Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
DATA ADAPTIVE COMPRESSION AND DATA ENCRYPTION USING KRONECKER PRODUCTS
Document Type and Number:
WIPO Patent Application WO/2016/029163
Kind Code:
A1
Abstract:
Digital files are compressed using a process including Schmidt decompositions of matrices using an algorithm, termed 'BSD' herein, which is based on an algebraic method generalizing QR decomposition. Software analyzes an input file and initially identifies a matrix M, with entries within a predefined set of integers, within the file. Next, essential entries are defined, extracted from M, that contain sufficient information to recover Musing BSD. The compressed file includes the essential entries and their positions within M. To achieve an encryption process, software encrypts the pattern matrix that includes the positions of the essential entries of M. To achieve a lossy compression, software identifies essential entries that contain sufficient information to recover an approximation to M for which the quality is determined by an error threshold. For a more efficient lossy compression, software uses singular value decomposition, BSD, and other signal processing of M.

Inventors:
BOUROUIHIYA ABDELKRIM (US)
Application Number:
PCT/US2015/046382
Publication Date:
February 25, 2016
Filing Date:
August 21, 2015
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
UNIV NOVA SOUTHEASTERN (US)
International Classes:
H04N21/262
Foreign References:
US20140222749A12014-08-07
US20100293210A12010-11-18
US20050031219A12005-02-10
US20030200097A12003-10-23
US20080109474A12008-05-08
Other References:
See also references of EP 3195604A4
Attorney, Agent or Firm:
BIANCO, Paul, D. (21355 East Dixie Highway Suite 11, Miami FL, US)
Download PDF:
Claims:
What is claimed is:

1. A method for encoding a digital data source, comprising:

using a computer executing software stored on non-transitory media, the software configured to:

identify an mn x pq matrix M, with entries within a predefined set of integers, within the digital data;

define a left essential matrix A;

define a right essential matrix B;

define a pattern matrix P for storing positions of essential entries; assign to matrix Me a starting value of M;

define a matrix Ae,

define a matrix Be,

assign a starting value to e;

a) select a non-zero entry de of Me,

b) store position (r, c) of the selected non-zero entry of Me at the eth column of P; c) select from Me two matrices Ae and Be having de as a common entry and for which Ae <S> Be/ de is a term in a BSD of M with respect to the shape (m, n, p, q); d) store in the m xp block of A the entries of M whose positions are the positions of the entries of Ae in Me,

e) store in the eth n x q block of B the entries of M whose positions are the positions of the entries of Be in Me, and

f) calculate the matrix Me+i = Me - Ae ® Be/ de, and if a predetermined error threshold between M and Me+i is reached, repeat steps (a)-(f) with Me+1, otherwise P, A, and B collectively represent encoded digital data corresponding to M; and g) transfer the encoded data to digital storage on at least one of the computer executing software or another computer, the encoded data comprising fewer data bytes than the source digital data, and representing at least one of all the information in the source digital data, and an approximation of all of the information in the source digital data.

2. The method of claim 1, wherein the non-zero number selected in step (a) corresponds to the first entry, with respect to the stereographic order, whose absolute value is the maximum of the absolute values of the entries of Me.

3. The method of claim 1, wherein the software is further configured, in step (c), to select from Me two matrices Ae and Be by:

i) calculating the Euclidean division of c by q and find the remainder j, and if remainder j is zero, replace j with q, then compute i=(c-j)/q+l; and

ii) calculating the Euclidean division of r by n and find the remainder 1, and if remainder 1 is zero, replace 1 with n, then compute k=(r-l)/n+l .

4. The method of claim 1, wherein the software is further configured to:

(i) carry out step (d) by, for each integer a between 1 and m, and each integer b between 1 and p, storing the entry of M at position (l+(a-l)n, j+(b-l)q) in left essential matrix A at position (a, (e-l)p+b), and storing the entry of Me at position (l+(a-l)n, j+(b-l)q) in matrix Ae at position (a, b); and

(ii) carry out step (e) by, for each integer a between 1 and n, and each integer b between 1 and q, storing the entry of M at position (l+(k-l)n, b+(i-l)q) in right essential matrix B at position (a, (e-l)q+b), and storing the entry of Me at position (l+(k-l)n, b+(i-l)q) in matrix Be at position (a, b).

5. The method of claim 1, wherein the predetermined stop value is infinity for a lossless encoding of the digital data.

6. The method of claim 1, wherein the predetermined stop value is a positive number for a lossy encoding of the digital data.

7. The method of claim 1, wherein the sum of the stored sizes of P, A, and B is less than the stored size of the digital data corresponding to M.

8. The method of claim 1 , wherein all of the entries of A and B are extracted from M and the entries of P are integers.

9. The method of claim 1, wherein the software is further configured to use P, A, and B to decode the digital data as matrix N, the software being thus further configured to:

define R as the number of columns in P;

define E as an mn x pq matrix;

define m as the number of rows in A;

define p as the number of columns in A divided by R;

define n as the number of rows of B;

define q as the number of columns in B divided by R;

assign a starting value to e;

extract an ordered pair (r, c) from the eth column of P;

carry out steps (c)-(f) in reverse, as steps (f)-(c), placing the entries of A and B in E at the same positions they occupy in M;

fill with zero all values of E which have not been assigned an entry from A or B;

assign to matrix Ee a starting value of E;

assign a starting value to e;

a) extract an ordered pair (r, c) from the eth column of P;

b) select de the entry of Ee at the position (r, c);

c) using the method of steps (c)-(f) of claim 1 , recover the matrices Ae and Be of claim 1 ; and

10. The method of claim 9, wherein in step (d), if N = M, the digital data has been decoded without a loss of data, and if N≠ M, N approximates the digital data.

11. The method of claim 1 , wherein the digital data corresponding to M is encrypted by encrypting P. 12. The method of claim 1 , wherein M is one of a plurality of matrices M identified within the digital data, and wherein steps (a)-(f) are performed for each of the plurality of matrices M, and wherein the digital data is encoded by the collective matrices of P, A, and B corresponding to the plurality of matrices M along with the size S of the matrix ME that represents the entire digital data.

13. A method for encoding a digital data source, comprising:

using a computer executing software stored on non-transitory media, the software configured to:

define R as the number of columns in P;

define E as an mn x pq matrix;

define m as the number of rows in A;

define p as the number of columns in A divided by R;

define n as the number of rows of B;

define q as the number of columns in B divided by R;

assign a starting value to e;

extract an ordered pair (r, c) from the eth column of P;

select de the entry of Ee at the position (r, c);

calculate the Euclidean division of c by q and find the remainder j, and if remainder j is zero, replace j with q, then compute i={c-j)lq+\;

calculate the Euclidean division of r by n and find the remainder /, and if remainder / is zero, replace / with n, then compute k={r-l)ln+\;

for each integer a between 1 and mn, and each integer b between 1 and pq, copy the entry of A at position (a, (e-l)p+b) into position {l+{a-\)n,j+{b-\)q) of E, and copy the entry of B at position (a, (e-l)q+b) into position (a+(k-l)n, b+{i-\)q) of E; fill with zero all values of E which have not been assigned an entry from A or B; assign to matrix Ee a starting value of E;

assign a starting value to e;

a) extract an ordered pair (r, c) from the eth column of P;

b) select de the entry of Ee at the position (r, c);

c) calculate the Euclidean division of c by q and find the remainder j, and if remainder j is zero, replace j with q, then compute i={c-j)lq+1; d) calculate the Euclidean division of r by n and find the remainder /, and if remainder / is zero, replace / with n, then compute k=(r-l)/n+l;

e) for each integer a between 1 and m, and each integer b between 1 and p, store the entry of Ee at the position (l+(a-l)n,j+(b-l)q) at the position (a, b) of matrix Ae, f) for each integer a between 1 and n, and each integer b between 1 and q, store the entry of Ee at the position (l+(k-\)n, b+(i-\)q) at the position (a, b) of matrix Be; and

h) transfer the encoded data to digital storage on at least one of the computer executing software or another computer, the encoded data comprising fewer data bytes than the source digital data, and representing at least one of all the information in the source digital data, and an approximation of all of the information in the source digital data.

14. The method of claim 13, wherein the software is further configured to use P, A, and B to decode the digital data as a matrix that approximates M.

15. The method of claim 13, wherein Mis one of a plurality of matrices M identified within the digital data, and wherein steps (a)-(f) are performed for each of the plurality of matrices M, and wherein the digital data is encoded by the collective matrices of P, A, and B corresponding to the plurality of matrices M along with the size S of the matrix ME that represents the entire digital data.

16. The method of claim 15, wherein each matrix M has a predetermined size, and wherein if the digital data is not evenly divisible by the predetermined size, a remaining partial matrix MP is padded with zeros to the predetermined size, and the are discarded when the digital data is decoded.

17. A method for encoding digital data, comprising:

using a computer executing software stored on non-transitory media, the software configured to:

identify an mn x pq matrix M, with entries within a predefined set of integers, within the digital data;

define an essential sequence S which is in two part Si and &;

define an essential matrix E;

define a pattern sequence PS for storing positions of n x q block matrices of E; define a matrix Ae,

define a matrix Be,

assign to matrix Me a starting value of M and assign a starting value to e;

a) select a non-zero entry de of Me,

b) store at the eth term of PS the position of the eth n x q block of Me that includes (r, c) the position of de,

c) store the eth n x q block matrix of M at the eth n x g block matrix of E, and following the lexicographic order, store the entries of the eth n x q block of M in ¾ d) calculate the Euclidean division of c by q and find the remainder j, and if the remainder is zero, replace j with q, then compute i=(c-j)/q+l; and calculate the Euclidean division of r by n and find the remainder /, and if the remainder is zero, replace / with n, then compute k=(r-l)/n+l;

e) for each integer a between 1 and m and each integer b between 1 and p, store at position (a, b) of matrix Ae the entry of Me at position (l+(a-l)n,j+(b-l)q), and following the lexicographic order, if position (l+(a-l)n,j+(b-l)q) of E is unfilled, store in it the entry of M located at the same position and store the same entry in sequence &; and

f) calculate the matrix Me+i = Me - Ae ® Be/ de, and if a predetermined error threshold between M and Me+i is reached, repeat steps (a)-(f) with Me+i, otherwise, collect Si and &to form a sequence S, which with PS and the shape (m, n, p, q) represent encoded digital data corresponding to M; and

g) transfer the encoded data to digital storage on at least one of the computer executing software or another computer, the encoded data comprising fewer data bytes than the source digital data, and representing at least one of all the information in the source digital data, and an approximation of all of the information in the source digital data.

