CSP beyond tractable constraint languages

The constraint satisfaction problem (CSP) is among the most studied computational problems. While NP-hard, many tractable subproblems have been identified (Bulatov 2017, Zhuk 2017) Backdoors, introduced by Williams, Gomes, and Selman (2003), gradually extend such a tractable class to all CSP instances of bounded distance to the class. Backdoor size provides a natural but rather crude distance measure between a CSP instance and a tractable class. Backdoor depth, introduced by Mählmann, Siebertz, and Vigny (2021) for SAT, is a more refined distance measure, which admits the parallel utilization of different backdoor variables. Bounded backdoor size implies bounded backdoor depth, but there are instances of constant backdoor depth and arbitrarily large backdoor size. Dreier, Ordyniak, and Szeider (2022) provided fixed-parameter algorithms for finding backdoors of small depth into the classes of Horn and Krom formulas. In this paper, we consider backdoor depth for CSP. We consider backdoors w.r.t. tractable subproblems CΓ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_\Gamma $$\end{document} of the CSP defined by a constraint language Γ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\Gamma }$$\end{document}, i.e., where all the constraints use relations from the language Γ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\Gamma }$$\end{document}. Building upon Dreier et al.’s game-theoretic approach and their notion of separator obstructions, we show that for any finite, tractable, semi-conservative constraint language Γ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\Gamma }$$\end{document}, the CSP is fixed-parameter tractable parameterized by the backdoor depth into CΓ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\varvec{\Gamma }}$$\end{document} plus the domain size. With backdoors of low depth, we reach classes of instances that require backdoors of arbitrary large size. Hence, our results strictly generalize several known results for CSP that are based on backdoor size.


Introduction
To face the NP-completeness of the Constraint Satisfaction Problem (CSP), much effort has been spent in identifying polynomial-time solvable subproblems [1].Tractability can be reached by 1. restricting the constraint language in terms of limiting the relations allowed to be used in constraints (e.g., [2][3][4][5][6]), 2. restricting the graphical structure of how constraints and variables interact (e.g., [7][8][9]), or 3. restricting both language and structure with hybrid restrictions (e.g., [10][11][12]).Some of the considered restrictions are gradual in the sense that they support an infinite chain of classes C 0 C 1 C 2 . . . of instances, where each C i can be solved in polynomial time.When the order of polynomial bound on the solving time remains the same for all the i = 0, 1, 2, . . .one speaks about fixed-parameter tractability (FPT) [13][14][15][16][17].Most structural restrictions like bounded treewidth or hypertree width are gradual by definition [9].In contrast, language restrictions tend to be categorical by definition, as either an instance belongs to a class C defined by a tractable constraint language or it does not.
However, by means of (strong 1 ) backdoors introduced by Williams, Gomes and Selman [18,19], one can build a chain . . . on top of such a class defined by a language.A CSP instance belongs to C i if there is a set of i variables, called a backdoor, such that all possible instantiations of these variables move the instance into the base class C .The size of a smallest backdoor provides a distance measure between the considered CSP instance and the base class.
The size of a smallest backdoor is a fundamental but still rather crude distance measure.Samer and Szeider [20] therefore proposed backdoor trees, where one counts the number of leaves of a decision tree ranging over all the variables of a backdoor; Ordyniak et al. [21] obtained further fixed-parameter tractability results for backdoor trees.A backdoor of size k over a Boolean domain can yield backdoor trees between k + 1 and 2 k leaves, and so it is more efficient to minimize the number of leaves than the size.Very recently, in the context of SAT, Mählmann, Siebertz, and Vigny [22] proposed the concept of backdoor depth, which extend backdoor trees by adding nodes where the tree branches into connected components.The advantage of considering backdoors of small depth relies on the observation that if an instance decomposes into multiple components, then each component can be treated independently.This way, one is allowed to use in total an unbounded number of backdoor variables.However, as long as the depth of the extended decision tree is bounded, one can still utilize it for efficiently solving the instance.In the context of graphs, similar ideas are used in the study of tree-depth [23,24] and elimination distance [25,26].
The challenging algorithmic question is to find a backdoor of small depth into a fixed base class, if it exists.Mählmann et al. [22] gave an FPT algorithm for SAT with respect to the base class C 0 consisting of formulas without variables; any bounded-depth backdoor into that class must contain all the variables of the instance.Already for this simple base class, there are instances of bounded backdoor depth that cannot be efficiently solved by other known methods.Previously [27], we extended this FPT result to bounded-depth backdoors into the classes of Horn (CNF formulas where each clause contains at most one positive literal), dual Horn (each clause contains at most one negative literal), and Krom formulas (each clause contains at most two literals).

