Exact Algorithms For Steiner Tree: Ondra Such y

Exact Algorithms for Steiner Tree
Ondra Suchý
Faculty of Information Technology,

Czech Technical University in Prague,
Prague, Czech Republic.
ondrej.suchy@fit.cvut.cz
New Developments in Exact Algorithms and Lower Bounds,

IIT Delhi, 13th December 2014
Ondra Suchý (FIT CTU Prague) Exact Algorithms for Steiner Tree IIT Delhi, 13.12.2014 1 / 41
Outline
1 Problem introduction and classical results

2 Exact algorithms for the general undirected case
3 Directed variants and algorithms for them
4 Algorithms for Steiner problems in sparse graphs
Steiner Tree
T
Steiner Tree
Given: Undirected graph G = (V , E ) and a set T ⊆ V
Find: A minimun size tree H = (V 0 , E 0 ) subgraph of G such that T ⊆ V 0 .
Steiner Tree
T
Steiner Tree
Steiner Tree
Steiner Tree T
The vertices in T are called terminals
The vertices in V \ T are called Steiner points
Denote n := |V |, m := |E |, and t := |T |
Steiner Tree
Steiner Tree T
The vertices in T are called terminals
The vertices in V \ T are called Steiner points
Denote n := |V |, m := |E |, and t := |T |
A minimum size:
Vertex cardinality: |V 0 | or rather |S| := |V 0 \ T | (default)
Edge cardinality: |E 0 | = |V 0 | − 1, this is equal to the above
Node weighted: Given w : V → N minimize w (S)
Edge weighted: Given w : E → N minimize w (E 0 )
Steiner Tree (default decision variant)
Given: Undirected graph G = (V , E ), a set T ⊆ V , and k ∈ N
Question: Is there a tree H = (V 0 , E 0 ), subgraph of G , such that T ⊆ V 0
and |V 0 \ T | ≤ k.
Importance
Fundamental problem of network design

Motivated by applications in, e.g.,
I VLSI routing
I phylogenetic tree reconstruction
I network routing
Several books devoted to Steiner Trees
I Dietmar Cieslik: Steiner minimal trees, Kluwer Academic, 1989
I Frank K. Hwang, Dana S. Richards, Pawel Winter: The Steiner tree
problem, North-Holland, 1992
I Alexandr O. Ivanov, Alexei A. Tuzhilin: Minimal networks — the
Steiner problem and its generalizations, CRC Press, 1994
I Hans J. Prömel, Angelika Steger: The Steiner Tree Problem — A Tour
through Graphs, Algorithms, and Complexity, Springer, 2002
11th DIMACS Implementation Challenge (ending two weeks ago)
devoted to Steiner Tree problems
Hardness and Approximability Results
NP-complete [Garey & Johnson 1979],
even on planar graphs [Garey & Johnson, SIAM J. Appl. Math 1977]
approximable to within O(log n) but
not within (1 − ε)(log t) unless NP ⊆ DTIME[N polylogn ]
[Klein & Ravi, Journal of Algorithms 1995];
The edge weighted variant is APX-complete even on complete graphs
with weights 1 and 2 [Bern, Plassmann, Inf. Proc. Lett. 1989]
There are many approximation results on various variants of Steiner
tree basically on every conference, e.g.,
I Bateni, Hajiaghayi, Marx: Approximation schemes for Steiner Forest on
planar graphs and graphs of bounded treewidth STOC 2010
I Bateni, Checkuri, Ene, Hajiaghayi, Korula, Marx: Prize-collecting
Steiner problems on planar graphs SODA 2011.
I STOC 2014, FOCS 2013, and elsewhere
An online compendium of approx. results: Hauptmann and Karpinski:
A Compendium on Steiner Tree Problems, University of Bonn.
Exact Algorithms
In this talk:
1 Parameterized algorithms — exponential in some presumably small
parameter:
I number of Steiner points in the solution k := |S| = |V 0 \ T |
I number of terminals t
I total cardinality of the tree t + k
2 Kernelizations
3 Exact exponential time algorithms — exponential in the “input size”
Trivial Algorithms
T
Simple algorithm:
Leaves of an optimal tree are terminals.
An optimal tree contains at most t vertices of degree at least 3.
Once the set T 0 of vertices of degree ≥ 3 is known, the optimal tree
can be computed as minimum spanning tree of T ∪ T 0 , where the
lengths are distances in G .
2
This gives nO(t) and O(2 3 n nO(1) ) = O(1.6181n ) algorithm for edge
weighted variant of Steiner Tree.
Parameterization by the Solution Size
Theorem
Steiner Tree is W[2]-hard with respect to the number k of Steiner
points in the solution.
We show a parameterized reduction from

