1. EDA (CS286.5b) Day 6 Partitioning: Spectral + MinCut

2. Today Alternate views of partitioning Two things we can solve optimally –(but don’t exactly solve our original problem) Techniques –Linear Placement w/ squared wire lengths –Network flow MinCut

3. Optimization Target Place cells In linear arrangement Wire length between cells: –distance=X i - X j –cost is sum of distance squared Want to pick X i ’s to minimize cost

4. Spectral Partitioning Minimize Squared Wire length -- 1D layout Start with connection array C (c i,j ) “Placement” Vector X for x i placement cost = 0.5*  (all I,j) (x i - x j ) 2 c i,j cost sum is X’BX –B = D-C –D=diagonal matrix, d i,i =  (over j) c i,j

5. Spectral Partition Constraint: X’X=1 –prevent trivial solution all x i ’s same Minimize cost=X’BX w/ constraint –minimize L=X’BX- (X’X-1) –  L/  X=2BX-2 X=0 –(B- I)X=0 –X => Eigenvector of B –cost is Eigenvalue

6. Spectral Partitioning X (x i ’s) continuous use to order nodes –real problem wants to place at discrete locations –this is one case where can solve LP problem (continuous x’s) then move back to closests discrete point form cut partition from ordering

7. Spectral Ordering Midpoint bisect isn’t necessarily best place to cut, consider: K (n/4) K (n/2)

8. Spectral Partitioning Options Can bisect by choosing midpoint –(not strictly optimizing for minimum bisect) Can relax cut critera –min cut w/in some  of balance With iteration, can reweigh connections to change cost model being optimized –linear –(distance) 1.X

9. Spectral Options (2) Ratio Cut –minimize (cut/|A||B|) idea tradeoff imbalance for smaller cut –more imbalance =>smaller |A||B| –so cut must be much smaller to accept Maybe: encourage “closeness” –by weighting connection between nodes –(have to allow some nodes to not be close)

10. Spectral vs. FM From Hauck/Boriello ‘96

11. Improving Spectral More Eigenvalues –look at clusters in n-d space –5--70% improvement over EIG1

12. Spectral Note Unlike KLFM, attacks global connectivity characteristics Good for finding “natural” clusters –hence use as clustering heuristic for multilevel algorithms (both types)

13. Spectral Theory There are conditions under which spectral is optimal [Boppana/FOCS28 (1987)] –B=A+diag(d) –g(G,d)=[sum(B)-n* (B S )]/4 –B S mapping Bx to closest point on S –h(G) = max d g(G,d) h(G) lower bound on cut size

14. Spectral Theory Boppana paper gives a probabilistic model for graphs –model favors graphs with small cuts –necessary since truly random graph has cut size O(n) –shows high likelihood of bisection being lower bound In practice –known to be very weak for graphs with large cuts

15. Max Flow MinCut

16. MinCut Goal Cut find minimum flow cut between a source and a sink –no balance guarantee

17. MaxFlow Set all edge flows to zero –F[u,v]=0 While there is a path from s,t –(breadth-first-search) –for each edge in path f[u,v]=f[u,v]+1 –f[v,u]=-f[u,v] O(|E|*cutsize) Our problem simpler than general case CLR

18. Directionallity Construct 1   

19. Fanout Construct 1     

20. Balanced Cut Pick a start node and a finish node Compute min-cut start to finish If halves sufficiently balanced, done else –collapse all nodes in smaller half into one node –pick a node adjacent to smaller half –collapse that node into smaller half –repeat from min-cut computation FBB -- Yang/Wong ICCAD’94

21. Observation Can use residual flow from previous cut when computing next cuts Consequently, work of multiple network flows is only O(|E|*final_cut_cost)

22. Picking Nodes Optimal: –would look at all s,t pairs –also look at all nodes to collapse into smaller –could formulate as branching search Probably be more clever: –I.e. everything in first group around s (or t) would be there regardless of which s or t picked as kernal –probably similar observations would reduce branching factor on picking node to collapse in

23. Picking Nodes Randomly pick –(maybe try several starting points) With small number of adjacent nodes, –could afford to branch on all

24. Approximation Can find 1/3 balanced cuts within O(log(n)) of best cut in polynomial time –algorithm due to Leighton and Rao –exposition in Hochbaum’s Approx. Alg. Bound cut size –Boppana -- lower bound h(G) –Boppana/Spectral cut -- one upper bound –Leighton/Rao -- upper bound

25. Min Cut Replication After have (once have) kernel sets for each half, determine replication set Given: P 1, P 2 Determine subset of P 1 to replicate into P 2 –(and vice-versa) Algorithm: Hwang+ElGamal / TR CAD v14n1p96

26. Min Cut Replication To figure out what nodes from P 1 to replicate into P 2 Start with P 1 Add nodes adjacent to P 1, call these I 1,2 Add a start node (connect to inputs of P 1 ) Add a target node (connect to I 1,2 ) Keep original edges contained in P 1 Keep edges from P 1 to I 1,2

27. Replication Picture start target P1P1 I 1,2 Replication set

28. Min Cut Replication Perform MaxFlow/MinCut in new graph nodes unreachable from start node form new replication set (those nodes go in P 2, reducing it’s IO) n.b. -- cut minimal, but number of replicated nodes not guaranteed to be minimum

29. Apply Example?

30. Big Ideas Divide-and-Conquer Techniques –flow based –numerical/linear-programming based Exploit problems we can solve optimally