Subspace Codes for Random Networks Based on Plu¨cker coordinates and Schubert Cells Anirban Ghatak Department of ECE Indian Institute of Science, Bangalore, India-560012 Email: [email protected] Abstract—Theconstructionofcodesintheprojectivespacefor furtheremployedinconjunctionwithalexicographicordering 3 error control in random networks has been the focus of recent of subspaces ([4]). The above technique and its variants ([6]) 1 research. The Plu¨cker coordinate description of subspaces has have resulted in some of the best-known constant dimension 0 been discussed in the context of constant dimension codes, as 2 well astheSchubertcell descriptionof certain codeparameters. subspace codes till date. Other important constructions of n In this paper we use this classical tool to reformulate some constant dimension subspace codes include those presented a standard constructions of constant dimension codes and give a in [7], [8], [9]. J unifiedframework.Wepresentageneralmethodofconstructing Restriction of codeword subspaces to a particular Grassman- non-constantdimensionsubspacecodeswithrespecttominimum 9 nian yields constant dimension subspace codes; a natural subspace distance or minimum injection distance as the union 2 generalizationwastoattempttheconstructionofnon-constant of constant dimension subspace codes restricted to selected Schubert cells. The selection of these Schubert cells is based dimension subspace codes overthe entire projective space. In ] T on the subset distance of tuples corresponding to the Plu¨cker this context, a new metric, namely the injection distance, was I coordinate matrices associated with the subspaces contained in definedin[10],whichwasshowninsomecases(forinstance, . the respective Schubert cells. In this context, we show that a s aworst-caseadversarialmodel)togiveamoreprecisemeasure c recent construction of non-constant dimension Ferrers-diagram of the performance of a non-constant dimension code than [ rank-metric subspace codes is subsumed in our framework. the subspace distance metric. In [11], subspace codes (non- 2 I. INTRODUCTION constant dimension) were constructed for the injection metric v A linear multicast network coding model is defined as based on the methodsin [3]. Boundson the size of projective 2 follows: A source node transmits n packets, each an m-ary space codes, analogousto classical codingtheoreticones, and 6 symboloverF ; each nodein the networktransmits F -linear some explicit constructions of non-constant dimension codes 3 q q 6 combinationsoftheincomingpacketsatthatnode,andatany were presented in [12]. . destination node the received packets may be represented as Anewparadigminrandomnetworkcodingwasintroducedby 1 0 the rows of an N ×m matrix Y = AX +DZ, where the the so-called orbit codes [13], which generalizes some of the 3 matricesX,Z are ofrespectivedimensions n×m,t×m over resultsin[7]anddescribesconstantdimensionsubspacecodes 1 F . The matrix X has n source packets as its rows, A is the as orbits of the action of suitable subgroups of the general q : v transfer matrix of the network, the rows of Z are the error lineargrouponaGrassmannian.Subsequentwork([14],[15], i packets,whichmayberandomorintroducedbyanadversary, [16]) characterized orbit codes which result from the action X and D the transfer matrix with respect to these error packets. of cyclic subgroups. Recently, in [17], the classical tool of r a In non-coherent or random network coding, it is generally Plu¨cker coordinates was employed to compute the so-called assumed that the source and the destination nodes have no ‘Plu¨ckerorbits’associatedwithanorbitcodeandthedefining knowledge of the network topology - the only assumption equations of Schubert varieties were used in the description is that the number of error packets, i.e. the parameter t, is of constant dimension subspace codes. bounded. The contributions of this paper are listed below. Themotivationforconstructionofsubspacecodesforrandom (1) We have used some classical results ([18]) to provide a networks over the projective space came from the seminal unifiedframeworkforsomestandardconstructionsofsub- work of Ko¨tter and Kschischang [1], where the problem of space codes (e.g. in [1], [2], [3]) in terms of the Plu¨cker error-and-erasure correction in random linear networks was coordinatesoftheconstructedsubspaces.Inparticular,we translated into that of transmission and recovery of linear show that a strong correspondence exists: the generat- subspacesofaprojectivespace.Subsequentresearch([2], [5]) ing matrices in both the so-called ‘lifting construction’ constructed subspace codes, with subspace distance ([1]) as and its generalization correspond exactly to the matrix the metric, basedon rank-metriccodes.A verysignificantde- of the independent Plu¨cker coordinates with appropriate velopmentin theconstructionofconstantdimensionsubspace constraints on the choice of indices. codes was the use of Ferrers diagram representation of rank- metric codes which fit into a row-reduced echelon form rep- (2) We show that the characteristic representation of a Schu- resentation of subspaces in a Grassmannian ([3]), which was bert cell in a projective space in terms of the indices corresponding to a non-zero Plu¨cker coordinate can be B. The Plu¨cker Coordinates of Subspaces used as an “identifying vector”([3], [4], [11]) for sub- A d-dimensional linear space L in Pn is a set of points d spaces in that Schubert cell. In particular, we show that P =(p(0),··· ,p(n)) ∈ Pn whose coordinatesp(j) satisfy a both the subspace distance and injection distance among system of n−d independent linear equations: the subspacesin a projectivespace can be boundedusing n the subset distance among the defining index sets of α p(j) = 0; i=1,··· ,n−d. ij corresponding Schubert cells. j=0 X (3) We give a construction of non-constant dimension sub- There exist d + 1 points P = (p (0),··· ,p (n)) ;i = i i i space