Set Cover
Given: A universe U, a family F of its subsets, and k ∈ N S
Question: Is there a subfamily F 0 ⊂ F, |F 0 | ≤ k such that F ∈F 0 F = U?
Set Cover was shown W[2]-hard with respect to k in Downey &

Fellows 1999
Hardness of Steiner Tree, continued
For an instance (U, F, k) of Set Cover consider the following
instance (G = (V , E ), T , k) of Steiner Tree :
V = U ∪ F ∪ {t0 } t0
E = {{u, F } | u ∈ F ∈ F}
∪{{t0 , F } | F ∈ F}
T = U ∪ {t0 } F
Steiner points 1-1 correspond
to the sets in F
we claim that
the instances are equivalent U
V = U ∪ F ∪ {t0 } t0
E = {{u, F } | u ∈ F ∈ F}
∪{{t0 , F } | F ∈ F}
T = U ∪ {t0 } F
to the sets in F
we claim that
V = U ∪ F ∪ {t0 } t0
E = {{u, F } | u ∈ F ∈ F}
∪{{t0 , F } | F ∈ F}
T = U ∪ {t0 } F
to the sets in F
we claim that
Corollary
If for any k ≥ 3 and ε > 0 Steiner Tree can be solved in time O(nk−ε )
then the Strong ETH fails.
follows from the results of Patrascu and Williams [SODA 2010], since
Dominating Set is a special case of Set Cover,
Parameterization by Terminals
Theorem [Dreyfus & Wagner, Networks 1971] and [Levin 1971]
Edge weighted Steiner Tree can be solved in time
O(3t · n + 2t · n2 + n(n log n + m)).
Proof:
The proof goes by dynamic programming.
Pick any terminal t0 and let T 0 = T \ {t0 }
For every nonempty X ⊂ T 0 and every v ∈ V we compute:
ST(X , v ) = minimum edge weight of a Steiner tree for (X ∪ {v })
Note that we allow v ∈ X

The answer is stored in ST(T 0 , t0 )
The trivial case: If X = {x} for some x ∈ T 0 then for every v ∈ V we
set ST({x}, v ) = distG (x, v ).
Dreyfus-Wagner Algorithm continued
Now suppose |X | ≥ 2
Look at the tree from v
Starting from v go along the tree until you reach either a vertex in X
or a vertex of degree at least 3. Let us call it u. Possibly u = v .
X
u
v
X0
Dreyfus-Wagner Algorithm continued
Now suppose |X | ≥ 2
Look at the tree from v
Starting from v go along the tree until you reach either a vertex in X
or a vertex of degree at least 3. Let us call it u. Possibly u = v .
If u ∈ X then we let X 0 = {u}.
Otherwise we let X 0 be the vertices in X in one connected component
of the tree with {u} removed.
In both cases we have ∅ =6 X0 ( X
and the tree can be split into three pieces
I the path from v to u (possibly trivial) X
I a tree for u and X 0 (possibly trivial) u
I a tree for u and X \ X 0 v
X0
D-W Algorithm Recurrence
6 X 0 ( X and the tree can be split into three pieces
We have ∅ =
I the path from v to u (possibly trivial)
I a tree for u and X 0 (possibly trivial)
X
I a tree for u and X \ X 0 u
v
The table can be computed

using the following recurrence: X0
ST(X , v ) = min(distG (v , u) + min (ST(X 0 , u) + ST(X \ X 0 , u)))

v ∈V ∅6=X 0 (X
Both X 0 and X \ X 0 are strictly smaller and nonempty
D-W Algorithm Recurrence
6 X 0 ( X and the tree can be split into three pieces
We have ∅ =
I the path from v to u (possibly trivial)
I a tree for u and X 0 (possibly trivial)
X
I a tree for u and X \ X 0 u
v
The table can be computed

