FindRoot cannot find obvious solution

533 views
Skip to first unread message

Mukhtar Bekkali

unread,
Apr 27, 2004, 5:13:06 AM4/27/04
to
Here is my code in Mathematica 5:

h1=1+(y-x)^2;
h2=1+(2/3-x^2)^2;
b=h2/(h1+h2);
f=x(1-x)^2;
s=f*D[b,y];
g[y_]:=NIntegrate[s,{x,0,1}];
FindRoot[g[y]==0,{y,0.35,0.45}]

The output is

NIntegrate::inum: Integrand H is not numerical at {x} = {0.5`}.
FindRoot::cvmit: Failed to converge to the requested accuracy or precision
within 100 iteration

and it gives me y->0.45 or the upper boundary. However, when I plot g[y] I
can see that the solution is somewhere around y=0.4. Using Solve instead of
FindRoot does not give me any solutions.

What am I doing wrong here? Thank you in advance. Mukhtar Bekkali


Jens-Peer Kuska

unread,
Apr 28, 2004, 7:06:33 AM4/28/04
to
Hi,

and what is with

h1 = 1 + (y - x)^2;
h2 = 1 + (2/3 - x^2)^2;
b = h2/(h1 + h2);
f = x(1 - x)^2;
s = f*D[b, y];

g[yy_?NumericQ] := NIntegrate[Evaluate[s /. y -> yy], {x, 0, 1}]

FindRoot[g[y] == 0, {y, 0.35, 0.45}]


Regards

Jens

"Mukhtar Bekkali" <mbek...@iastate.edu> schrieb im Newsbeitrag
news:c6l872$ire$1...@smc.vnet.net...

Oleksandr Pavlyk

unread,
Apr 28, 2004, 7:12:47 AM4/28/04
to
Hi Mukhtar,

Attributes[NIntegrate]
{HoldAll, Protected}

So s is not evaluated until later.
Your code would look much cleaner if
you kept arguments of your functions
explicitly

h1[x_, y_] = 1 + (y - x)^2;
h2[x_] = 1 + (2/3 - x^2)^2;
b[x_, y_] = h2[x]/(h1[x, y] +
h2[x]);
s[x_, y_] = FullSimplify[
x*(1 - x)^2*Evaluate[
D[b[x, y], y]]];

g[y_] := NIntegrate[s[x, y],
{x, 0, 1}]

Now, unlike in your code, Plot[g[y],{y,0.2, 0.8}]
does plotting without prolifirating bunch of errors.
Furhermore, it produces correct result for
FindRoot:

FindRoot[g[y], {y, 0.35,
0.45}, WorkingPrecision ->
16]

<< Error messages NItegrate::ploss
suppressed >>

{y -> 0.399370682489405571438\
7173794`16.}

If those messages annoy you, like they do for me,
you can work with the interpolated function

ig = Interpolation[
Table[{y, g[y]},
{y, 0.2, 0.8, 0.005}],
InterpolationOrder -> 5];

Then the result is spit out without errors.

FindRoot[ig[y], {y, 0.35,
0.45}, WorkingPrecision ->
16]

{y -> 0.399370682489405632471\
6472443`16.}

So now I've got the question for the group.

It seems to me like a general problem, having
a function f[x], defined via Integral
(or Sum, or Product)

f[x] = Integrate[ f[x,y],{ y, domain}]

i can plot the function, and see it's zero approximately,
but when fed to FindRoot, I get a bunch of errors.
I do not understand why, but that aside, the only clean
way out is to interpolate f[x] first, and then use FindRoot.

If anybody on the list has a better workaround, please share :).
Your insight would be greatly appreciated.

Regards,
Sasha

Anton Antonov

unread,
Apr 28, 2004, 7:13:48 AM4/28/04
to
Mukhtar Bekkali wrote:

>Here is my code in Mathematica 5:
>
>h1=1+(y-x)^2;
>h2=1+(2/3-x^2)^2;
>b=h2/(h1+h2);
>f=x(1-x)^2;
>s=f*D[b,y];
>g[y_]:=NIntegrate[s,{x,0,1}];
>FindRoot[g[y]==0,{y,0.35,0.45}]
>
>The output is
>
>NIntegrate::inum: Integrand H is not numerical at {x} = {0.5`}.
>FindRoot::cvmit: Failed to converge to the requested accuracy or precision
>within 100 iteration
>
>and it gives me y->0.45 or the upper boundary. However, when I plot g[y] I
>can see that the solution is somewhere around y=0.4. Using Solve instead of
>FindRoot does not give me any solutions.
>
>What am I doing wrong here? Thank you in advance. Mukhtar Bekkali
>
>
>

You can try this:


In[158]:=
Clear[f, g, s, y, x]


h1 = 1 + (y - x)^2;
h2 = 1 + (2/3 - x^2)^2;
b = h2/(h1 + h2);

f = x*(1 - x)^2;


s = f*D[b, y];

g[s_, (y_)?NumericQ] := NIntegrate[s, {x, 0, 1}];
FindRoot[g[s, y] == 0, {y, 0.35, 0.45}]


Out[165]=
{y -> 0.3993706824894056}


The reason why NIntegrate gives "inum"messages is explained at
http://support.wolfram.com/mathematica/mathematics/numerics/nsumerror.html

Anton Antonov
Wolfram Research, Inc.

