# Solving Partition Problems with Linear Programming

Yesterday I promised to show how to solve partition problems with integer linear programming. Let us take the problem of yesterday: We want to split the numbers from 1 to 30 in 10 triplets (x,y,z) such that x+y+z=0 modulo 31. We will use linear problems of the following type.

$$Ax=b, \quad x_i \in \{0,1\}, \quad c^T x \to \text{max.}$$

There will be one variable in x for each possible triplet. The value of this variable determines if the triplet is in the selection (1) or not (0). So the j-th column of A refers to the triplet with number j. The i-th row of A is a linear constraint which makes sure that the number i is used only in one triplet. A contaoins only 0-1-values, and

$$a_{i,j} = 1 \quad\Leftrightarrow\quad i \in T_j = \{n_{1,j},n_{2,j},n_{3,j}\}.$$

The constraints will be

$$\sum_j a_{i,j} x_j = 1.$$

We just need any feasible point and could use any target function. Alternatively, we could use „less equal“ in the previous line and maximize the number of triplets used.

How to do that in EMT? We first store all possible triplets into one matrix.

>function makeT (n:index=31) ...
$T=[];$for i=1 to n-2;
$for j=i+1 to n-1;$       k=n-mod(i+j,n);
$if k!=i and k!=j then T=T_[i,j,k]; endif;$   end;
$end;$return T
$endfunction >makeT(7) 1 2 4 1 4 2 1 6 7 2 4 1 2 5 7 3 4 7 3 5 6 3 6 5 5 6 3 >T=makeT(31);  Then we define the matrix A. >function makeA (T) ...$v=sort(unique(flatten(T)));
$A=zeros(cols(v),rows(T));$for i=1 to cols(v);
$for j=1 to rows(T);$       if any(v[i]==T[j]) then A[i,j]=1; endif;
$end;$end;
$return A;$endfunction
>makeT(5)
1             4             5
2             3             5
>fraction makeA(makeT(5))
1         0
0         1
0         1
1         0
1         1
>A=makeA(T);
>size(A)
[31,  405]


The function makeA() can take any matrix with partitions in its rows. The first command simply selects the numbers that appear in any partition. In our case, A is simply the vector [1,…,30].

Now we need to solve the problem with integer linear programming. We use the LPSOLVE library for this (given to EMT by the developers).

>function solveP (A,T) ...
$x=ilpsolve(A,ones(rows(A))',ones(cols(A)),$         vlb=zeros(cols(A)),vub=ones(cols(A)),>max);
$return T[nonzeros(x')];$  endfunction
>solveP(A,T)
1             5            25
2             7            22
3            11            17
4             6            21
8             9            14
10            23            29
12            24            26
13            19            30
15            20            27
16            18            28


The return value of the function ilpsolve() is a 0-1-vector. We want to print the triplets which are marked by 1 in this vector. The variables vlb and vub are lower and upper bounds for the variables. Interestingly, the solution works without these restrictions. Nevertheless, more restrictions usually mean shorter calculations.

# Modulo Arithmetic with Euler Math Toolbox

Recently, I came across the problem to compute expressions in the modulo arithmetic with integer numbers. Here is an example: We like to compute expressions modulo 31. Since 31 is prime the numbers 0 to 30 with addition and multiplication modulo 31 will be a field. Each number a between 1 and 31 will have a unique multiplicative inverse b, such that ab=1 modulo 31. E.g.

>mod(27*23,31)
1


The inverse of 27 is 23 modulo 31.

How can one find the inverse element? The fastest way uses Euclid’s algorithm. If you compute the greatest common divider of 27 and 31 using Euclid’s algorithm you can derive x and y such that 27x+31y=1. Then x (taken modulo 31) will be the inverse of 27 modulo 31. Here is the computation for these numbers:

$$31=27+4, \quad 27=6 \cdot 4 + 3, \quad 4 = 3 + 1,$$

and by computing backwards

$$1 = 4 – 3 = 7 \cdot 4 – 27 = 7 \cdot 31 – 8 \cdot 27.$$

Note -8=23 modulo 31. We get our result 23.