codes which is an extension of a construction 0, ··· , d, which span the d-space L . Consider the (d + d in [13] to the non-constant dimension case - for both 1)× (n +1) matrix P = (p (j)); denote as D(j ,··· ,j ) i 0 d subspace distance and injection distance. The codes are the determinant of the (d+1)×(d+1) submatrix formed describedasunionsofSchubertcellsselectedtoguarantee by those columns of P indexed by the sequence of integers a minimum subspace or injection distance. Moreover,we j ≤ ··· ≤ j , 0 ≤ j ≤ n. It follows that at least one of 0 d k show that the Khaleghi-Kschischang construction [11] the N+1 such determinants, where N := n+1 −1, must for non-constant dimension Ferrers-diagram rank-metric d+1 be non-zero. Written in a lexicographic order on the column subspacecodesissubsumedin ourproposedconstruction indices, (··· ,D(j ,··· ,j ),···) determines(cid:0)a p(cid:1)oint in PN; 0 d if the constant dimension subcode within each Schubert these determinants constitute the Plu¨cker coordinates of the cell is a suitable rank-metric code. d-dimensional linear (projective) subspace L . d Theorganisationofthispaperisasfollows:SectionIIisa re- Theorem 2.2: ([18], [21]) There exists a natural bijective vision of pertinent mathematicalresults regardingthe Plu¨cker correspondencebetween the d-spaces in Pn and the points of coordinate description of projective subspaces and Schubert PN whose Plu¨cker coordinatessatisfy the following relations: cells. In Section III, we reformulate standard subspace code d+1 constructionsin the language of Plu¨cker coordinatesand give (−1)αD(j0,··· ,jd−1,kα)D(k0,··· ,kˇα,··· ,kd+1)=0. a unified framework. The next section (Section IV) gives α=0 the general construction of non-constant dimension subspace X (1) codes. We conclude with a summary of the results and a where kˇα denotes omission of kα and 0≤jβ, kγ ≤n. discussion on the directions of ongoing and future research ItiseasilyverifiedthatthePlu¨ckercoordinatesofanyd-space in this context. inPn indeedsatisfytheaboverelations.Provingtheconverse, i.e. any point of PN whose coordinates satisfy the above relations corresponds to a unique d-plane in Pn, offers some II. PRELIMINARYCONCEPTS interesting insights [18]. For instance, it can be shown that any d-subspace of the projective space Pn has an associated In this section we briefly review and organize some math- matrix representation having the (d+1)×(d+1) identity ematical concepts which will be used in interpreting recent matrix as a sub-matrix. The following proposition spells this results in random network code constructions and lead to the out in detail. formulation of new coding schemes. Proposition 1: [18] There is a natural bijective correspon- dence between the set of points of PN whose coordinates A. Projective Space over a Field satisfy the quadraticrelationsin Theorem2.2, with the condi- tionD(k ,··· ,k )6=0,andtheaffine(d+1)(n−d)spaceof We begin with the definition of a projective space over an 0 d (d+1)×(n+1)matrices(p (j)), i=0,··· ,d, j =0,··· ,n, arbitrary field. i such that the (d + 1) × (d + 1) sub-matrix p (k ), i = Definition 2.1: [19] The projective space associated with a i α vector space V over a field K is the set P(V) of lines in V. 0,··· ,d, α = 0,··· ,d is the identity. Moreover, such a matrix p (j) correspondsto the point of PN with coordinates Moreover, the projective space P(V) is associated with the i D(j ,··· ,j )=Det(p (j ));suchapointof PN corresponds K-vector space V in the following canonical way [19]: 0 d i β to a (d+1)×(n+1) matrix with entries: Considertheequivalencerelation ∼onthesetV∗ :=V \{0}, where, for x,y ∈ V∗, x ∼ y if and only if y = ax, a ∈ pi(j)=D(k0,··· ,ki−1,j,ki+1,··· ,kd)/D(k0,··· ,kd). K, a 6= 0. Then the canonical map P : V∗ 7−→ P(V) (2) associates each vector x with the projective point Kx. It follows from the proposition that the set of points of PN, The dimension of P(V) is defined as Dim(P(V)) := whose co-ordinates satisfy the quadratic relations, is covered Dim(V) − 1; P(0) is empty, and its dimension is −1. by N + 1 copies of the affine (d + 1)(n − d) space; this Therefore, a d-dimensional subspace or a d-space of the n- is the Grassmann manifold or the Grassmannian denoted by dimensionalprojectivespace Pn(K) overa field K is a setof G(d+1,n+1). In the next section we show that the above pointswhoserepresentativevectorsalongwiththezerovector, proposition is the key to the interpretation of some well- form a d+1-dimensional subspace of the n+1-dimensional knownconstructionsforconstantdimensioncodesforrandom K-vector space V, where Pn(K)=P(V) [20]. networks in terms of the Plu¨cker coordinates of subspaces. Henceforth we term the matrix (p (j)) as the Plu¨cker coordi- Definition 3.1: [2] Let X be an l×m matrix over F . We i q nate matrix of the subspace whose representation as a point define the ‘lifting’ map as follows: of PN is given by: (··· ,p (j),···) in lexicographic order. i I:Fl×m −→G(l,l+m); X 7−→I(X)=h[I X]i q C. Schubert Cells and their Representation where I denotes the l ×l identity matrix and hMi denotes Let G(d,n) be the Grassmannian of d-dimensional sub- the row-space of the matrix M. The subspace I(X) is called spacesof an n-dimensionalK-vectorspace. For U ∈G(d,n), a ‘lifting’ of the matrix X. thereexistsauniquelydeterminedbasis{v1,v2,··· ,vd}, vi ∈ IfC isarank-metriccode(e.g.