Contribution.
In this paper, we provide the first positive algorithmic results for utilizing backdoors of bounded depth for CSP.Our main technical result covers all base classes C described by a finite semi-conservative .As our main result, we show the following (a formal statement is Corollary 22).
For any finite, tractable, semi-conservative constraint language , the CSP is fixedparameter tractable parameterized by the smallest depth of a backdoor into C plus the domain size of the instance.
Thus, we indeed have a chain C = C 0 C 1 C 2 . . . on top of any such class C , where C i contains instances with a backdoor of depth i, and where the order of the polynomial-time algorithm for solving C i is of the same order as the polynomial that bounds the solving time for C .
Backdoor depth can capture and exploit structure in CSP instances that is not captured by any other known method.In the following, we list here some known CSP parameters that admit fixed-parameter tractable CSP solving.For each of these parameters, there are CSP instances for which the parameter can be arbitrarily large, but where -backdoor depth is bounded by a constant: • backdoor size [28]; • backdoor depth for SAT [22,27]; • backdoor size into heterogeneous and scattered base classes [28,29]; • backdoor treewidth [30].
Please refer to the full version of a related paper [27, Appendix A] for a comprehensive comparison of the above notions.We closely follow the approach we introduced there.On a high level, we construct backdoors by simultaneously computing an upper bound in the form of an approximate backdoor and a lower bound, using so-called obstructions, i.e., parts of the instance that can be proven to be "far away" from the base class.As in our work on SAT, we use two types of obstructions: 1. a slightly modified version of the obstructions trees that have been introduced by Mählmann et al. [22] and 2. a new variant of separator obstructions that we introduced for SAT [27] to allow the handling of base classes that admit arbitrary long paths (in the incidence graph of a CNF formula).
This new variant of separator obstructions is tailor-made for CSP and base classes defined via finite constraint languages.It allows us to improve the algorithm's efficiency, from a tripleexponential run-time dependency to a double-exponential run-time dependency on backdoor depth.We present our results using the game-theoretic framework for backdoor depth that we introduced for SAT [27], which greatly simplifies the presentation of our algorithm.
A preliminary version of this paper has appeared at CP 2022 [31].The main feature of the current version is that it contains all the proofs, whereas due to space limitations, the conference version of the paper was missing several proofs of crucial lemmas and theorems.