Jens-Peer Kuska

unread,
Apr 29, 2004, 12:42:05 AM4/29/04
to
Hi,

and what is with

h1 = 1 + (y - x)^2;


h2 = 1 + (2/3 - x^2)^2;
b = h2/(h1 + h2);

f = x(1 - x)^2;


s = f*D[b, y];

g[yy_?NumericQ] := NIntegrate[Evaluate[s /. y -> yy], {x, 0, 1}]

FindRoot[g[y] == 0, {y, 0.35, 0.45}]


Regards

Jens

"Mukhtar Bekkali" <mbek...@iastate.edu> schrieb im Newsbeitrag
news:c6l872$ire$1...@smc.vnet.net...

Oleksandr Pavlyk

unread,
Apr 29, 2004, 12:48:09 AM4/29/04
to
Hi Mukhtar,

Attributes[NIntegrate]
{HoldAll, Protected}

Regards,
Sasha

Anton Antonov

unread,
Apr 29, 2004, 12:49:09 AM4/29/04
to
Mukhtar Bekkali wrote:

>Here is my code in Mathematica 5:
>
>h1=1+(y-x)^2;
>h2=1+(2/3-x^2)^2;
>b=h2/(h1+h2);
>f=x(1-x)^2;
>s=f*D[b,y];
>g[y_]:=NIntegrate[s,{x,0,1}];
>FindRoot[g[y]==0,{y,0.35,0.45}]
>
>The output is
>
>NIntegrate::inum: Integrand H is not numerical at {x} = {0.5`}.
>FindRoot::cvmit: Failed to converge to the requested accuracy or precision
>within 100 iteration
>
>and it gives me y->0.45 or the upper boundary. However, when I plot g[y] I
>can see that the solution is somewhere around y=0.4. Using Solve instead of
>FindRoot does not give me any solutions.
>
>What am I doing wrong here? Thank you in advance. Mukhtar Bekkali
>
>
>

You can try this:


In[158]:=
Clear[f, g, s, y, x]

h1 = 1 + (y - x)^2;
h2 = 1 + (2/3 - x^2)^2;
b = h2/(h1 + h2);

f = x*(1 - x)^2;


s = f*D[b, y];

Maxim

unread,
Apr 29, 2004, 3:23:08 AM4/29/04
to
I think that with all these nice workarounds it went unnoticed that
there *was* something wrong with Mukhtar's example. Consider:

In[1]:=
Module[
{y = Cos[x], f},
f[x_] := y;
FindRoot[f[x], {x, 1, 2},
Compiled -> True, MaxIterations -> 15,
StepMonitor :> Print["step: ", {x, f[x]}]]
]

Out[1]=
{x -> 1.5708}

In[2]:=
Module[
{y = Cos[x], f},
f[x_?NumericQ] := y;
FindRoot[f[x], {x, 1, 2},
Compiled -> True, MaxIterations -> 15,
StepMonitor :> Print["step: ", {x, f[x]}],
EvaluationMonitor -> 0]
]


FindRoot::cvmit: Failed to converge to the requested

accuracy or precision within 15 iterations.

Out[2]=
{x -> 2.}

This is similar to the issue that was discussed at

http://forums.wolfram.com/mathgroup/archive/2003/Dec/msg00401.html

Just as with the Plot examples, f[1] here is not numerical because x
in Cos[x] will remain unbound. But if we evaluate Block[{x=1}, f[x]],
then the result will be numerical, and this is how FindRoot works. And
then there's the question of how it will interact with Compiled->True.

But here we have a new twist, because, according to Trace, FindRoot
does obtain numerical values evaluating f[x] in both cases, so it's
not clear what happens in the second example, when we use
f[x_?NumericQ].

Oh, and EvaluationMonitor/StepMonitor won't help us here, because they
perform true delayed evaluation, while what FindRoot seems to do is to
evaluate the first argument once and then to perform all computations
on what it obtains (why does it have HoldAll attribute then? Simple:
to add extra confusion if FindRoot is used inside Module or With and
there is a name conflict). As an additional bonus, if you omit
EvaluationMonitor in the second example, then StepMonitor won't print
anything either. Lovely.