18. The method of claim 17, wherein the software is configured to de-encode the encoded digital data, by being further configured to:

define R as the number the number of terms in PS;

define E as an mn x pq matrix;

use the first Rnq terms of S to build R n x q block matrices of E, use PS to identify the positions of the R block matrices, and delete from S the used terms, then call essential position any filled position of E;

assign to matrix Me a starting value of E, and assign a starting value to e;

a) build a matrix Be whose entries are the entries of the eth n x q block matrix of Me, and following the lexicographic order compute (r, c) the position in Me of the first entry de in Be whose absolute value is equal to the maximum value of the absolute values of the entries of Be, then store de at the eth term of a sequence D;

b) calculate the Euclidean division of c by q and find the remainder j, and if the remainder is zero, then replace j with q, then compute i=(c-j)/q+l, and calculate the Euclidean division of r by n and find the remainder /, and if the remainder is zero, replace / with n, then compute k=(r-l)/n+l, and store i the e* term of a sequence I,j the eth term of a sequence J k the e* term of a sequence K, and / the e* term of a sequence L;

c) build an an m χ p matrix Ae as follows, for each (a, b), where 1 < a < m and I < b < p, if (l+(a-l)n,j+(b-l)q) is an essential position, then fill position (a, b) of Ae with the entry of Me located at position (l+(a-l)n,j+(b-l)q), fill this position of E with the first term of S and mark (J+(a-l)n,j+(b-l)q) as an essential position of E;

d) fill with zero the other positions of Ae;

f) compute Me+i = Me - Ae ® Be/ de, and if e < R, then repeat steps (a)-(d) with Me+i, and if e = R, then assign to matrix Ee a starting value of E, and assign a starting value to e; g) compute i, j, k, and / that are, respectively, the e term of I, the e term J the 6th term of K, and the eth term of L;

h) build an m x p matrix Ae and an n x g matrix B e are built as follows, for each integer a between 7 and m and each integer b between 1 and p , place at position (a, b) of Ae the entry of Ee at position (l+(a-l)n,j, and assign the 6th n x q block matrix of Ee to Be,

k) if e < R, then compute Ee+i = Ee - Ae ®Be / de, where de is the 6th term of D, and repeat steps (g) and (h) with Ee+i, and if e = R, then compute

N= Ai ®Bi /di + A2 ®Bi / d2 + ... + AR ®BR /dR,

wherein the matrix N is the matrix of the output file.

19. The method of claim 17, wherein lossless compression is achieved by stopping iterative computation of steps (a)-(d) in step (f) after the Schmidt rank of M is reached.

20. The method of claim 17, wherein error is measured by using at least one of PSNR, PEVQ, and SSIM.

Description:
DATA ADAPTIVE COMPRESSION AND DATA ENCRYPTION

USING KRONECKER PRODUCTS

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims the benefit of related U.S. Patent Application No. 62/040,674, filed

August 22, 2014, the contents of which are incorporated herein by reference in their entirety.

FIELD OF THE DISCLOSURE

The disclosure relates to a system and method for reducing binary file size, and in particular, to file compression using matrices.

BACKGROUND OF THE DISCLOSURE

Singular Value Decomposition (SVD) yields a method to decompose each mn x pq matrix M into a sum with a minimum number of terms, each of which is the Kronecker product of an m x p matrix by and n * q matrix. This decomposition is known as a Schmidt decomposition of M. We shall say that Mis decomposed with respect to the decomposition shape (m, n, p, q).

Assuming that M represents a digital file, dropping some terms from the decomposition of M and using the other terms to build a matrix that approximates M leads to a lossy compression of the digital file. In addition to this compression method, there is another compression method based on SVD known as compression with SVD. Every compression method based on SVD has an energy-compaction property which causes the method to be useful for compression. With SVD, singular values and vectors are to be stored to construct the output file. These values and entries are not necessarily integers even if all entries in the original matrix are integers. Thus, storing in a computer the singular values and vectors without losing too much information requires per pixel a much larger amount of memory space than the amount of memory space occupied by a pixel in the original file. Therefore, the compression ratio with SVD is not as desirable in comparison with ratios achieved by other existing compression methods, such as JPEG [see references 4-6, below]. Other compressions schemes based on algebra include algorithms based on QR and QLP decompositions, [1A]. SUMMARY OF THE DISCLOSURE

In an embodiment of the disclosure, a method for encoding digital data comprises using a computer executing software stored on non-transitory media, the software configured to identify an mn x pq matrix M, with entries within a predefined set of integers, within the digital data; define a left essential matrix A; define a right essential matrix B; define a pattern matrix P for storing positions of essential entries; assign to matrix Me a starting value of M; define a matrix Ae, define a matrix Be, assign a starting value to e; a) select a non-zero entry de of Me, b) store the position (r, c) of the selected non-zero entry of Me at an eth column of P; c) select from Me two matrices Ae and Be having de as a common entry and for which Ae ® Be/ de is a term in a BSD of M with respect to the parameters m, n, p, and q; d) store in the m xp block of A the entries of M whose positions are the positions of the entries of Ae in Me, e) store in the 0th n x q block of B the entries of M whose positions are the positions of the entries of Be in Me, and f) calculate the matrix Me+i = Me - Ae ® Be/ de, and if a predetermined error threshold between M and Me+i is reached, repeat steps (a)-(f) with Me+i, otherwise, P, A, and B collectively represent encoded digital data corresponding to M; and g) transfer the encoded data to digital storage on at least one of the computer executing software or another computer, the encoded data comprising fewer data bytes than the source digital data, and representing at least one of all the information in the source digital data, and an approximation of all of the information in the source digital data.

In a variation thereof, the non-zero number selected in step (a) corresponds to the first entry, with respect to the stereographic order, whose absolute value is the maximum of the absolute values of the entries of Me.

In a further variation thereof, the software is configured, in step (c), to select from Me two matrices Ae and Be by i) calculating the Euclidean division of c by q and find the remainder j, and if remainder j is zero, replace j with q, then compute i=(c-j)/q+l; and ii) calculating the Euclidean division of r by n and find the remainder 1, and if remainder 1 is zero, replace 1 with n, then compute k=(r-l)/n+l .

In a yet further variation thereof, the software is further configured to (i) carry out step (d) by, for each integer a between 1 and m, and each integer b between 1 and p, storing the entry of M at position (l+(a-l)n, j+(b-l)q) in left essential matrix A at position (a, (e-l)p+b), and storing the entry of Me at position (l+(a-l)n, j+(b-l)q) in matrix Ae at position (a, b); and (ii) carry out step (e) by, for each integer a between 1 and n, and each integer b between 1 and q, storing the entry of M at position (l+(k-l)n, b+(i-l)q) in right essential matrix B at position (a, (e-l)q+b), and storing the entry of Me at position (l+(k-l)n, b+(i-l)q) in matrix Be at position (a, b).

In other variations thereof, the predetermined stop value is infinity for a lossless encoding of the digital data; the predetermined stop value is a positive number for a lossy encoding of the digital data; the sum of the stored sizes of P, A, and B is less than the stored size of the digital data corresponding to M; and/or all of the entries of A and B are extracted from M and the entries of P are integers.

In another variation thereof, the software is further configured to use P, A, and B to decode the digital data as matrix N, the software being thus further configured to define R as the number of columns in P; define E as an mn x pq matrix; define m as the number of rows in A; define p as the number of columns in A divided by R; define n as the number of rows of B; define q as the number of columns in B divided by R; assign a starting value to e; extract an ordered pair (r, c) from the eth column of P; carry out steps (c)-(f) in reverse, as steps (f)-(c), placing the entries of A and B in E at the same positions they occupy in M; fill with zero all values of E which have not been assigned an entry from A or B; assign to matrix Ee a starting value of E; assign a starting value to e; a) extract an ordered pair (r, c) from the eth column of P; b) select de the entry of Ee at the position (r, c); c) using the method of steps (c)-(f) of claim 1, recover the matrices Ae and Be of claim 1; and d) if e < R, compute Ee+i = Ee - AE <8> Be/de and repeat (a)-(d) with Ee+i, and if e = R, compute N= Ai <8> Bi + A2 <8> B2 + ...+ AR <8> BR.

In other variations thereof, in step (d), if N = M, the digital data has been decoded without a loss of data, and if N≠ M, N approximates the digital data; and/or the digital data corresponding to M is encrypted by encrypting P.

In yet another variation thereof, the software is further configured to use P, A, and B to decode the digital data as a matrix that approximates M; wherein M is one of a plurality of matrices M identified within the digital data, and wherein steps (a)-(f) are performed for each of the plurality of matrices M, and wherein the digital data is encoded by the collective matrices of P, A, and B corresponding to the plurality of matrices M along with the size S of the matrix ME that represents the entire digital data.

In another embodiment of the disclosure, a method for encoding digital data, comprises using a computer executing software stored on non-transitory media, the software configured to define R as the number of columns in P; define E as an mn x pq matrix; define m as the number of rows in A; define p as the number of columns in A divided by R; define n as the number of rows of B; define q as the number of columns in B divided by R; assign a starting value to e; extract an ordered pair (r, c) from the eth column of P; select de the entry of Ee at the position (r, c); calculate the Euclidean division of c by q and find the remainder j, and if remainder j is zero, replace j with q, then compute i=(c-j)lq+1; calculate the Euclidean division of r by n and find the remainder /, and if remainder / is zero, replace / with n, then compute k=(r-l)/n+l; for each integer a between 1 and mn, and each integer b between 1 and pq, copy the entry of A at position (a, (e-l)p+b) into position (l+(a-l)n,j+(b-l)q) of E, and copy the entry of B at position (a, (e- \)q+b) into position (a+(k-l)n, b+(i-\)q) of E; fill with zero all values of E which have not been assigned an entry from A or B; assign to matrix Ee a starting value of E; assign a starting value to e; a) extract an ordered pair (r, c) from the eth column of P; b) select de the entry of Ee at the position (r, c); c) calculate the Euclidean division of c by q and find the remainder j, and if remainder j is zero, replace j with q, then compute i={c-j)l q+\; d) calculate the Euclidean division of r by n and find the remainder /, and if remainder / is zero, replace / with n, then compute k={r-l)ln+\; e) for each integer a between 1 and m, and each integer b between 1 and p, store the entry of Ee at the position (l+(a-l)n,j+(b-l)q) at the position (a, b) of matrix Ae; f) for each integer a between 1 and n, and each integer b between 1 and q, store the entry of Ee at the position (l+(k-l)n, b+{i-\)q) at the position (a, b) of matrix Be; and g) if e < R, compute Ee+i = Ee - Ae ® Be/de and repeat (a)-(g) with Ee+i, and if e = R, compute N= Ai ® Bi + A2 ® B2 + ...+ AR ® BR; and h) transfer the encoded data to digital storage on at least one of the computer executing software or another computer, the encoded data comprising fewer data bytes than the source digital data, and representing at least one of all the information in the source digital data, and an approximation of all of the information in the source digital data.

In a variation thereof, the software is further configured to use P, A, and B to decode the digital data as a matrix that approximates M; wherein M is one of a plurality of matrices M identified within the digital data, and wherein steps (a)-(f) are performed for each of the plurality of matrices M, and wherein the digital data is encoded by the collective matrices of P, A, and B corresponding to the plurality of matrices M along with the size S of the matrix ME that represents the entire digital data. In another variation thereof, each matrix M has a predetermined size, and wherein if the digital data is not evenly divisible by the predetermined size, a remaining partial matrix MP is padded with zeros to the predetermined size, and are discarded when the digital data is decoded.

In a further embodiment of the disclosure, a method for encoding digital data, comprises using a computer executing software stored on non-transitory media, the software configured to use SVD to find Ma, an mn x pq matrix, with the lowest Schmidt rank R for which PSNR(Ma, M) > a predetermined value; quantize Ma to find a matrix M whose entries are integers; define a left essential matrix A; define a right essential matrix B; define a pattern matrix P for storing positions of essential entries; assign to matrix Me a starting value of M; define a matrix Ae; define a matrix Be; assign a starting value to e; a) select a non-zero entry de of Me; b) store the position (r, c) of the selected non-zero entry of Me at an eth column of P; c) select from Me two matrices Ae and Be having de as a common entry and for which Ae ® Be/ de is a term in the Schmidt decomposition of M with respect to the parameters m, n, p, and q; d) store in the 0th m xp block of A the entries of M whose positions are the positions of the entries of Ae in Me, e) store in the eth n x q block of B the entries of M whose positions are the positions of the entries of Be in Me, f) calculate the matrix Me+i = Me - Ae ® Be/ de, and if e < R , repeat steps (a)-(f) with Me+i, and if e = R, then P, A, and B collectively represent encoded digital data corresponding to M.

