How do you calculate the sum of combinations of 1000 dice rolls?












3












$begingroup$


For two dice rolls we can calculate the number of combinations for each summed total:




  1. Rolling a 2: one chance (1&1)

  2. Rolling a 3: two chances (2&1)(1&2)

  3. Rolling a 4: three chances (3&1)(1&3)(2&2)

  4. Rolling a 5: four chances (4&1)(1&4)(3&2)(2&3)

  5. Rolling a 6: five chances (5&1)(1&5)(4&2)(2&4)(2&2)

  6. Rolling a 7: six chances (6&1)(1&6)(5&2)(2&5)(4&3)(3&4)

  7. Rolling an 8: five chances (6&2)(2&5)(5&3)(3&5)(4&4)

  8. Rolling a 9: four chances (6&3)(3&6)(5&4)(4&5)

  9. Rolling a 10: three chances (6&4)(4&6)(5&5)

  10. Rolling an 11: two chances (6&5)(5&6)

  11. Rolling a 12: one chance (6,6)


How do we go about this for n dice rolls?
For example how do we find the total number of values which will sum to 150 if we roll 100 die?










share|cite|improve this question











$endgroup$








  • 2




    $begingroup$
    For $n$ large we use approximations. For example, the sum $Y$ of the $1000$ rolls in the title has a close to normal distribution. The normal approximation would also be adequate for practical purposes even in the case $n=100$. Of course the sum will not be $50$.
    $endgroup$
    – André Nicolas
    Apr 3 '15 at 15:15












  • $begingroup$
    Answering your 2nd question, if you really need this partitioning, you'll need a generating function, but I doubt there'll be some nice form; it's better to follow Andre's suggestion and use CLT
    $endgroup$
    – Alex
    Apr 3 '15 at 15:17






  • 1




    $begingroup$
    How would you deal with this, if for example instead of having the sides valued 1-6, they are now values 0-5?
    $endgroup$
    – Adam
    Apr 3 '15 at 15:33










  • $begingroup$
    @AndréNicolas: actually there is a general formula, as reported in my answer: can you help to find the asymptotic for large $n$ ($m$)?
    $endgroup$
    – G Cab
    Nov 17 '16 at 18:04










  • $begingroup$
    @Alex: the ogf exists and have quite a simple formula (see my answer).
    $endgroup$
    – G Cab
    Nov 17 '16 at 18:05
















3












$begingroup$


For two dice rolls we can calculate the number of combinations for each summed total:




  1. Rolling a 2: one chance (1&1)

  2. Rolling a 3: two chances (2&1)(1&2)

  3. Rolling a 4: three chances (3&1)(1&3)(2&2)

  4. Rolling a 5: four chances (4&1)(1&4)(3&2)(2&3)

  5. Rolling a 6: five chances (5&1)(1&5)(4&2)(2&4)(2&2)

  6. Rolling a 7: six chances (6&1)(1&6)(5&2)(2&5)(4&3)(3&4)

  7. Rolling an 8: five chances (6&2)(2&5)(5&3)(3&5)(4&4)

  8. Rolling a 9: four chances (6&3)(3&6)(5&4)(4&5)

  9. Rolling a 10: three chances (6&4)(4&6)(5&5)

  10. Rolling an 11: two chances (6&5)(5&6)

  11. Rolling a 12: one chance (6,6)


How do we go about this for n dice rolls?
For example how do we find the total number of values which will sum to 150 if we roll 100 die?










share|cite|improve this question











$endgroup$








  • 2




    $begingroup$
    For $n$ large we use approximations. For example, the sum $Y$ of the $1000$ rolls in the title has a close to normal distribution. The normal approximation would also be adequate for practical purposes even in the case $n=100$. Of course the sum will not be $50$.
    $endgroup$
    – André Nicolas
    Apr 3 '15 at 15:15












  • $begingroup$
    Answering your 2nd question, if you really need this partitioning, you'll need a generating function, but I doubt there'll be some nice form; it's better to follow Andre's suggestion and use CLT
    $endgroup$
    – Alex
    Apr 3 '15 at 15:17






  • 1




    $begingroup$
    How would you deal with this, if for example instead of having the sides valued 1-6, they are now values 0-5?
    $endgroup$
    – Adam
    Apr 3 '15 at 15:33










  • $begingroup$
    @AndréNicolas: actually there is a general formula, as reported in my answer: can you help to find the asymptotic for large $n$ ($m$)?
    $endgroup$
    – G Cab
    Nov 17 '16 at 18:04










