ~ 10/9 }

    [...]so perhaps there's some hidden pattern at large here.


    None that I know of, perhaps just a coincidence. Regrettably, no one commented on those nice results I found, which is a pity as the last two seem to me rather remarkable. Another nice one I also found is:\[ \Gamma \cosh \left (\sqrt[6]{\frac{250}{57}}\phi \right ) \]where \(\phi\) is the Golden Ratio and which evaluates to  6.2000000002  on my HP-71B.
          
    Additional comments:  The first square root (95888) can be used to generate an integer perfect square nearly as awesome as the ones shown in LOL the First, namely

                309657875727390472 = 958879999999999999999993958468209

    with no less than 18 consecutive '9's in the middle ! I'd say this must be the smallest perfect square with that many consecutive '9's (or any other diigit) in the middle, by far.

          
  4. What does this HP-71B user-defined function compute ?

          10  DEF FNC(M,N)=M!/N!/(M-N)!


    Another trick question. At first sight it would seem that the  "!"  symbol stands for the factorial (which is the case for standard mathematical notation,) so this user-defined function would seem to be computing  m!/n!/(m-n)! , which is the number of combinations of m things taken n at a time without repetition.

    However, this is not the case here because the HP-71B uses the keyword FACT for the factorial while  "!"  is the comment delimiter, so everything in the line after the first  "!"  is considered a comment and of course it won't be executed at all.

    This means that the line reduces to  10 DEF FNC(M,N)=M , which just returns the value passed in the M argument. As J-F Garnier said: "It doesn't do much, I'm afraid."

          
  5. The HP-71B does not allow for variable names beginning with a 2-letter or more prefix, but executing this assignment in a program or from the keyboard ...

          FORM=STOP

    ... doesn't result in a Syntax Error so what gives ?


    Yet another trick question. The HP-71B parser doesn't generally care for spaces between keywords, etc., so it interprets this line as  FOR M=S TO P , which just begins a FOR-NEXT loop using scalar numeric variables M, S and P, all of it perfectly legal syntax.  Smile

    Additional comments:  JoJo1973 nailed this one, despite not owning a 71B. His experience with a non-HP, non-calc machine (a Commodore 64 no less) served him well to understand what was happening. As he rethorically asked: "FORM=STOP is a variable assignment or the beginning of a loop: FOR M = S TO P ?".

          
  6. Write code to compute to full accuracy this function S(x) given the argument x. Use your code to compute S(42): \[ S(x) = \frac{1}{\sum_{n=0}^{\infty} \binom {2 n}{n}^3 \frac{{nx \space + \space 5}}{2^{12 n \space + \space 4}}} \]My original solution is this 3-line HP-71B user-defined function:

          1  DEF FNS(X) @ S=0 @ T=0 @ N=0 @ REPEAT @ S=S+T
          2  T=COMB(2*N,N)^3*(X*N+5)/2^(12*N+4) @ N=N+1
          3  UNTIL S=S+T @ FNS=1/S @ END DEF

             >DESTROY ALL
             >FNS(42)

                 3.14159265358
           { fully correct save 1 ulp }

    This is yet another marvelous Ramanujan's series. The correct value of this series is exactly \(\pi\) and has the awesome BBP formula-like property that, as the denominators are exactly 16 · 2n, this can be used to compute the second block of n binary digits of \(\pi\) without having to compute the first n binary digits.

    Additional comments:  John Keith provided a nice RPL solution for this mini-challenge and he also added an interesting observation but he never posted the value produced by his program and that's not Ok. Remember: you must mandatorily post both code AND results, one without the other won't do, certainly not when also posting the result is so extremely little extra work.

    Juan14 provided algebraic manipulations to achieve extra speed, a working 35-step RPN program to be run in Free42 Decimal and the result accurate to 34 digits save 1 ulp. Not bad. Plus he was amazed by yet another unexpected appearance of \(\pi\).



That's all, the party's over now and it was enjoyable even if few people actually attended so let's call it a wrap. Thanks to everyone who viewed this thread and/or contributed to it.

This will be my last challenge-oriented SRC for an indefinite period of time so I really hope you enjoyed it while it lasted. All good things ...

Bye.
V.