CSP
Let D be a set and n and n be non-negative integers.An n-ary relation on D is a subset of D n .For a tuple t ∈ D n , we denote by t[i], the i-th entry of t, where 1 ≤ i ≤ n.For two tuples t ∈ D n and t ∈ D n , we denote by t • t , the concatenation of t and t .
An instance of a constraint satisfaction problem (CSP) I is a triple V , D, C , where V is a finite set of variables over a finite set (domain) D, and C is a set of constraints.We assume that D is given explicitly as a list of all domain values.A constraint c ∈ C consists of a scope, denoted by V (c), which is an ordered list of a subset of V , and a relation, denoted by R(c), which is a |V (c)|-ary relation on D; |V (c)| is the arity of c.To simplify notation, we sometimes treat ordered lists without repetitions, such as the scope of a constraint, like sets.For a variable v ∈ V (c) and a tuple t ∈ R(c), we denote by t[v], the i-th entry of t, where i is the position of v in V (c).For a CSP instance I = V , D, C we sometimes denote by V (I), D(I), and C(I), its set of variables V , its domain D, and its set of constraints C, respectively.We usually assume, w.l.o.g, that each variable in V (I) appears in the scope of at least one constraint in C(I).The size |I| of a CSP instance I is the sum of the sizes of its constraints, where the size of a constraint of arity a with t tuples and domain size We denote the assignment τ : {x} → D with τ (x) = q simply by x = q.
A constraint c ∈ C(I) of arity a is tautological if it contains all the |D| a possible tuples.Obviously, removing a tautological constraint from a CSP instance does not change its satisfiability.We denote by I[τ ] the CSP instance with variables V \ V , domain D, and constraints C[τ ], where C[τ ] contains all non-tautological constraints c[τ ] for every c ∈ C. We would like to point out that the removal of tautological constraints is important in the context of backdoor depth as it makes the notion more powerful.
Let τ 1 : V 1 → D and τ 2 : V 2 → D be two assignments.We say that the two assignments The incidence graph of a CSP instance I is the bipartite graph G I whose vertices are the variables and constraints of I, and where a variable x and a constraint c are adjacent if and only if x ∈ V (c).Via incidence graphs, graph theoretic concepts directly translate to CSP instances.For instance, we say that I is connected if G I is connected, and I is a connected component of I if D(I ) = D(I), and where V (I ) and C(I ) are maximal subsets of V (I) and C(I), respectively, such that G I is connected.Conn(I) denotes the set of connected components of I. Occasionally, we will also consider the primal graph of a CSP instance I, which has as vertex set V (I), and has pairs of variables adjacent if they appear together in the scope of a constraint.
A constraint language over a domain D is a set of relations over D. D( ) is the set of all the elements appearing in the relations in .We denote by arity( ) the maximum arity of any relation in .C denotes the class of CSP instances I with the property that for each c ∈ C(I ) we have R(c) ∈ .CSP( ) refers to the CSP with instances restricted to C .A constraint language is tractable if CSP( ) can be solved in polynomial time, is linear-time tractable if CSP( ) can be solved in linear time.
is semi-conservative (or 1-conservative) [32,33], if for each q ∈ D( ) one can express with the unary constraint x = q; more precisely, there is a satisfiable instance I q of CSP( ) and a variable x ∈ V (I q ) such that for each solution τ of I q we have τ (x) = q.Semiconservative constraint languages are very natural, as one would expect in any reasonable practical settings that the unary relations are present.Indeed, some authors (e.g., [3]) even define the CSP so that every variable can have its own set of domain values, making (semi-) conservativeness a built-in property.
A constraint language is closed under assignments if for every constraint c with R(c) ∈ and every assignment τ , it holds that R(c[τ ]) ∈ .For a constraint language we denote by * the smallest constraint language that contains and is closed under assignments.Lemma 1 ([33]) If a semi-conservative constraint language is tractable, then * is also tractable.
We would like to point out that the original definition of a backdoor by Williams et al. [18,19] assumes the base class to be closed under assignments.Hence, it is natural to assume this property in the context of backdoor depth, directly or indirectly by means of semiconservativeness of the considered language.

Backdoors
Backdoors are defined relative to some fixed base class C of instances of the problem under consideration (i.e., CSP), for which satisfiability and membership in C are polynomial-time decidable.In the context of CSP, we define a C-backdoor set of a CSP instance I as a set B ⊆ V (I) of variables such that I[τ ] ∈ C for every τ : B → D(I ).For a constraint language , we usually denote the base class C by itself.Thus, for example, instead of C -backdoors, we talk of -backdoors.If we know a C-backdoor set B of I, we can reduce the satisfiability of I to the satisfiability of |D(I )| |B| CSP instances in C. The challenging problem is to find a C-backdoor set of a given instance that reduces the satisfiability problem to instances from C.