  • $begingroup$
    @Alex: the ogf exists and have quite a simple formula (see my answer).
    $endgroup$
    – G Cab
    Nov 17 '16 at 18:05














3












3








3


1



$begingroup$


For two dice rolls we can calculate the number of combinations for each summed total:




  1. Rolling a 2: one chance (1&1)

  2. Rolling a 3: two chances (2&1)(1&2)

  3. Rolling a 4: three chances (3&1)(1&3)(2&2)

  4. Rolling a 5: four chances (4&1)(1&4)(3&2)(2&3)

  5. Rolling a 6: five chances (5&1)(1&5)(4&2)(2&4)(2&2)

  6. Rolling a 7: six chances (6&1)(1&6)(5&2)(2&5)(4&3)(3&4)

  7. Rolling an 8: five chances (6&2)(2&5)(5&3)(3&5)(4&4)

  8. Rolling a 9: four chances (6&3)(3&6)(5&4)(4&5)

  9. Rolling a 10: three chances (6&4)(4&6)(5&5)

  10. Rolling an 11: two chances (6&5)(5&6)

  11. Rolling a 12: one chance (6,6)


How do we go about this for n dice rolls?
For example how do we find the total number of values which will sum to 150 if we roll 100 die?










share|cite|improve this question











$endgroup$




For two dice rolls we can calculate the number of combinations for each summed total:




  1. Rolling a 2: one chance (1&1)

  2. Rolling a 3: two chances (2&1)(1&2)

  3. Rolling a 4: three chances (3&1)(1&3)(2&2)

  4. Rolling a 5: four chances (4&1)(1&4)(3&2)(2&3)

  5. Rolling a 6: five chances (5&1)(1&5)(4&2)(2&4)(2&2)

  6. Rolling a 7: six chances (6&1)(1&6)(5&2)(2&5)(4&3)(3&4)

  7. Rolling an 8: five chances (6&2)(2&5)(5&3)(3&5)(4&4)

  8. Rolling a 9: four chances (6&3)(3&6)(5&4)(4&5)

  9. Rolling a 10: three chances (6&4)(4&6)(5&5)

  10. Rolling an 11: two chances (6&5)(5&6)

  11. Rolling a 12: one chance (6,6)


How do we go about this for n dice rolls?
For example how do we find the total number of values which will sum to 150 if we roll 100 die?







probability






share|cite|improve this question















share|cite|improve this question













share|cite|improve this question




share|cite|improve this question








edited Apr 3 '15 at 15:18







Adam

















asked Apr 3 '15 at 15:12









AdamAdam

165




165








  • 2




    $begingroup$
    For $n$ large we use approximations. For example, the sum $Y$ of the $1000$ rolls in the title has a close to normal distribution. The normal approximation would also be adequate for practical purposes even in the case $n=100$. Of course the sum will not be $50$.
    $endgroup$
    – André Nicolas
    Apr 3 '15 at 15:15












  • $begingroup$
    Answering your 2nd question, if you really need this partitioning, you'll need a generating function, but I doubt there'll be some nice form; it's better to follow Andre's suggestion and use CLT
    $endgroup$
    – Alex
    Apr 3 '15 at 15:17






  • 1




    $begingroup$
    How would you deal with this, if for example instead of having the sides valued 1-6, they are now values 0-5?
    $endgroup$
    – Adam
    Apr 3 '15 at 15:33










  • $begingroup$
    @AndréNicolas: actually there is a general formula, as reported in my answer: can you help to find the asymptotic for large $n$ ($m$)?
    $endgroup$
    – G Cab
    Nov 17 '16 at 18:04










  • $begingroup$
    @Alex: the ogf exists and have quite a simple formula (see my answer).
    $endgroup$
    – G Cab
    Nov 17 '16 at 18:05














  • 2




    $begingroup$
    For $n$ large we use approximations. For example, the sum $Y$ of the $1000$ rolls in the title has a close to normal distribution. The normal approximation would also be adequate for practical purposes even in the case $n=100$. Of course the sum will not be $50$.
    $endgroup$
    – André Nicolas
    Apr 3 '15 at 15:15












  • $begingroup$
    Answering your 2nd question, if you really need this partitioning, you'll need a generating function, but I doubt there'll be some nice form; it's better to follow Andre's suggestion and use CLT
    $endgroup$
    – Alex
    Apr 3 '15 at 15:17