Of course, we could program that in EMT or Maxima. Maxima contains already a function for the inverse modulo a number.

>&inv_mod(27,31)
23


We also know the small Lemma of Fermat.

$$a^{p-1} = 1 \text{ mod } p.$$

Thus

$$a \cdot a^{p-2} = 1 \text{ mod } p.$$

This can also be used to compute the inverse. To get the result in EMT we need a trick which I am going to explain below. But in Maxima we can compute very large powers, albeit very slowly.

>&mod(27^29,31)
23


For single computations or small numbers, symbolic computation is okay. But assume we need more speed. The computation above does not work in EMT, because the power is too large. Python, by the way, has infinite integers too. So the computation works there too.

>>> print 27**29
323257909929174534292273980721360271853387
>>> print 27**29 % 31
23


One of the reasons of this posting is to show how this can be done in EMT and other languages without an infinite integer arithmetic, and moreover much faster. The trick is the fact that we can as well take the modulo in each step of the multiplication. Here is the first simple code.

>mod(27^29,31) // WRONG RESULT!
12
>function map powmod (a,n,p) ...
$b=1;$  loop 1 to n; b=mod(b*a,p); end;
$return b$  endfunction
>powmod(27,29,31) // CORRECT RESULT!
23


However, this is still not the fastest way. For powers, we can use this trick.

$$x^{2y} = \left(x^y\right)^2, \quad x^{2y+1} = \left(x^y\right)^2 \cdot x.$$

This leads to the following code. By the way, the function matrixpower() of EMT uses the same technique. To understand the code, note that it is not efficient to use a mapping function recursively. Thus we call it from another function.