Backdoor depth
Component backdoor trees generalize backdoor trees as considered by Samer and Szeider [20] by allowing an additional type of nodes, component nodes, where the current instance is split into connected components.More precisely, let C be a class of CSP instances (called the base class) and I a CSP instance.A component C-backdoor tree for I is a pair (T , ϕ), where T is a rooted tree and ϕ is a mapping that assigns each node t a CSP instance ϕ(t) such that the following conditions are satisfied: (b) Conn(ϕ(t)) = {I 1 , . . ., I k } for k ≥ 2 and t has exactly k children t 1 , . . ., t k with ϕ(t i ) = I i ; in this case we call t a component node.
Thus, a backdoor tree as considered by Samer and Szeider [20] is just a component backdoor tree without component nodes.The depth of a backdoor is the largest number of variable nodes on any root-leaf path in the tree.The C-backdoor depth depth C (I) of an instance I into a base class C is the smallest depth over all component C-backdoor trees of I .If C is defined in terms of a constraint language , we simply write depth (I).
Alternatively, we can define C-backdoor depth recursively as follows: depth where L(T ) denotes the set of leaves of T , using induction on d and m.The statement holds if d = 0 or m ≤ 1.We show that it also holds for larger d and m.If the root is a variable node, then it has δ children c 1 , . . ., c δ , and the subtree rooted at any of these children represents a component backdoor tree for the CSP instance ϕ(c i ) of depth d − 1.Therefore, by the induction hypothesis, we obtain that s i = ∈L(T i ) |ϕ( )| ≤ δ d−1 m, for every subtree T i rooted at c i .Consequently, ∈L(T ) |ϕ( )| = 1≤i≤δ s i ≤ δδ d−1 m = δ d m, as required.If, on the other hand, the root is a component node, then its children, say c 1 , . . ., c k , are labeled with CSP instances of sizes m 1 + • • • + m k = m.Therefore, for every subtree T i of T rooted at c i , we have that T i is a component backdoor tree of depth d for ϕ(c i ), which using the induction hypothesis implies that ∈L(T i ) |ϕ( )| ≤ δ d m i .Hence, we obtain ∈L(T ) |ϕ( )| ≤ δ d m in total.To solve the CSP instance I, we first solve all CSP instances associated with the leaves of T .Because, as shown above, their total size is at most δ d m, this can be achieved in time O((δ d m) c ), because C can be solved in time O(n c ) for some constant c ≥ 1 and input size n.Let us call a leaf true/false if and only it is labeled by a satisfiable/unsatisfiable CSP instance, respectively.We now propagate the truth values upwards to the root, considering a component node as the logical and of its children, and the a variable node as the logical or of its children.I is satisfiable if and only if the root of T is true.We can carry out the propagation in time linear in the number of nodes of T , which is linear in the number of leaves of T , i.e., at most δ d m.
Apart from lifting the approach from SAT to CSP, our tailor-made separator obstructions also allow us to obtain a more efficient algorithm.As our first order of business, we state an equivalent formulation of backdoor depth using connector-splitter games, as we previously introduced for SAT [27], allowing us to greatly simplify the exposition of our algorithm.Definition 3 Let be a finite constraint language that is closed under assignments and let I = V , D, C be a CSP instance.We denote by Game(I, ) the so-called -backdoor depth game on I.The game is played between two players, the connector and the splitter.The positions of the game are CSP instances.At first, the connector chooses a connected component of I to be the starting position of the game.The game is over once a position in the base class C is reached.We call these positions the winning positions (of the splitter).In each round the game progresses from a current position J to a next position as follows.
• The splitter chooses a variable v ∈ V (J ).
• The connector chooses an assignment τ : {v} → D and a connected component J of J [τ ].The next position is J .
In the (unusual) case that a position J contains no variables anymore but J is still not in C , the splitter loses.For a position J , we denote by τ J the assignment of all variables assigned up to position J .
The following observation follows easily from the definitions of the game and the fact that the (strategy) tree obtained by playing all possible plays of the connector against a given strategy for the splitter forms a component backdoor tree and vice versa.In particular, the splitter choosing a variable v at position J corresponds to a variable node and the subsequent choice of the connector for an assignment τ of v and a component of J [τ ] corresponds to a component node (and a subsequent variable or leaf node) in a component backdoor tree.

Observation 4 The splitter has a strategy for the game Game(I, ) to reach within at most d rounds a winning position if and only if I has a -backdoor of depth at most d.
Backdoor depth games mean that we no longer have to explicitly construct a backdoor.Instead, in Section 6, we compute winning strategies for the splitter, which appear to be easier to reason about.Such a strategy can then be automatically converted into a backdoor algorithm (Lemma 6).
We start by describing these so called splitter-algorithms and how they can be turned into an algorithm to compute backdoor depth.The algorithms will have some auxiliary internal state that they modify with each move.Formally, a splitter-algorithm for a game Game(I, ), where is a finite constraint language that is closed under assignments, is a procedure that • gets as input a (non-winning) position J of the game, together with an internal state • and returns a valid move for the splitter at position J , together with an updated internal state.
Suppose we have a game Game(I, ) and some additional input S. For a given strategy of the connector, the splitter-algorithm plays the game as one would expect: In the beginning, an internal state is initialized with S (if no additional input is given, it is initialized empty).Whenever the splitter should make its next move, the splitter-algorithm is queried using the current position and internal state and afterwards the internal state is updated accordingly.