using the following recurrence: X0
ST(X , v ) = min(distG (v , u) + min (ST(X 0 , u) + ST(X \ X 0 , u)))

v ∈V ∅6=X 0 (X
Both X 0 and X \ X 0 are strictly smaller and nonempty

Running time:
Each vertex of T 0 can be either in X 0 , in X \ X 0 , or in T 0 \ X
There are 3t−1 n2 evaluations of the recurrence.
One can save by precomputing the second minimum.
The running time follows
Improvements of the D-W Algorithm
The 1971 D-W algorithm achieves time
O(3t · n + 2t · n2 + n(n log n + m))
This can be improved to O(3t · n + 2t (n log n + m))
by computing the distances more cleverly on demand
[Erickson, Monma, Veinott Mathematics of Operations Research
1987]
In 2007 Fuchs, Kern, and Wang [Math. Meth. Oper. Res.] improved
this to O(2.684t nO(1) ) and
−1
Mölle, Richter, and Rossmanith [STACS 2006] to O((2 + ε)t nf (e ) )
later the above two groups together [Theory Comput. Syst. 2007]
improved the exponent to O(( − εln ε )−δ ) for any 1/2 < δ which gives,
e.g., O(2.5t n14.2 ) or O(2.1t n57.6 )
By using subset convolution and Möbius inversion, one can get to a
running time of Õ(2t n2 + nm) for the node weighted case with
bounded weights [Björklund, Husfeldt, Kaski, Koivisto STOC 2007]
Improvements of the D-W Algorithm
The 1971 D-W algorithm achieves time
O(3t · n + 2t · n2 + n(n log n + m))
This can be improved to O(3t · n + 2t (n log n + m))
by computing the distances more cleverly on demand
[Erickson, Monma, Veinott Mathematics of Operations Research
1987]
In 2007 Fuchs, Kern, and Wang [Math. Meth. Oper. Res.] improved
this to O(2.684t nO(1) ) and
−1
Mölle, Richter, and Rossmanith [STACS 2006] to O((2 + ε)t nf (e ) )
later the above two groups together [Theory Comput. Syst. 2007]
improved the exponent to O(( − εln ε )−δ ) for any 1/2 < δ which gives,
e.g., O(2.5t n14.2 ) or O(2.1t n57.6 )
By using subset convolution and Möbius inversion, one can get to a
running time of Õ(2t n2 + nm) for the node weighted case with
bounded weights [Björklund, Husfeldt, Kaski, Koivisto STOC 2007]
All these algorithms take exponential space.
Polynomial Space Algorithms
First polynomial space algorithm: O(6t nO(log t) ) for edge weighted

variant [Fomin, Grandoni, Kratsch ESA 2008 ]
Combining with, e.g., D&W for t < log n one obtains
O ∗ (2O(t log t) )-time polynomial space algorithm for the edge weighted
variant.
Finally, Nederlof [ICALP 2009] gave a 2t nO(1) -time polynomial space
algorithm for edge weighted variant with bounded weights
It is pretty simple and based on the inclusion-exclusion principle
Further Improvements
Can we hope for an algorithm running in (2 − ε)t nO(1) -time?
This would imply an algorithm for Set Cover with running time
(2 − ε)|U| |F|O(1) (by the presented reduction)
t0
Set Cover Conjecture (SeCoCo) [Cygan, Dell, Lokshtanov, Marx,

Nederlof, Okamoto, Paturi, Saurabh, Wahlström CCC 2012]
There is no (2 − ε)|U| |F|O(1) -time algorithm for Set Cover.
OPEN: Is there any relation between SeCoCo and SETH?