[23]),thesubspacecodeI(C)= Kn, of the form: {I(C)|C ∈C} is called the ‘lifting’ of C. v =(∗,··· ,∗,1,0,··· ,0), A generalization of the lifting construction was achieved in 1 v =(∗,··· ,∗,0,∗,··· ,∗,1,0,··· ,0), [3] and [11], wherematricesin the row reducedechelonform 2 . . (RREF) have been used in conjunction with Ferrers diagram . representation of rank-metric codes to construct the desired v =(∗,··· ,∗,0,∗,··· ,∗,0,∗,··· ,∗,1,0,··· ,0). d subspaces.Wenowbrieflystatethisconstructionthroughsome (d−1)zeros relevant definitions ([3], [11]). The subspace U is uniquely determined by the placement | {z } Definition 3.2: A Ferrers diagram [22] is a graphical of the terminal 1’s in the v ’s, indexed by a sequence of i representation of integer partitions as an array of dots such integers 1 ≤ b < b < ··· < b ≤ n, and the elements 1 2 d that the i-th row has exactly the same number of dots as the of K which occupy the ∗ positions. All the subspaces which i-thelementinthepartition.Anm×n Ferrersdiagramhas m are parametrized by a particular tuple β = (b ,b ,··· ,b ), 1 2 d dots in the rightmost column and n dots in the topmost row. denotingthepositionsofthetrailing1’sinabasisoftheabove Definition 3.3: Let F be a l×m Ferrers diagram. A rank- form,aresaidtoconstitutetheSchubertCellS ofdimension: β metric code C is called a Ferrers diagram rank-metric code d (FDRMcode)associatedwithF ifallthecodewordsarel×m d+1 d =(b −1)+···+(b −d)= b − . matrices in which the positions of the non-zero entries match β 1 d i 2 i=1 (cid:18) (cid:19) thoseofthedotsinF.Inthiscase,thecodewordmatricesare X said to fit the Ferrers diagram F. AnalternativedescriptionofSchubertcellsinvolvessubspaces Definition 3.4: Let v ∈{0,1}n+1 be a vector of weight k. whose basis vectors form a d × n matrix in row reduced The profile matrix of v, denoted P , is a k×(n+1) matrix echelonform,with the definingd-tuple indexingthe positions v in RREF such that: of the leading 1’s instead. The Schubert cell constituted of such subspaces, say, S with α=(a ,a ,··· ,a ) and the a (i) The leading coefficients of the rows of P appear in the α 1 2 d i v satisfy 1≤a <a <···<a ≤n, has dimension: columns indexed by supp(v). 1 2 d (ii) The entries in P other than 0’s and 1’s are dots. d = (n−a −(d−1))+···+(n−a ) v α 1 d Extracting the dots of P forms a Ferrers diagram which is d v d = nd− a − . denoted by F(v). i 2 Definition 3.5: Let v ∈ {0,1}n+1 be a vector of weight i=1 (cid:18) (cid:19) X k and let X ∈ Fl×m, l ≤ k, m ≤ n−k +1, be a matrix Inboththecases,themaximumdimensionn(n−d)isachieved q which fits F(v). Then the (generalized)lifting of X (by v) is for the tuple (1,··· ,d) and the minimum dimension is 0 for defined as I (X)=hP (X)i⊆G(l,l+m), where P (X) is v v v the tuple (n−d+1,··· ,n). The second characterization of the profile matrix of v with the dots replaced by the entries Schubert cells is relevant to the construction of non-constant of X. dimension subspace codes for random networks. As before, if C is a rank-metric code, then I (C) = v {I (C)|C ∈C} is called the generalized‘lifting’of C, which III. A UNIFIEDFRAMEWORK FORLIFTING v is an FDRM subspace code associated with F(v). CONSTRUCTIONS USINGPLU¨CKERCOORDINATES In this section, we reformulate two well-known methods B. Unified Framework Using Plu¨cker Coordinates forconstructingconstantdimensionalsubspacecodesinterms Tointerprettheaboveconstructionsofconstantdimensional of the Plu¨cker coordinates of certain subspaces, providing a subspace codes in terms of the Plu¨cker coordinates of the unified framework for both of them. associated subspaces, we provethe followingpropositionas a consequence of Theorem 2.2 and Proposition 1. A. Overview of the Lifting Constructions Proposition 2: If X ∈ F(d+1)×(n+1) is in row-reduced q The Lifting Construction was first proposed in [1] to yield echelon form (RREF), then the Plu¨cker coordinate matrix constant dimensional Reed-Solomon-like codes for error and associated with the subspacehXi⊆G(d+1,n+1)coincides erasure correction in random networks. In a subsequent work with X, upto multiplication by a unit. [2], constant dimension subspace codes were constructed by Proof: Suppose the columns k ,k ,··· ,k of X contain the 0 1 d ‘lifting’ rank-metric codes as follows. leading elements of the rows. As D(k ,··· ,k ) 6= 0, by 0 d Proposition 1, hXi is identified with a point of PN, where is: N= n+1 −1, corresponding to a (d+1)×(n+1) matrix 1 0 1 0 1 0 d+1 with entries: 0 1 1 0 1 0 (cid:0) (cid:1) Iv(X)=0 0 0 1 1 0 pi(j)=D(k0,··· ,ki−1,j,ki+1,··· ,kd)/D(k0,··· ,kd), 0 0 0 0 0 1 Thepositionsofthenon-zeroentriesinv areindexedas:k = which is the associated Plu¨cker coordinate matrix. 0 0,k =1,k =3,k =5. Hence the matrix corresponding to Now we give the following interpretations of the lifting 1 2 3 thePlu¨ckercoordinateD(0,1,3,5)isthe4×4identitymatrix construction and its generalization in terms of the Plu¨cker and so, D(0,1,3,5) = 1. As k−l = 1 and m = n−k+1, coordinates. boththematrixX′ andtheaugmentedmatrixA(X)aregiven Lifting Construction: In the setting of Proposition 2, set by: d = l −1,n = l +m−1 and 0 ≤ k < ··· < k ≤ n, 0 d 1 1 with D(k ,··· ,k ) 6= 0. Let X ∈ Fl×m, for instance, a 0 d q 1 1 codewordmatrixofanl×mrank-metriccodeoverFq.Selecta X′ =A(X)=0 1 subspaceofG(d+1,n+1)whosePlu¨ckercoordinatessatisfy 0 0 the following conditions: k = i,i = 0,··· ,d, i.e. we have i D(0,··· ,l −1) = 1 and all other coordinates are obtained Consider the four coordinates: D(2,1,3,5),D(0,2,3,5), from those of the form D(j ,··· ,j ), 0 ≤ j ≤ n, at most D(0,1,2,5),D(0,1,3,2),involvingthe first column of A(X) 0 d β one of the j ∈/ {0,··· ,l −1}, and that particular column arising out of the column of P with index 2. We have: β v is from the matrix X in ascending order. This subspace is 1 0 0 0 1 1 0 0 precisely the lifted subspace I(X)=h[I X]i. 1 1 0 0 0 1 0 0 GeneralizedLiftingConstruction:Thegeneralizedliftingcon- Det =1, Det =1 0 0 1 0 0 0 1 0 struction may similarly be described as the selection of sub- 0 0 0 1 0 0 0 1 spaceswithparticularconstraintsontheirPlu¨ckercoordinates. , Given a vector v as described, set k = d + 1 and let k , i=0,··· ,d, 0≤k ≤n,denotethepositionsofnon-zero 1 0 1 0 1 0 0 1 i i componentsofv inascendingorder.If X ∈Fl×m beamatrix 0 1 1 0 0 1 0 1 q Det =0, Det =0. which fits F(v), where l ≤ k, m ≤ n−k+1, (e.g. X is a 0 0 0 0 0 0 1 0 codewordofaFDRMcodeassociatedwiththeFerrersdiagram 0 0 0 1 0 0 0 0 F(v)) construct the augmented matrix A(X)∈Fk×n−k+1 in q which is in accordance