#codingexercise matrix addition
Int [, ] addition ( int [,] left, int [,] right)
{
If (left == null || right == null) return null;
Int lc  = left.GetLength (0);
Int lr   = left.GetLength (1);
Int rc  = right.GetLength (0);
Int rr  = right.GetLength (1);
If (lc != rc || lr != rr) return null;
Var ret = new int [lr, lc]();
For (int r = 0; r < lr;  r++)
{
   For (int  c = 0; c < lc; c++)
     {
           Ret [r,c] += left [r,c] + right [r,c];
      }
}
Return ret;
}
#codingexercise matrix subtraction

Int [, ] Subtraction ( int [,] left, int [,] right)

{

If (left == null || right == null) return null;

Int lc  = left.GetLength (0);

Int lr   = left.GetLength (1);

Int rc  = right.GetLength (0);

Int rr  = right.GetLength (1);

If (lc != rc || lr != rr) return null;

Var ret = new int [lr, lc]();

For (int r = 0; r < lr;  r++)

{

   For (int  c = 0; c < lc; c++)

     {

           Ret [r,c] += left [r,c] - right [r,c];

      }

}

Return ret;

}

In continuation of our discussion on the convergence of steepest descent method, we will see that there is instant convergence even with a set of eigenvectors. For a symmetric matrix, there exists a set of n orthogonal eigenvectors of A. As we can scale eigenvectors arbitrarily, each of them is chosen as unit-length and the error term is expressed as a linear combination  of this eigenvector.  Then we see that the residual can be expressed as the sum of the eigenvector components. We saw that when the set has only one eigenvector, the convergence is achieved in one step by choosing the inverse of the eigenvalue. Here all vectors have a common eigenvalue and therefore leads again to a onestep convergence.

The lowest value of the function is at the minimum of the paraboloid.

#codingexercise
int[,] Transpose (int[,] matrix, int row, int col)
{
   var ret = new int[col, row]
   for (int i = 0; i < row; i++)
     for (int j = 0; j < col; j++)
     {
           ret[j, i] = matrix[i,j]
     }
    return ret;
}

Assert (matrix = Transpose(Transpose(matrix));

We will continue to look at the convergence of Steepest Descent method. Let's first consider the case where ei is an eigenvector with eigenvalue Lambda-e. Then the residual which is A applied to the negative eigenvector ei, is also an eigenvector.  Since we discussed earlier that the position in the next step can be determined from the position in the current step based on step length and direction, we see that when we express the residuals as eigen vectors, the next residual turns out to be zero. This implies that it takes only one step to converge to the exact solution.  The current position lies on one of the axes of the ellipsoid, and so the residual points directly to the center of the ellipsoid. Choosing gradient as  negative eignenvalue gives us instant convergence.

#codingexercise
Matrix multiplication
Int [, ] multiply ( int [,] left, int [,] right)
{
If (left == null || right == null) return null;
Int lc  = left.GetLength (0);
Int lr   = left.GetLength (1);
Int rc  = right.GetLength (0);
Int rr  = right.GetLength (1);
If (lc != rr) return null;
Var ret = new int [lr, rc]();
For int r = 0; r < lr;  r++;
   For int  c = 0; c < rr; c++
     {
         For ( int k = 0; k < lc; k++)
                     Ret [r,c] += left [r,k] × right [k, r];
}
Return ret;