  • 1




    $begingroup$
    How would you deal with this, if for example instead of having the sides valued 1-6, they are now values 0-5?
    $endgroup$
    – Adam
    Apr 3 '15 at 15:33










  • $begingroup$
    @AndréNicolas: actually there is a general formula, as reported in my answer: can you help to find the asymptotic for large $n$ ($m$)?
    $endgroup$
    – G Cab
    Nov 17 '16 at 18:04










  • $begingroup$
    @Alex: the ogf exists and have quite a simple formula (see my answer).
    $endgroup$
    – G Cab
    Nov 17 '16 at 18:05








2




2




$begingroup$
For $n$ large we use approximations. For example, the sum $Y$ of the $1000$ rolls in the title has a close to normal distribution. The normal approximation would also be adequate for practical purposes even in the case $n=100$. Of course the sum will not be $50$.
$endgroup$
– André Nicolas
Apr 3 '15 at 15:15






$begingroup$
For $n$ large we use approximations. For example, the sum $Y$ of the $1000$ rolls in the title has a close to normal distribution. The normal approximation would also be adequate for practical purposes even in the case $n=100$. Of course the sum will not be $50$.
$endgroup$
– André Nicolas
Apr 3 '15 at 15:15














$begingroup$
Answering your 2nd question, if you really need this partitioning, you'll need a generating function, but I doubt there'll be some nice form; it's better to follow Andre's suggestion and use CLT
$endgroup$
– Alex
Apr 3 '15 at 15:17




$begingroup$
Answering your 2nd question, if you really need this partitioning, you'll need a generating function, but I doubt there'll be some nice form; it's better to follow Andre's suggestion and use CLT
$endgroup$
– Alex
Apr 3 '15 at 15:17




1




1




$begingroup$
How would you deal with this, if for example instead of having the sides valued 1-6, they are now values 0-5?
$endgroup$
– Adam
Apr 3 '15 at 15:33




$begingroup$
How would you deal with this, if for example instead of having the sides valued 1-6, they are now values 0-5?
$endgroup$
– Adam
Apr 3 '15 at 15:33












$begingroup$
@AndréNicolas: actually there is a general formula, as reported in my answer: can you help to find the asymptotic for large $n$ ($m$)?
$endgroup$
– G Cab
Nov 17 '16 at 18:04




$begingroup$
@AndréNicolas: actually there is a general formula, as reported in my answer: can you help to find the asymptotic for large $n$ ($m$)?
$endgroup$
– G Cab
Nov 17 '16 at 18:04












$begingroup$
@Alex: the ogf exists and have quite a simple formula (see my answer).
$endgroup$
– G Cab
Nov 17 '16 at 18:05




$begingroup$
@Alex: the ogf exists and have quite a simple formula (see my answer).
$endgroup$
– G Cab
Nov 17 '16 at 18:05










2 Answers
2






active

oldest

votes


















2












$begingroup$

Note that the answer to your problem in general, can be given as follows.

Let us define:
$$
eqalign{
& {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
{rm 1} le {rm integer};y_{,j} le r + 1 hfill cr
y_{,1} + y_{,2} + ; cdots ; + y_{,m} = s + m hfill cr} right. = cr
& = {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
{rm 0} le {rm integer};x_{,j} le r hfill cr
x_{,1} + x_{,2} + ; cdots ; + x_{,m} = s hfill cr} right. = cr
& = N_b (s,r,m) cr}
$$

then we have the formula




$$
N_b (s,r,m)quad left| {;0 le {rm integers };s,m,r} right.quad = sumlimits_{left( {0, le } right),,k,,left( { le ,{s over r}, le ,m} right)} {left( { - 1} right)^k left( matrix{
m hfill cr
k hfill cr} right)left( matrix{
s + m - 1 - kleft( {r + 1} right) cr
s - kleft( {r + 1} right) cr} right)}
$$




Also consider that the o.g.f. on the parameter $s$ is:




$$
F_b (x,r,m) = sumlimits_{0, le ,,s,,left( { le ,mr} right)} {N_b (s,r,m);x^{,s} } = left( {1 + x^{,1} + x^{,2} + ; cdots ; + x^{,r} } right)^{,m} = left( {{{1 - x^{,r + 1} } over {1 - x}}} right)^{,m}
$$




For more details have a look to the answers to this other post.



When the number of rolls ($m$) takes large values, the formula above becomes impractical and we shall resort to an asymptotic approximation.