  
All My Articles & other Materials here:  Valentin Albillo's HP Collection
 
Visit this user's website Find all posts by this user
Quote this message in a reply
04-20-2024, 06:27 PM
Post: #19
RE: [VA] SRC #017 - April 1st, 2024 Spring Special
(04-19-2024 09:51 PM)Valentin Albillo Wrote:        
...So, the fractional part's limit is symbolically identified as:

          1 - \(\gamma\)

which is as can be expected because the Laurent series expansion of the \(\Gamma\) function near zero is:

         [Image: Gamma%20Laurent.jpg]

and you can see the  - \(\gamma\)  term there.


Additional comments:  Solver extraordinaire Gerson W. Barbosa tackled this mini-challenge extensively, correctly produced the limit to great accuracy and identified it as the Euler-Mascheroni \(\gamma\) constant, but he also tried to significantly improve the accuracy and said (my bold):
    "Here we also notice that the mantissas of the errors, the second column in the table, appear to tend to another constant. Regardless of any attempt to identify it, we can try to use it to get a few more correct digits, [...] Now we have about 50% more correct digits, 23"
but I think that his reasoning is circular because he's using the actual, externally-obtained 34-digit \(\gamma\) value (0.5772156649015328606065120900824024, as seen in his "L5th" program,) to compute the second column of the table in order to obtain the second constant (0.98905599...,) and use it to improve the result. In other words, he's using the real \(\gamma\) to improve the accuracy of his computed \(\gamma\), i.e. a circular way of proceeding.

As for the second constant's (0.98905599...) identification, it's mainly the O(z) term in the Laurent expansion above, namely:
\[ \frac{1}{12} (6 \gamma^2 + \pi^2) z\]

Hello, Valentín,

Yes, indeed that was a circular procedure. But then my interest had shifted to using these values of the Γ function to get γ on the calculator to full accuracy using less bytes than just writing the constant itself. Perhaps I should have used another label instead "L5th" for that one.

But I have managed to do it also in a non-circular way, as you can see from this later post:

(04-09-2024 05:19 PM)Gerson W. Barbosa Wrote:  But there’s a better way to get those extra digits. We only have to use the GAMMA function twice:

00 { 29-Byte Prgm }
01▸LBL "E_M"
02 1
03 -11
04 10↑X
05 GAMMA
06 LASTX
07 R↓
08 FP
09 -
10 R↑
11 +/-
12 GAMMA
13 FP
14 -
15 2
16 ÷
17 END


This will return

0.577215664901532860606565

Anyway, it is possible to get about the same accuracy with only one evaluation of the Γ function near zero without resorting to circular proceeding. We can ignore the higher-order terms and easily isolate γ in the Laurent series expansion you have mentioned above:

\[\gamma\approx\frac{1-\sqrt{1-\frac{z^2\pi^2}{6}+2\left({z\Gamma\left({z}\right)-1}\right)}}{z}\]

      N              (1-√(1-z²π²/6+2(zΓ(z)-1)))/z; z=10⁻ᴺ        
    -------------------------------------------------------
      1              0.585903128471302463502247124180713        
      2              0.57730596205274564964950936551113      
      3              0.5772165719234095045247918598371         
      4              0.577215673975865776232977270554       
      5              0.57721566499228031030482893493        
      6              0.5772156649024403392246767465        
      7              0.577215664901541935396815027        
      8              0.57721566490153295135441953       
      9              0.5772156649015328615139945        
     10              0.577215664901532860615600     
     11              0.57721566490153286060691        
     12              0.5772156649015328606039        
   --------------------------------------------------------



Best regards,

Gerson.


-----------------------------------------------------------

00 { 42-Byte Prgm }
01▸LBL "gamma"
02 +/-
03 10↑X
04 PI
05 RCL× ST Y
06 X↑2
07 6
08 ÷
09 RCL ST Y
10 GAMMA
11 RCL× ST L
12 LN
13 E↑X-1
14 STO+ ST X
15 -
16 1
17 X<>Y
18 -
19 SQRT
20 +/-
21 1
22 +
23 X<>Y
24 ÷
25 END
Find all posts by this user
Quote this message in a reply
Post Reply 




User(s) browsing this thread: 2 Guest(s)