>function powmodrek (a,n,p) ...
$if n==0 then return 1$  elseif n==1 then return a
$elseif mod(n,2)==0 then$     b=powmodrek(a,n/2,p); return mod(b*b,p);
$else$     b=powmodrek(a,(n-1)/2,p); return mod(mod(b*b,p)*a,p);
$endfunction >function map powmod (a:number,n:nonnegative integer,p:index) ...$  return powmodrek(a,n,p);
$endfunction >powmod(27,29,31) 23 >powmod(27,1:30,31) [27, 16, 29, 8, 30, 4, 15, 2, 23, 1, 27, 16, 29, 8, 30, 4, 15, 2, 23, 1, 27, 16, 29, 8, 30, 4, 15, 2, 23, 1]  The last computation is not effective, of course. If we need to compute all powers we should use a simple loop instead. But I included that example to show that powmod() can map to vector input. As you see, the powers of 27 do not cycle through all the number 1 to 30. For other base numbers, they do. >powmod(3,1:30,31) [3, 9, 27, 19, 26, 16, 17, 20, 29, 25, 13, 8, 24, 10, 30, 28, 22, 4, 12, 5, 15, 14, 11, 2, 6, 18, 23, 7, 21, 1]  The second reason for this posting is to demonstrate the solution to another problem. We want to partition the numbers 1 to 30 into triplets with a+b+c=0 modulo 31. This is surprisingly easy. We have for all q different from 1 $$1+q+q^2 = \dfrac{q^3-1}{q-1}$$ Thus by the Lemma of Fermat $$3^0 + 3^{10} + 3^{20} = \dfrac{3^{30}-1}{3^{10}-1} = 0 \text{ mod }31.$$ Our triples can be found easily from this. Just multiply the equation with powers of 3. >for k=0 to 9; ... > powmod(3,k,31)+"+"+powmod(3,10+k,31)+"+"+powmod(3,20+k,31)+"=0", ... >end; 1+25+5=0 3+13+15=0 9+8+14=0 27+24+11=0 19+10+2=0 26+30+6=0 16+28+18=0 17+22+23=0 20+4+7=0 29+12+21=0  It is also possible to find such a partition with any desired property using linear programming and LPSOLVE in EMT. This should be the topic of another posting, however. # Euler Math Toolbox and Matlab Recently, this site got more traffic than usual. So I thought I welcome all of you in the new year (which is already three weeks old now). And I provide some new content for you to read. For me, the selected topic is important. I welcome very much any discussion about it. I am currently writing a requested paper for an on-line journal about Euler Math Toolbox, so I could benefit from your input. What is the difference between Euler Math Toolbox and Matlab? Or rather, what is the intended difference? And are there alternatives to both? First of all the following points are obvious to me. • Matlab is profiling itself in the professional market. It is called the „industry standard“, and it is advertised as a must-know to students. Matlab tries very much to serve the needs of the industry. The best example is its simulation toolbox which connects Matlab to hardware. Matlab is backed up by a group of commercial programmers. • Euler Math Toolbox is for education and research. It is not intended to be used as a professional tool in industry. EMT is backed up by open software. • Both systems fall short for specialized applications. Most professional or research software is compiled in a general programming language and does things that a one-for-all software cannot do. • We need to educate our students in general programming, not in software systems, especially not in commercial systems. With a good background in programming any software can be adapted easily. • That said, a tool like Euler Math Toolbox or Matlab can be useful for quick computations or to mathematical demos. And for this, both systems are equally well equipped with an advantage for EMT. • We should have free and open, reliable numerical libraries for our main computer languages. If you are interested in the primary differences between EMT and Matlab, here is a list. • EMT is open and free, Matlab is commercial with a considerable pricing. • The user interface of EMT is notebook oriented similar to Maple or Mathematica, Matlab is command oriented. Both have scripts, of course, and are similar in this respect. • Both systems can do symbolic computations, EMT with the free Maxima, Matlab with an additional package you need to buy. • EMT can use many open external programs like