Kernelization
Theorem [folklore / Dom, Lokshtanov, Saurabh ACM Transactions on

Algorithms 2014 / ICALP 2009]
There is no polynomial kernel for Steiner Tree parameterized by t + k
unless NP⊆coNP/poly, which would imply the collapse of the Polynomial
Hierarchy to the third level.
Proof:
By the framework of Bodlaender, Downey, Fellows, and Hermelin
[J. Comput. Syst. Sci. 2009/ ICALP 2008] we have to show that
Steiner Tree is compositional with this parameterization
Consider instances (G1 , T1 , k1 ), . . . , (Gs , Ts , ks )
We may assume that |T1 | = |T2 | = . . . = |Ts | = t and
k1 = k2 = . . . = ks = k
We denote T1 = {t11 , . . . , tt1 }, etc.
Kernel Lower Bound continued
Ss
We let G = i=1 Gi
For each j ∈ {1, . . . , t} we add to G a new vertex tj and connect it to
the vertices tj1 , tj2 , . . . , tjs by paths of length k + 2.
T1 T2 T3
We let T = {t1 , . . . , tt } and k 0 = t · (k + 2) + k.
Ss
We let G = i=1 Gi
T1 T2 T3
Ss
We let G = i=1 Gi
Kernel Lower Bound continued 2
We claim that (G , T , k 0 ) is a yes instance if and only if

∃i ∈ {1, . . . , s} such that (Gi , Ti , k) is a yes-instance.
If Gi is a yes-instance, then use the solution and add the paths to it.
A solution for G , must contain at least one path for each tj
Hence, exactly one path.
If they ended in different Gi ’s, then the graph would be disconnected.
Exponential Time Algorithms
the fast exponential algorithms are obtained by combining branching

for large values of t and FPT algorithms for small t
the fastest for weighted case is based on the algorithm of Mölle et al.
achieving O(1.42n ) in exponential space
The only paper devoted to such algorithms is by Fomin, Grandoni,
Kratsch, Lokshtanov, Saurabh [Algorithmica 2009 / ESA 2008 ]
It uses more involved branching, quasi FPT algorithm, and analyses
the running time by Measure & Conquer.
It is polynomial space and originally achieved running time O(1.59n )
for the weighted case and O(1.55n ) for the cardinality case.
Plugging in the O ∗ (2t ) algorithm of Nederlof, the running time can
be improved to O(1.36n ) for the cardinality case.
Steiner Problems in Directed Graphs
In directed graphs “to be connected” can mean several things:
Connect one distinguished root by directed paths
to all other terminals
Directed Steiner Tree (DST)
Connect all terminals among each other

Strongly Connected
Steiner Subgraph (SCSS)
Connect given terminal vertex pairs

in given directions
Directed Steiner Network (DSN)
1−
DSN hard to approximate within O(2log n ) unless
NP ⊆ TIME (2polylog (n) ) [Dodis, Khanna STOC 1999]
Parameterization by Steiner Points—Directed
With respect to the number k of Steiner points in the solution:
The hardness reduction is easy to modify to show that both
Directed Steiner Tree and Strongly Connected Steiner
Subgraph are W[2]-hard.
It is enough to orientate all the edges “towards the universe”
t0
Parameterization by Steiner Points—Directed
With respect to the number k of Steiner points in the solution:
The hardness reduction is easy to modify to show that both
Directed Steiner Tree and Strongly Connected Steiner
Subgraph are W[2]-hard.
It is enough to orientate all the edges “towards the universe”
and add backward arcs.
Hence Directed Steiner Network is also W[2]-hard.
t0
Parameterization by Terminals—Directed
With respect to the number t of terminals:

All the FPT algorithms for Steiner Tree can be adapted to solve
Directed Steiner Tree.
Feldman, Ruhl [FOCS 1999/ SIAM J. Comput. 2006]:
I SCSS can be solved in O(mn2t−3 + n2t−2 log n) time
I DSN can be solved in O(mn4t−2 + n4t−1 log n) time,
Guo, Niedermeier, S.[ISAAC 2009/SIAM J. Disc. Math. 2011]:
I SCSS is FPT with respect to t for the augmentation case — add arcs
to existing graph to achieve the requested connectivity
I SCSS is equivalent to TSP on the terminals if the graph is complete
and the edge weights are 1 and 2.
I SCSS is W[1]-hard for the cardinality case
Hardness of SCSS
Theorem [Guo, Niedemeier,S. 2009]