Maxim Rytin
m...@prontomail.com

Gary.A...@frb.gov

unread,
Apr 29, 2004, 3:24:09 AM4/29/04
to

the following seems to provide the correct solution.
but, you will get a warning that NIntegrate is having trouble because the
integral is approximately zero.
i think that the code in the original posting may have applied the
Replacement Rule for the y before applying the D function.


exprGen=
Simplify[
With[{h1 = 1 + (yVar - xVar)^2,
h2 = 1 + (2/3 - xVar^2)^2},
With[{b = h2/(h1 + h2),f = xVar(1 - xVar)^2},
expr=(f*D[b, yVar])]]]


g[yArg_?NumberQ]:=
NIntegrate[(expr)/.yVar->yArg,{xVar,0,1}]


FindRoot[g[y] == 0, {y, 0.35, 0.45}]


"Jens-Peer Kuska"
<ku...@informatik.uni- To: math...@smc.vnet.net
leipzig-de> cc:
Subject: Re: FindRoot cannot find obvious solution
04/28/2004 06:56 AM


Hi,

and what is with

h1 = 1 + (y - x)^2;


h2 = 1 + (2/3 - x^2)^2;
b = h2/(h1 + h2);

f = x(1 - x)^2;


s = f*D[b, y];

g[yy_?NumericQ] := NIntegrate[Evaluate[s /. y -> yy], {x, 0, 1}]

FindRoot[g[y] == 0, {y, 0.35, 0.45}]


Regards

Jens

"Mukhtar Bekkali" <mbek...@iastate.edu> schrieb im Newsbeitrag
news:c6l872$ire$1...@smc.vnet.net...

> Here is my code in Mathematica 5:
>
> h1=1+(y-x)^2;
> h2=1+(2/3-x^2)^2;
> b=h2/(h1+h2);
> f=x(1-x)^2;
> s=f*D[b,y];
> g[y_]:=NIntegrate[s,{x,0,1}];
> FindRoot[g[y]==0,{y,0.35,0.45}]
>
> The output is
>
> NIntegrate::inum: Integrand H is not numerical at {x} = {0.5`}.

> FindRoot::cvmit: Failed to converge to the requested accuracy or
precision

Andrzej Kozlowski

unread,
Apr 29, 2004, 7:53:41 PM4/29/04
to

On 29 Apr 2004, at 16:05, Maxim wrote:

> I think that with all these nice workarounds it went unnoticed that
> there *was* something wrong with Mukhtar's example. Consider:
>
> In[1]:=
> Module[
> {y = Cos[x], f},
> f[x_] := y;
> FindRoot[f[x], {x, 1, 2},
> Compiled -> True, MaxIterations -> 15,
> StepMonitor :> Print["step: ", {x, f[x]}]]
> ]
>
> Out[1]=
> {x -> 1.5708}
>
> In[2]:=
> Module[
> {y = Cos[x], f},
> f[x_?NumericQ] := y;
> FindRoot[f[x], {x, 1, 2},
> Compiled -> True, MaxIterations -> 15,
> StepMonitor :> Print["step: ", {x, f[x]}],
> EvaluationMonitor -> 0]
> ]
>
>

> FindRoot::cvmit: Failed to converge to the requested

> accuracy or precision within 15 iterations.
>
> Out[2]=
> {x -> 2.}
>
> This is similar to the issue that was discussed at
>
> http://forums.wolfram.com/mathgroup/archive/2003/Dec/msg00401.html
>
> Just as with the Plot examples, f[1] here is not numerical because x
> in Cos[x] will remain unbound. But if we evaluate Block[{x=1}, f[x]],
> then the result will be numerical, and this is how FindRoot works. And
> then there's the question of how it will interact with Compiled->True.
>
> But here we have a new twist, because, according to Trace, FindRoot
> does obtain numerical values evaluating f[x] in both cases, so it's
> not clear what happens in the second example, when we use
> f[x_?NumericQ].
>
> Oh, and EvaluationMonitor/StepMonitor won't help us here, because they
> perform true delayed evaluation, while what FindRoot seems to do is to
> evaluate the first argument once and then to perform all computations
> on what it obtains (why does it have HoldAll attribute then? Simple:
> to add extra confusion if FindRoot is used inside Module or With and
> there is a name conflict). As an additional bonus, if you omit
> EvaluationMonitor in the second example, then StepMonitor won't print
> anything either. Lovely.
>
> Maxim Rytin
> m...@prontomail.com
>
>
>