Povray, Python, C, Scilab, Latex. For Matlab, you can buy powerful external packages to do industry strength computations. • Matlab has an optional compiler. In EMT, you need to write C or Python for more speed. • Matlab is used all around the world, and you will easily find someone with the same problem as you have. EMT has a community, but it is way smaller. There might be more differences that I forgot to mention. But all in all it is clear to me that EMT deserves to be used in education much more than it is right now. Matlab, on the other hand, should be removed from general education. It should be reserved for specialized courses done by engineers who have used the program for their work. Alternatives? There are many. • R is a statistical package featuring a matrix language as well. It is free and has a big community. The programming in R is not as easy to learn as they claim, but to get statistics done it is not necessary. The interface is ugly and cumbersome to use. • Maple, Mathematica or other algebra software are commercial packages in the higher price range. They offer more advanced features than Maxima and often yield better results. But they are not designed to be used as numerical software. • Geogebra tries to be the knife for all purposes. It is a beautiful software for schools and has a large community. Clearly, it falls short in the area of numerical and other programming. Actually, it is a whole different category of software. • Python is probably the best alternative to big software packages. There are very nice packages for plotting and for numerical computations. The symbolic package falls short in comparison, however. The Sage project relies completely on Python and has a web interface which clearly is the proper way in the future. • General programming langes are the way to do software education in universities for engineers and math students. Most courses have an obligatory course in Java, and rightly so. Open for discussion. # Pure Symbolic Functions in Maxima and EMT I like to demonstrate a few tricks with Maxima and EMT. Assume you want to generate the Vandermonde matrix, and check the identity $$\det \begin{pmatrix} 1 & x_0 & \ldots & x_0^n \\ \vdots & \vdots & & \vdots \\ 1 & x_n & \ldots & x_n^n \end{pmatrix} = \prod\limits_{0 \le i<j \le n} (x_j-x_i)$$ You can generate this matrix easily in EMT for specific values. >n=3; v=0:3; V=v'^(0:3) 1 0 0 0 1 1 1 1 1 2 4 8 1 3 9 27 >det(V) 12 It is a bit more difficult to check the right hand side of the equation. Of course, you can simply write a function >function vprod (v:vector) ...$p=1; n=length(v);
$for i=1 to n;$   for j=i+1 to n;
$p=p*(v[j]-v[i]);$   end;
$end;$return p;
$endfunction >vprod(v) 12 It is also possible with Matrix tricks. We generate all differences in a matrix, square the matrix, set the diagonal to 1, take the total product of all matrix elements and take the 4-th root. >(prod(prod(setdiag((v-v')^2,0,1))'))^(1/4) 12 But that is not the general case with variables. To generate this we use purely symbolic functions in EMT. These functions evaluate in Maxima at the time when they are used, not when they are defined. >function mrow (k,v,n) &&= create_list(v[k]^m,m,0,n) m create_list(v , m, 0, n) k >function Vand (v,n) &&= apply(matrix,create_list(mrow(k,v,n),k,0,n)) apply(matrix, create_list(mrow(k, v, n), k, 0, n)) >&Vand(v,2) [ 2 ] [ 1 v v ] [ 0 0 ] [ ] [ 2 ] [ 1 v v ] [ 1 1 ] [ ] [ 2 ] [ 1 v v ] [ 2 2 ] This is almost self explaining, once you know that „matrix(a,b,c)“ generates a matrix in Maxima with rows „a,b,c“. The creation takes place for the specific symbolic v when „Vand(v,2)“ is called. Maxima prints the result with indices. But it is a string of the following form. >V &= Vand(v,2); >""+V matrix([1,v[0],v[0]^2],[1,v[1],v[1]^2],[1,v[2],v[2]^2]) Since EMT vectors start with index 1 by default, we can only evaluate this expression for zero based vectors (unless we modify the functions above a little bit). >v=0:2; zerobase v; >V() 1 0 0 1 1 1 1 2 4 Maxima can factor the