To this regard note that each variable $x_j$ is a discrete uniform variable with support $[0,r]$, therefore with mean and variance given by $$
mu = {r over 2}quad sigma ^{,2} = {{left( {r + 1} right)^{,2} - 1} over {12}}
$$

The sum of $m$ such variables tends very quickly to be Normally distributed with mean $m mu$ and variance $msigma ^2$, that is




$$
p(s,r,m) = {{N_{,b} (s,r,m)} over {left( {r + 1} right)^{,m} }};; to ;{cal N}left( {m{r over 2},;m{{left( {r + 1} right)^{,2} - 1} over {12}}} right)
$$




Refer also to this related post.






share|cite|improve this answer











$endgroup$





















    0












    $begingroup$

    This topic is called 'convolutions' in probability and computer science.
    Because of the intensive and repetitive computation necessary, finding
    exact probabilities of sums on n > 2 dice is usually done by computer
    algorithm and several examples are available by googling 'convolutions discrete probability dice'; some have nice pictures, even if you ignore the code. Also, I found one well-written book chapter that shows formulas in mathematical rather than program form:
    www.dartmouth.edu/~chance/teaching_aids/books_articles/probability_book/Chapter7.pdf



    The 'envelope' of the PDF for two dice is formed by two straight lines. For three
    dice the envelope is made of pieces of three parabolas, one for each tail, and an inverted
    one in the center, with a result that is already suggestive of the shape of a normal density.
    As n increases, envelopes very rapidly become very close to a normal PDF. As pointed out in comments, convergence is to be expected because of
    the Central Limit Theorem. But the rate of convergence is astonishingly fast.



    So for $n = 1000$, a normal distribution with the appropriate mean 1000(7/2) and variance 1000(35/12)
    should give an extremely good approximations to normal, even reasonably far into the tails (remembering, of course, that 1000 rolls of a die cannot give values less than 1000 or
    greater than 6000).



    Dice with any number of sides could be treated in the same way for $n ge 2$.
    A coin is also a 'two-sided die', and multiple coin tosses are binomial, which converges to normal. Even unfair dice would work, but some configurations of unfairness might converge more slowly to normal. Convolution algorithms work for unfair dice as well as fair ones, but results may be more difficult to summarize.






    share|cite|improve this answer











    $endgroup$













      Your Answer





      StackExchange.ifUsing("editor", function () {
      return StackExchange.using("mathjaxEditing", function () {
      StackExchange.MarkdownEditor.creationCallbacks.add(function (editor, postfix) {
      StackExchange.mathjaxEditing.prepareWmdForMathJax(editor, postfix, [["$", "$"], ["\\(","\\)"]]);
      });
      });
      }, "mathjax-editing");

      StackExchange.ready(function() {
      var channelOptions = {
      tags: "".split(" "),
      id: "69"
      };
      initTagRenderer("".split(" "), "".split(" "), channelOptions);

      StackExchange.using("externalEditor", function() {
      // Have to fire editor after snippets, if snippets enabled
      if (StackExchange.settings.snippets.snippetsEnabled) {
      StackExchange.using("snippets", function() {
      createEditor();
      });
      }
      else {
      createEditor();
      }
      });

      function createEditor() {
      StackExchange.prepareEditor({
      heartbeatType: 'answer',
      autoActivateHeartbeat: false,
      convertImagesToLinks: true,
      noModals: true,
      showLowRepImageUploadWarning: true,
      reputationToPostImages: 10,
      bindNavPrevention: true,
      postfix: "",
      imageUploader: {
      brandingHtml: "Powered by u003ca class="icon-imgur-white" href="https://imgur.com/"u003eu003c/au003e",
      contentPolicyHtml: "User contributions licensed under u003ca href="https://creativecommons.org/licenses/by-sa/3.0/"u003ecc by-sa 3.0 with attribution requiredu003c/au003e u003ca href="https://stackoverflow.com/legal/content-policy"u003e(content policy)u003c/au003e",
      allowUrls: true
      },
      noCode: true, onDemand: true,
      discardSelector: ".discard-answer"
      ,immediatelyShowMarkdownHelp:true
      });


      }
      });














      draft saved

      draft discarded


















      StackExchange.ready(
      function () {
      StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmath.stackexchange.com%2fquestions%2f1218764%2fhow-do-you-calculate-the-sum-of-combinations-of-1000-dice-rolls%23new-answer', 'question_page');
      }
      );

      Post as a guest















      Required, but never shown

























      2 Answers
      2






      active

      oldest

      votes








      2 Answers
      2






      active

      oldest

      votes









      active

      oldest

      votes






      active

      oldest

      votes









      2












      $begingroup$

      Note that the answer to your problem in general, can be given as follows.