}

We resume our discussion on matrix operations. In this post, we will pick up Jacobii iteration. The Jacobii method also solves the same linear equation as before. The matrix A is split into two parts: D whose diagonal elements are identical to those of A, and whose off-diagonal elements are zero and E whose diagonal elements are zero and whose off diagonal elements are identical to those of A. A is the sum of D and E. The usefulness of D is that it is a diagonal matrix and so it can be inverted.
As compared to the steepest descent method, this does not have to look for a value that converges to zero.Using the linear equation and the split of the matrix into the diagonal specific matrices, we can now write equation as an iteration where the next step is calculated based on a constant times the previous step taken together with another constant. Each iteration leads closer to the final solution which is called a stationary point because the next iteration at the solution will be the same as the previous step. Let us take a closer look at what each iteration does. If we express each iterate (current step) as the sum of the exact solution and the error term, then the equation becomes a simple translation in terms of previous step. In other words, each iteration does not affect the correct part of the iterate but only the error term as the iterations converge to the stationary point. Hence the initial vector x0 has no effect on the inevitable outcome. And it also does not affect the number of iterations required to converge to a given tolerance.
We will next discuss spectral radius which determines the speed of convergence in this method.Suppose that vj is the eigenvector of B with the largest eigenvalue. The spectral radius of B  is this largest eigenvalue. Now the error can be expressed as a linear combination of eigenvectors, one of which will be in the direction of vj. and this one will be the slowest to converge because the eigenvalue is proportional to the steepness of the corresponding slope.  This follows from the eigenvectors being aligned along the axes of the paraboloid describing the quadratic form of the function. The rate of convergences depends on this spectral radius and it is in turn dependent on A. Unfortunately, this technique does not apply to all A. However, this method illustrates the usefulness of eigenvectors. The eigenvector paths of each successive error term determine the path of convergence.They converge normally at the rate defined by their eigenvalues.
#coding exercise
Given the matrix in the coding exercise described earlier, print the elements in  a single dimensional sorted order
List<int> GetSortedElements(int[,] matrix, int row, int col)
{
 // assuming parameter validation
  var ret = new List<int>();
  int i = 0;
  int j = 0;
  var front = new List<int>(); // column corresponding to each row of the front
  for (int i = 0; i < row; i++) { front.add(0); } // initialize to first column index for each row
  while (i < row && j < col)
 {
  // extract min of allfront
  var candidate = new List<int>();
  for (int k = 0; k< row; k++)
        candidate.add((front[k] < col ? matrix[k, front[k]] :  INT_MAX);
  int min = candidate.min();
  i = candidate.IndexOf(min);
  front[i] = front[i] + 1;
  j = front[i];
  ret.Add(min);
}
return ret; 
}
23 34 45 46 68 89
27 35 47 69 97 110
32 46 48 65 98 112

To continue on the coding exercise, there are several other ways to find a number in the matrix like we have described. The one we described had a binary chop method to search for the number in each row of the matrix. The binary chop operation on a single row takes log N time and if there are n rows we are repeating to this to a total complexity of O(N log N) times. We make no use of the sorted order of numbers in the column in this approach and the convergence is linear to the number of rows.
We could go binary chopping across rows as well because the elements are sorted there too and iterate down the columns. That would however put us at no better benefit over the one we have already. It would be a pure chance if one were to work better than the other unless we knew which one of the two - row or column is smaller.
Another technique that does not recognize the sorted order of the numbers and instead uses the divide and conquer approach - is to partition the matrix. One example of this would be to split the matrix into two or more sub-matrices and then recursively apply the solution to each of the sub-matrices. We could easily eliminate a matrix if the number we are looking for lies outside the right bottom entry in the matrix since this is the largest per the description of the matrix.
If we choose for example a point in the center of the matrix, say mid = (row /2, col/2) and then get four sub matrices of different sizes, then we can easily exclude regions of matrices pretty quickly. By excluding submatrices we attempt to converge onto the number in the bigger matrix logarithmically. This is way better but it involves parallel and recursive computation and consequently more resources. The approach is still log N. However note the termination condition of this matrix. It is not that we can say that the bottom right number is greater than the search number and the top left number is smaller than the search number, the number can be checked along the diagonal. To the contrary, if we take the example [ 53 78
                                                       87  92 ] and looking for a number 62, we see that this is not sufficient to go diagonally. We have to exhaust the rows one after another anyways. While it is easier to exclude the matrix by looking at the top left and bottom right numbers in the matrix, the same cannot be said for when a number is bound within them. In such a case there is additional time complexity of going linearly from row to row which may take O(N) time anyways.
This leads us to a potentially faster solution by enumerating the elements of the array in sorted order until we reach the search number. In the example above, we traverse row order or read column order which ever gives us the next higher number and compare it with the previous position. For example,
if we took a matrix
23 34 45 46 68 89
27 35 47 69 97 110
32 46 48 65 98 112
then we start from the top left and progress down the next higher number as if we were snaking our way up the progression. In the above example, we progress down the number from the top left and compare the row order as well as the column order and keeping track of the starting point we return to the number adjacent to the starting point. For example we progress down from 23, 27, 32 and come back to 34 after we have exhausted column order. Similarly we start from 34 and note the next number 45 and traverse down the column to read 34, 35 until we encounter a number that is higher than 45. Notice that the front for the propagation is a diagonal which in this case is  46, 47 and 46. If we keep track of the elements coming out of the front, we can easily sort all the entries in the matrix in a linear order. Each row will contribute one number to the front and we update the front on this row as we include the numbers in our sorted list.. We therefore burn the matrix by the contents across  the front and this is another way to solve the problem in O(N) time. Note that if all the numbers in the front are higher than the desired number, we stop right then instead of burning through the rest of the matrix. Thus there is some optimization possible as well
Lastly, I want to add that there is perhaps modifications possible where instead of keeping track of the front, we simply update the start and the end for each row based on the findings from the preceding row. Thus we optimize our original solution to not repeat the entirety of each row but to take triangularly smaller portions as we go down the rows one after another.