with Proposition 2; and the Plu¨cker two steps as follows: coordinate matrix coincides with the lifted matrix I (X). v (1) To each column of X append k−l zeros from below to It follows that both the lifting construction and its general- form X′ ∈Fk×m. ization have precise interpretations in terms of the Plu¨cker q (2) Add n−k−m+1 all-zerocolumnsamongthe columns coordinatedescription of subspaces. In fact this interpretation ofX′preservingtheorderofthosecolumnsin P toform provides a unified framework as the encoding procedure is v A(X). shown to be equivalentto a choice of a set of column indices from the underlying (codeword) matrices and computing the SelectasubspaceofG(d+1,n+1)whosePlu¨ckercoordinates Plu¨cker coordinates for those particular indices. In the lifting satisfy the following constraints: D(k ,··· ,k ) = 1 and the 0 d construction the ‘pivotal’ set of indices {k ,··· ,k } was remaining coordinates are obtained from those of the form 0 d constrained to be just the set {0,··· ,d}. Hence it actually D(j ,··· ,j ), 0 ≤ j ≤ n, where at most one of the 0 d β limits the choice of subspaces to the so-called principal j ∈/ {k ,··· ,k }, and that particular column is from the β 0 d Schubert cell, with a corresponding matrix representation augmented matrix A(X) in ascending order. This subspace is having an identity matrix in the leftmost d+1 columns. In the lifted subspace I (X)=hP (X)i. v v the generalized lifting construction, the (d + 1) × (d + 1) We now give a simple example illustrating the Plu¨cker coor- identity sub-matrix is interspersed with columns from the dinate description of the generalized lifting construction. underlying (codeword) matrix; hence, other Schubert cells Example Let v = 110101∈ {0,1}6, i.e. n = 5,k = 4. Then are also exploited. But, in both cases the lifted matrix still the profile matrix of v and the associated Ferrers diagram are coincides with the (d+1)×(n+1) matrix formed by the given by: ‘independent’ Plu¨cker coordinates in the setting of Theorem 2.2 and Proposition 1, as has been outlined earlier. 1 0 • 0 • 0 • • 0 1 • 0 • 0 Pv =0 0 0 1 • 0 F(v)=• • IV. A CONSTRUCTION OFNON-CONSTANTDIMENSIONAL • SUBSPACE CODES 0 0 0 0 0 1 The construction of non-constant dimensional subspace In this case, l =3<k =4, but m=2=n−k+1. Now if codes has been addressed in [3] and [11]: in the former, X ∈F3×2 isamatrixwhich‘fits’ F(v),thentheliftedmatrix by ‘puncturing’ a constant dimension code that results from 2 a multilevel code construction; in the latter, by partitioning (1) Selection of Schubert cells whose constituent subspaces the projective space followed by a two-level selection of across the cells are at a prescribed minimum subspace subspaces. In this section we present a construction of non- distance from each other; constant dimension subspace codes which is an extension of (2) Selection of subsets of subspaces in each Schubert cell the construction of orbit codes for random networks [13] for which have the prescribed minimum subspace distance constant dimension codes. The Khaleghi-Kschischang con- among them. struction [11] also employs the so-called injection distance In [11], the selection of subsets within a Schubert cell is metric,ratherthanthesubspacedistancemetric,astheformer achieved via lifting of Ferrers diagram rank-metric (FDRM) sometimes outperforms the latter in a worst-case adversarial codes, while a profile vector selection algorithm is imple- model. So we first outline our construction for subspace mented to choose appropriate Schubert cells which also sup- distance and then discuss the modifications required with portthelargestliftedFDRMcodes.Asdiscussedinpreceding injection distance as the metric. sections, Schubert cells can be characterized by the tuples which indexthe positionsof leading1’s of a matrix in RREF. V. CODES WITHSUBSPACE DISTANCE ASMETRIC Following Proposition 3, the selection of a Schubert cell is First we describe the selection of subspaces in terms of translated into the selection of a tuple of column indices in theirPlu¨ckercoordinates.RecallfromProposition1that every the Plu¨cker coordinate matrix of a representative subspace. d-space in the projective n-space has a non-zero Plu¨cker For the selection of subspaces within a Schubert cell we coordinate D(k ,··· ,k ) such that the (d + 1) × (d + 1) propose a method exploiting the characteristic representation 0 d submatrix of the coordinate matrix formed by the columns ofsubspacesinthe Schubertcell.First westate some relevant indexed by k , i=0,··· ,d, is the identity. Let U, V be two definitions. i different subspaces, of possibly different dimensions d , d , Definition 5.1: Let the set of basis vectors of any subspace u v with the tuples {ku} and {lv} as the column indices of the of a Schubert cell Sα have a d×n matrix representation in identity submatrix in each case. Then the symmetric distance RREF with the positions of the leading 1’s indexed by the between the tuples {ku} and {lv} givesa lowerboundon the tuple {α} = {a1,a2,··· ,ad}. The Schubert cell matrix Aα subspace distance ds(U,V) as follows. corresponding to the Schubert cell Sα is the d×n matrix in Proposition 3: Let U, V be subspaces of Pn of projective RREF in which the columnsindexedby the tuple {α} form a dimensionssandtrespectively;andlet{k ,k ,··· ,k }, 0≤ d×d identity matrix and all other columns are zero. 0 1 s kα ≤ n, and {l0,l1,··· ,lt}, 0 ≤ lβ ≤ n, be the tuples Definition 5.2: Given a Schubert cell Sα characterized by specifying the identity submatrices of the respective Plu¨cker the tuple {α} = {a1,a2, ··· ,ad}, 1 ≤ a1 < ··· < coordinate matrices. Then ds(U,V) ≥ ∆({kα},{lβ}), where ad ≤n, the complementary Schubert cell Sαc is characterized ∆({u},{v}) is the symmetric distance between the sets by the tuple {β} = {b1,··· ,b(n−d)} = {1,2,··· ,n} \ {u}, {v}. {a1,a2,··· ,ad}. Proof: Without loss of generality assume that the Plu¨cker Evidently the Schubert cell