SCSS is W[1]-hard with respect to k + t for the cardinality case.
We reduce
Multicolored Clique(MCC)
Given: A graph G = (V , E ), k ∈ N and a coloring c : V → {1, . . . , k}.
Decide: Is there a clique in G taking exactly one vertex of each color?
Parameter: k
MCC is W[1]-hard [Pietrzak, JCSS 2003]
we use the edge representation strategy by Fellows, Hermelin,
Rosamond and Vialette [Theor. Comp. Sci. 2009]
SCSS Hardness Proof ctd t1,2 t2,1
uv ab vu ba
u a v b
t1 t2
For each color i introduce a vertex ti and for each vertex v of color i
we introduce an oriented triangle ti , v , v 0
For each pair of colors i 6= j, introduce two vertices ti,j and tj,i .
For each edge uv with c(u) = i and c(v ) = j, introduce two triangles
0 and t , x , x 0
ti,j , xuv , xuv j,i vu vu
0 , x ), (x 0 , x ), (x 0 , x ), (x 0 , x )
add arcs (xuv v u uv vu u v vu
k
let T = {t1 , . . . , tk } ∪ {ti,j | i 6= j} and k 0 = 2k + 2

2
SCSS Hardness Proof Continued 2 t1,2 t2,1
uv ab vu ba
We claim that the instance (G , k, c)
of MCC is equivalent to
the constructed instance of SCSS
u a v b
t1 t2
Selecting a vertex or edge in G correspond to selecting the vertices of

the corresponding triangle in the constructed digraph.
If there is a mcc, then we can connect the terminals (easy to check).
If the terminals are connected, then for each color there is one
triangle selected, otherwise the terminal would be isolated.
The same for pairs of colors.
Exactly one triangle selected for each terminal.
The selected edges must be between selected vertices.
The selected vertices form a clique.
SCSS Running Time Lower Bound
Corollary
SCSS cannot be solved in no(t/ log t) time unless ETH fails.
the reduction can be done the same way with Paritioned

Subgraph Isomorphism
the lower bound follows from Marx [FOCS 2007 / Theory of
Computing 2010]
Sparse graphs
Sparse graph classes often studied:

Graphs of bounded treewidth
planar graphs
Kh -minor free — graphs that do not contain Kh as a minor
Kh -topological minor free — for topological minor, one can only
contract edge if one of the endpoints has degree 2
d-degenerate ⇔ every subgraph has vertex of degree at most d
Sparse directed = sparse underlying undirected
Graphs of Bounded Treewidth
Classical dynamic programming gives 2O(tw log tw ) · n-time algorithm

for Steiner Tree
2
This was slightly improved to O(Btw +2 · tw · n), where Bk is the Bell
number, by Chimani, Mutzel, Zey [IWOCA 2011/ J. of Disc. Algor.
2012]
There is a 3tw · nO(1) randomized algorithm, with no false positives
and probability of false negative at most 12 [Cygan, Nederlof,
Pilipczuk, Pilipczuk, van Rooij, Wojtaszczyk FOCS 2011]
It is based on a technique called “Cut & Count”
Cut and Count for Steiner Tree
We want to know, whether there is a connected subgraph of given

size containing T .

size containing T .
We count the number of all subgraphs of given size containing T .
We do it in a way that the connected ones are counted once, while
the others are counted an even number of times.
If there was an odd number of connected ones, then the total number
will be also odd.

size containing T .
We count the number of all subgraphs of given size containing T .
We do it in a way that the connected ones are counted once, while
the others are counted an even number of times.
If there was an odd number of connected ones, then the total number
will be also odd.
We use random weights on vertices to ensure that, with high
probability, there is a unique Steiner tree of minimum weight.
Achieving Unique Solution
Isolation Lemma [Mulmuley, Vazirani, Vazirani Combinatorica 1987]

6 F ⊆ 2U . For each u ∈ U, choose a weight w (u) ∈ {1, . . . , N}
Let ∅ =
uniformly and independently at random. Then the probability that a set of
minimum weight in F is unique is at least 1 − |U|
N .
We use N = 2n.
We guess the minimum weight, i.e., we compute the parity of the
number of subgraphs for each possible weight.
Cut and Count — What We Count
We pick an arbitrary terminal t1 ∈ T .