I really fail to see what point are you trying to make? Certainly, your
use of Module has nothing to do with the problem which is entirely due
to the setting Compile->True. If you set Compile->False, which is only
natural since the function defined in this way obviously can't be
compiled, then everything, including StepMonitor works exactly as it
should. Why on earth you keep insisting that Mathematica should behave
correctly when you give her nonsensical input to work on?


y = Cos[x];
f[(z_)?NumericQ] := y;
FindRoot[f[x], {x, 1, 2}, Compiled -> False,


MaxIterations -> 15, StepMonitor :>
Print["step: ", {x, f[x]}]]


"step: "{2., -0.4161468365471424}


"step: "{1.5649043758915777, 0.005891916813448233}


"step: "{1.570978574535018, -0.00018224773911250164}


"step: "{1.570796325773051, 1.0218455583390503*^-9}


"step: "{1.5707963267948966, 6.123233995736766*^-17}


{x -> 1.5707963267948966}


Andrzej Kozlowski
Chiba, Japan
http://www.mimuw.edu.pl/~akoz/

Maxim

unread,
May 8, 2004, 1:46:47 AM5/8/04
to
The recommendation to simply set "EvaluateNumericalFunctionArgument"
to False ( http://support.wolfram.com/mathematica/mathematics/numerics/nsumerror.html
) can be a bad idea. The reason is that with the default settings for
FindRoot the jacobian is still pre-evaluated. Hence it may happen that
we are evaluating one function and using the derivative of an entirely
different function:

In[1]:=
Developer`SetSystemOptions["EvaluateNumericalFunctionArgument" ->
False];
Module[{f = If[TrueQ[# > 0], #, -#] &},
FindRoot[f[x] == 2, {x, 1}]
]

FindRoot::lstol: The line search decreased the step size to within
tolerance specified by AccuracyGoal and PrecisionGoal but was
unable to find a sufficient decrease in the merit function.
You may need more than MachinePrecision digits of working
precision to meet these tolerances.

Out[2]=
{x -> 1.}

One way is to use RuleDelayed for Jacobian:

In[3]:=
Developer`SetSystemOptions["EvaluateNumericalFunctionArgument" ->
False];
Module[{f = If[TrueQ[# > 0], #, -#] &},
FindRoot[f[x] == 2, {x, 1}, Jacobian :> {{f'[x]}}]
]

Out[4]=
{x -> 2.}

This method isn't guaranteed to work either, because the first step in
evaluation of f'[x] is to evaluate f', and the question of how
Mathematica does that is equally unclear:

In[5]:=
Module[{f, g},
f = If[TrueQ[# > 0], #, -#] &;
g[x_] := If[TrueQ[x > 0], x, -x];
{f', g'}
]

Out[5]=
{If[TrueQ[#1 > 0], 1, -1] &, -1 &}

The output for g' agrees with what the documentation says: g' is
evaluated as Evaluate[D[g[#],#]]& (actually the reference for
Derivative says D[f[#]&,{#,n}], which doesn't make any sense. So I
assume it was supposed to mean Evaluate[D[f[#],{#,n}]]&). But f' is
evaluated differently, because there are special rules for the
derivatives of pure functions (or maybe specifically for control
structures like If).

Because of this, the safest way is still to use f[x_?NumericQ], even
with "EvaluateNumericalFunctionArgument"->False, to block the
evaluation of f'...and then to hope that it doesn't give some
unexpected effect in combination with Compiled->True (which is the
default setting for all numerical functions that use Compiled).

Maxim Rytin
m...@prontomail.com

Andrzej Kozlowski

unread,
May 9, 2004, 3:08:12 AM5/9/04
to
On 8 May 2004, at 14:24, Maxim wrote:


>
> Because of this, the safest way is still to use f[x_?NumericQ], even
> with "EvaluateNumericalFunctionArgument"->False, to block the
> evaluation of f'...and then to hope that it doesn't give some
> unexpected effect in combination with Compiled->True (which is the
> default setting for all numerical functions that use Compiled).
>


There seem to be no point in using
"EvaluateNumericalFunctionArgument"->False together with f[x_?NumericQ]
since using the first option you loose all the performance benefits of
the new approach in Mathemtica 5.0 (which you can see in the examples
on the Wolfram support page). And secondly: I have not observed any
problems with using f[x_?NumericQ] except when one chooses to do
something unnatural like:

Module[
{y = Cos[x], f},
f[x_?NumericQ] := y;
FindRoot[f[x], {x, 1, 2},
Compiled -> True, MaxIterations -> 15,
StepMonitor :> Print["step: ", {x, f[x]}],
EvaluationMonitor -> 0]
]

Just changing it to


Module[
{ f},
f[x_?NumericQ] := Cos[x];


FindRoot[f[x], {x, 1, 2},
Compiled -> True, MaxIterations -> 15,
StepMonitor :> Print["step: ", {x, f[x]}]]
]

makes everything work perfectly. This seems to be the case with almost
every "problem" you have reported for as long as I can remember. (To
deal with the remaining ones, like why 1`2 == 2 gives True, etc,
requires only careful reading of the documentation).