determinant of any Vandermonde determinant of specific size. >&factor(determinant(Vand(v,3))) (v - v ) (v - v ) (v - v ) (v - v ) (v - v ) (v - v ) 1 0 2 0 2 1 3 0 3 1 3 2 This is in accordance with the theory. Let me try the following matrix. $$\begin{pmatrix} (x_0-x_0)^n & \ldots & (x_0-x_n)^n \\ \vdots & & \vdots \\ (x_n-x_0)^n & \ldots & (x_n-x_n)^n \end{pmatrix}$$ >function mrow (k,v,n) &&= create_list((v[k]-v[m])^n,m,0,n) n create_list((v - v ) , m, 0, n) k m >function Vand (v,n) &&= apply(matrix,create_list(mrow(k,v,n),k,0,n)) apply(matrix, create_list(mrow(k, v, n), k, 0, n)) >&Vand(v,2) [ 2 2 ] [ 0 (v - v ) (v - v ) ] [ 0 1 0 2 ] [ ] [ 2 2 ] [ (v - v ) 0 (v - v ) ] [ 1 0 1 2 ] [ ] [ 2 2 ] [ (v - v ) (v - v ) 0 ] [ 2 0 2 1 ]  I am not sure if you are aware of the following formula. >&factor(determinant(Vand(v,3))) 2 2 2 2 2 9 (v - v ) (v - v ) (v - v ) (v - v ) (v - v ) 1 0 2 0 2 1 3 0 3 1 2 (v - v ) 3 2  The factor 9 is interesting. For n=4, it is 96, and for n=5, it is 2500. Do you guesswork. The factorization of the case n=4 takes some seconds. I did not try higher degrees. Of course, it is easy to compute this numerically, even for very large matrices. >function d1(v) := det((v'-v)^(cols(v)-1)) >d1(0:10) 1.43679021499e+072  # December Problem I found the following problem on the page of the wonderful Science Blogs. It is in German. Let me translate it. Bernd writes two different numbers between 1 and 1000 on two cards of paper. Anna selects one, and after seeing it has to decide if she wants to keep it or change to the other card. The player with the higher card wins. This is a case of a two person game with the following tactics for each player. • Bernd: He can write any combination (i,j) with 1 <= i < j <= N. • Anna: She can decide to change up to a number A with 1 <= A <= N. Note that we simplified the situation for Anna. It does not make sense to change with 3, but not with a smaller number. Weather or not this simplification makes the outcome worse for Anna remains to be investigated. Now, Bernd can select any of his tactics (i,j) with probability p(i,j), and Anna can select any of her tactics with probability q(A). The problem to select the optimal strategy (i.e. the distribution of probabilities) for each player can be solved with optimization. I have done similar problems with EMT on this page. I solved the problem above for (N=10) with a bit of code. However, I want to spoil the solution here. If you want to try on your own, do not read further. • Strategy for Bernd: Bernd should select the pairs (1,2),…,(N-1,N) at random with equal probability. • Strategy for Anna: Before each game, Anna should select a number A from 1 to N-1 at random, and switch if her number is less or equal A. Let us prove that this is optimal. Assume, Anna selects tactics A. So she will change cards at any number smaller or equal A. Then Bernd will lose 1/(N-1) on average with his strategy. This is so, since the only case that makes a difference for him is the pair (A,A+1), where Anna wins all the time. So, no matter what tactics A Anna selects, the worst case is an average loss of 1/(N-1). Only if Anna would select A=N, he would do better. Now assume, Bernd selects tactics (i,j). Then Anna, with her strategy, will win in any case if i<=A<j. In all other cases, the average result is 0. We need to find the worst case for Anna based on her strategy. That is indeed the case, if Bernd selects a pair (i,i+1) and it is equal to a win of 1/(N-1). It is now clear that there cannot be a better strategy for either player. Anna cannot win more than Bernd loses on average. With a bit more care in the proof, it is also clear that it would not help Anna to add more strategies. Assume, she rejects some numbers, and accepts others, instead of rejecting the numbers 1 to A altogether for some A. Then Bernds tactics will lose less or equal 1/(N-1) on average. For an example, assume Anna rejects 1 and 3. Then (1,2) will be a loss for Bernd, (2,3) will be a win, and (3,4) will be a loss, all other cases are 0 on average. This is a very nice problem. I could not find a reference in the net, however. But it should be known. Let me add the code for EMT. >v1=flatten(dup(1:10,10)); >v2=flatten(dup((1:10)',10)); >i=nonzeros(v1<v2); v1=v1[i]; v2=v2[i]; >A=(1:10)'; >M=-((v1<=A)&&(A<v2)); >function game (R) ...$## Maximizes G such that R'.q>=G, sum q_i = 1, q_i>=0.