Definition 5
We say a splitter-algorithm implements a strategy to reach for a game Game(I, ) and input S within at most d rounds a position and internal state with some property if and only if initializing the internal state with S and then playing Game(I, ) according to the splitter-algorithm leads-no matter what strategy the connector is usingafter at most d rounds to a position and internal state with said property.
Using the following observation converts splitter-algorithms into algorithms for backdoors.It builds backdoors by always trying out all the next moves of the connector.

Lemma 6
Assume there exists a function f (d, ) and a splitter-algorithm that implements a strategy to reach for each game Game(I, ) and non-negative integer d within at most f (d, ) rounds either a winning position or (an internal state representing) a proof that the -backdoor depth of I is larger than d.Further assume this splitter-algorithm always takes at most O(|I|) time to compute its next move.Then there exists an algorithm that, for a CSP instance I, a finite constraint language that is closed under assignments, and a non-negative integer d in time at most

) either returns a component -backdoor tree of depth at most f (d, )
or concludes that the -backdoor depth of I is larger than d.
Proof Assume we are given an instance I, a finite constraint language that is closed under assignments and a non-negative integer d.We compute a component -backdoor tree of depth at most f (d, ) by starting at the root and then iteratively expanding the leaves, using the splitter-algorithm to compute the next variable to place into the backdoor.For each position we reach, we store the internal state of the splitter-algorithm in a look-up table, indexed by the position.This way, we can easily build the backdoor tree in a depth-first or breadth-first way.If we encounter at any time an internal state representing a proof that the backdoor depth of I is larger than d, we can abort.If this is not the case, then we are guaranteed that every leaf represents a winning position and therefore an instance accepted by C .We have therefore found a component backdoor tree of depth at most f (d, ).
Without loss of generality, we can assume that I is connected.We need to expand the root node I of the tree f (d, ) times.We show that expanding a node J in our tree i times takes time at most |D(I )| 2i c|J | for some constant c.To expand it a node J , we run the splitter-algorithm in time c|J | to get the next variable.Then we enumerate all |D(I )| many assignments to this variable.The instance splits after an assignment into some components J 1 , . . ., J k with For improved readability, we may present splitter-algorithms as continuously running algorithms that periodically output moves (via some output channel) and always immediately as a reply get the next move of the connector (via some input channel).Such an algorithm can easily be converted into a procedure that gets as input a position and internal state and outputs a move and a modified internal state: The internal state encodes the whole state of the computation, (e.g., the current state of a Turing machine together with the contents of the tape and the position of the head).Whenever the procedure is called, it "unfreezes" this state, performs the computation until it reaches its next move and then "freezes" and returns its state together with the move.
Our main result is an approximation algorithm (Theorem 21) that either concludes that there is no backdoor of depth d, or computes a component backdoor tree of depth at most 2 O(d) .Using Lemma 6, we see that this is equivalent to a splitter-algorithm that plays for 2 O(d) rounds to either reach a winning position or a proof that the backdoor depth is larger than d.
Here and in the following, we say that a constraint is -bad for a finite constraint language if its relation is not in ; otherwise we say that the constraint is -good.Note that if is closed under assignments, then a -good constraint remains -good even after assigning additional variables and a conversely a constraint that is -bad in some subinstance obtained by assigning some variables is also -bad in the original instance.
Our proofs of high backdoor depth come in the form of so-called obstruction trees, which have first been introduced by Mählmann et al. [22].These are trees in the incidence graph of a CSP instance.Their node set therefore consists of both variables and constraints.Obstruction trees of depth d describe parts of an instance for which the splitter needs more than d rounds to win the backdoor depth game.For depth zero, we simply take a single -bad constraint that is not allowed by the base class.Roughly speaking, an obstruction tree of depth d > 0 is built from two "separated" obstruction trees T 1 , T 2 of depth d − 1 that are connected by a path.Because the two obstruction trees are separated but in the same component, we know that for any choice of the splitter (i.e., choice of a variable v), there is a response of the connector (i.e., an assignment of v and a component) in which either T 1 or T 2 is whole.Then the splitter needs by induction still more than d − 1 additional rounds to win the game.