Andrzej

> The recommendation to simply set "EvaluateNumericalFunctionArgument"
> to False (
> http://support.wolfram.com/mathematica/mathematics/numerics/
> nsumerror.html
> ) can be a bad idea. The reason is that with the default settings for
> FindRoot the jacobian is still pre-evaluated. Hence it may happen that
> we are evaluating one function and using the derivative of an entirely
> different function:
>
> In[1]:=
> Developer`SetSystemOptions["EvaluateNumericalFunctionArgument" ->
> False];
> Module[{f = If[TrueQ[# > 0], #, -#] &},
> FindRoot[f[x] == 2, {x, 1}]
> ]
>
> FindRoot::lstol: The line search decreased the step size to within
> tolerance specified by AccuracyGoal and PrecisionGoal but was
> unable to find a sufficient decrease in the merit function.
> You may need more than MachinePrecision digits of working
> precision to meet these tolerances.
>

> Out[2]=


> {x -> 1.}
>
> One way is to use RuleDelayed for Jacobian:
>
> In[3]:=
> Developer`SetSystemOptions["EvaluateNumericalFunctionArgument" ->
> False];
> Module[{f = If[TrueQ[# > 0], #, -#] &},
> FindRoot[f[x] == 2, {x, 1}, Jacobian :> {{f'[x]}}]
> ]
>

> Out[4]=
> {x -> 2.}
>

Maxim

unread,
May 11, 2004, 5:37:52 AM5/11/04
to
Andrzej Kozlowski <ak...@mimuw.edu.pl> wrote in message news:<c7klcs$2kn$1...@smc.vnet.net>...

> makes everything work perfectly. This seems to be the case with almost
> every "problem" you have reported for as long as I can remember. (To
> deal with the remaining ones, like why 1`2 == 2 gives True, etc,
> requires only careful reading of the documentation).

Actually this was discussed in a different thread (
http://forums.wolfram.com/mathgroup/archive/2004/Apr/msg00615.html ),
so I'll paste your answer from there:

--------paste----------
Andrzej Kozlowski wrote:

Besides it is well known that Mathematica's high precision arithmetic
model, based on approximate interval arithmetic, works well only for
numbers whose uncertainty is much smaller than the number itself. This
is the case in all applications I have ever heard of. For very low
precision numbers like your examples, it is known to give excessively
pessimistic or wrong answers, but so what?
---------end-----------

The statement that arbitrary-precision arithmetics in Mathematica is
based on interval arithmetics is inaccurate; perhaps we might say that
since we have an estimate for the error, then we also implicitly have
an interval, but this just leads to confusion with true interval
arithmetics which is implemented for Interval objects.

Instead, arbitrary-precision computations in Mathematica use
significance arithmetics and the notion of condition number to keep
track of precision. For example, squaring a number can lose one binary
digit, so after calculating the square Mathematica subtracts Log[10,2]
from the precision of the input -- it doesn't use Interval to
represent bignums and this squaring isn't performed on any intervals.

Next, the suggestion that the uncertainty should be much smaller
(exactly how much that would be?) than the number itself seems to be
rather vacuous -- if the uncertainty and the number are of the same
order, that would mean that no digits are known for certain, that is,
the precision is zero.

I wouldn't say that true interval arithmetics is 'known' to give wrong
answers -- that would totally defy the purpose of intervals; the whole
point of using them for approximate calculations is to obtain bounds
which are certain to encapsulate the possible error (so 'pessimistic'
and 'wrong' are more like antinomes -- pessimistic means that those
bounds might be too wide).

But now, as far as I understand from your last post, you changed your
mind and instead of saying that those examples just gave meaningless
output due to low precision (which doesn't seem like a wise thing for
Interval to do anyway), you claim that, after reading documentation,
you can offer some explanation other than "so what?". Then please
elaborate on why 1`2==2.2, 1`2==2.3 and
IntervalMemberQ[Sin[Interval[1`2]],0] work the way they do. I'd like
to hear the explanation, that's why I was asking the question in the
first place.