matrix for Sαc is given by the coordinate matrices M(U), M(V), of respective dimensions (n−d)×n matrix Acα in RREF with the columns indexed (s+1)×(n+1)and (t+1)×(n+1),are in RREF, keeping by {β} forming an (n−d)×(n−d) identity matrix and the the columns of the identity submatrices intact. Now remaining columns all zero. Let S be a Schubert cell with the associated d×n Schubert α ∆({kα},{lβ})=|{kα}\{lβ}|+|{lβ}\{kα}|. cell matrix Aα. Also denote the d×n RREF ‘matrix’ rep- resentation of the S with asterisks indicating arbitrary field α Let a := |{k }\{l }|, b := |{l }\{k }|. Then there are α β β α elements (vide Subsection II-C) as M(S ). α s + 1 − a = t + 1 − b matching positions in the matrices Consider the set of d × (n − d) matrices such that the M(U), M(V) where there are leading 1’s; hence we have columns of these matrices correspond to those columns in- Dim(U ∩V)≤(1/2)(s+t−(a+b)). But then dexed by the set: β = {b1,b2,··· ,bn−d} := {1,2,··· ,n}\ {a ,a ,··· ,a }inascendingorder,havingnon-zeroelements d (U,V) := DimU +DimV −2Dim(U ∩V) 1 2 d s only in the positions of the asterisks in M(S ). Select an α ≥ s+t−2((1/2)(s+t−(a+b))) additive group G of these matrices such that each element of = a+b. the group has rank at least r/2. For each G∈G construct the d×n matrix C(G) as follows: TheaboveresultisageneralreformulationofLemma2in[3], C(G)=Aα+G·Acα. which relates the subspace distance between the subspaces of The constantdimension subspace code constructedwithin the a projective space with the Hamming distance between their Schubert cell S is then given by: α binary profile vectors. C ={hC(G)i|G∈G}. Following [11], a two-step strategy for constructing non- α constant dimension subspace codes over a projective space The minimum subspace distance of the code satisfies d ≥ min is outlined below: r, as borne out in the following theorem. Theorem 5.3: Let G , G belong to an additive group of Definition 6.1: [10] The injection distance between sub- 1 2 d × (n − d) matrices with minimum rank r/2. Then the spaces U, V of Pn is given by: subspaces spanned by the matrices C(G ) and C(G ) within 1 2 d (U,V):=max{DimU,DimV}−Dim(U ∩V) the Schubert cell S are at a minimum subspace distance I α d ≥r, where C(G ), i=1,2, are defined as above. =Dim(U +V)−min{DimU,DimV}. min i Proof: The d×n matrices C = C(G ), i = 1,2, have the i i The injection distance between two subspaces is thus related matricesG , G ,whichbelongtotheadditivegroupandhave 1 2 to the subspace distance between them as follows ([10]): minimumrankr/2, asthed×(n−d)submatricesconstituted 1 1 of the columns indexed by b1,b2,··· ,bn−d. Then we have: dI(U,V)= ds(U,V)+ |DimU −DimV| (3) 2 2 d (C , C )=2rank C1 −2d=2rank C1−C2 −2d≥r. It is evident that, upto a scaling factor, the injection distance S 1 2 C2 C2 coincides with the subspace distance for constant-dimension (cid:20) (cid:21) (cid:20) (cid:21) subspace codes. It was observed in [10] that the injection The last step follows from the fact that the non-zero columns distance gives a more precise measure of the subspace code of the difference matrix C −C constitute an element of the 1 2 performance when a single error packet can cause simultane- additive group of matrices of minimum rank r/2. ous error and erasure. However, in a less pessimistic channel Hence we can describe our two-step construction of non- modelthantheworst-caseadversarialchannel,forinstance,in constantdimensionsubspacecodeswitha minimumsubspace the case where a single error packet can introduce either an distance d as follows. min error or an erasure but not both, the subspace distance is still (1) Selection of Schubert cells: Select the largest subset S of appropriate. Schubertcells subjectto the followingcondition:Forany B. ConstructionofNon-constantDimensionSubspaceCodes: S ,S ∈S, we have ∆({α},{β})≥d ; α β min First we prove a counterpart of Proposition 3, with the (2) Selection of subspaces in each Schubert cell: The sub- same setting and notation, for the injection distance d (U,V) spaces within a Schubert cell are chosen such that their I between subspaces U, V of Pn. spanning matrices satisfy the following condition: The Proposition 4: Let U, V be subspaces of Pn of projective submatrix indexed by the characteristic columns of the dimensionssandtrespectively;andlet{k ,k ,··· ,k }, 0≤ complementary Schubert cell matrix belongs to an addi- 0 1 s k ≤ n, and {l ,l ,··· ,l }, 0 ≤ l ≤ n, be the tuples tive group of minimum rank ⌊d /2⌋. α 0 1 t β min specifying the identity submatrices of the respective Plu¨cker The non-constant dimension subspace code is, therefore, the coordinate matrices. Then d (U,V) ≥ ⌊1∆({k },{l })⌋, union: C , where the union is over all the Schubert cells I 2 α β α where∆({u},{v})isthesymmetricdistancebetweenthesets S ∈S. α {u}, {v}. S Proof: From the proof of Proposition 3, we obtain: VI. CODES WITHINJECTIONDISTANCE ASMETRIC Dim(U∩V)≤(1/2)(s+t−(a+b)),wherea:=|{k }\{l }| α β and b:=|{l }\{k }|. So we have: β α Inthissectionwefirstintroducethenotionofinjectiondis- tance anddiscussits applicabilityvis-a`-vissubspacedistance. 