In another embodiment of the disclosure, a method for encoding digital data, comprises using a computer executing software stored on non-transitory media, the software configured to: identify an mn x pq matrix M, with entries within a predefined set of integers, within the digital data; define an essential sequence S which is in two part Si and &; define an essential matrix E; define a pattern sequence PS for storing positions of n x q block matrices of E; define a matrix Ae, define a matrix Be, assign to matrix Me a starting value of M and assign a starting value to e; a) select a non-zero entry de of Me, b) store at the eth term of PS the position of the eth n x q block of Me that includes (r, c) the position of de, c) store the eth n x q block matrix of M at the eth n x q block matrix of E, and following the lexicographic order, store the entries of the eth n x q block of M in Si; d) calculate the Euclidean division of c by q and find the remainder j, and if the remainder is zero, replace j with q, then compute i=(c-j)/q+l; and calculate the Euclidean division of r by n and find the remainder /, and if the remainder is zero, replace / with n, then compute k=(r-l)/n+l; e) for each integer a between 1 and m and each integer b between 1 and p, store at position (a, b) of matrix Ae the entry of Me at position (l+(a-l)n,j+(b-l)q), and following the lexicographic order, if position (l+(a-l)n,j+(b-l)q) of E is unfilled, store in it the entry of M located at the same position and store the same entry in sequence &; and f) calculate the matrix Me+i = Me - Ae ® Be/ de, and if a predetermined error threshold between M and Me+i is reached, repeat steps (a)-(f) with Me+i, otherwise, collect Si and &to form a sequence S, which with PS and the shape (m, n, p, q) represent encoded digital data corresponding to M; and g) transfer the encoded data to digital storage on at least one of the computer executing software or another computer, the encoded data comprising fewer data bytes than the source digital data, and representing at least one of all the information in the source digital data, and an

approximation of all of the information in the source digital data.

In a variation thereof, the software is configured to de-encode the encoded digital data, by being further configured to: define R as the number the number of terms in PS; define E as an mn x pq matrix; use the first Rnq terms of S to build R n x q block matrices of E, use PS to identify the positions of the R block matrices, and delete from S the used terms, then call essential position any filled position of E; assign to matrix Me a starting value of E, and assign a starting value to e; a) build a matrix Be whose entries are the entries of the eth n x q block matrix of Me, and following the lexicographic order compute (r, c) the position in Me of the first entry de in Be whose absolute value is equal to the maximum value of the absolute values of the entries of Be, then store de at the eth term of a sequence D; b) calculate the Euclidean division of c by q and find the remainder j, and if the remainder is zero, then replace j with q, then compute i=(c- j)/q+l, and calculate the Euclidean division of r by n and find the remainder /, and if the remainder is zero, replace / with n, then compute k=(r-l)/n+l, and store i the eth term of a sequence I,j the e* term of a sequence J k the eth term of a sequence K, and / the e* term of a sequence L; c) build an an m χ p matrix Ae as follows, for each (a, b), where 1 < a < m and 1 < b < p, if (l+(a-l)n,j+(b-l)q) is an essential position, then fill position (a, b) of A with the entry of Me located at position (l+(a-l)n, j+(b-l)q), fill this position of E with the first term of S and mark (l+(a-l)n,j+(b-l)q) as an essential position of E; d) fill with zero the other positions of Ae; f) compute Me+i = Me - Ae ® Be/ de, and if e < R, then repeat steps (a)-(d) with Me+i, and if e = R, then assign to matrix Ee a starting value of E, and assign a starting value to e; g) compute k, and / that are, respectively, the eth term of I, the eth term J the eth term of K, and the eth term of L; h) build an m χ p matrix Ae and an n x q matrix Be are built as follows, for each integer a between I and m and each integer b between 1 and p, place at position (a, b) of Ae the entry of Ee at position (l+(a-l)n,j, and assign the e n x q block matrix of Ee to Be; k) if e < R, then compute Ee+i = Ee - Ae ®Be / de, where de is the eth term of D, and repeat steps (g) and (h) with Ee+i, and if e = R, then compute N= Ai ®Bi / di + A2 ®B2 /d2 + ... + AR ®BR /da, wherein the matrix N is the matrix of the output file.

In a further variation thereof, lossless compression is achieved by stopping iterative computation of steps (a)-(d) in step (f) after the Schmidt rank of M is reached.

In a still further variation thereof, error is measured by using at least one of PSNR, PEVQ, and SSIM. BRIEF DESCRIPTION OF THE DRAWINGS

A more complete understanding of the present disclosure, and the attendant advantages and features thereof, will be more readily understood by reference to the following detailed description when considered in conjunction with the accompanying drawings wherein:

FIG. 1A is a diagram of an embodiment of BSD Lossless Compression in accordance with the disclosure, wherein the algorithms diagrammed are executed from top to bottom and left to right, and wherein letters in circles throughout the drawings refer to logical flow patterns between the various algorithms;

FIG. IB is a diagram of an embodiment of BSD Lossless Decompression in accordance with the disclosure;

FIG. 1C is a diagram of an embodiment of BSD Lossy Compression in accordance with the disclosure;

FIG. 2A is a diagram of an embodiment of BSD Lossy Compression using SVD and blocks, in accordance with the disclosure;

FIG. 2B is a diagram of an embodiment of BSD Decompression, where SVD and blocks were used, in accordance with the disclosure;

FIG. 3 A is a diagram of an embodiment of BSD Compression and Encryption in accordance with the disclosure;

FIG. 3B is a diagram of an embodiment of BSD Decompression in accordance with the disclosure; and

FIG. 4 depicts computing equipment, part or all of which can be used to carry out the disclosure. DETAILED DESCRIPTION OF THE DISCLOSURE

As required, detailed embodiments are disclosed herein; however, it is to be understood that the disclosed embodiments are merely examples and that the systems and methods described below can be embodied in various forms. Therefore, specific structural and functional details disclosed herein are not to be interpreted as limiting, but merely as a basis for the claims and as a representative basis for teaching one skilled in the art to variously employ the present subject matter in virtually any appropriately detailed structure and function. Further, the terms and phrases used herein are not intended to be limiting, but rather, to provide an understandable description of the concepts.

The terms "a" or "an", as used herein, are defined as one or more than one. The term plurality, as used herein, is defined as two or more than two. The term another, as used herein, is defined as at least a second or more. The terms "including" and "having," as used herein, are defined as comprising (i.e., open language). The term "coupled," as used herein, is defined as "connected," although not necessarily directly, and not necessarily mechanically. Herein, reference to 'the software' refers to electronic execution of instructions embodying the methods of the disclosure, the methods of the disclosure thereby being carried out by one or more electronic processors, including, for example, parallel processors. The term 'memory value', or size in memory, may be considered to also indicate the size of the referenced element within non- volatile storage.

In an embodiment of the disclosure, digital files are compressed using a method which includes Schmidt decompositions of matrices. The disclosure provides for performing lossless and lossy compressions of any digital file that can be represented by a matrix M. The lossless compression is achieved by an apparatus that stores some entries of M to recover M. The lossy compression is achieved by an apparatus that combines methods of the disclosure with other techniques that include low rank approximation and compressing by blocks. In another aspect, an apparatus executes software configured to execute the methods of the disclosure to compresses a file, and in an embodiment, enables a user to choose a key code that is needed at the

decompression process. Accordingly, the present invention enables data compression and data security. The disclosure provides a method to find Schmidt decompositions of matrices. Herein, this method will be referred to as BSD (Bourouihiya Schmidt Decomposition). A BSD of a matrix M uses the entries of M and elementary operations to compute each entry of each matrix involved in the decomposition of M. The disclosure uses this fact to enable lossless and lossy compressions of digital files. In an embodiment, a lossless compression is almost impossible with SVD, however by combining BSD and SVD, the invention enables lossy compression while solving the storage problems encountered by compressing with SVD, and with a very low cost in term of error. Compression in accordance with the disclosure is competitive with popular existing methods, such as JPEG or JPEG2000.

The present disclosure provides the following embodiments: the use of BSD to achieve lossless or lossy compressions of digital files; the use of BSD, low rank approximation with SVD, compressing by blocks, and other techniques to achieve efficient lossy compressions of digital files, including enabling a user to choose from a multitude of parameters to adapt the compression method to a specific application; and encryption of a digital file, wherein a user can choose a key code that is required for decryption at the decompression process, wherein the encryption methods of the disclosure can be used with the lossy and lossless compressions of the disclosure.

Lossless compression of digital data or files is accomplished using BSD, as follows, although this method may also be used to achieve lossy compression:

Each mn x pq-matrix M can be written as

M = Ai ®Bi + Ai ®B2+ ... + AR ®BR, (4.1)

where Ak is an m χ p -matrix and Bk is an n x q-matrix. In the art, Equality (4.1) is known as a Schmidt decomposition of M and R is the (Schmidt) rank if the number of terms in (4.1) is the minimum possible. We shall denote rank (M) = R. If M is decomposed using BSD, then each entry in Ak or Bk is computed using entries extracted from M and the elementary operations, for each k=l,...,R.

Example 1

Consider the 6 x 6-matrix M whose entries are integers between 0 and 255. Thus, 36 Bytes is the memory space required to store M in a computer. In this example, BSD is used to decompose M with respect to (3, 2, 2, 3), i.e., into a sum of matrices, for which each term is the Kronecker product of a J χ 2 matrix by a 2 χ 3 matrix.

where

The entries of Ai, A2, Bi, and B2 are computed using the essential sequence S=(9, 12, 21, 24, 33, 36, 28, 29, 30, 34, 35, 1, 4, 13, 16, 25, 2, 3, 7, 8) whose terms are called essential entries of M (these entries are bolded and underlined in M). The methodology for selecting essential entries is provided elsewhere herein. To recover M, we need to store the decomposition shape (3, 2, 2, 3), the essential sequence S, and a pattern matrix that includes the position (6, 6) (used to locate Ai and Bi) and the position (1, 1) (used to locate A2 and Bi). Since there are 20 essential entries, the decomposition shape includes 4 entries, and the pattern matrix includes 4 entries, then 28 Bytes is the memory space required to store the compressed file. Thus, a lossless compression of Mis achieved, from 36 Bytes to 28 Bytes. A mathematical proof for the BSD algorithm is provided elsewhere herein.

If the Schmidt decomposition of M was carried out using SVD, then φ

Every entry in the matrice Νιι, N21, N12, or N22 is not an integer between 0 and 255 and can only be computed approximately. Therefore, storing these matrices does not achieve a lossless compression of M. This is because, storage of real numbers greatly increases the size of the compressed file.

Using BSD for image compression, it is unlikely that storing the compressed file requires a memory space that is more than the memory space required to store the original file. As a practical matter, most compressions are substantially smaller than the original file using the methods of the disclosure.

Embodiment 1

With reference to FIG. 1 A, in an embodiment of the disclosure, lossless compression of digital files is achieved with BSD. As explained elsewhere herein, lossy compression can be achieved with a similar method.

Assume that the input file is represented by an mn x pq matrix M that will be decomposed with respect to the decomposition shape (m, n, p, q). In the course of the compression and decompression processes, we shall define and compute the following matrices and sequence.

P: The pattern matrix of M.

E: The essential matrix of M, this is an mn x pq matrix.