$## Returns {q,G}.$    m=rows(R); n=cols(R);
$M=R|-1_ones(1,n); b=zeros(m,1)_1; eq=ones(m,1)_0;$    c=zeros(1,n)|1; restr=ones(1,n)|0;
$res=simplex(M,b,c,eq,restr,>max);$    return {res[1:n]',res[-1]};
$endfunction >{pB,res}=game(M); fraction res -1/9 >k=nonzeros(pB); fraction (pB[k]_v1[k]_v2[k])' 1/9 1 2 1/9 2 3 1/9 3 4 1/9 4 5 1/9 5 6 1/9 6 7 1/9 7 8 1/9 8 9 1/9 9 10 >{pA,res}=game(-M'); fraction res 1/9 >fraction pA'|(1:10)' 1/9 1 1/9 2 1/9 3 1/9 4 1/9 5 1/9 6 1/9 7 1/9 8 1/9 9 0 10 This looks a bit difficult. Here are some explanations. The vectors v1 and v2 simply contain the pairs (i,j) with 1 <= i < j <= N = 10. The trick to get all pairs is to create matrices with 1 to 10 in each row resp. column and to „flatten“ them to vectors. Then we select only the pairs with i < j using the nonzeros() function of EMT. The matrix M is then constructed in the following way: The element in row A and column (i,j) is -1 (a loss for Bernd) if i <= A < j. This is the matrix of average outcomes if Anna selects A and Bernd selects (i.j). The game() function calls the simplex algorithm for the following problem. $$Mp \ge G, \quad \sum_k p_k = 1, \quad G \to \text{Max.}$$ The probabilities p(k) are then maximizing the worst case G, for all rows of M (tactics of Anna). The result is p and G. We print the nonzero elements of p along with the corresponding i and j. The same is finally done with the problem seen from Anna’s side. By the way, the fact that both results must be equal (besides the sign) is an application of duality in optimization. Really a nice problem! # Linear Optimization with Maxima and EMT Maxima and EMT do of course have implementations of the Simplex algorithm. The algorithm in Maxima seems to work differently for floating point numbers and for fractions. Here is an example in fractions. We solve a problem in standard form. $$Ax=b, \quad c^Tx \to \text{Min.}$$ Then Maxima has the function linear_program(). >&load(simplex); >A &:= [1,1,-1,0;2,-3,0,-1;4,-5,0,0] 1 1 -1 0 2 -3 0 -1 4 -5 0 0 >b &:= [1,1,6] [1, 1, 6] >c &:= [1,-2,0,0] [1, -2, 0, 0] >&linear_program(A,b,c) 13 19 3 [[--, 4, --, 0], - -] 2 2 2 Since we have carefully defined the data as numerical and symbolic variables, we can use the same data for the simplex() algorithm in EMT. >x=simplex(A,b',c,eq=0,>min); fraction x' [13/2, 4, 19/2, 0] Note „eq=0“, which sets all equations (in each row of A) to equalities. The default is „eq=-1“ which means inequalilites of the form „<=“. The solution would then be different. There is another form of the command in Maxima. >&minimize_lp(x-2*y, ... > [x+y-z=1,2*x-3*y-v=1,4*x-5*y=6,x>=0,y>=0,z>=0,v>=0]) 3 19 13 [- -, [v = 0, z = --, y = 4, x = --]] 2 2 2 Of course, the simplex() command and the Maxima commands can do a lot more. EMT does also have efficient packages for integer programming. Start with the tutorial to learn more about this. With only two variables, we can even draw the feasible set. With three variables, we can use Povray to do this. In the examples of EMT, you find the following plot. # Povray in Euler Math Toolbox EMT can do a lot of 3D plots, but it is not an OpenGL application. This means that it has restrictions concerning hidden surfaces and intersecting objects. But if you install Povray you can do a lot more. Note that you need to put the path to the installation into your system path, or set the „povengine“ variable as described in the tutorial about Povray in EMT. In this tutorial you will find a lot of examples to get you started. Combining EMT with Povray is a fascinating application of EMT. The first simple function is pov3d(). This function is similar to plot3d(). Here is an example. >load