2d (U,V) := 2max{DimU,DimV}−2Dim(U ∩V) I Next we modify the construction of non-constant dimension ≥ DimU +DimV −2Dim(U ∩V) subspace codes presented earlier to adapt it for the injection ≥ a+b. distance metric. As ∆({k },{l })=a+b, the proposition follows. α β A. The Adversarial Channel Model and Injection Distance Alternatively, from Equation 3 and Proposition 3, we have: between Subspaces: 2d (U,V)≥∆({k },{l })+|s−t| I α β Recallthatinthelinearmulticastnetworkcodingmodelthe and the proposition follows. received packets at a destination are expressed as the rows of As suggested above, to capture the effect of the injection a matrix: Y =AX+DZ, where the matrix A is the network distancemetricin ourframework,we modifythedefinitionof transfermatrixseen bythesourcepackets(rowsof X)andD the symmetric distance between two sets. isthetransfermatrixseenbytheerrorpackets(rowsof Z).In Definition 6.2: For two finite sets S ,S of respective the general random network coding problem, the matrix A is 1 2 cardinalitiess ,s ,themodifiedsymmetricdistanceisdefined unknownto the receiverexceptfora lowerboundon the rank 1 2 as follows: ofthematrix.Intheworst-caseadversarialmodelinvestigated in[10],thefollowingadditionalassumptionsaremade: ∆ (S ,S )=#{S \S }+#{S \S }+|s −s | m 1 2 1 2 2 1 1 2 (1) The adversary knows the matrix X and can choose the where#S denotesthecardinalityofthesetS and‘||’denotes transfer matrix A while respecting the rank constraint; the absolute value. (2) The matrix D is arbitrarily chosen by the adversary. Proposition5:Themodifiedsymmetricdistanceisametric. Proof: Itisevidentthat∆ (S ,S )=∆(S ,S )+|s −s |. where δ is the minimum rank distance of C. m 1 2 1 2 1 2 r Symmetry and the fact that ∆ (S ,S ) ≥ 0 with equality if Define V := F d×(n−d). The kernel of the surjective map: m 1 2 q and only if S = S follow immediately. Triangle inequality π : V → V/C is a subspace of dimension d(n −d) − D. 1 2 is satisfied because ∆() is a metric and for positive integers The largest FDRM code which matches M(S ) is the largest α s ,s ,s we have: |s −s |+|s −s |≥|s −s |. subcode C′ ⊆C subject to the constraint that all its codeword 1 2 3 1 2 1 3 2 3 Nowwereformulatethetwo-stepconstructionofnon-constant matriceshavezerosin d(n−d)−D fixedpositions.Invoking α dimension subspace codes with injection distance as the the standard vector-space isomorphism between F m×n and q metric. From Propositions 4 and 5, the selection criterion F mn where m,n are positive integers, any codeword c ∈C′ q of Schubert cells for the injection distance metric is read- may be expressed as a d(n−d)-dimensional vector over F q ily obtained from the subspace distance case. Moreover, as with non-zero elements in D coordinates i ,··· ,i . Let α 1 Dα the second step is essentially a construction of a constant- φ:V →F d(n−d)−Dα be the surjective map which sets zero q dimensionsubspacecodewithinaSchubertcell,thetransition the coordinates i ,··· ,i . Now consider the linear map: 1 Dα from subspace distance to injection distance involves only a Φ:V −→V/C×F d(n−d)−Dα; v 7−→(π(v),φ(v)). scaling factor. Hence we may state the construction of non- q constantdimensionsubspace codeswith a minimuminjection Evidently, DimΦ(V) ≤ 2d(n −d)−D −D . The largest α distance δmin as follows. FDRM code is the kernel of the map Φ, and hence, by the (1) Selection of Schubert cells: Select the largest subset S of rank-nullity theorem, has dimension: Schubertcells subjectto the followingcondition:Forany κ≥d(n−d)−(2d(n−d)−D−D ) α S ,S ∈S, ∆ ({α},{β})≥2δ ; α β m min =D +max{d,n−d}(min{d,n−d}−δ +1)−d(n−d) α r (2) Selection of subspaces in each Schubert cell: The sub- =D −max{d,n−d}(δ −1). α r spaces within a Schubert cell are chosen such that their spanning matrices satisfy the following condition: The submatrix indexed by the characteristic columns of the A. Selection of Schubert Cells: complementary Schubert cell matrix is an element of an It follows from Theorem 7.1 that the size of the FDRM additive group of minimum rank δ . min non-constant dimension subspace codes depends directly on Thenon-constantdimensionsubspacecodeistheunion: Cα, the dimension of the chosen Schubert cells. In [11] the over all the Schubert cells Sα ∈S, as defined earlier. relation between the so-called profile vectors associated with S each Schubert cell and the dimension of the FDRM codes VII. NON-CONSTANTDIMENSION FERRERS-DIAGRAM constructedhasbeen used in an algorithmforselection of the RANK-METRICSUBSPACE CODES Schubertcells. We presenta moredirectapproach,basingour In the proposed construction of non-constant dimension choice on the dimension of each Schubert cell, as follows. subspace codes, if the additive group G of d × (n − d) Recall from Subsection II-C that the dimension of a Schubert matrices is chosen to be a rank-metric code [23], we have a cellS inthesecondrepresentationcharacterizedbythetuple α so-callednon-constantdimensionFerrers-diagramrank-metric {α} is given by: D = nd(α)− d(α)a − d(α) , where code (FDRM) subspace code. We show that the Khaleghi- α i=1 i 2 d(α) := |{α}| = d is the length of the tuple. Let A denote Kschischang construction [11], which has resulted in some P (cid:0) (cid:1) theset ofallsuch distincttuplesα inlexicographicorder.For of the best-known subspace codes in terms of code rate, is each α∈A define the choice function: subsumed in our framework when the additive group G is a rank-metric code. F(S )=D −max{d(α),n−d(α)}(δ −1). (4) α α r In the proposed construction, the FDRM codeword matrices The selection of Schubert cells proceeds via a greedy algo- havenon-zeroelementsinpositionswhichmatchtheasterisks rithm, similar to that in [11], which selects at each stage the in the ‘matrix’ M(S ), where S is the Schubertcell charac- α α tuple {α} with the largest F(S ) and discards all the tuples terizedbythe tuple {α}={a ,a , ··· ,a }.First we provea α 1 2 d β ∈ A such that ∆({α},{β}) < d for the subspace dis- lowerboundonthedimensionofsuchanFDRMcode,similar min tancemetricor⌊1∆({α},{β})⌋+⌊1|#{α}−#{β}|⌋<δ to Theorem 3 in [11], in this general setting. 