S: Essential sequence, built using the pigeonhole principle (it should be understood that using the pigeonhole principle is optional, as detailed elsewhere, herein).

With reference to FIG. 1 A, the following sets of steps describe a BSD of M. In the following, we describe the eth set where e is an integer between 1 and the rank of M that will be computed during the compression process. At this set of steps, an mn x pq matrix Me is defined. At 100, the compression process starts with Mi=M.

Step 1: The software computes d, the maximum value of the absolute values of the entries of Me. Following the lexicographic order, position (r, c) in Me of the first entry de whose absolute value is equal to d is stored at the eth column of the pattern matrix P (102).

Step 2: The software performs the Euclidean division of c by q and find the reminder j. If the reminder is zero, then j is replaced with q. The software then computes i=(c-j)/q+l .

Step 3: The software performs the Euclidean division of r by n and find the reminder /. If the reminder is zero, then / is replaced with n. The software then computes k=(r-l)/n+l.

Step 4: For each integer a between 1 and m and each integer b between 1 and p, the entry of Me at position (l+(a-l)n,j+(b-l)q) is stored at position (a, b) of the matrix Ae (104).

Following the lexicographic order, if position (l+(a-l)n,j+(b-l)q) of E is unfilled, we store in it the entry of M located at the same position; and we store the same entry in the sequence S (106). This entry is then called an essential entry of M.

Step 5: For each integer a between 1 and n and each integer b between 1 and q, the entry of

Me at position (a + (k-1) n,b + (i-1) q) is stored at position (a, b) of the matrix Be (108).

Following the lexicographic order, if position (a+(k-l)n, b+(i-l)q) of E is unfilled, we store in it the entry of M located at the same position (106); and we store the same entry in the sequence S (108). This entry is then called an essential entry of M.

Step 6: The software computes the matrix Me+i = Me - Ae ® Be/ de. (110) (As detailed elsewhere herein, the disclosure proves the fact that rank {Me -Ae ® Be / de) = rank (Me)-1)

If Me+i = 0, then the lossless compression of Mis achieved (112).

If Me+i≠ 0, then Steps 1-6 are performed with Me+i (114).

The process stops after R sets of steps, where R is the Schmidt rank of M.

With reference to FIG. 1C, the lossy compression process includes sets of the same steps 1- 5, while an error threshold is chosen at Step 1 (140); and the second statement of Step 6 is replaced with a statement that stops the process when the error threshold is exceeded (142).

The decomposition shape (m, n, p, q), the pattern matrix P, and the essential sequence S constitute the compressed file. There are other possible embodiments of BSD, including for example replacing step 1 with "Following the lexicographic order, position (r, c) of the first nonzero entry de of Me is stored at the 6th column of the pattern matrix P." Various embodiments can lead to different compression results.

The lossless compression described in this embodiment assumes that the size of the data matrix fits with the shape of the Schmidt decomposition. In other cases, the data matrix is augmented to fit the shape and the added entries are disregarded at the decompression process. Notice, that there are shapes (such as (m, n, 1, 1)) that fit with the size of every matrix. Embodiment 1 Decompression

With reference to FIG. IB, in the decompression process, the quadruple (m, n, p, q), the matrix P, and the sequence S constitute the input (120).

Step 1: The software computes R that is equal to the number of columns in P (122).

Step 2: Starting with e=l and finishing with e=R, the apparatus performs the following substeps for each e (124).

Step 2.1: The apparatus extracts the ordered pair (r, c) from the eth column of P.

Step 2.2: The software performs the Euclidean division of c by q and find the reminder j. If the reminder is zero, then j is replaced with q. The software then computes i=(c-j)/q+l .

Step 2.3 : The software performs the Euclidean division of r by n and find the reminder /. If the reminder is zero, then / is replaced with n. The software then computes k=(r-l)/n+l.

Step 2.4. Following the lexicographic order of the set { (a,b): 1 < a < m and 1 < b < p }, the apparatus stores the first terms of S that are enough to fill in E every unfilled position (l+(a- l)n, j+(b-l )q); S is then replaced with a new sequence that includes all terms of S but the terms used in Step 2.4 for e. After that, the old sequence is deleted and the new sequence is called S.

Step 2.5. Following the lexicographic order of the set { (a,b): 1 < a < n and 1 < b < q }, the apparatus stores the first terms of S that are enough to fill in E every unfilled position (a + (k-1) n,b + (i-1) q); S is then replaced with a new sequence that includes all terms of S but the terms used in Step 2.5 for e. After that, the old sequence is deleted and the new sequence is called S.

Step 3: The software fills with 0 each position in E that is not filled after step 2 (126). Step 4: At this step, an mn x pq matrix Ee is defined, for each e=l,...,R. The process starts with Ei = E. The software extracts the ordered pair (r, c) from the eth column of P, extracts the entry de of Ee at position (r,c), and performs the following sub-steps (128).

Step 4.1: The software performs the Euclidean division of c by q and find the reminder j. If the reminder is zero, then j is replaced with q. The software then computes i=(c-j)/q+l .

Step 4.2: The software performs the Euclidean division of r by n and find the reminder /. If the reminder is zero, then / is replaced with n. The software then computes k=(r-l)/n+l.

Step 4.3 : For each integer a between 1 and m and each integer b between 1 and p, the entry of Ee at the position (l+(a-l)nj+(b-l)q) is a is placed at position (a, b) of Ae.

Step 4.4: For each integer a between 1 and n and each integer b between 1 and q, the entry of Ee at the position (a+(k-l)n, b+(i-l)q) is placed at position (a, b) of Be.

If e < R, then the software computes Ee+i = Ee - Ae ®Be / de and redo sub-Steps 4.1-4.4 (130).

If e = R, then the software performs Step 5.

Step 5: The software collect the matrices found in Step 4 as follows:

N= Ai ®Bi /di + A2 ®Bi /di + ... + AR ®BR /dR.

The matrix N is the matrix of the output file. If the compression is lossless, then N = M (132). If the compression is lossy, then N approximates M.

Assume that a digital file is represented by M whose entries are integers between 0 and 255. Thus, every pixel in the file requires 1 byte to be stored in a computer. To store the

decomposition shape we need 4 bytes. The terms of the essential sequence S are extracted from M, and hence, the memory space required to store S is less than the memory space required to store M. The number of entries of P is less than the minimum of the numbers 2mp and 2nq, while M counts mnpq entries. Pratically, the pattern matrix P occupies less than 1% of the memory space occupied by M. Thus, compression with BSD has a lossless compression ratio that is more than 0.99. For example, if m=16, n=24, p=8, and q=16, the memory size of the matrix Mis 49,152 bytes. Meanwhile, to store the decomposition shape (16,24,8,16) and the pattern matrix P, we need no more than 260 bytes, that is less than 0.53% of the size of M. Thus, the

compression ratio is more than 0.994. In many cases, there are several choices for the decomposition shape (m, n, p, q). Some of these choices are more appropriate to adapt to some type of files to be compressed. For example, to decommpose a 256 x 256 matrix, we can chose from 49 decomposition shapes.

The disclosure uses algebraic computations that include the Kronecker product. This can be distinguished from lossless compression methods based on statistical models and bit sequences that include Huffman coding and arithmetic coding, which are logical and statistical. The disclosure provides algebraic lossless compression that can be applied to digital data.

Embodiment 2

With reference to FIG. 2A, the second embodiment uses BSD, low rank approximation, and compressing by blocks, to achieve efficient lossy compressions of digital files. To measure the quality of the output file, Peak Signal to Noise Ratio (PSNR) is used. It should be understood that while PSNR is referenced herein as a measurement of quality or error, there are many different types of measurements of error which can be substituted, for example root means squared, PEVQ (Perceptual Evaluation of Video Quality, which is advantageous for video data), and SSIM (Structural Similarity Index Method, which, like PSNR, is advantageous for images), where the other measurement is selected as best suits the type of original data file used, or as otherwise deemed suitable for a particular purpose.

Assume that the input file is an image represented by a matrix M whose entries are integers in [0, 255], for which we want the entries of the compressed file and output file to be also integers in [0, 255]. Thus, each entry can be stored with one byte in a digital device.

For the compression process, the inputs include M, Er, a number that sets a threshold error, and the parameters m, n, p, and q that define the block matrices. We assume that the size of Mis several times bigger than the size of an mn x pq -matrix. In the process of compression, M will be split into mn x pq-block. matrices. In the course of the compression process the apparatus builds a pattern matrix P, an essential sequence S, and a sequence of ranks R.

Step 1 : If the height of M is not divisible by mn or the width of M is not divisible by pq, then columns or rows of zeroes are added to Mto reach the divisibility (160). The added parts are then disregarded at the decompressing process.

Step 2: The software splits the image into blocks. Each block is an mn x pq -matrix. Let's say that h is the number of these blocks (162). BSD will be used with respect to (m,n,p,q).

Step 3: For each e=l,...,h, the software performs the following steps for the 6th block Me. Step 3.1: Using SVD, the software computes the matrix De that has the lowest rank possible Re among all matrices for which the PSNR with Mis bigger or equal to Er (164).

Step 3.2: The software rounds up the matrix De and obtains a matrix Qe whose entries are integers. (Most of the entries of Qe are in [0, 255], but few entries maybe out of that range) (166).

Step 3.3: Using Embodiment 1, the apparatus compresses Qe and computes the pattern matrix Pe and the essential sequence Se (168).

Step 3.4: The software stores the rows of the eth pattern matrix in the rows (2e-l) and 2e of the pattern matrix P; the terms of Se in S (successively after the last term of S), and Re in R (170).

The quadruple (m, n, p, q), the matrix P, the sequence S, the sequence R, and the size of M constitute the compressed file (172). Embodiment 2 Decompression

With reference to FIG. 2B, for the decompression process, the input is the compressed file that includes the quadruple (m, n, p, q), the matrix P, the sequence S, the sequence R, and the size of M (180).

Step 1. The software computes h as the number of terms of R (182).

Step 2. For each e=l,..., h, the apparatus performs the following steps to find the eth block of the output matrix.

Step 2.1. The software extracts from the input matrices the matrix Pe as the rows (2e-l) and 2e of the pattern matrix P and extracts from R the number Re (184).

Step 2.2. With the matrix Pe, the number Re, and the sequence S, the apparatus uses steps 2-5 of the decompression process of embodiment 1 to build a block matrix Ne that approximates the original block Me (186). The pigeonhole principle is used to replace S with a new sequence that includes all terms of S but the terms used in Step 2.2. After that the old sequence is deleted and the new sequence is called S (188).

Step 3. The software collects the h blocks built in Step 2 as one matrix N (190). The software then uses the size of M and a quantization of Nto build a matrix Nq whose entries are integers in [0, 255], Nq has the size of M, and Nq approximates M (192). Compressing by blocks or tiling allow parallel processing that significantly reduces the compression and decompression time.

Embodiment 3

With reference to FIG. 3A, in a third embodiment, software of the disclosure encrypts digital files, in addition to performing compression. After compressing a file, the software enables a user to choose a key code that is needed at the decompression process. The encryption software can be used with the lossy or a lossless compression methods of the disclosure.

In an example of the encryption process, assume that the input file is represented by an mn x pq-matrix M.

Step 1. The software compresses the file using the lossless compression method described in the first embodiment. The outputs are then the essential sequence S and the pattern matrix P (200).

Step 2. The user chooses a key code K, a number that can be of one to hundreds of digits (202).

Step 3. This step can be any one-to-one algorithm implemented in the software to convert the matrix P to a matrix L, using the key code (204).