Definition 7
Let I be a CSP instance and be a constraint language that is closed under assignments.We inductively define -obstruction trees T of increasing depth.
• Let c be a -bad constraint of I.The set T = {c} is a -obstruction tree in I of depth 0.
• Let T 1 be a -obstruction tree of depth i in I. Let β be a partial assignment of the variables in I. Let T 2 be an -obstruction tree of depth i in ) is contained both in a constraint of T 1 and T 2 .Let further P be a path (in the incidence graph) connecting T 1 and T 2 in I.
Note that the idea behind the partial assignment β in Definition 7 is that it allows some variables to appear in both T 1 and T 2 as long as those can be assigned in such a way (using β) that T 2 is still a -obstruction tree of depth i in I[β]; please also refer to Fig. 1 for an illustration of a -obstruction tree.
The following central lemma now shows the most crucial property of obstruction trees, namely, that they can be used to obtain lower bounds for the backdoor depth of a CSP instance.
Lemma 8 Let I be a CSP instance and be a constraint language that is closed under assignments.If there is a -obstruction tree of depth d in I, then the -backdoor depth of I is at least d + 1.
To show Lemma 8, we will need the following three auxiliary lemmas.

Lemma 9 Let I be a CSP instance and be a constraint language that is closed under assignments. Let β be a partial assignment of the variables in I and T be a -obstruction tree of depth d in I[β]. Then, T is also a -obstruction tree of depth d in I.
Proof We use induction on d.If d = 0, then there is a -bad constraint c of I[β] such T = {c}.Therefore, c is also a -bad constraint in I and T is a -obstruction tree of depth 0 in I.
Towards showing the induction step, let d > 0. Then there is a -obstruction tree T 1 in Moreover, there is a path P connecting T 1 and T 2 in I.Because of the induction hypothesis T 1 is a -obstruction tree in I of depth d − 1 and T 2 is a -obstruction tree in I[β ] of depth d − 1.Moreover, no variable of I[β ] is contained both in a constraint of T 1 and a constraint of T 2 , which implies that T is a -obstruction tree in I of depth d; this is because the variables in var(β) are neither in T 1 nor in T 2 .
Fig. 1 A -obstruction tree of depth 2 consisting of T 1 (given in blue), T 2 (given in red), and the path P (given in green).Note that T 1 , T 2 , and P are disjoint but their constraints can have common variables as long as each of these variables allows for an assignment that keeps T 2 intact.In the figure, the constraints of T 1 , T 2 , and P share the variable v and setting v to 0 leaves T 2 intact.Rectangles represent constraints, which can be either -good or -bad, filled rectangles represent -bad constraints, and circles represent variables

Lemma 10 Let I be a CSP instance and be a constraint language that is closed under assignments. A set T is a -obstruction tree of depth d in I if and only if there is a component I of I such that T is a -obstruction tree of depth d in I .
Proof This follows because all variables and constraints belonging to T induce a connected subgraph of I.
Lemma 11 Let I be a CSP instance and be a constraint language that is closed under assignments.Let T be a -obstruction tree of depth d in I, v ∈ V (I) be a variable not contained in any constraint of T , and τ be an assignment to that variable.Then, T is a -obstruction tree of depth d in I[τ ].
Proof Because v does not appear in any constraint of T all constraints of T in I are still present in I[τ ] and have the same scope as in I; this also implies that every such constraint is -good ( -bad) in I if and only if it is in I[τ ].Moreover, because every variable of T is contained in some constraint of v, it also follows that v is not contained as a variable in T .Therefore, T has the same variables and constraints in I as in I[τ ] and moreover all constraints remain the same, which shows the lemma.
We are now ready to show Lemma 8.

