13 replies to this topic

[DelMonte]

Hi,

For some reason fx-9860GII seems to have some problems with poisson distribution.

When dealing with big numbers 9860 gives wrong answers, when certain limit is crossed.

Here's my test scenario:

Function PoissonCD(X,MEAN)

"X" "MEAN" "CORRECT ANSWER" "9860's ANSWER"
9995 10000 0,48272 0,48272
9996 10000 0,48670 0,48670
9997 10000 0,49069 0,49069
9998 10000 0,49468 0,49468
9999 10000 0,49867 0,80853 (wrong answer)
10000 10000 0,50266 0,80900 (wrong answer)


Has anyone else noticed this?

Rgs,

DelMonte

[TovAre]

Hi,

For some reason fx-9860GII seems to have some problems with poisson distribution.

When dealing with big numbers 9860 gives wrong answers, when certain limit is crossed.

Here's my test scenario:

Function PoissonCD(X,MEAN)

"X" "MEAN" "CORRECT ANSWER" "9860's ANSWER"
9995 10000 0,48272 0,48272
9996 10000 0,48670 0,48670
9997 10000 0,49069 0,49069
9998 10000 0,49468 0,49468
9999 10000 0,49867 0,80853 (wrong answer)
10000 10000 0,50266 0,80900 (wrong answer)


Has anyone else noticed this?

That is interesting, I've never noticed (I don't normally usel values that large) but I'm getting the same error with firmware version 02.00.0200

It's also bug because the manual says it should deal fine with 9 digits and display a warning if you used values which might cause invalid answers.


ttp://keisan.casio.com/ calculates this correctly (as do firmware 01.03.000 ... emulated), so they do know how.


Have you asked Casio about it yet?

Cheers,

Tov Are

[DelMonte]

Have you asked Casio about it yet?

I sent the same question to their support yesterday, but I suppose it takes at least a few working days to get a reply.

One interesting detail:

I also have Casio ClassPad 330 and it seems to have two "options":

You can use an "official" function= poissonCDf(0,9999,10000) to get a result 0,49867, which is correct, but it takes about 2,5 minutes to get that answer!

or you can use an undocumented syntax (skip the lower boundary) poissonCDf(9999,10000) to get an incorrect result 0,80853, but it's calculated immediately;-) (with poissonCDf(9998,10000) you get a correct result).

Rgs,

DelMonte

[DelMonte]

Have you asked Casio about it yet?

I sent the same question twice. First one was sent on May 9th and the second one on May 14th, because I didn't receive any reply to the first one.

But today I got two very different answers from Casio support. Both from Casio European Technical Center, but from two different persons:

1st answer:

With regards to your question we can offer you the following information:
Please update your fx-9860G to the new version 2.0. This version shows the correct.

2nd answer:

We forwarded your inquiry to the resposible division for a detailed check. As it can take time
to receive the answer we like to apologise and ask for your patience.
As soon as we get the answer we'll forward it to you.


For some reason the second answer seems to be a better one. At least that service agent did bother to read my message and also understood the problem. The first answer has some flaws:

"Please update your fx-9860G to the new version 2.0"

I have a fx-9860GII, not 9860G. I also have the latest version of the firmware: 02.00.0200
I also included that information in my message.

"This version shows the correct."

No, it doesn't. Of course it takes almost 15 seconds to test it and find out that it doesn't "show the correct".

This was the first time I contacted Casio's support and I have to say I'm a bit disappointed.

Edited by DelMonte, 18 May 2010 - 03:16 PM.

[TovAre]

I sent the same question twice. First one was sent on May 9th and the second one on May 14th, because I didn't receive any reply to the first one.

But today I got two very different answers from Casio support. Both from Casio European Technical Center, but from two different persons:
[...]
This was the first time I contacted Casio's support and I have to say I'm a bit disappointed.

I contacted support via a form on www.casio-intl.com and got a response from Casio Japan on a different matter. First response I got was an autoreply appologizing that the office was closed due to national hollidays, then the reply below.