The encrypted file includes the sequence S, the matrix L, the shape (m, n, p, q), and the user key code K (206).

There are numerous possible embodiments which can be used to implement an algorithm for step 3. In an example, assume that the user chose a number K = KiK2...Kh with h digits and h is less or equal to the number of rows of P. For each e=l, ..., h, the software shifts the entries of the 6th row of P in a circular way by Ke places. Other known and hereinafter developed methods can be used to convert P to L using the key code.

Embodiment 3 Decompression / Decryption

For the decryption process, the input includes S, L, K, and (m, n, p, q) (210).

Step 1. The software uses K, L, and the inverse of the algorithm of the third step in the encryption process to recover P (212).

Step 2. The software uses the lossless decompression method, described in the first embodiment, with S, P, and (m, n, p, q) to recover M (214). The user can use an encryption apparatus executing software of the present disclosure, to secure her/his files without the involvement of a third-party. The choice by the user of the encrypting algorithm for Step 3 in the encryption process can be a part of the key code. The decomposition shape (m, n, p, q) can be also part of the key code. All these options render access of secured data by unauthorized people more difficult.

In one embodiment, for example useful within a corporate environment, software of the disclosure can be configured to generate an encrypting algorithm for each employee. An administrator manager will have access to the encrypting algorithm of each employee, while the employee can choose a key code that locks and unlocks her/his files. If one employee stops working for the company and the company wants the former employee to become unable to access the company files, the administartor has only to change the encrypting algorithm for that employee. No employee will be required to reset her/his key code.

The lossless compression method used in the first step of the encryption process can be replaced by any lossy compression method described in the first or second embodiment.

Embodiment 4

In an embodiment of the disclosure, lossless compression of digital files is achieved with BSD. As explained elsewhere herein, lossy compression can be achieved with a similar method.

Assume that the input file is represented by an mn x pq-matrix M that will be decomposed with respect to the decomposition shape (m, n, p, q). In the course of the compression and decompression processes, we shall define and compute the following matrices.

P: The pattern matrix of M, this is a 2 x R-matrix.

E: The essential matrix of M, this is an mn x pq-matrix.

A: The left essential matrix of M, this is an m x Rp-matrix.

B: The right essential matrix of M, this is an n x Rq-matrix.

The following steps describe a BSD decomposition of M. In the following, we describe the eth set, where e is an integer between / and the rank of M that will be computed during the compression process. In this set of steps, an mn x pq-matrix Me is defined. The compression process starts with Mi=M. Step 1: The software computes the maximum value d of the absolute values of the entries of Me. Following the stereographic order, the position (r, c) of the first entry of Me that is equal to d is stored at the e* column of the pattern matrix P.

Step 2: The software performs the Euclidean division of c by q and find the reminder j. If the reminder is zero, then j is replaced with q. The software then computes i=(c-j)/q+l .

Step 3: The software performs the Euclidean division of r by n and find the reminder 1. If the reminder is zero, then 1 is replaced with n. The software then computes k=(r-l)/n+l .

Step 4: For each integer a between 1 and m and each integer b between 1 and p, the entry of M at the position (l+(a-l)n,j+(b-l)q) is a left essential entry. This entry is stored at the position (a, (e-l)p+b) of the left essential matrix A. The entry of Me at the position (l+(a-l)n, j+(b-l)q) is stored at the position (a, b) of the matrix Ae.

Step 5: For each integer a between 1 and n and each integer b between 1 and q, the entry of M at the position (a+(k-l)n, b+(i-l)q) is a right essential entry. This entry is stored at the position (a, (e-1) q+b) of the right essential matrix B. The entry of Me at the position (a + (k-1) nb + (i-l) q) is stored at the position (a, b) of the matrix Be.

Step 6: The software computes the matrix Me+i = Me - Ae ® Be. (As detailed elsewhere herein, the disclosure proves the fact that rank (Me -Ae ® Be) = rank (Me)-1)

If Me+i = 0, then the lossless compression of M is achieved.

If Me+i≠ 0, then Steps 1-6 are performed with Me+i.

The lossy compression process includes sets of the same steps 1-5, while the second statement of Step 6 is replaced with an error threshold that stops the process. P, A and B constitute the compressed file.

Embodiment 4 Decompression

In the decompression process, the matrices P, A and B constitute the input.

Step 1: The software computes R the number of columns in P, m the number of rows of A, p the number of columns of A divided by R, n the number of rows of B, and q the number of columns of B divided by R.

Step 2: For each e=l,...,R, the apparatus extracts the ordered pair (r, c) from the eth column of P and performs the following sub-steps to fill the essential matrix E with the essential entries that are left from A and right from B. Step 2.1: The software performs the Euclidean division of c by q and find the reminder j. If the reminder is zero, then j is replaced with q. The software then computes i=(c-j)/q+l .

Step 2.2: The software performs the Euclidean division of r by n and find the reminder 1. If the reminder is zero, then 1 is replaced with n. The software then computes k=(r-l)/n+l .

Step 2.3: For each integer a between 1 and mn and each integer b between 1 and pq, the entry of A at the position (a, (e-l)p+b) is placed at the position (l+(a-l)n, j+(b-l)q) of E; and the entry of B at the position (a, (e-l)q+b) is placed at the position (a+(k-l)n, b+(i-l)q) of E.

Step 3: The software fills with 0 the positions in E corresponding to the positions of nonessential entries in M.

Step 4: At this step, an mn x pq-matrix Ee is defined, for each e=l ... R. The process starts with Ei = E. The software extracts the ordered pair (r, c) from the eth column of P and performs the following sub-steps.

Step 4.1: The software performs the Euclidean division of c by q and find the reminder j. If the reminder is zero, then j is replaced with q. the software then computes i=(c-j)/q+l .

Step 4.2: The software performs the Euclidean division of r by n and find the reminder 1. If the reminder is zero, then 1 is replaced with n. The software then computes k=(r-l)/n+l .

Step 4.3: For each integer a between 1 and m and each integer b between 1 and p, the entry of Ee at the position (l+(a-l)n,j+(b-l)q) is a is placed at the position at (a, b) of the matrix Ae.

Step 4.4: For each integer a between 1 and n and each integer b between 1 and q, the entry of Ee at the position (a+(k-l)n, b+(i-l)q) is placed at the position (a, b) of the matrix Be.

If e < R, then the software computes Ee+i = Ee - Ae ® Be and redo sub-Steps 4.1-4.4.

If e = R, then the software performs Step 5.

Step 5: The software collect the matrices found in Step 4 as follows:

N= Ai ® Bi + A2 ® B2 + ...+ AR <8> BR.

The matrix N is the matrix of the output file. If the compression is lossless, then N = M. If the compression is lossy, then N approximates M.

The essential matrix E built at the decompression process can be built at the compression process and replaces the left and right essential matrices A and B. In this case, the compressed file is formed by E and P.

The entries of E are the essential entries of the original matrix M and the non-essential entries can be any number chosen. Thus, using other lossless compression on E will automatically lead to a compressed file that has a memory value less than the memory value of the compressed file if the same lossless compression method is used on M.

In practice, the pattern matrix P can occupy less than 1% of the memory space occupied by M. Thus, compression with BSD has a lossless compression ratio that is more than 0.99. For example, if m=16, n=24, p=8, and q=16, the memory size of the matrix Mis 49,152 bytes.

Meanwhile, to store the decomposition shape (16,24,8,16) and the pattern matrix P, we need no more than 260 bytes, that is less than 0.53% of the size of M. Thus, the compression ratio is more than 0.994.

In many cases, there are several choices for the parameters m, n, p and q. Some of these choices are more appropriate to adapt to some type of files to be compressed.

Embodiment 5

The second embodiment uses BSD, low rank approximation, and compressing by blocks, to achieve efficient lossy compressions of digital files. To measure the quality of the output file, Peak Signal to Noise Ratio (PSNR) is used.

Assume that the input file is an image represented by a matrix M whose entries are integers in [0, 255], for which we want the entries of the compressed file and output file to be also integers in [0, 255].

For the compression process, the inputs include M, Er, a number that sets a threshold error, and the parameters m, n, p, and q that define the block matrices. We assume that the size of M is several times bigger than the size of an mn x pq-matrix. In the process of compression, M will be split into mn x pq-block matrices.

Step 1: If the height of M is not divisible by mn or the width of M is not divisible by pq, then columns or rows of zeroes are added to M to reach the divisibility. The added parts are then disregarded at the decompressing process.

Step 2: The software splits the image into blocks. Each block is an mn x pq-matrix. Let's say that h is the number of these blocks.

Step 3: For each e=l,...,h, the software performs the following steps for the eth block Me.

Step 3.1: Using SVD, the software computes the matrix De that has the lowest rank possible Re among all matrices for which the PSNR with M is bigger or equal to Er. Step 3.2: The software rounds up the matrix De and obtains a matrix Qe whose entries are integers. (Most of the entries of Qe are in [0, 255], but few entries are out of that range).

Step 3.3: Using BSD of Me with Re sets of the steps described in 4.1.2, the software decompresses Me and, as in 4.1.2, builds the left essential, right essential, and pattern matrices of

Step 3.4: The software stores the rows of the eth pattern matrix in the rows (2e-l) and 2e of a matrix P, the pattern matrix; the rows of the e* left essential matrix in the rows ((e-l)m +1) to em of a matrix A, the left essential matrix; the rows of the e* right essential matrix in the rows ((e-l)n +1) to en of a matrix B, the right essential matrix, and Re in a sequence of numbers R. The matrices A, B, P, and R along with the size S of the original file constitute the compressed file.

Embodiment 5 Decompression

For the decompression process, the input is the compressed file that includes A, B, P, R, and S.

Step 1. The software extracts, from the input matrices, h, m, n, p, and q.

Step 2. For each e=l,...,h, the apparatus performs the following steps to find the e* block of the output matrix.

Step 2.1. The software extracts from the input matrices, the matrices Ae, Be, and Pe and the number Re.

Step 2.2. The software uses similar to the decompression process described in 4.1.4 to build a block matrix Ne that approximates the block Me.

Step 3. The software collects the h blocks built in Step 2 and forms a matrix, and then the software uses S to build N a matrix that approximates M. If we want the entries of the output to be an integer in [0, 255], we may need to quantize N.

Compressing by blocks or tiling allow parallel processing that significantly reduces the compression and decompression time.

Testing

Using Embidiment 5, the inventor has tested the compression and encryption methods of the disclosure. Some of those testing results are as follows. The test images evaluated are color images in RGB format. Each image is represented by three matrices. Before compressing using BSD, a matrix was formed with the three matrices one next to the other in a horizontal orientation.

To compare the quality of the output images, Peak Signal to Noise Ratio (PSNR) is usedwhich is commonly used to measure the quality of reconstruction of lossy compression. The larger is the value for PSNR, the better is the quality.

SVD can generally achieve a very good PSNR, albeit with the disadvantages described herein, and which are known. The tests show that BSD leads to a PSNR that is at least very close to the PSNR achieved by SVD, while the compression ratio for BSD is significantly better than the compression ratio for SVD.

The first two images "Splash" and "Lenna" are included in the set of test-images of University of South California linked at http://sipi.usc.edu/database/database.php.

The second two images "Deer" and "Spider Web" are part of the set of test-images offered at http://www.imagecompression.info/.

The third two images "kodimOl" and "kodim23" are included in Kodak test-images at http://rOk.us/graphics/kodak/.

Each image is compressed using PNG, JPEG, BSD, and SVD. The trial version of

PhotoPad Image Editor was used to achieve the PNG compression and the JPEG compression with its highest quality 100. The compressed files with BSD are stored in PNG format.