      Let us define:
      $$
      eqalign{
      & {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
      {rm 1} le {rm integer};y_{,j} le r + 1 hfill cr
      y_{,1} + y_{,2} + ; cdots ; + y_{,m} = s + m hfill cr} right. = cr
      & = {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
      {rm 0} le {rm integer};x_{,j} le r hfill cr
      x_{,1} + x_{,2} + ; cdots ; + x_{,m} = s hfill cr} right. = cr
      & = N_b (s,r,m) cr}
      $$

      then we have the formula




      $$
      N_b (s,r,m)quad left| {;0 le {rm integers };s,m,r} right.quad = sumlimits_{left( {0, le } right),,k,,left( { le ,{s over r}, le ,m} right)} {left( { - 1} right)^k left( matrix{
      m hfill cr
      k hfill cr} right)left( matrix{
      s + m - 1 - kleft( {r + 1} right) cr
      s - kleft( {r + 1} right) cr} right)}
      $$




      Also consider that the o.g.f. on the parameter $s$ is:




      $$
      F_b (x,r,m) = sumlimits_{0, le ,,s,,left( { le ,mr} right)} {N_b (s,r,m);x^{,s} } = left( {1 + x^{,1} + x^{,2} + ; cdots ; + x^{,r} } right)^{,m} = left( {{{1 - x^{,r + 1} } over {1 - x}}} right)^{,m}
      $$




      For more details have a look to the answers to this other post.



      When the number of rolls ($m$) takes large values, the formula above becomes impractical and we shall resort to an asymptotic approximation.

      To this regard note that each variable $x_j$ is a discrete uniform variable with support $[0,r]$, therefore with mean and variance given by $$
      mu = {r over 2}quad sigma ^{,2} = {{left( {r + 1} right)^{,2} - 1} over {12}}
      $$

      The sum of $m$ such variables tends very quickly to be Normally distributed with mean $m mu$ and variance $msigma ^2$, that is




      $$
      p(s,r,m) = {{N_{,b} (s,r,m)} over {left( {r + 1} right)^{,m} }};; to ;{cal N}left( {m{r over 2},;m{{left( {r + 1} right)^{,2} - 1} over {12}}} right)
      $$




      Refer also to this related post.






      share|cite|improve this answer











      $endgroup$


















        2












        $begingroup$

        Note that the answer to your problem in general, can be given as follows.

        Let us define:
        $$
        eqalign{
        & {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
        {rm 1} le {rm integer};y_{,j} le r + 1 hfill cr
        y_{,1} + y_{,2} + ; cdots ; + y_{,m} = s + m hfill cr} right. = cr
        & = {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
        {rm 0} le {rm integer};x_{,j} le r hfill cr
        x_{,1} + x_{,2} + ; cdots ; + x_{,m} = s hfill cr} right. = cr
        & = N_b (s,r,m) cr}
        $$

        then we have the formula




        $$
        N_b (s,r,m)quad left| {;0 le {rm integers };s,m,r} right.quad = sumlimits_{left( {0, le } right),,k,,left( { le ,{s over r}, le ,m} right)} {left( { - 1} right)^k left( matrix{
        m hfill cr
        k hfill cr} right)left( matrix{
        s + m - 1 - kleft( {r + 1} right) cr
        s - kleft( {r + 1} right) cr} right)}
        $$




        Also consider that the o.g.f. on the parameter $s$ is:




        $$
        F_b (x,r,m) = sumlimits_{0, le ,,s,,left( { le ,mr} right)} {N_b (s,r,m);x^{,s} } = left( {1 + x^{,1} + x^{,2} + ; cdots ; + x^{,r} } right)^{,m} = left( {{{1 - x^{,r + 1} } over {1 - x}}} right)^{,m}
        $$




        For more details have a look to the answers to this other post.



        When the number of rolls ($m$) takes large values, the formula above becomes impractical and we shall resort to an asymptotic approximation.