(What I liked about it was that it was courtious, it gave a good explanation while still being open to further input.).

Dear customer,

Sorry for our delay in replying.

Our R&D says;

A data sort work is effected in case calculating median or quartile point（Q1, Q3）inside fx-9860g2 that is why upper limit is fixed for a frequency. And it is set to 65535 at this unit.

Moreover, only whole numbers are sorted at above said sort work, that is why any frequency including decimal or fraction resulted acalculia (not computable) at this unit.

We would appreciate it if you kindly let us have concrete examples; to calculate median or quartile point（Q1, Q3）with a frequency greater than 2^16-1 or with a frequency including decimal or fraction for our further study.

We really sorry for our delayed reply even if your inquiry was sent to us during our irregularly continuous holidays.

Sincerely,

[DelMonte]

I contacted support via a form on www.casio-intl.com and got a response from Casio Japan on a different matter.

Thx for the tip! I'll give www.casio-intl.com a try...

[DelMonte]

Finally an answer from Casio:

"As we promised you about two weeks ago we forwarded your inquiry to our parent company in Japan. We received now the answer which we like to send you without any further comment.

Quote:
As the characteristic of the Poisson distribution, it is designed to make the precision of λ (average) <100.

Although old version had expected accuracy, the latest OS could not keep the same accuracy because the improved algorithms for other distribution had an influence.
We would like to apologize for inconvenience which our operation manual could not give enough explanation about the clear usage condition.

In addition, we feel that in the case of distribution calculation by using mentioned data, it would be better to use "Normal distribution".

Poisson distribution is appropriate distribution calculation when you want to see the rare case for a definite period of time at a fixed place.
Therefore, it would be better to use this calculation under the condition of λ (average) <25.

In fact, under the condition of λ (average) <10, it shows the result which is close to "Normal distribution".

We would like to review and improve the algorithms of the precision of calculation under the condition of λ＞100 as one of important themes in the future.

:unquote"

So I think that's it. Have to use TI-Nspire instead for those calculations...

[TovAre]

I was using casios graphing calculator and guess what? its standard deviation numbers ARE ALL WRONG. made me fail a fucking midterm.

Do you have an example where you got an incorrect answer?

They cannot be all wrong, try entering 1,2,3,4,5,6,8,9 into the list. Press F1 for 1Var, then see if the standard deviation you get is 2.73861278

You can also double check calculations using wolfram alpha:

http://www.wolframal...2,3,4,5,6,7,8,9

Regards,

Tov Are

[TovAre]

sorry!
http://www.google.co...ved=0CCkQ9QEwAw

According to wikipedia it's a less commonly used estimator for the standard deviation and says:

" But this estimator, when applied to a small or moderately sized sample, tends to be too low: it is a biased estimator."

IMHO, you might be able to pull of a complaint? Try something like:

"I realize I made a mistake on the exam and I think it happened because I hang out on math-forums on the internet and read a lot of extra material. This made it easier for me to make that sort of mistake ... it was just a lot more more intuitive to use the more common formula with Bessel's correction due to the small sample size given on the exam ... I mean wouldn't you?"

http://en.wikipedia....ndard_deviation

Cheers

[TovAre]

Hi,

For some reason fx-9860GII seems to have some problems with poisson distribution.

When dealing with big numbers 9860 gives wrong answers, when certain limit is crossed.

Here's my test scenario:

Function PoissonCD(X,MEAN)

"X" "MEAN" "CORRECT ANSWER" "9860's ANSWER"
9995 10000 0,48272 0,48272
9996 10000 0,48670 0,48670
9997 10000 0,49069 0,49069
9998 10000 0,49468 0,49468
9999 10000 0,49867 0,80853 (wrong answer)
10000 10000 0,50266 0,80900 (wrong answer)


Has anyone else noticed this?

Rgs,

DelMonte

This appeares to be fixed now in version 2.01

Cheers,