We actually count the parity of the number of pairs of
I a subgraph (X , F ) with T ⊆ X of given size and particular weight and
I a cut (X1 , X2 ) such that
F X1 ∩ X2 = ∅,
F X1 ∪ X2 = X
F t1 ∈ X1
F there is no edge with t1 X2
one endpoint in X1 and one in X2
X1
For connected subgraphs only the cut X1 = X is possible.

For a disconnected subgraph, there are 2cc(X )−1 possible cuts.
The number can only become odd if there is a Steiner tree of the
given size.
Cut and Count—Running Time
t1 X2
X1
There are three possible states of a vertex: in X1 , in X2 , outside X

Correctness of the cut can be checked in polynomial times
Using principle of inclusion and exclusion, we get to 3tw · tw O(1) · n
for given size and weight
The weights are between 1 and 2n2 , sizes between 1 and n
Total running time 3tw · nO(1)
t1 X2
X1

In the same paper: If there is a constant ε > 0 and an (3 − ε)pw · nO(1)
algorithm for Steiner Tree, then the Strong ETH fails.
t1 X2
X1

In the same paper: If there is a constant ε > 0 and an (3 − ε)pw · nO(1)
algorithm for Steiner Tree, then the Strong ETH fails.
There are more recent approaches which are deterministic and apply
to weighted or counting variants [Bodlaender, Cygan, Kratsch,
Nederlof ICALP 2013] [Fomin, Lokshtanov, Saurabh SODA 2014]
2
OPEN: On directed graphs nothing better than O ∗ (2tw ) is known
Steiner Tree in Planar Graphs
Consider Steiner Tree in planar graphs parameterized by the

solution size k
We can contract any edge connecting two terminals
Hence, every other vertex on any path is Steiner
Thus, the graph has diameter ≤ 2k
The treewidth is bounded by O(k) and hence it is FPT wrt k.
Recently Pilipczuk, Pilipczuk, Sankowski, van Leeuwen [STACS 2013]
2/3
presented an algorithm running in time O(2O(((k+t) log(k+t)) ) n)
√
And even more recently this was improved to O(2O( (k+t) log(k+t)) n)
[Pilipczuk, Pilipczuk, Sankowski, van Leeuwen FOCS 2014]
OPEN: Is there a subexponential algorithm parameterized only by t or
only by k?
Directed Steiner Tree in Sparse Graphs
Directed Steiner Tree in sprase graphs studied by Jones,

Lokshtanov, Ramanujan, Saurabh, S. [ESA 2013]
In directed planar graphs we can only contract strongly connected
components formed by terminals.
Hence we cannot bound the treewidth of the graph.
Moreover, after contracting the strongly connected the graph may no
longer be sparse.
In fact, DST is W[2]-hard with respect to k even on 2-degenerate
graphs.
DST Hardness in Degenerate Graphs
Observation
DST is W[2]-hard with respect to k even on 2-degenerate graphs.
Use the reduction from setcover, but replace each terminal by a cycle
of vertices of degree three.
Subdivide the edges in these cycles.
All new vertices are terminals
Directed Steiner Tree in Sparse Graphs continued
D[T ] arbitrary
O ∗ (3hk+o(hk) )-time on Kh -minor free digraphs
(the algorithm is based on a novel branching rule in combination with
the Nederlof’s algorithm)
O ∗ (f (h)k )-time on Kh -topological minor free digraphs
(using the decomposition theorem of Grohe and Marx [STOC 2012])
D[T ] acyclic
O ∗ (3hk+o(hk) )-time on Kh -topological minor free digraphs
O ∗ (3dk+o(dk) )-time on d-degenerate graphs
I DST is FPT wrt k on o(log n)-degenerate graph classes
I FPT algorithm for undirected Steiner Tree on d-degenerate
k
o( )
For any constant c > 0, no f (k)n log k -time algorithm on graphs of
degeneracy c log n unless ETH fails.
I no O ∗ (2o(d)f (k) )-time algorithm unless ETH fails
no O ∗ (2f (d)o(k) )-time algorithm unless ETH fails
Kernelization in Sparse Graphs
Steiner Tree does not have a polynomial kernel with respect to

k + t even on 2-degenerate graphs, unless NP⊆coNP/poly
[Cygan, Pilipczuk, Pilipczuk, Wojtaszczyk Disc. App. Math. 2012]
Steiner Tree has an O((k + t)142 )-size kernel on planar graphs
[Pilipczuk, Pilipczuk, Sankowski, van Leeuwen FOCS 2014]
a polynomial kernel also exists on bounded genus graphs and for the
edge weighted variant
OPEN: Kernel on planar graphs with respect to only k or only t?
OPEN: Does it have a kernel on Kh -minor free or
Kh -topological-minor free graphs?
OPEN: Improve the size of the kernel.
SCSS in Sparse Graphs
SCSS in planar graphs was recently studied by Chitnis, Hajiaghayi,