        To this regard note that each variable $x_j$ is a discrete uniform variable with support $[0,r]$, therefore with mean and variance given by $$
        mu = {r over 2}quad sigma ^{,2} = {{left( {r + 1} right)^{,2} - 1} over {12}}
        $$

        The sum of $m$ such variables tends very quickly to be Normally distributed with mean $m mu$ and variance $msigma ^2$, that is




        $$
        p(s,r,m) = {{N_{,b} (s,r,m)} over {left( {r + 1} right)^{,m} }};; to ;{cal N}left( {m{r over 2},;m{{left( {r + 1} right)^{,2} - 1} over {12}}} right)
        $$




        Refer also to this related post.






        share|cite|improve this answer











        $endgroup$
















          2












          2








          2





          $begingroup$

          Note that the answer to your problem in general, can be given as follows.

          Let us define:
          $$
          eqalign{
          & {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
          {rm 1} le {rm integer};y_{,j} le r + 1 hfill cr
          y_{,1} + y_{,2} + ; cdots ; + y_{,m} = s + m hfill cr} right. = cr
          & = {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
          {rm 0} le {rm integer};x_{,j} le r hfill cr
          x_{,1} + x_{,2} + ; cdots ; + x_{,m} = s hfill cr} right. = cr
          & = N_b (s,r,m) cr}
          $$

          then we have the formula




          $$
          N_b (s,r,m)quad left| {;0 le {rm integers };s,m,r} right.quad = sumlimits_{left( {0, le } right),,k,,left( { le ,{s over r}, le ,m} right)} {left( { - 1} right)^k left( matrix{
          m hfill cr
          k hfill cr} right)left( matrix{
          s + m - 1 - kleft( {r + 1} right) cr
          s - kleft( {r + 1} right) cr} right)}
          $$




          Also consider that the o.g.f. on the parameter $s$ is:




          $$
          F_b (x,r,m) = sumlimits_{0, le ,,s,,left( { le ,mr} right)} {N_b (s,r,m);x^{,s} } = left( {1 + x^{,1} + x^{,2} + ; cdots ; + x^{,r} } right)^{,m} = left( {{{1 - x^{,r + 1} } over {1 - x}}} right)^{,m}
          $$




          For more details have a look to the answers to this other post.



          When the number of rolls ($m$) takes large values, the formula above becomes impractical and we shall resort to an asymptotic approximation.

          To this regard note that each variable $x_j$ is a discrete uniform variable with support $[0,r]$, therefore with mean and variance given by $$
          mu = {r over 2}quad sigma ^{,2} = {{left( {r + 1} right)^{,2} - 1} over {12}}
          $$

          The sum of $m$ such variables tends very quickly to be Normally distributed with mean $m mu$ and variance $msigma ^2$, that is




          $$
          p(s,r,m) = {{N_{,b} (s,r,m)} over {left( {r + 1} right)^{,m} }};; to ;{cal N}left( {m{r over 2},;m{{left( {r + 1} right)^{,2} - 1} over {12}}} right)
          $$




          Refer also to this related post.






          share|cite|improve this answer











          $endgroup$



          Note that the answer to your problem in general, can be given as follows.

          Let us define:
          $$
          eqalign{
          & {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
          {rm 1} le {rm integer};y_{,j} le r + 1 hfill cr
          y_{,1} + y_{,2} + ; cdots ; + y_{,m} = s + m hfill cr} right. = cr
          & = {rm No}{rm .},{rm of},{rm solutions},{rm to};left{ matrix{
          {rm 0} le {rm integer};x_{,j} le r hfill cr
          x_{,1} + x_{,2} + ; cdots ; + x_{,m} = s hfill cr} right. = cr
          & = N_b (s,r,m) cr}
          $$

          then we have the formula




          $$
          N_b (s,r,m)quad left| {;0 le {rm integers };s,m,r} right.quad = sumlimits_{left( {0, le } right),,k,,left( { le ,{s over r}, le ,m} right)} {left( { - 1} right)^k left( matrix{
          m hfill cr
          k hfill cr} right)left( matrix{
          s + m - 1 - kleft( {r + 1} right) cr
          s - kleft( {r + 1} right) cr} right)}
          $$




          Also consider that the o.g.f. on the parameter $s$ is:




          $$
          F_b (x,r,m) = sumlimits_{0, le ,,s,,left( { le ,mr} right)} {N_b (s,r,m);x^{,s} } = left( {1 + x^{,1} + x^{,2} + ; cdots ; + x^{,r} } right)^{,m} = left( {{{1 - x^{,r + 1} } over {1 - x}}} right)^{,m}
          $$




          For more details have a look to the answers to this other post.