2 2 min for the injection distance metric. The next stage proceeds Theorem 7.1: The dimension κ of the largest FDRM code by applying the same choice function to the modified set whose codewords have non-zero elements precisely in the A:=A\B, where B is the set of the discarded tuples {β} at positions of the asterisks in M(S ) satisfies: κ ≥ D − α α the previous stage. The selection algorithm terminates when max{d,n−d}(δ −1),whereδ istheminimumrankdistance r r A:=∅.ThechoicefunctioninEquation4,whichis basedon of the rank-metric code and D , the number of asterisks in α thelowerboundonthedimensionoftheFDRMsubspacecode M(S ), is the dimension of the Schubert cell S . α α given in Theorem 7.1, ensures that at every stage a Schubert Proof: It is known [23] that the dimension of a linear cell which supports the largest FDRM code is chosen among maximumrank-distance(MRD)codeC,whosecodewordsare the contendingcells. The largest numberof subspaces belong d×(n−d) matrices over F , is given by: q to the principal Schubert cell of the Grassmannian G(d,n), D =max{d,n−d}(min{d,n−d}−δ +1) with d = ⌊n/2⌋ or d = ⌈n/2⌉; for these values of d, the r negative term in the choice function is also minimum for a q dI ds N LB(ds) KK(ds) LB(dI) KK(dI)) 2 2 4 9 15.1515 15.1518 15.3238 15.6245 given δr. Hence the algorithm proceedsby choosing either of 2 2 4 10 20.1534 20.1534 20.1967 20.3294 these subspaces (different when n odd) in the first step. 2 2 4 12 30.1556 30.1557 30.1998 30.3346 2 3 6 13 28.0030 28.0030 28.0134 28.0263 3 2 4 7 8.0131 8.0145 8.0464 8.1331 B. Comparison with the Khaleghi-Kschischang algorithm: 3 2 4 8 12.0135 12.0135 12.0160 12.0311 The above algorithm and that stated in [11] differ in 4 2 4 7 8.0030 8.0032 8.0142 8.0522 4 2 4 8 12.0031 12.0031 12.0034 12.0068 two aspects, one of which can be explained as a general TABLEI reformulation and the other, a slight modification. COMPARISONOFLOWERBOUNDSONTHERATESOFOURFDRM (i) Binary Profile Vectors vs. Characteristic Tuples: In [11], SUBSPACECODESWITHTHEACTUALRATESOFCODESIN[11] the selection of Schubert cells is based on the so-called binary profile vectors associated with each cell. The Ferrersdiagramrepresentationof eachprofile vectorthen It is seen that the calculated lower bound is not simply gives the dimension and the positions of non-zeroentries a computation of the choice function in each case and then of the appropriate rank-metric code. In our construction summing the values in the exponent.The justification for this we simply use the characteristic tuple representation of modification is as follows: The choice function F(S ) for α each Schubert cell and the rank-metric code is likewise a Schubert cell S ∈ S is zero or negative when it does α specifiedbytheso-calledSchubertcell“matrix”(withthe not support a rank-metric code of the given dimensions with asterisks in the “basis vectors” replacing the dots in the theprescribedminimumrank-distance.However,theparticular Ferrers diagrams) - a direct formulation. Schubert cell still maintains the inter-cell minimum subspace (ii) The Choice Function for the Greedy Algorithm: The or injection distance and so contributes at least one subspace following scoring function is used to select a new profile codeword to the entire code. To give a simple illustration, vector v in [11]: consider the following construction of a non-constant dimen- sionFDRMsubspacecodeinP4(F )withminimumsubspace n i 2 score(v,δ )= v¯v −max{wt(v),η(v)}(δ −1), distance 4. A set S for this code consists of Schubert cells r i j r specified by the tuples {1,2},{3,4},{1,2,3,4,5}, with cor- i=1j=1 XX (5) responding choice function values of 3,−1,−5, respectively. where η(v) = n−(wt(v)+mint∈s(v)t)+1, wt(v) the ForthesecondtupleM(Sα)hasonly2asteriskswhilethelast (Hamming) weight of the binary vector v ∈ {0,1}n, one has none, being the 5×5 identity matrix, obviously not v¯ the complement of v and s(v) the support of v, enoughto supportthe requiredrank-metriccodein each case. i.e. the set whose elements denote the positions of the However,theycanstill contributeasingle subspacecodeword 1’s in v. It is evident that the sum term in the above each. scoring function computes D , the number of asterisks InTable1wehavegivenafewlowerboundsonthecoderates α in the ‘matrix’ representation of the chosen Schubert in our construction with respect to given minimum subspace cell in our framework, which, in our choice function distance (column LB(d )) and minimum injection distance s (Equation 4), is the exact formula based on the char- (column LB(d )) for some small field sizes q and projective I acteristic tuple. Rewriting the second term in Equation dimensions N −1. It is seen that they are comparable with 4 as max{d(α),n − (d(α) + min {a }) + 1}(δ − 1) those of the actual codes constructed in [11], which are also i i r makes it equivalent to the second term of Equation 5. presented in the columns ‘ KK(d )’ (minimum injection dis- I It modifies the parameter n − d(α) by subtracting the tance)and‘KK(d )’(minimumsubspacedistance).Infact,for s numberof leftmost all-zero columns in M(S ), which is subspace distance the rates of the constructed codes actually α not necessary for making a choice in our framework. coincide with the computed lower bound in half of the cases, while the other values show only marginalimprovement.The C. Numerical Results: difference between the rates of the constructed codes and Wecomputelowerboundsontheratesofcodesconstructed the computed lower bounds is more pronounced for injection for a given minimum subspace distance (using symmetric distance. distance) and for a given minimum injection distance (using the modified symmetric distance), from the values of the VIII. CONCLUSION choice function (Equation 4) for each selected Schubert We have established a unified framework for the construc- cell. For a selected set of Schubert cells S, a lower bound tion of constant-dimensionrandomnetworkcodes in terms of on the code rate of a non-constant dimension FDRM theclassicaldescriptionofthesubspacesofa projectivespace subspace code over a finite field Fq is given by the formula: usingthePlu¨ckercoordinates.Furtherwehavegivenageneral logq( Sα∈SqF(Sα)) where constructionofnon-constantdimensionsubspacecodes,which exploits a basic representation of Schubert systems, for both P F(Sα), if F(Sα)≥0 the subspace distance and the injection distance metrics. We F(S )= α (0 otherwise. havedemonstratedthatourframeworksubsumestheconstruc- tion of [11], which is corroborated by the proximity of the [4] N. Silberstein and T. Etzion, “Enumerative Coding for Grassmannian estimated lower bounds for our