povray Functions to generate Povray files for the open Raytracer. >pov3d("x^2+y^3",axiscolor=red,angle=20°, ... > look=povlook(red,0.2),level=-1:0.5:1,zoom=3.8); As you see, you can use an expression, set angles, zoom, and levels just as in plot3d(). The new parameter is „look“, which makes the plot a transparent red. The output will cast shadows from the light source. Moreover, there is a coordinate axis. You can control all that in the plv3d() function. The graphics is not in the graphics window, but in the text window, just like inserted graphics from EMT. Moreover, the graphics file is in the working directory of EMT. By default, this is the directory „Euler“ in your user directory („C:\Users\Name\Euler“). By default, it is a 450×450 PNG. But you can do much more in EMT with Povray. EMT can handle Povray is by Povray objects stored as Povray commands in strings. For some large objects files are used. If you generate a Povray scene with these objects, you need to start the generator with povstart() and finish with povend(). That command will also start the Povray program. Here is an example. >povstart(angle=-10°,center=[0.5,0.5,0.5],zoom=7); >x=0:0.02:1; y=x'; z=x*y; vx=-y; vy=-x; vz=1; >mesh=povtriangles(x,y,z,"",vx,vy,vz); >cl=povdisc([0.5,0.5,0],[1,1,0],2); ... >ll=povdisc([0,0,1/4],[0,0,1],2); >writeln(povdifference(mesh,povunion([cl,ll]),povlook(green))); >writeln(povintersection([mesh,cl],povlook(red))); ... >writeln(povintersection([mesh,ll],povlook(gray))); >writeln(povpoint([1/2,1/2,1/4],povlook(gray),size=2*defaultpointsize)); >writeAxes(0,1,0,1,0,1,d=0.015); ... >povend(w=600,h=600); Let me explain the commands one by one. First, we start the new scene with povstart(). This generates a Povray file in the user directory. It does also set the camera angle, the point to look at, and the camera zoom. It can set a lot more. You need to read the documentation for Povray in EMT for more information. We now generate objects. The povtriangles() functions generates a file containing the description of a mesh and a string containing the Povray command for this mesh. The string „mesh“ simply contains „object { mesh1 }“. We then generate more strings containing commands for Povray objects. E.g., cl is a disk which is later used to intersect the green surface with in height 1/2. >cl cylinder { <0.492929,0.492929,0>, <0.507071,0.507071,0>, 2 texture { pigment { color rgb <0.470588,0.470588,0.470588> } } finish { ambient 0.2 } } The function difference() then is a command computing the difference of the mesh „mesh“ and the disk „cl“. >povdifference(mesh,povunion([cl,ll]),povlook(green)) difference { object { mesh1 } union { cylinder { <0.492929,0.492929,0>, <0.507071,0.507071,0>, 2 texture { pigment { color rgb <0.470588,0.470588,0.470588> } } finish { ambient 0.2 } } cylinder { <0,0,0.24>, <0,0,0.26>, 2 texture { pigment { color rgb <0.470588,0.470588,0.470588> } } finish { ambient 0.2 } } texture { pigment { color rgb <0.470588,0.470588,0.470588> } } finish { ambient 0.2 } } texture { pigment { color rgb <0.0627451,0.564706,0.0627451> } } finish { ambient 0.2 } } We use writeln() to write this string to the Povray file which has been opened in EMT with the open() command. That is the main trick to generate Povray scenes in EMT. Finally, povend() generates the file. This time, we specify a larger image which can be found in the user directory of EMT. There is a lot more you can do with this combination of EMT and Povray. In fact, you can do everything, Povray is capable of. You may need to define new functions for this. Have a look at the tutorial and the visit the links at the end for more examples. # Random Dices This is just an answer to a question that came up in a discussion. Can you simulate 2 dice throws and display the frequencies in EMT? Actually, this is quite simple. We demonstrate it with 20 repetitions first. The trick in a matrix language is always to do all random stuff beforehand, and then to apply matrix functions to get the desired statistics. In our