Maxim Rytin
m...@prontomail.com

Andrzej Kozlowski

unread,
May 13, 2004, 12:15:38 AM5/13/04
to
On 11 May 2004, at 18:20, Maxim wrote:

> Andrzej Kozlowski <ak...@mimuw.edu.pl> wrote in message
> news:<c7klcs$2kn$1...@smc.vnet.net>...
>

>> makes everything work perfectly. This seems to be the case with almost
>> every "problem" you have reported for as long as I can remember. (To
>> deal with the remaining ones, like why 1`2 == 2 gives True, etc,
>> requires only careful reading of the documentation).
>

I wrote "approximate" interval arithemtic. Mathematica's model of
significnace arithemtic is an approximation to interval aithmetic.
Mathemaitca uses an approximate model because real interval arithemtic
implemented via Interval is far too slow for everyday use. An
approximate model rather naturally suffers form weaknesses: in
Mathematica's case one of them is dealing with low precision numbers.
However, if you ned to do that you can always use true interval
arithemtic. As for your other question I will just cite from the
documentation for Equal:

Approximate numbers are considered equal if they differ in at most
their last eight binary digits (roughly their last two decimal digits).

Andrzej

Andrzej Kozlowski

unread,
May 14, 2004, 12:30:49 AM5/14/04
to
On 13 May 2004, at 23:26, Daniel Lichtblau wrote:
>
> The issue of 1`2==2.2 is a bit dicey. I believe the most useful
> interpretation is roughly as Andrzej suggests, that the model breaks
> down at sufficiently low precision. Yes, there are "contradictions" is
> the logic that lead to the equality above, but to me they are not
> terribly compelling.
>
>
> Daniel
>
>
At least in this respect Mathemaitca's behaviour agrees with the
documentation: since

RealDigits[1`2,2]

{{1,0,0,0,0,0,0},1}

has only 7 digits it does "differ in at most last eight binary digits"
from 2.

This is a bit more tricky

1.`2 == 0


True

Sin[1`2]==0

False

The key point is clearly the fact that here Sin slightly increases
precision:


Precision[Sin[1`2]]


2.1924

Looking at RealDigits[Sin[1`2],2] is not quite convincing, since
Mathematica still returns only 7 digits:


RealDigits[Sin[1`2],2]


{{1,1,0,1,1,0,0},0}

but I suspect that the statement in the documentation about "last eight
binary digits" is also only approximate and that this is actually a
border line case (just over the border line).

In any case, these are clearly pathologies of significance arithmetic
and without importance.

Maxim

unread,
May 14, 2004, 12:32:51 AM5/14/04
to
Andrzej Kozlowski <ak...@mimuw.edu.pl> wrote in message news:<c7uspa$q7s$1...@smc.vnet.net>...

In other words, using true interval arithmetic with low precision
numbers is safe? But that was exactly what I did in the examples with
Interval:

IntervalMemberQ[Interval[1`2],0]


IntervalMemberQ[Sin[Interval[1`2]],0]

How does your explanation clarify this inconsistency? As to Equal,
let's count digits:

In[1]:=
RealDigits[1`2, 2, 8, 1]
RealDigits[1`2, 2, 9, 1]

Out[1]=
{{0,1,0,0,0,0,0,0},2}

Out[2]=
{{0,1,0,0,0,0,0,0,0},2}

There's some uncertainty as to how many digits to take: by default
RealDigits returns 7 digits, but we can ask it to give us 8 digits;
all the next ones will be Indeterminate. I added leading zero just for
alignment with RealDigits[2.2,2] and RealDigits[2.3,2], which give

In[3]:=
RealDigits[2.2,2]
RealDigits[2.3,2]

Out[3]=
{{1,0,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,
1,0,0,1,1,0,0,1,1,0,0,1,1,0,1,0},2}

Out[4]=
{{1,0,0,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0,
0,1,1,0,0,1,1,0,0,1,1,0,0,1,1,0},2}

If the eight-digits rule applies here, 1`2 is either equal to anything
(eight digits is all we know) or unequal both to 2.2 and 2.3 by the
leading digit. Perhaps some more 'careful reading of documentation' is
required?

Maxim Rytin
m...@prontomail.com

Andrzej Kozlowski

unread,
May 14, 2004, 9:10:21 PM5/14/04
to

On 14 May 2004, at 13:12, Maxim wrote:

> Andrzej Kozlowski <ak...@mimuw.edu.pl> wrote in message

> news:<c7uspa$q7s$1...@smc.vnet.net>...

> In other words, using true interval arithmetic with low precision
> numbers is safe? But that was exactly what I did in the examples with
> Interval:
>
> IntervalMemberQ[Interval[1`2],0]

> IntervalMemberQ[Sin[Interval[1`2]],0]
>

> How does your explanation clarify this inconsistency?


You are not actually using "Interval arithmetic" here. At best,you are
using a mixture of Interval arithmetic and significance arithmetic. You
can certainly make accurate error estimates using interval arithmetic
with exact numbers.