construction with the actual Space”, IEEETrans.Information Theory,57(2011),365–374. [5] M.GadouleauandZ.Yan,“Constant-rank codesandtheirconnectionto coderates givenforsmallparametersin [11]. Somequestions constant-dimension codes”, IEEETrans.Information Theory,56(2010), regarding the above constructions remain open: 3207–3216. [6] A-L.Trautmann and J. Rosenthal, “New improvements on the echelon- (i) For the non-constant dimension FDRM subspace codes, Ferrersconstruction”, Proc.MTNS,Budapest (2010),5–9. will an exhaustive search algorithm yield a significantly [7] A.KohnertandS.Kurz,“Constructionoflargeconstantdimensioncodes better code-rate than the lower bound based on a greedy with a prescribed minimum distance”, in MMICS, vol. 5393 of Lecture NotesinComputer Science, Springer(2008),31–42. search? Is there a systematic search procedure to achieve [8] F. Manganiello, E. Gorla and J. Rosenthal, “Spread codes and spread the best possible code-rate? decoding in network coding”, Proc. IEEE International Symposium on InformationTheory,Toronto(2008),851–855. (ii) Is it possible to obtain a set (or better still, a group) of [9] M. Gadouleau and A. Goupil. “A matroid framework for noncoherent subspaces within a selected Schubert cell larger than the random network communications”, IEEETrans.Information Theory, 57 largest Ferrers-diagram rank-metric code, which satisfies (2011),1031–1045. [10] D. Silva and F. R. Kschischang, “On metrics for error correction in the minimum subspace/injectiondistance condition? This networkcoding”,IEEETrans.InformationTheory,55(2009),5479-5490. problemis a constant-dimensionsubspace code construc- [11] A. Khaleghi and F. R. Kschischang, “Projective space codes for the tion problem with the added constraint of the Schubert injectionmetric”,Proc.11thCanadianWorkshoponInformationTheory, Ottawa(2009),9–12. cell structure. [12] T. Etzion and A. Vardy, “Error-correcting codes in projective space”, So far there have been several approaches to the decoding IEEETrans.Information Theory,57(2011),1165–1173. [13] A-L.Trautmann,F.ManganielloandJ.Rosenthal,“Orbitcodes-anew problem of constant dimension codes in general ([2]), of conceptintheareaofnetworkcoding”, Proc.IEEEInformation Theory spread codes in particular,([24]) and employinglist decoding Workshop,Dublin(2010),1–4. methods ([25], [26], [27]). Decoding of ‘punctured’ non- [14] A-L. Trautmann and J. Rosenthal, “A complete characterization of irreduciblecyclicorbitcodes”, Proc.WCC2011-7thIntern’lWorkshop constant dimension codes is discussed in [3]. A possible onCodingandCrypto.(2011),219–228. decoding strategy of our non-constant dimension subspace [15] F. Manganiello, A-L. Trautmann and J. Rosenthal, “On conjugacy codes and those in [11] involves two steps: classes of subgroups of the general linear group and cyclic orbit codes”,Proc.IEEEISIT,(2011),1916–1920. (1) Estimating the Schubert cell to which the transmitted [16] A-L. Trautmann, F. Manganiello, Michael Braun and J. Rosenthal, subspace belongs. In [11], so-called binary asymmetric “Cyclic orbitcodes”, Online, arXiv.1112.1238v1. [17] J. Rosenthal and A-L. Trautmann, “A complete characterization of codeswith minimumasymmetricdistance greaterthan or irreducible cyclic orbit codes and their Plu¨cker embedding”, Designs, equaltotheminimuminjectiondistanceandbinarycodes CodesandCrypto.(Special IssueonCodingandCryptography) (2012). with minimum Hamming distance greater than or equal [18] S. L. Kleiman and D. Laksov, “Schubert calculus”, The American Mathematical Monthly, 79(1972),1061–1082. totheminimumsubspacedistancehavebeenusedtoesti- [19] P.Samuel,Projective Geometry,Springer-Verlag, NY,USA,1988. mate theprofilevectorcorrespondingtothe Schubertcell [20] J. W. P. Hirschfeld, Projective Geometries over Finite Fields, The containing the transmitted subspace. In our construction, Clarendon Press,OxfordUniversity Press,NewYork,2nded.,1998. [21] W.V.D.HodgeandD.Pedoe,MethodsofAlgebraicGeometry,Vol.II, a similar procedure, identifying each characteristic tuple CambridgeUniversity Press,Cambridge, 1952. with acodewordofa binaryasymmetriccodeorthatofa [22] J.H.vanLintandR.M.Wilson,ACourseinCombinatorics,Cambridge binary code with the same relations on the injection and University Press,NewYork,USA,1992. [23] E. M. Gabidulin, “Theory of codes with maximal rank distance”, subspace distances, would suffice forestimating the tuple ProblemsofInformation Transmission(inRussian),21(1985),1–12. corresponding to the Schubert cell. [24] E.Gorla, F.Manganiello andJ.Rosenthal, “Analgebraic approach for decodingspreadcodes”, Online,arXiv.1107.5523v3. (2) EstimatingthetransmittedsubsetwithintheSchubertcell, [25] H. Mahadavifar and A. Vardy, “List-decoding of Subspace Codes and which is the decoding problem for rank-metric constant- Rank-MetricCodesuptoSingletonBound”,Online,arXiv.1202.0866v1. dimension subspace codes. [26] V. Guruswami, S. Narayanan and C. Wang, “List decoding subspace codesfrominsertionsanddeletions”, Proc.ITCS,Cambridge,MA,USA, Hence another fruitful direction of research will be to devise (2012)183–189. an efficient general decoding algorithm for non-constant di- [27] A-L. Trautmann, N. Silberstein and J. Rosenthal, “List Decod- ing of Lifted Gabidulin Codes via the Plu¨cker Embedding”, Online, mension projective space codes. arXiv.1301.2165v1. ACKNOWLEDGMENT The authoris gratefulto Dilip P. Patil, Sumanta Mukherjee and Smarajit Das for their assistance. REFERENCES [1] R. Ko¨tter and F. R. Kschischang, “Coding for errors and erasures in randomnetworkcoding”,IEEETrans.Inf.Theory,54(2008),3579–3591. [2] D. Silva, F. R. Kschischang and R. Ko¨tter, “A rank-metric approach to error-control in random network coding”, IEEE Trans. Information Theory,54(2008),3951–3967. [3] T.EtzionandN.Silberstein, “Error-correcting codes inprojective space via rank-metric codes and Ferrers diagrams”, IEEE Trans. Information Theory,55(2009),2909–2919.