The output file for SVD is the file that is represented by the matrix Q in the compression process described in Embodiment 5.

The PSNRs for JPEG and BSD are almost the same. The compressed files for SVD and BSD have almost the same number of pixels, but the memory value per pixel for SVD is much larger than the memory value per pixel for BSD.

The difference between the PSNR for SVD and the PSNR for BSD is between 0.03 dB and 0.57 dB. For Splash, Deer, and Spider, the compression ratios for BSD is between 2.6 and 4 times the compression ratios for JPG. For Lenna, kodimOl, and kodim23 the compression ratios are slightly improved by BSD vs JPG.

In each of the compression apparatuses of the present invention, other image processing techniques can be implemented to enhance the image quality and lead to further compression. These techniques include entropy encoding, area image compression, and YCbCr conversion.

Embodiment 6

In this embodiment of the disclosure, lossless compression of digital files is achieved using the BSD method of the disclosure. The algorithm used in this embodiment is different from the algorithm used in embodiment 1, but the two algorithms are based on BSD, as disclosed herein, and which can carry out Schmidt decompositions. As explained elsewhere herein, lossy compression can be achieved with a similar method.

Assume that the input file is represented by an mn x pq matrix M that will be compressed using a Schmidt decomposition with respect to the decomposition shape (m, n, p, q). In the course of the compression and decompression processes, we shall define and compute the following matrices and sequence.

PS: The pattern sequence of M.

E: The essential matrix of M, this is an mn x pq matrix.

S: Essential sequence, built using the pigeonhole principle. S shall have two parts Si and Si. The following sets of steps describe the compression process. We describe the e set where e is an integer between 1 and the rank of M that will be computed during the compression process. At this set of steps, an mn x pq matrix Me is defined. The compression process starts with M1=M

Step 1: The software computes d, the maximum value of the absolute values of the entries of Me. Following the lexicographic order, let (r, c) be the position in Me of the first entry de whose absolute value is equal to d. Position e of the n x q block matrix of M that includes position (r, c) is stored at the eth term of the pattern sequence PS.

Step 2: The eth n x q block matrix of M is stored at the eth n x q block matrix of E; and following the lexicographic order, the entries of the eth n x q block matrix of M are stored in Si.

Step 3: The software performs the Euclidean division of c by q and find the remainder j. If the remainder is zero, then j is replaced with q. The software then computes i=(c-j)/q+l .

Step 4: The software performs the Euclidean division of r by n and find the remainder /. If the remainder is zero, then / is replaced with n. The software then computes k=(r-l)/n+l.

Step 5: For each integer a between 1 and m and each integer b between 1 and p, the entry of Me at position (l+(a-l)n,j+(b-l)q) is stored at position (a, b) of the matrix Ae. Following the lexicographic order, if position (l+(a-l)n,j+(b-l)q) of E is unfilled, we store in it the entry of M located at the same position, and we store the same entry in the sequence &.

Step 6: For each integer a between 1 and n, and each integer b between 1 and q, the entry of Me at position (a + (k-1) n,b + (i-1) q) is stored at position (a, b) of the matrix Be.

Step 7: The software computes the matrix Me+i = Me - Ae ® Be/ de. (As detailed elsewhere herein, the disclosure proves the fact that rank {Me -Ae ® Be / de) = rank (Me)-1)

If Me+i = 0, then the lossless compression of M is achieved.

If Me+i≠ 0, then Steps 1-7 are performed with Me+i.

The process stops after R sets of steps, where R is the Schmidt rank of M.

The lossy compression process includes sets of the same steps 1-6, while the second statement of Step 7 is replaced with a statement that includes an error threshold that stops the process.

The decomposition shape (m, n, p, q), the pattern sequence PS, and the essential sequence S constitute the compressed file, where S is a sequence whose first part is Si and second part is &. The essential sequence S has R(mp + nq -R) terms extracted from M and PS has R terms, while the matrix M has mpnq entries.

Embodiment 6 Decompression

In the decompression process, the quadruple (m, n, p, q), and the sequences PS and S constitute the input.

Step 1: The apparatus computes R that is equal to the number of terms in PS.

Step 2: The first Rnq terms of S are used to build R n x q block matrices of an mn x pq matrix E (the essential matrix). The positions of the R block matrices are given by PS. The first Rnq terms of S are then deleted from S. The positions of E that are filled after Step 2 are hereafter called essential positions.

Step 3: Starting with e=l and finishing with e=R, the apparatus performs the following sub-steps. At this step, mn x pq matrices Me are defined. The process starts with Mi=E.

Step 3.1: The apparatus builds the matrix Be whose entries are the entries of the eth n x q block matrix of Me. The apparatus computes d, the maximum value of the absolute values of the entries of Be. Following the lexicographic order, the apparatus computes (r, c) the position in Me of the first entry de in Be whose absolute value is equal to d. The entry de is stored at the eth term of a sequence D.

Step 3.2: The software performs the Euclidean division of c by q and finds the remainder j. If the remainder is zero, then j is replaced with q. The number j is stored at the eth term of a sequence J. The software then computes i=(c-j)/q+l. The number i is stored at the eth term of a sequence /.

Step 3.3: The software performs the Euclidean division of r by n and finds the remainder /. If the remainder is zero, then / is replaced with n. The number / is stored at the eth term of a sequence L. The software then computes k=(r-l)/n+l. The number k is stored at the eth term of a sequence K.

Step 3.4: At this sub-step an an m χ p matrix Ae is built. For each ordered (a, b), where 1 < a < m and 1 < b < ρ, ϊΐ (l+(a-l)n,j+(b-l)q) is an essential position, then the position (a, b) of Ae is filled with the entry of Me located at position (l+(a-l)n,j+(b-l)q). The other positions of Ae can be filled with any value. Step 3.5: If e = R, the apparatus performs Step 4. If not, the apparatus computes the matrix Me+i = Me - Ae ® Be/ de and performs the sub-steps of Step 3 for Me+i.

Step 4: Starting with e=l and finishing with e=R, the apparatus computes j, the eth term of J and /, the 6th term of L. Following the lexicographic order, if (l+(a-l)n, j+(b-l)q) is not an essential position of E, where 1 < a < m and 1 < b < p, then this position is filled with the first term of S. This term is then deleted from S and the position (l+(a-l)n,j+(b-l)q) becomes a new essential position of E.

Step 5: At this step, an mn x pq matrix Ee is defined, for each e=l,...,R. The process starts with Ei = E. Starting with e=l and finishing with e=R, the apparatus performs the following sub-steps.

Step 5.1: The apparatus computes k, and / that are, respectively, the eth term of 7, the eth term J the eth term of K, and the eth term of L.

Step 5.2: At this sub-step, an m χ p matrix Ae and an n x q matrix Be are built. For each integer a between I and m and each integer b between I and p, the entry of Ee at the position (l+(a-l)n,j+(b-l)q) is placed at position (a, b) of Ae. The entries of Be are the entries of eth n x q block matrix of Ee.

If e < R, then an electronic processing apparatus of the disclosure carrying out this method computes Ee+i = Ee - Ae ®Be / de, where de is the eth term of D, and performs sub-steps 5.1 and If e = R, then the apparatus performs sub-step 5.3.

Step 5.3: The apparatus collects the matrices found in sub-step 5.2 as follows:

N= Ai ®B1 /d1 + A2 ®B2 / d2 + ... + An ®BR /dR.

The matrix N is the matrix of the output file. If the compression is lossless, then N = M and R is the Schmidt rank of M. If the compression is lossy, then N approximates M and R is less than the Schmidt rank of M.

Additional Examples

In the following additional examples, 'the apparatus' refers to an apparatus, including a processor, executing software in accordance with the disclosure. Consider the 6 x 6-matrix A/ whose entries are integers between 0 and 255. Thus, the memory value of A/ is 36 Bytes.

In Section 1, I use BSD to achieve a lossless compression of Λ/. We shall use BSD to decompose Λ/ into a sum of matrices, for which each term is the Kronecker product of a 3 x 2-matrix by a 2 x 3- matrix. My previous summary includes a diagram that describes the lossless compression, the lossy compression, and the encrypting algorithm.

Since, rank(Af) = 2, the lossless compression is achieved in two steps.

In section 2, I use BSD and SVD to achieve a lossy compression of Λ/. I approximate M with a matrix A whose rank is equal to 1. Thus, the lossy compression is achieved in 1 step.

Lossless Compression with BSD (The first Embodiment)

.1 Compression Process

1. The input is the matrix A/.

Step 1.1 The entry 36 has the biggest absolute value and it is positioned at (6,6)= ( 6th row, 6th column). The position (6,6) is stored in the first row of the pattern matrix P.

Step 1.2 The first left essential entries are chosen from A/:

Step 1.3 The first right essential entries are chosen from A/:

Step 1.4 The apparatus computes

i

positioned at (1, 1) in the matrix Mi. The position (1,1) is stored in the second row of the pattern matrix P.

Step 2.2 The second left essential entries are chosen from Λ/:

Step 2.3 The second right essential entries are chosen from A/:

Step 2.4 The apparatus computes Mi - LE2 ®RE2/(1 -9§§) to find the zero-matrix. This means that rank (A/) = 2 and the decomposition of M is achieved.

The compressed file includes the pattern matrix P, the left essential matrix LE, and the right essential matrix RE:

The matrices P, LE, and RE have 28 entries. Thus, we compressed the memory value of A/ from 36 bytes to 28 bytes.

The entries 28, 1, 36, and 9 are part of both the left and right essential entries. We then can reduce the memory value of the compressed file from 28 bytes to 24 bytes. Decompression Process

The inputs are the matrices P, LE, and RE.

The number of rows of P is 2. It is equal to the rank of the original matrix A/. It is also equal to the number of steps to recover A/.

The number of columns of LE is 4 and the number of columns of RE is 6. This means that the original matrix A/ was decomposed into a sum of matrices, for which each term is the Kronecker product of a 3 x 2- matrix by a 2 x 3- matrix. Therefore, we can recover the matrices LE\, RE\ , LE?, and RE2.

Step 1.1 Using the first row of P, that is (6, 6), the apparatus puts the entries of LE\ and RE\ in a matrix E at the same positions they occupy in the matrix A/.

Step 1.2 Using the second row of P, that is (1, 1), the apparatus puts the entries of LEi and REi in the matrix E at the same positions they occupy in the matrix A/.

Step 1.3

The matrix E is called the essential matrix and has the size of A/. The entries of E that are not essential entries of A/ can be 0 or any number we want:

\ \ \ 2 2

3 i 3

Step 2.2 Using the position (1, 1), the second term of the decomposition

of Λ/ is A2 ® B2, where

Using Theorem 2.4 of the paper "Inverting the tensor product of bounded operators on Hilbert Spaces" , we can prove that

Thus, the matrix M is recovered. Lossy Compression (The second Embodiment)

Compression Process

The input is the matrix Λ/ and T: a PSNR threshold to determine the quality of the output. We take T = 44 for this example.

Step 1.1 Using the low rank approximation theorem (EckartYoungMirsky

theorem), the apparatus computes N the matrix with the lowest rank among the matrices that approximate M with a PSNR greater or equal to T. The algorithm to achieve the low rank approximation uses SVD:

/ 4.8677 6.8704 7.2192 7.5680

6.2138 8.9633 9.3121 9.6609

15.6689 16.6759 17.5226 18.3692

20.0021 21.7559 22.6025 23.4492

26.4702 26.4815 27.8260 29.1705

33.7904 34.5485 35.8930 37.2374 /

Step 1.2 The matrix JV is quantized to obtain a matrix Q with integer entries:

V

Step 1.3 Using, with Q, the same steps used with M for the lossless compression, we find the pattern matrix P— (6, 6), the left essential matrix LE, and the right essential matrix RE:

2. The compressed file consists of P, LE, and RE. The memory value of the three matrices is 14. The compression ratio is then 36/14 ~ 2.57.