bool Exists ( int[,] matrix, int number)
{
  // assuming parameter validation
  // we go row wise and do a binary chop

  int start = 0;
  int end = matrix.GetLength(0);
  for (int row = 0; row < matrix.GetLength(0); row++)
  {
       start  = 0;
       If ( BinaryChop(matrix, row, number,  start, ref end))
        Return true;
  }
 return false;
}


bool BinaryChop(int[,] matrix, int row, int number, int start, ref int end)
{
  while (start < end)
  {
       Int mid = checked ( (start + end) / 2);
       if (matrix[row, mid] == number) return true;
       if (matrix[row, mid] < number)
             start = mid + 1
       else
             end = mid  - 1
  }
  if (matrix[row, start] == number) return true;
  if (matrix[row, end] == number) return true;
 return false;
}

In today's post we will continue our discussion on matrix factorization. We reviewed the method of steepest descent earlier. If we carefully note the iteration step, we were computing the cost of steepest descent with two matrix vector products - one to compute the cost of the current residual and another to compute the step length. We can eliminate one because we can write the equation in terms of the residual and not in terms of the steps. And although we have to compute the starting residual, each subsequent iteration depends only on the previous value. However the disadvantage with this translation is that we lose the feedback from the value of the current step.
At this stage it's probably helpful to review eigenvector. An eigenvector is a non-zero vector that does not rotate when B is applied to it (except perhaps to point in precisely the opposite direction). Since there is scaling and not skewing, we can take the effect of eigenvector as some scalar constant applied to B. This scalar constant denoted by Lambda is called the eigenvalue of B.
As we will see, this is now relevant to our iterations. In each iteration, we depend on applying B to a vector over and over again which either grows it or shrinks it as we proceed along the search line. In particular if we use an eigenvalue of -0.5, the vector converges to zero which we found useful in our zigzag path.
Moreover, if B is symmetric, then there exists a set of n linearly independent eigenvectors of B, denoted by v1, v2, .. vn. This set is not unique and each of them has a corresponding eigenvalue.
Note that the vector applied to B need not be an eigenvector. But we can take a set of n eigenvectors to form the basis and any n-dimensional vector can be expressed as a linear combination of these eigenvectors. If B were to be unsymmetric, then finding a set of n independent eigenvectors is difficult. These matrices are known as defective and we won't speculate why they are called so.  The eigenvalues of a positive-definite matrix are all positive.

#codingexercise
Given a symmetric matrix of increasing numbers from left to right and top to bottom, find the presence of a number
// there are many ways, i just want to find something easy
bool Exists ( int[,] matrix, int number)
{
  // assuming parameter validation
  // we go row wise and do a binary chop
  for (int row = 0; row < matrix.GetLength(0); row++)
  {
       If (BinaryChop(matrix, row, number)) return true;
  }
 return false;
}


bool BinaryChop(int[,] matrix, int row,  int number)
{
  int start = 0;
  int end = matrix.GetLength(0);

  while (start < end)
  {
       Int mid = checked ( (start + end) / 2);
       if (matrix[row, mid] == number) return true;
       if (matrix[row, mid] < number)
             start = mid + 1
       else
             end = mid  - 1
  }
  if (matrix[row, start] == number) return true;
  if (matrix[row, end] == number) return true;
 return false;
}

In my previous post, I mentioned an example of LDAP login. There are quite a few open source OpenID, OAuth packages as well. I was looking at one in particular called passport-oauth which is written in Node.js. As we know OAuth 2.0 is used for authentication and authorization with the use of scope for resource access granularity. This package eases authorization by making it provider specific instead of using the OAuth framework directly. It's interface is cleaner to use in that it uses a provider and OAuth strategy and returns a token.
var passport = require('passport') , OAuthStrategy = require('passport-oauth').OAuthStrategy; passport.use('provider', new OAuthStrategy({ requestTokenURL: 'https://www.provider.com/oauth/request_token', accessTokenURL: 'https://www.provider.com/oauth/access_token', userAuthorizationURL: 'https://www.provider.com/oauth/authorize', consumerKey: '123-456-789', consumerSecret: 'shhh-its-a-secret' callbackURL: 'https://www.example.com/auth/provider/callback' }, function(token, tokenSecret, profile, done) { User.findOrCreate(..., function(err, user) { done(err, user); }); } ));