IntervalMemberQ[Interval[{1-1/100,1+1/100}],0]
False
and so on.

I have already answered the other question in another posting. This
happens because Sin increases Precision (and by the way, it's not yet
another "paradox"). No inconsistency.


As I wrote in my other posting, the actual rule is probably based on
Precision and not exactly the number of digits, which only roughly
corresponds to it. The exact rule for identifying "close" approximate
numbers is not given because nobody needs it.
In any case, as I wrote before, significance arithmetic does not work
well with low precision numbers and that is the price one pays for it
being manageable fast. If you need error estimates for low precision
computations use(exact!) interval arithmetic.

Andrzej Kozlowski

unread,
May 15, 2004, 4:06:23 AM5/15/04
to
Since my other posting somehow did not make it or at least I have not
seen it I thought I should clarify just this one point, though I
consider all this exchange totally without any practical significance.


>> But that was exactly what I did in the examples with
>> Interval:
>>
>> IntervalMemberQ[Interval[1`2],0]

>> IntervalMemberQ[Sin[Interval[1`2]],0]
>>

>> How does your explanation clarify this inconsistency?
>
>

Firstly, as I wrote , this has nothing to do with interval arithmetic
an even nothing to do with Interval.
It is simply this:


1`2==0

True


Sin[1`2]==Sin[0]

False

Why is that? Well, the answer is here:


Precision[1`2]


1.9999999999999998


Precision[Sin[1`2]]


2.1924023244417263

Sin increases Precision and 2 digits of precision seem to be the
borderline for this problem. I can't tell with complete confidence
exactly what the "identifying" rule is, but I do not need to know nor
particularly want to know.

Andrzej

On 15 May 2004, at 09:58, Andrzej Kozlowski wrote:

>
> On 14 May 2004, at 13:12, Maxim wrote:
>
>> Andrzej Kozlowski <ak...@mimuw.edu.pl> wrote in message

>> news:<c7uspa$q7s$1...@smc.vnet.net>...

>> In other words, using true interval arithmetic with low precision
>> numbers is safe? But that was exactly what I did in the examples with
>> Interval:
>>
>> IntervalMemberQ[Interval[1`2],0]

>> IntervalMemberQ[Sin[Interval[1`2]],0]
>>

Maxim

unread,
May 17, 2004, 3:29:38 AM5/17/04
to
Andrzej Kozlowski <ak...@mimuw.edu.pl> wrote in message news:<c83qlt$li3$1...@smc.vnet.net>...

> On 14 May 2004, at 13:12, Maxim wrote:
>
> > In other words, using true interval arithmetic with low precision
> > numbers is safe? But that was exactly what I did in the examples with
> > Interval:
> >
> > IntervalMemberQ[Interval[1`2],0]
> > IntervalMemberQ[Sin[Interval[1`2]],0]
> >
> > How does your explanation clarify this inconsistency?
>
>
> You are not actually using "Interval arithmetic" here. At best,you are
> using a mixture of Interval arithmetic and significance arithmetic. You
> can certainly make accurate error estimates using interval arithmetic
> with exact numbers.
>
>
> IntervalMemberQ[Interval[{1-1/100,1+1/100}],0]
> False
> and so on.
>
> I have already answered the other question in another posting. This
> happens because Sin increases Precision (and by the way, it's not yet
> another "paradox"). No inconsistency.
>

But nobody says that Interval[Sin[1`2]] and Sin[Interval[1`2]] should
be identical; it's no surprise that Sin[Interval[{0,Pi}]] is not
Interval[{Sin[0],Sin[Pi]}].

Here's what the question was about (it seems that I have to start with
that). We create Interval[1`2]; numbers -0.29 and 2.2 belong to the
interval, therefore taking Sin of the interval we must obtain an
interval which includes at least the range from -0.28 to 1.0;
otherwise it would be like squaring Interval[{0,2}] and still getting
Interval[{0,2}] as the answer.

Instead, Sin[Interval[1`2]] returns an interval which includes (by
IntervalMemberQ) only numbers from 0.12 to 1.0.

It's quite plausible that, while the interval bounds are rounded
outwards as they should be, the increased precision 'eats up' this
rounding and in the end we get an interval which is narrower than
necessary. But to say that there is no inconsistency in the strict
mathematical sense is incorrect.

We've been through that before: if equality is determined by the
precision of the number (that would be explanation #4 or so by now?),
then only numbers in the range from 1-0.01 to 1+0.01 should be
considered as equal to 1`2.