          When the number of rolls ($m$) takes large values, the formula above becomes impractical and we shall resort to an asymptotic approximation.

          To this regard note that each variable $x_j$ is a discrete uniform variable with support $[0,r]$, therefore with mean and variance given by $$
          mu = {r over 2}quad sigma ^{,2} = {{left( {r + 1} right)^{,2} - 1} over {12}}
          $$

          The sum of $m$ such variables tends very quickly to be Normally distributed with mean $m mu$ and variance $msigma ^2$, that is




          $$
          p(s,r,m) = {{N_{,b} (s,r,m)} over {left( {r + 1} right)^{,m} }};; to ;{cal N}left( {m{r over 2},;m{{left( {r + 1} right)^{,2} - 1} over {12}}} right)
          $$




          Refer also to this related post.







          share|cite|improve this answer














          share|cite|improve this answer



          share|cite|improve this answer








          edited Jan 7 at 18:11

























          answered Nov 17 '16 at 17:57









          G CabG Cab

          18.1k31237




          18.1k31237























              0












              $begingroup$

              This topic is called 'convolutions' in probability and computer science.
              Because of the intensive and repetitive computation necessary, finding
              exact probabilities of sums on n > 2 dice is usually done by computer
              algorithm and several examples are available by googling 'convolutions discrete probability dice'; some have nice pictures, even if you ignore the code. Also, I found one well-written book chapter that shows formulas in mathematical rather than program form:
              www.dartmouth.edu/~chance/teaching_aids/books_articles/probability_book/Chapter7.pdf



              The 'envelope' of the PDF for two dice is formed by two straight lines. For three
              dice the envelope is made of pieces of three parabolas, one for each tail, and an inverted
              one in the center, with a result that is already suggestive of the shape of a normal density.
              As n increases, envelopes very rapidly become very close to a normal PDF. As pointed out in comments, convergence is to be expected because of
              the Central Limit Theorem. But the rate of convergence is astonishingly fast.



              So for $n = 1000$, a normal distribution with the appropriate mean 1000(7/2) and variance 1000(35/12)
              should give an extremely good approximations to normal, even reasonably far into the tails (remembering, of course, that 1000 rolls of a die cannot give values less than 1000 or
              greater than 6000).



              Dice with any number of sides could be treated in the same way for $n ge 2$.
              A coin is also a 'two-sided die', and multiple coin tosses are binomial, which converges to normal. Even unfair dice would work, but some configurations of unfairness might converge more slowly to normal. Convolution algorithms work for unfair dice as well as fair ones, but results may be more difficult to summarize.






              share|cite|improve this answer











              $endgroup$


















                0












                $begingroup$

                This topic is called 'convolutions' in probability and computer science.
                Because of the intensive and repetitive computation necessary, finding
                exact probabilities of sums on n > 2 dice is usually done by computer
                algorithm and several examples are available by googling 'convolutions discrete probability dice'; some have nice pictures, even if you ignore the code. Also, I found one well-written book chapter that shows formulas in mathematical rather than program form:
                www.dartmouth.edu/~chance/teaching_aids/books_articles/probability_book/Chapter7.pdf



                The 'envelope' of the PDF for two dice is formed by two straight lines. For three
                dice the envelope is made of pieces of three parabolas, one for each tail, and an inverted
                one in the center, with a result that is already suggestive of the shape of a normal density.
                As n increases, envelopes very rapidly become very close to a normal PDF. As pointed out in comments, convergence is to be expected because of
                the Central Limit Theorem. But the rate of convergence is astonishingly fast.



                So for $n = 1000$, a normal distribution with the appropriate mean 1000(7/2) and variance 1000(35/12)
                should give an extremely good approximations to normal, even reasonably far into the tails (remembering, of course, that 1000 rolls of a die cannot give values less than 1000 or
                greater than 6000).