Note that the provider need not be a third party referral site, it can be a corporate OTP provider. However the service implementation  could unify both LDAP and OTP so that the auth is in one place
This will simply be
Def login (user, password): # ldap
Def login (user, token):  # OTP
Now let's look at the OTP provider. That grants a token as well. It takes a request and generates a secret. The secret along with a PIN is used as the credential to be validated. The site's request and response is important here for the service implementation to decide what to look up.
The request initiated by the service provider to the identity provider and the SAML contains both information the type of request and the callback.
as in https://openam.example.com:8443/opensso/SSORedirect/metaAlias/internal/externalidpv2?SAMLRequest=&relaystate=<callbackURL>
The SAMLRequest is base64 encoded SAML message usually including both a <saml:AuthnStatement> which describes the act of authentication at the identity provider and a <saml:AttributeStatement> which asserts a multivalued attribute associated with this principal (aka subject)
The same could perhaps be done via OAuth using password grant
curl -i -k -X POST -u user:password -d {"grant_type"="password", "username="whoisthis", "password"="OTPCode", "scope"="what"}
 https://openam.example.com:8443/openam/oauth2/access_token
{
    "expires_in": 599,
    "token_type": "Bearer",
    "refresh_token": "f6dcf133-f00b-4943-a8d4-ee939fc1bf29",
    "access_token": "f9063e26-3a29-41ec-86de-1d0d68aa85e9"
}

$ curl https://openam.example.com:8443/openam/oauth2/tokeninfo\
?access_token=f9063e26-3a29-41ec-86de-1d0d68aa85e9
{
    "mail": "demo@example.com",
    "scope": [
        "mail",
        "cn"
    ],
    "cn": "demo",
    "realm": "/",
    "token_type": "Bearer",
    "expires_in": 577,
    "access_token": "f9063e26-3a29-41ec-86de-1d0d68aa85e9”

}

We continue our discussion on conjugate gradient methods. First we will discuss the steepest descent method because in our description earlier, we hinted towards excluding certain descent but we were not clear about the choice and the step length. When a matrix is positive definite, the function f(x) is a paraboloid. We want to reduce the function. In the method of the steepest descent, we start at an arbitrary point x0 and slide down to the bottom of the paraboloid. We take a series of steps by choosing a direction and a step length. We choose the direction in which f decreases most quickly, which is the direction opposite f'(x) - the gradient. When we take this step, there will be an error which is the distance from the solution. If we convert this quantity into a vector with the matrix, then this vector called the residual gives the direction of the steepest descent.
Next we choose the step length This is where we see an interesting observation. We want to choose the step length that minimizes f  along a line. We are restricted to choosing a point on the intersection of the vertical plane and the paraboloid. We choose a point based on new = old + step-length-alpha scaled residual. We can pick a residual but we don't know what step length to take on it. At the bottom of the paraboloid, the gradient is zero and the function cannot be minimized any more. So the step length is zero and the iteration terminates. The step length minimizes the function when the directional derivative on this search line is equal to zero. The directional derivative can be expressed in terms of the gradient transpose and the residual taken together. Setting it to zero which is saying their dot product is zero means that the two vectors are orthogonal. As we progress down the search line the rate of increase of f are the projections of the gradient onto the search line and the projection is zero at the minimum. We can conveniently pick the next direction as the orthogonal at the minimum along a search line. Thus we find a zig zag path during the iteration which appears because each gradient is orthogonal to the previous gradient.

#codingexercise:

LDAP is the triangle in which we want to put all devices and applications in. How do we authenticate a user using LDAP?

def load_user(id):
   ld = ldap.initialize('ldap://localhost:1389')
   name = id
   ld.simple_bind_s()
   basedn = "ou=people,dc=example,dc=com"
   filter = "(|(cn=\*" + name + "\*)(sn=\*" + name + "\*))"
   results = ld.search_s(basedn, ldap.SCOPE_SUBTREE, filter)
   import io
   from contextlib import redirect_stdout
   with io.StringIO() as buf, redirect_stdout(buf):
           ldif_writer = ldif.LDIFWriter(buf)
           for dn, entry in results:
                ldif_writer.unparse(dn, entry)
                output = buf.getvalue()
                uname=""
                uid=""
                if id in output:
                    for line in output.splitlines(True):
                          if "dn:" in line:
                              kvp = line.split().split(',')
                              uid = str([i.replace("uid=", "") for i in kvp if 'uid=' in i])
                          if "cn:" in line:
                              uname = line.replace("cn: ", "")
                     if uname and uid:
                         return Users(uname, uid, True)
    ld.unbind_s()
    return Users('Anonymous',  'unknown', False)