Maxim Rytin
m...@prontomail.com

Andrzej Kozlowski

unread,
May 18, 2004, 4:26:39 AM5/18/04
to

On 17 May 2004, at 16:21, Maxim wrote:

> Andrzej Kozlowski <ak...@mimuw.edu.pl> wrote in message

> news:<c83qlt$li3$1...@smc.vnet.net>...


>> On 14 May 2004, at 13:12, Maxim wrote:
>>
>>> In other words, using true interval arithmetic with low precision
>>> numbers is safe? But that was exactly what I did in the examples with
>>> Interval:
>>>
>>> IntervalMemberQ[Interval[1`2],0]
>>> IntervalMemberQ[Sin[Interval[1`2]],0]
>>>
>>> How does your explanation clarify this inconsistency?
>>
>>
>> You are not actually using "Interval arithmetic" here. At best,you are
>> using a mixture of Interval arithmetic and significance arithmetic.
>> You
>> can certainly make accurate error estimates using interval arithmetic
>> with exact numbers.
>>
>>
>> IntervalMemberQ[Interval[{1-1/100,1+1/100}],0]
>> False
>> and so on.
>>
>> I have already answered the other question in another posting. This
>> happens because Sin increases Precision (and by the way, it's not yet
>> another "paradox"). No inconsistency.
>>
>

> But nobody says that Interval[Sin[1`2]] and Sin[Interval[1`2]] should
> be identical; it's no surprise that Sin[Interval[{0,Pi}]] is not
> Interval[{Sin[0],Sin[Pi]}].
>
> Here's what the question was about (it seems that I have to start with
> that). We create Interval[1`2]; numbers -0.29 and 2.2 belong to the
> interval, therefore taking Sin of the interval we must obtain an
> interval which includes at least the range from -0.28 to 1.0;
> otherwise it would be like squaring Interval[{0,2}] and still getting
> Interval[{0,2}] as the answer.
>
> Instead, Sin[Interval[1`2]] returns an interval which includes (by
> IntervalMemberQ) only numbers from 0.12 to 1.0.
>
> It's quite plausible that, while the interval bounds are rounded
> outwards as they should be, the increased precision 'eats up' this
> rounding and in the end we get an interval which is narrower than
> necessary. But to say that there is no inconsistency in the strict
> mathematical sense is incorrect.
>

> We've been through that before: if equality is determined by the
> precision of the number (that would be explanation #4 or so by now?),
> then only numbers in the range from 1-0.01 to 1+0.01 should be
> considered as equal to 1`2.
>
> Maxim Rytin
> m...@prontomail.com
>
>
>

This is argument is clearly not getting anywhere, and as it concerns
matters of so little significance that nobody else is likely to pay
attention to it, this is defintely my last response.

I never wrote that if equality is determined by the precision of the
number in the sense of interval arithmetic, because significance
arithemtic only approximates interval arithmetic. And by interval
arithmetic I mean what everybody except you means by interval
arithmetic: exact interval arithemtic. I do not know what the hybrid
one inplemented in Matheamtica is for, it seems ot me just a byproduct
of something.

The actual situation is this:

1. Mathematica has a fully implemented exact interval arithemtic. It is
fully logically self-consistent (barring bugs) and the fact that such
arithmetic must be in certian ways quite unlike ordinary arithmetic .
It is in fact "pessimistic" in the sense I described in another thread
that you started by claiming that a perfectly sensible answer was
wrong. (Since you seem to know very little about these matters I
suggest looking at 97 of "Structure and Interpretation of Computer
Programs" by Abelson and Sussman and solving problem 2.16. I would
certainly be interested in seeing your solution.) In any case, exact
interval arithmetic can be used to give bounds on errors when working
with low precision numbers. Fortunately most normal people do not do
this very often so the slowness is not a terrible problem.

2. Mathematica has also significance arithmetic, which is an
approximation to interval arithmetic. It is fast and works well in the
sort of situations that most people most of the time find themselves
in. It achieves this speed by "bending corners", in such a way that
computations with low precision numbers are generally very unreliable
and can give paradoxical results. In particular it produces the effect
you have re-discovered. (If you bothered to check the archives you
would have discovered that exactly the same discussion, including the
example with 1`2 was held here already a few years ago). I believe
that what i have written roughly corresponds to what actually happens
but in any case it does not matter at all. The reason why it does not
mater is that
1. You have been told that significance arithmetic is unreliable in
such cases
2. As Richard Fateman pointed out the last time this issue was
discussed (at least in my memory) question like:

1`2 == 2

do not make sense and there is never any reason to ask them.

3. If you can devise a system of arithmetic that has the advantages of
significance arithmetic without its defects than I am sure everyone
will be impressed and maybe even Wolfram will want to hire you.

Reply all
Reply to author
Forward
0 new messages