3. The matrix Q is the output of the compression with SVD. We have PSNR(M,Q)= 44.4510.

2.2 Decompression Process

The steps are similar to the steps of the lossless decompression

1. The inputs are the matrices P, LE, and RE.

2. The number of rows of P is 1. It is equal to the rank of the matrix JV.

It is also equal to the number of steps to build the output matrix that approximates Λ/.

3. The number of columns of LE is 2 and the number of columns of RE is 3. This means that the original matrix M was decomposed into a sum of matrices, for which each term is the Kronecker product of a 3 x 2- matrix by a 2 x 3-matrix. Using the first row of P, that is (6, 6), the apparatus puts the entries of LE and RE in a matrix E at the same positions they occupy in the matrix

Q- The matrix E is called the essential matrix and has the size of M. The entries of E that are not essential entries of M can be 0 or any number we want:

Before quantization, the output is the matrix

V

After quantization, we obtain the matrix

The matrix A is the output of the lossy compression of M. We have PSNR(A/, A) = 44.4510.

For this example, PSNR(M, A) = PSNR(M, Q), but, in general PSNR(A/, .4) is little bit less than PSNR(A/, Q).

Step 5 is not necessary because the matrix A was constructed in one step (this step consists of items 4-7). In general, E is used if more than one steps is needed. For example, in the lossless compression of Λ/, the matrix M is recovered using two steps. The tensor product of bounded operators on Hilbert spaces plays an important role in mathematics and its applications. The applications include composite quantum systems in quantum mechanic, control theory, statistics, signal processing, computer computing, and elsewhere [10, 13, 14]. For the finite dimensional case, the tensor product of operators is equivalent to the Kronecker product of matrices, with which is related an important inverse problem. That is decomposing an m1m2 x n1 n2 matrix Λ/ into a sum with a minimum number of terms, each of which is the Kronecker product of an m1 x η1 matrix and an m2 x n2 matrix. This is the so called Schmidt decomposition of M. The number of terms in this decomposition is the Schmidt rank of A/ that we shall denote rank® (A/). The classical method to find Schmidt decompositions for matrices is using SVD. We shall call this method SSVD.

Assume that an m1m2 x n1 n2 matrix A/ has an SSVD for which rank® (A/) = r. Therefore, all matrices involved in the decomDosition of A/ have a total number of entries equals to r( m1n1 + m2n2). Meanwhile, M has m1m2 n1 n2 entries. Thus, storing the matrices involved in the decomposition of M achieves a lossless compression of A/ if r( m1n1 + m2n2) < m1m2nin2. If s < r, Eckart- Young-Mirsky theorem uses SSVD to approximate A/ with an s Schmidt rank matrix N leading to the lowest error possible when A/ is approximated with an s Schmidt rank matrix [8]. Hence, storing the matrices involved in the decomposition of N achieves a lossy compression of A/ if s( m1n1 + m2n2) < m1m2 x n1 n. 2 There is another compression method based on SVD known as compression with SVD [2, 12]. To compress with SVD, singular values and singular vectors of A/ are to be stored.

Digital data is represented by matrices whose entries are in a certain class, i.e., a finite set of integers. However, the singular values and entries of the singular vectors are not necessarily integers for a matrix with integers entries. Therefore, a compression with SSVD or SVD leads to an output, for which each entry is more likely to be an irrational number. To store in a digital device an irrational number without significant loss of information, we need a memory space much larger than the space required to store an integer. Thus, with SSVD or SVD a lossless compressions of a digital file is almost impossible and a lossy compression leads to a compression ratio that cannot compete with ratios achieved by other existing compression methods such as JPEG. This paper generalizes a result in [4] that provides inverse formulas for the tensor product of bounded operators on separable Hilbert spaces. These formulas are then used to develop an algorithm, with several versions, to find finite Schmidt decompositions, if any, of bounded operators on the tensor product of separable Hilbert spaces. To the best of the author's knowledge, this is the first algorithm to find Schmidt decompositions of bounded operators on the tensor product of infinite dimensional separable Hilbert spaces. For matrices, unlike with SSVD the new algorithm is practical and do not require numerical computations related to spectral decompositions. Indeed, each entry in a matrix that is part of a term in the decomposition of a matrix Λ/ is computed using some entries of M combined with the four elementary operations: +,— , x, ÷. This leads to one of the applications of the theory of this paper. That is a new lossless compression method based on SVD, with which the storage problem for digital data is solved. This is because, the new method leads to a compressed file whose entries are extracted from the original file.

In addition to compression of digital data, the theory of this paper has applications in operator theory. In particular, properties of an operator expressed with a finite Schmidt decomposition are reflected by the operators involved in the decomposition. For example, if

is the decomposition of a compact operator, then the operator Fit is compact, for each

In section 2, we collect some definitions and properties that are necessary to state our results. In section 3, we construct inverse formulas for the tensor product of bounded operators in separable Hilbert spaces, we state a Schmidt decomposition theorem for bounded operators on the tensor product of Hilbert spaces, and we describe an algorithm to find Schmidt decompositions. In section 4, we present Schmidt decomposition algorithms for matrices; and we show how lossless compressions are possible with these algorithms. 2 Preliminaries and Notations

For this paper, all Hilbert spaces are assumed to be separable. Most of the definitions and results in this section can be found in [11].

LetH and K be two Hilbert spaces. We denote by B(H, K ) the space of bounded operators fromH to K. We denote by L2(H, K ) the space of Hilbert- Schmidt operators fromH to K. We denote by % 6 H, denote by x" the linear form onH defined by x, that is

Notice that (Ax)' = Ax', for each complex number A. Riesz representation theorem states thaHt and Ή' are identified by the ant i- linear isometry: x→ i\ and so Ή. and Ή" are identified by a linear isometry, where H" denotes the bidual of H- We shall take the last identification as an equality. Thus,

For each linear mapping F : H— > /C, we define the transpose of F as follows: r

For two Hilbert spaces H1 and H 2 th e Hilbert space €ΆΙΪ be interpreted as the Hilbert space L2(H, K ) · This interpretation is based on the identification of xi ® X2€ "Wi ® W2 with a rank one operator in the following way:

x x

From now on, we shall take this identification as an equality.

Let K\ and K2 be two Hilbert spaces. The tensor product of two operators can be defined as 3 Schmidt Decompositions for Bounded Operators

Definition 3.1. We say that F€ B(Hi ® W2,£i ® Κ½) has a Schmidt rank r, if r is the minimum number such that F can be written in the form

F

The following Proposition generalizes results published by the author in [4].

AC Χ

where /

/ /

C

P

f

f /

K Κ i j i

Proof. Statement (i) is straightforward,

/

i K

VU

/

V i

V

Statement (tit) is an easy consequence of (it).

(iv) The identity

u A

and statement (tit) imply the inequality

u

and so Pu.v is continuous.

Statement (v) is straightforward. □

( ' / /C C

V r F

; and

V F u u

and so rank®(F) = 1.

Statement (ri) is a straightforward consequence of statement (i). □ C

φ

u / 0 /C / { /C r

∑ £ u

The last sum implies ranks (F) < r.

(ii) To finish the proof of Theorem 3.6, it suffices to prove by induction on n the following claim.

For n = 0, we prove a stronger statement, that is each FMS

i s independent. For that, assume F

/

uO O j u

X] £ φ U

C ]

where cj, ..., <ν 6 C. Therefore, we can rewrite (3-3) as

k F r * C

a

/ /

ir AC

T j j where

Therefore, equality (F(u), v) = 1 implies the equality:

(3.7) and so u (3.8)

Subtracting (3.6) from (3.8), we obtain

F a [ n

au φ

On the right hand side of (3.11) we relabel the first double sum and we switch and relabel the last double sum to obtain

Γ Γ

where we agree that is zero any sum whose upper bound is less than the lower bound. Consequently, we obtain

F—

+ Δ }2 ® (3.12)

Using Theorem 3.6, rank¾(F— 2?u.r(F)) > r— 1, while the Schmidt rank of the right hand side of (3.12) is less or equal to r— 1. Therefore, we obtain the desired result: rank#(F— PU.„(F)) = r— 1. □

Corollary 3.8. Let F€ B(H\ ® Ητ, Κχ ® £2). Let V = {Vn} be a sequence of mappings defined on B(Hi ® ΉΊ, K\ ® IC2) as follows.

• V0 = I.

• Assume that T>Q, Z>„_i are constructed, for some n > 0.

/. // 2>„_i(F) ≠ 0, we choose u" = uf ® uj 6 ¾i ® ¾2 and vn = Vi ® € K.\ ® fC2 such that

(Ρη_,(Γ)(Μ-), ι;η) = 1

and we let

Vn = (I - Vu-,v-)Vn-l. f ( )

Proof. Theorem (3.7) states that rank®(PiF) = rank®(F)— 1. An easy induction implies

Hence, statements (t) and (ii) are equivalent.

Notice that

\

f i f

£ C

is obtained using Corollary 3.8 with the same notations. Therefore,

i

; / ( //

( C

4 Schmidt Decompositions for Matrices

4.1 BSD Algorithms

£ X

d /C J A

H H

α

% a

U

u / { }

To prove (iii) and (iv), we use similar to the arguments used to prove (t) and (it)

□ j

φ then

Using Theorem 3.7, we have

®( / /lk tt/( Jfi)) e( )

?

/

φ u

/

/

\

Φ / x M / / /

x

The algorithm stops after r steps if and only if . In this case, we say that the Schmidt decomposition of is obtained following the pattern matrix Thus, to find the Schmidt decomposition of M\, we can use one of many canonical BSD algorithms, each of which is determined by its pattern matrix. However, the pattern matrix cannot be defined prior to the algorithm, because at each step the chosen entry must be nonzero. A canonical BSD algorithm can be rather defined by stating how to choose the nonzero entry, e.g., choosing the first nonzero entry at each step.

Corollary 4.3. Let / Proof. We let Mi = M and we use the notations of the above algorithm.

The entries of A/n and A/21 are extracted from Mi , and so each entry of M

/ /

/

/

/ (

As a consequence of Corollary 4.3, if each entry of AJ takes a rational number, then so does each entry of each matrix involved in a canonical BSD of A/.

/

with respect to (3, 2, 2, 3).

Using the entry A/(l, 1) = 1, a first term in the decomposition is A/n ® A/2i, where

/

t

/

w

Lossless Compression with BSD

/

/

Hence, for each the entries of are extracted from

Definition 4.5. Aih denotes an ni x mi matrix whose entries are the entries of Λ/ positioned at the positions of the entries of Mih in Λ/Λ; and A2h denotes an n2 x m2 matrix whose entries are the entries of A/ positioned at the positions of M2h in Α/¾. i

is called an essential entry and its position in A/ is called an essential position. /

entries at the essential positions of A/ coincide with the essential entries of A/. The matrix E can take any value at a position that is not an essential position of A/. y

Aih excluding any entry of A/ that is previously stored in S. We call 5, the essential sequence of M.

Lemma 4.6. Following P, the sum of the first r terms in a canonical BSD of E provides the canonical BSD of M following P. n /

/ / hi A/ A

Proposition 4.7. (i) The matrix M can be recovered using the matrices An, ond the pattern matrix P.