                Dice with any number of sides could be treated in the same way for $n ge 2$.
                A coin is also a 'two-sided die', and multiple coin tosses are binomial, which converges to normal. Even unfair dice would work, but some configurations of unfairness might converge more slowly to normal. Convolution algorithms work for unfair dice as well as fair ones, but results may be more difficult to summarize.






                share|cite|improve this answer











                $endgroup$
















                  0












                  0








                  0





                  $begingroup$

                  This topic is called 'convolutions' in probability and computer science.
                  Because of the intensive and repetitive computation necessary, finding
                  exact probabilities of sums on n > 2 dice is usually done by computer
                  algorithm and several examples are available by googling 'convolutions discrete probability dice'; some have nice pictures, even if you ignore the code. Also, I found one well-written book chapter that shows formulas in mathematical rather than program form:
                  www.dartmouth.edu/~chance/teaching_aids/books_articles/probability_book/Chapter7.pdf



                  The 'envelope' of the PDF for two dice is formed by two straight lines. For three
                  dice the envelope is made of pieces of three parabolas, one for each tail, and an inverted
                  one in the center, with a result that is already suggestive of the shape of a normal density.
                  As n increases, envelopes very rapidly become very close to a normal PDF. As pointed out in comments, convergence is to be expected because of
                  the Central Limit Theorem. But the rate of convergence is astonishingly fast.



                  So for $n = 1000$, a normal distribution with the appropriate mean 1000(7/2) and variance 1000(35/12)
                  should give an extremely good approximations to normal, even reasonably far into the tails (remembering, of course, that 1000 rolls of a die cannot give values less than 1000 or
                  greater than 6000).



                  Dice with any number of sides could be treated in the same way for $n ge 2$.
                  A coin is also a 'two-sided die', and multiple coin tosses are binomial, which converges to normal. Even unfair dice would work, but some configurations of unfairness might converge more slowly to normal. Convolution algorithms work for unfair dice as well as fair ones, but results may be more difficult to summarize.






                  share|cite|improve this answer











                  $endgroup$



                  This topic is called 'convolutions' in probability and computer science.
                  Because of the intensive and repetitive computation necessary, finding
                  exact probabilities of sums on n > 2 dice is usually done by computer
                  algorithm and several examples are available by googling 'convolutions discrete probability dice'; some have nice pictures, even if you ignore the code. Also, I found one well-written book chapter that shows formulas in mathematical rather than program form:
                  www.dartmouth.edu/~chance/teaching_aids/books_articles/probability_book/Chapter7.pdf



                  The 'envelope' of the PDF for two dice is formed by two straight lines. For three
                  dice the envelope is made of pieces of three parabolas, one for each tail, and an inverted
                  one in the center, with a result that is already suggestive of the shape of a normal density.
                  As n increases, envelopes very rapidly become very close to a normal PDF. As pointed out in comments, convergence is to be expected because of
                  the Central Limit Theorem. But the rate of convergence is astonishingly fast.



                  So for $n = 1000$, a normal distribution with the appropriate mean 1000(7/2) and variance 1000(35/12)
                  should give an extremely good approximations to normal, even reasonably far into the tails (remembering, of course, that 1000 rolls of a die cannot give values less than 1000 or
                  greater than 6000).



                  Dice with any number of sides could be treated in the same way for $n ge 2$.
                  A coin is also a 'two-sided die', and multiple coin tosses are binomial, which converges to normal. Even unfair dice would work, but some configurations of unfairness might converge more slowly to normal. Convolution algorithms work for unfair dice as well as fair ones, but results may be more difficult to summarize.







                  share|cite|improve this answer














                  share|cite|improve this answer



                  share|cite|improve this answer








                  edited Apr 4 '15 at 0:09

























                  answered Apr 4 '15 at 0:02









                  BruceETBruceET

                  35.2k71440




                  35.2k71440






























                      draft saved

                      draft discarded




















































                      Thanks for contributing an answer to Mathematics Stack Exchange!


                      • Please be sure to answer the question. Provide details and share your research!

                      But avoid



                      • Asking for help, clarification, or responding to other answers.

                      • Making statements based on opinion; back them up with references or personal experience.


                      Use MathJax to format equations. MathJax reference.


                      To learn more, see our tips on writing great answers.




                      draft saved


                      draft discarded














                      StackExchange.ready(
                      function () {
                      StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmath.stackexchange.com%2fquestions%2f1218764%2fhow-do-you-calculate-the-sum-of-combinations-of-1000-dice-rolls%23new-answer', 'question_page');
                      }
                      );

                      Post as a guest















                      Required, but never shown





















































                      Required, but never shown














                      Required, but never shown












                      Required, but never shown







                      Required, but never shown

































                      Required, but never shown














                      Required, but never shown












                      Required, but never shown







                      Required, but never shown







                      Popular posts from this blog

                      Mario Kart Wii

                      What does “Dominus providebit” mean?

                      File:Tiny Toon Adventures Wacky Sports JP Title.png