case sum(W)‘ does the job. It sums all rows of W and returns a vector with of sums. Then getmultiplicities() extracts the statistics. In our case, we know that the numbers 2 to 12 can appear in the sum. If we don’t know that, sort(unique()) extracts a sorted set of unique elements in the data. >n=20; >W=intrandom(n,2,6) 6 2 4 1 5 3 1 3 5 6 3 3 1 2 6 4 5 6 3 2 1 3 5 4 6 3 3 3 6 1 3 3 6 2 6 6 4 4 5 3 >sum(W)' [8, 5, 8, 4, 11, 6, 3, 10, 11, 5, 4, 9, 9, 6, 7, 6, 8, 12, 8, 8] >f=getmultiplicities(2:12,sum(W)') [0, 1, 2, 2, 3, 1, 5, 2, 1, 2, 1]  We change n to 20000, and see the expected distribution. >columnsplot(f,lab=2:12):  # About Multi-Variate Functions in EMT You define a vector valued function of two variables in EMT as follows. >function f([x,y]) := [x^2+y^2-10,x*y-1] >broyden("f",[1,2]) [0.317837, 3.14626]  Then f can be used in the form f(x,y) or f(v) with a 1×2 vector v. The Broyden algorithm needs the second form. This can also be done for a symbolic function. >function f([x,y]) &= [x^2+y^2-10,x*y-1] 2 2 [y + x - 10, x y - 1] >function Df([x,y]) &= jacobian(f(x,y),[x,y]) [ 2 x 2 y ] [ ] [ y x ] >newton2("f","Df",[1,2]) [0.317837, 3.14626]  The function f can now work numerically just as above. As a symbolic function, it is defined in Maxima too. There, it can only be used as f(x,y). Likewise for the derivative of f defined in Df. The Newton algorithm needs this derivative matrix to do its job. Note, that the Broyden and the Newton algorithms work for row and for column vectors alike. It depends on the start vector. E.g., if you start with a row vector, the algorithms expect the functions to return row vectors. You can also use the following trick. >function f(v) &= [v[1]^2+v[2]^2-10,v[1]*v[2]-1] 2 2 [v + v - 10, v v - 1] 2 1 1 2 >function Df(v) &= jacobian(f(v),[v[1],v[2]]) [ 2 v 2 v ] [ 1 2 ] [ ] [ v v ] [ 2 1 ] >newton2("f","Df",[1,2]) [0.317837, 3.14626]  The first definition is a valid definition of a numerical function. But for Maxima, the symbolic function can also be defined this way. Maxima interprets v[1] as a variable, as you see in the definition of Df. Of course, you can no longer write f(x,y). # Visualizing the Newton Method in EMT In a post by squareCircleZ I found a visualization of the Newton method with JavaScript. This looks like a modern and sleek way of providing demos to students. I have some doubts about the depth of the insights gained by this click and try demos. But, for a first clue, they seem to be okay. However, I’d much prefer this to be done with a little bit of programming in EMT. Here is the code. I am going to explain it below. >function f(x) &= x^4+x-1 4 x + x - 1 >function df(x) &= diff(f(x),x) 3 4 x + 1 >function newstep (x) &= x - f(x)/df(x) 4 x + x - 1 x - ---------- 3 4 x + 1 >&ratsimp(newstep(x)) 4 3 x + 1 -------- 3 4 x + 1 >plot2d("f",0,2); >function vstep (x) ...$  plot2d([x,x],[0,f(x)],style="--",color=blue,>addpoints,>add);
$x1=newstep(x);$  plot2d([x,x1],[f(x),0],style="-",color=blue,>add);
$return x1;$endfunction
>x=1.9;
>loop 1 to 8; x=vstep(x); end:

It is all quit straightforward. Using Maxima, we compute the derivative or our symbolic function f(x), and then the Newton step. The simplification with ratsimp() is only to show that the original form can be simplified quite substantially. We plot the function, and define a function vstep() which draws two lines (one with endpoints) and returns the new value of the Newton algorithm. Then we do this 8 times starting from x=1.9. Note the colon : after the last statement, which inserts the graphics into the notebook.

If you save all that in an EMT notebook it should provide a nice environment for own experiments by the students. With a little bit of thought, we can do a lot more now. E.g., we could try to experimentally prove the the order of the algorithm is quadratic. This involves logarithms and quite a bit of thought.

By the way, Maxima can also give us the formula for the root of this particular function. As you will see, it is not very useful.

>\$solve(f(x),x)[2]