(it) Using the fact that M is decomposed with respect to (n ι, π2, mi, m2), the matrix M can be recovered using the essential sequence S and the pattern matrix P.

(Hi) The number of entries of S is less or equal to

(iv) The values of the entries of the matrices An , A2\, A\r, A2r as well as the values of the terms of S are in the class of the values of the entries of M.

Proof, (i) Given Αη, Α2χ..., A\r, A&, and the pattern matrix P, we build an nm2 x m\rn2 matrix E as follows.

/

&

for each 6 between 1 and n2 and each a between 1 and m2. An entry of E that is not defined by (4.4) nor (4.5) can take an arbitrary value. Looking at definition 4.5, we conclude that E is an essential matrix of Λ/, and so, by Lemma 4.6, M is the sum of the first r terms of a BSD of E following P.

(ii) Assume that (nj , n2l mi, m2), the sequence $, and the matrix P are given. To build E, we shall use the pigeonhole principle and the stereographic order, i.e, we start with the successive positions of the first row, we then follow with the successive positions of the second row, and so on, to fill the η\η2 χτηχτη2 matrix E with the terms of 5. 5

s Assume Steps 1 to h-1 are performed for some h between 1 and r— 1 Step h Let i,j, k, and I be defined by (4.3).

1. Using some of the first ηχτηχ terms that come after the first Sh-\ terms of $, we fill E at the positions j

that are unfilled. Let t be equal to the first t terms of S used until the first part of Step h.

2. Using some of the first n2m2 terms that come after the first t terms of S, we fill E at the positions

that are unfilled. Let s/, be equal to the first sf, terms of 5 are used until Step h.

The process stops after a maximum of r steps. After that, any unfilled position of E can be filled with an arbitrary value.

It is clear that E is an essential matrix of A/, and so, by Lemma 4.6, A/ is the sum of the first r terms of a BSD of E following P. J

where k, and / are defined by (4.3). Therefore, the number of entries of 5 is less or equal than ( )

n

In the followings, we state for each of the methods SVD, SSVD, BSD(i), and BSD(ii) the number of entries needed to recover an nin2 x m1m2 matrix M whose entries are in a certain class of integers. We assume that rank(Af) = R and rank® (A/)— r.

The parameters (3,2,2,3), the pattern matrix P = [(1, 1), (3, 2)], and the essential sequence S = (1, 4, 13, 16, 25, 28, 2, 3, 7, 8, 9, 5, 14, 17, 26, 29, 15, 19, 20, 21) are what we need to recover M.

We first build the 6 x 6 essential matrix E.

Step 1 The first row of P is (1, 1).

1. The positions to first fill in E are (1, 1), (1, 4), (3, 1), (3, 4), (5, 1), (5, 4), in which we store the entries 1, 4, 13, , 16, 25, 28.

2. The positions to second fill in E are (1, 2), (1, 3), (2, 1), (2, 2), (2, 2), in which we store the entries 2, 3, 7, 8, 9. Notice that the pigeonhole principle eliminates the position (1, 1) that is already filled.

Step 2 The second row of P is (3, 2).

1. The positions to third fill in E are (1, 5), (3, 2), (3, 5), (5, 2), (5, 5), in which we store the entries 5, 14, 17, 26, 29. Notice that the pigeonhole principle eliminates the position (1, 3) that is already filled.

2. The positions to fourth fill in E are (3, 3), (4, 1), (4, 2), (4, 3), in which we store the entries 15, 19, 20, 21. Notice that the pigeonhole principle eliminates the position (3, 1) and (3, 2) that are already filled.

Thus, we obtain

filling the empty entries with arbitrary values and decomposing E following P, we obtain Λ/.

Therefore, the compression of M is of 28 entries whose values are integers between 1 and 36.

(b) Using SVD, we find the Schmidt decomposition

We conclude that the compression of At includes 24 real numbers that are not integers and probably irrational.

(c) The SVD of Λ/ leads to the singular values approximately equal to 127.2064 and 4.9526 and the singular vectors which are appraxi mati vely the columns of the following matrix:

We conclude that the compression of M includes 26 real numbers that are not integers and probably irrational.

(d) In the following table, we summarize the results for each compression. The error is the Frobenius norm of the difference between M and the output of the decompression computed using MATLAB.

An integer between 0 and 255 requires 1 byte to be stored in a computer. Thus, M requires 36 bytes to be stored, while the compressed file with BSD requires 28 bytes, and so a lossless compression is achieved.

MATLAB yields nonzero errors for SSVD and SVD. This means that using 4 bytes to store each entry of the compressed file with SSVD or SVD is not enough to exactly recover Λ/. Thus, to recover Λ/ we need more than 96 bytes with SSVD and more than 104 bytes with SVD. Therefore, no lossless compression is achieved.

In [6], the author introduces a lossy compression method based on BSD hybridized with SSVD. Applied to an image, the compression ratio with the new method is much higher than the compression ratio with SVD or SSVD, while the decompression outputs have the same quality. In some cases where the image is from standard test images, it was also noticed that the compression ratio with the new method are more than twice the compression ratio with JPEG. The author has submitted a patent application [7] that protect both the lossless compression with BSD and the lossy compression with methods based on BSD hybridized with SSVD.

Example Computer System

FIG. 4 illustrates the system architecture for a computer system 700, such as a general purpose computer, or a series of interconnected computers, or a computer containing a plurality of processors, on which or with which the disclosure may be implemented. The exemplary computer system of FIG. 4 is for descriptive purposes only. Although the description may refer to terms commonly used in describing particular computer systems, the description and concepts equally apply to other systems, including systems having architectures dissimilar to FIG. 4. Computer system 700 executes software stored on non-transitory media, the software configured to received inputs from other computing devices or from human interface devices.

Computer system 700 includes at least one central processing unit (CPU) 705, or server, which may be implemented with a conventional microprocessor, a random access memory (RAM) 710 for temporary storage of information, and a read only memory (ROM) 715 for permanent storage of information. A memory controller 720 is provided for controlling RAM 710.

A bus 730 interconnects the components of computer system 700. A bus controller 725 is provided for controlling bus 730. An interrupt controller 735 is used for receiving and processing various interrupt signals from the system components.

Mass storage may be provided by DVD ROM 747, or flash or rotating hard disk drive 752, for example. Data and software, including software 400 of the disclosure, may be exchanged with computer system 700 via removable media such as diskette, CD ROM, DVD, Blu Ray, or other optical media 747 connectable to an Optical Media Drive 746 and Controller 745.

Alternatively, other media, including for example a media stick, for example a solid state USB drive, may be connected to an External Device Interface 741, and Controller 740. Additionally, another computing device can be connected to computer system 700 through External Device Interface 741, for example by a USB connector, BLUETOOTH connector, Infrared, or WiFi connector, although other modes of connection are known or may be hereinafter developed. A hard disk 752 is part of a fixed disk drive 751 which is connected to bus 730 by controller 750. It should be understood that other storage, peripheral, and computer processing means may be developed in the future, which may advantageously be used with the disclosure. User input to computer system 700 may be provided by a number of devices. For example, a keyboard 756 and mouse 757 are connected to bus 730 by controller 755. An audio transducer 796, which may act as both a microphone and a speaker, is connected to bus 730 by audio controller 797, as illustrated. It will be obvious to those reasonably skilled in the art that other input devices, such as a pen and/or tablet, Personal Digital Assistant (PDA), mobile/cellular phone and other devices, may be connected to bus 730 and an appropriate controller and software, as required. DMA controller 760 is provided for performing direct memory access to RAM 710. A visual display is generated by video controller 765 which controls video display 770. Computer system 700 also includes a communications adapter 790 which allows the system to be interconnected to a local area network (LAN) or a wide area network (WAN),

schematically illustrated by bus 791 and network 795.

Operation of computer system 700 is generally controlled and coordinated by operating system software, such as a Windows system, commercially available from Microsoft Corp., Redmond, WA. The operating system controls allocation of system resources and performs tasks such as processing scheduling, memory management, networking, and I/O services, among other things. In particular, an operating system resident in system memory and running on CPU 705 coordinates the operation of the other elements of computer system 700. The present disclosure may be implemented with any number of commercially available operating systems.

One or more applications, such as an HTML page server, or a commercially available communication application, may execute under the control of the operating system, operable to convey information to a user.

All references cited herein are expressly incorporated by reference in their entirety. It will be appreciated by persons skilled in the art that the present disclosure is not limited to what has been particularly shown and described herein above. In addition, unless mention was made above to the contrary, it should be noted that all of the accompanying drawings are not to scale. There are many different features to the present disclosure and it is contemplated that these features may be used together or separately. Thus, the disclosure should not be limited to any particular combination of features or to a particular application of the disclosure. Further, it should be understood that variations and modifications within the spirit and scope of the disclosure might occur to those skilled in the art to which the disclosure pertains. Accordingly, all expedient modifications readily attainable by one versed in the art from the disclosure set forth herein that are within the scope and spirit of the present disclosure are to be included as further embodiments of the present disclosure. References:

[1] Abdelkrim Bourouihiya, The tensor Product of Frames, Sampling theory in signal and Image processing, Vol. 7, No. 1 (2008), pp. 65-76.

[1A] H. Cheng, Z. Gimbutas, P.-G. Martinsson, V. Rokhlin, On the compression of low rank matrices, SIAM J. Sci. Comput., 26 (2005), pp. 1389-1404.

[2] G. Eckart, G. Young, The approximation of one matrix by another of lower rank,

Psychometrika, 1, 1936, pp. 211-218.

[3] Horn, Roger A.; Johnson, Charles R., Topics in Matrix Analysis, 1991, Cambridge University Press.

[4] Satish K. Singh,and Shishir Kumar. Mathematical transforms and image compression: A review. Maejo Int. J. Sci. Technol. 2010, 4(02), 235-249.

[5] S. O. Aase, J. H. Husoy and P. Waldemar, A critique of SVD-based image coding systems. IEEE International Symposium on Circuits and Systems on VLSI 1999, Vol. 4, Orlando, Florida, USA, pp.13-16.

[6] B. Arnold and A. Mclnnes. An investigation into using singular value decomposition as a method of image compression. College of Redwood, University of Canterbury, New Zealand. Technical Report (2000).

[7] H. C. Andrews and C. L. Paterson, Singular value decomposition (SVD) image coding. IEEE Trans. Comm 1976, 24, 425-432.

[7a] G. H. Golub and C. Reinsels, Singular value decomposition and least square solutions, Numer.Math., 1970, 14, 403-420. 1976, 24, 425-432.

[8] V. Singh , Recent Patents on Image Compression - A Survey

http://www.benthamscience.com/open/rptsp/articlesA^002/47 RPTSP.pdf

[9] Julie Kamm and James G. Nagy, kronecker product and SVD approximations in image restoration, Linear Algebra and its Applications 284, (1998) , 177-192

[10] Jain, Anil K. (1989), Fundamentals of Digital Image Processing, Prentice Hall. [11] Kadison, Richard V.; Ringrose, John R. (1997), Fundamentals of the theory of operator algebras. Vol. I, Graduate Studies in Mathematics 15, Providence, R.I.: American Mathematical Society.

[12] Steeb, Willi-Hans, Matrix Calculus and Kronecker Product with Applications and C++ Programs, 1997, World Scientific Publishing.

[12a] Steeb, Willi-Hans, Matrix Calculus and Kronecker Product with Applications, 2011, World Scientific Publishing.