Marx [SODA 2014]
√
They show a 2O(t log t) · nO( t) -time algorithm and also
√
that an f (k) · no( t) algorithm would imply that ETH fails.
OPEN: Can the algorithm for Directed Steiner Network be
also somehow speed up on planar graphs?
Some Open Problems
The main open problems are (not repeating all already mentioned)
Subexponential algorithm for Steiner Tree with respect to only t
Improvement to the kernels for planar graphs
generalizing the result to the higher connectivity settings
(first results for SCSS obtained by Chitnis, Esfandiari, Hajiaghayi,
Khandekar, Kortsarz, Seddighin [IPEC 2014])
Thank you for your attention!

Exact Algorithms For Steiner Tree: Ondra Such y

Uploaded by

Copyright:

Available Formats

Exact Algorithms For Steiner Tree: Ondra Such y

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Exact Algorithms For Steiner Tree: Ondra Such y

Uploaded by

Copyright:

Available Formats

Exact Algorithms for Steiner Tree

Faculty of Information Technology,

New Developments in Exact Algorithms and Lower Bounds,

1 Problem introduction and classical results

Fundamental problem of network design

We show a parameterized reduction from

Set Cover was shown W[2]-hard with respect to k in Downey &

ST(X , v ) = minimum edge weight of a Steiner tree for (X ∪ {v })

Note that we allow v ∈ X

The table can be computed

ST(X , v ) = min(distG (v , u) + min (ST(X 0 , u) + ST(X \ X 0 , u)))

Both X 0 and X \ X 0 are strictly smaller and nonempty

The table can be computed

ST(X , v ) = min(distG (v , u) + min (ST(X 0 , u) + ST(X \ X 0 , u)))

Both X 0 and X \ X 0 are strictly smaller and nonempty

First polynomial space algorithm: O(6t nO(log t) ) for edge weighted

Set Cover Conjecture (SeCoCo) [Cygan, Dell, Lokshtanov, Marx,

OPEN: Is there any relation between SeCoCo and SETH?

Theorem [folklore / Dom, Lokshtanov, Saurabh ACM Transactions on

We let T = {t1 , . . . , tt } and k 0 = t · (k + 2) + k.

We let T = {t1 , . . . , tt } and k 0 = t · (k + 2) + k.

We let T = {t1 , . . . , tt } and k 0 = t · (k + 2) + k.

We claim that (G , T , k 0 ) is a yes instance if and only if

the fast exponential algorithms are obtained by combining branching

Connect all terminals among each other

Connect given terminal vertex pairs

With respect to the number t of terminals:

Theorem [Guo, Niedemeier,S. 2009]

Selecting a vertex or edge in G correspond to selecting the vertices of

the reduction can be done the same way with Paritioned

Sparse graph classes often studied:

Classical dynamic programming gives 2O(tw log tw ) · n-time algorithm

We want to know, whether there is a connected subgraph of given

We want to know, whether there is a connected subgraph of given

We want to know, whether there is a connected subgraph of given

Isolation Lemma [Mulmuley, Vazirani, Vazirani Combinatorica 1987]

We pick an arbitrary terminal t1 ∈ T .

For connected subgraphs only the cut X1 = X is possible.

There are three possible states of a vertex: in X1 , in X2 , outside X

There are three possible states of a vertex: in X1 , in X2 , outside X

There are three possible states of a vertex: in X1 , in X2 , outside X

Consider Steiner Tree in planar graphs parameterized by the

Directed Steiner Tree in sprase graphs studied by Jones,

Steiner Tree does not have a polynomial kernel with respect to

SCSS in planar graphs was recently studied by Chitnis, Hajiaghayi,

You might also like