Proof of Lemma 8
Assume there exists a -obstruction tree of depth d in I.We will show that the connector has a strategy for the game Game(I, ) to reach within i ≤ d rounds a position J such that there is a -obstruction tree in J of depth d − i.This allows the connector to reach after d rounds a position that contains a -obstruction tree of depth 0, i.e., a -bad constraint.Thus, the splitter has no strategy to win the game after at most d rounds and by Observation 4, the statement of this lemma is proven.We show the claim by induction on i.The claim trivially holds for i = 0.So suppose that i > 0 and let J be the position reached by the connector after i − 1 rounds.Then, by the induction hypothesis, J contains a -obstruction tree T of depth at least Since the depth is at least one, there further exist a -obstruction tree T 1 of depth d − i in J , a partial assignment β of the variables in J , and a -obstruction tree T 2 of depth d − i in ) is contained both in a constraint of T 1 and T 2 .Now, let v be the next variable chosen by the splitter.We distinguish the following cases: 123 1. Assume v is not contained in any constraint of T 1 .Then, as the connector, we assign v an arbitrary value.Let τ be this assignment.By Lemma 11, T 1 is a -obstruction tree of depth d − i in J [τ ].We choose the connected component J containing T 1 in J [τ ].By Lemma 10 T 1 is also a -obstruction tree of depth d − i in J . 2. Assume v is not contained in any constraint of T 2 .By Lemma 9, T 2 is a -obstruction tree of depth d − i in J .We proceed analogously to the previous case, and reach a position in which T 2 is a -obstruction tree of depth d − i. 3. Otherwise, v occurs both in a constraint of T 1 and T 2 .Since T is a -obstruction tree in J , it follows by Definition 7 that v / ∈ V (J [β]) and therefore v ∈ V (β).The connector can now choose the assignment v = β(v) for v.Because of Lemma 9, T 2 is a -obstruction tree of depth containing T 2 .By Lemma 10, T 2 is also a -obstruction tree in this component.
Our splitter-algorithm will construct obstruction trees of increasing depth by a recursive procedure (Lemma 20) that we outline now.We say a splitter-algorithm satisfies property i if it reaches in each game Game(I, ) within g C (i, d) rounds (for some function g C (i, d)) either 1. a winning position, or 2. a position J and a -obstruction tree T of depth i in I such that no variable in V (J ) occurs in a constraint of T , or 3. a proof that the -backdoor depth of I is at least d.
A splitter-algorithm satisfying property d + 1 then directly implies our main result, the approximation algorithm for backdoor depth, using Lemmas 8 and 6.Assume we have a strategy satisfying property i − 1, let us describe how to use it to satisfy property i.If at any point we reach a winning position, or a proof that the -backdoor depth of I is at least d, we are done.Let us assume this does not happen, so we can focus on the much more interesting case 2).
We use property i − 1 to construct a first tree T 1 of depth i − 1, and reach a position J 1 .We use it again, starting at position J 1 to construct a second tree T 2 of depth i − 1 that is completely contained in position J 1 .Since T 1 and T 2 are in the same component of F, we can find a path P connecting them.Let β be the assignment that assigns all the variables the splitter chose until reaching position J 1 .Then T 2 is an obstruction tree not only in J 1 but also in I [β].In order to join both trees together into an obstruction of depth i, we have to show, according to Definition 7 that no variable v ∈ V (I[β]) occurs both in a constraint of T 1 and T 2 .Since no variable in V (J 1 ) occurs in a constraint of T 1 (property i − 1), and T 2 was built only from J 1 , this is the case.The trees T 1 and T 2 are "separated" and can be safely joined into a new obstruction tree T of depth i (details also in proof of Lemma 20).
Finally, we need to ensure is that we reach a position J such that no variable in V (J ) occurs in a constraint of T .This then guarantees that T is "separated" from all future obstruction trees that we may want to join it with to satisfy property i + 1, i + 2 and so forth.
It is important to note here, that the exact notion of "separation" between obstruction trees plays a crucial role and is one of the main differences between the approaches used by Mählmann et al. [22] and Dreier et al. [27].The former solve the separation problem in a "brute-force" manner: If we translate their approach to the language of splitter-algorithms, then the splitter simply selects all variables that occur in a clause of T .For their base classthe class Null of formulas without variables-there are at most 2 O(d) variables that occur in an obstruction tree of depth d.Thus, in only 2 O(d) rounds, the splitter can select all of them, fulfilling the separation property.This completes the proof for the base class Null.
However, already for backdoor depth to Krom formulas (or equivalently backdoor depth to some finite constraint language of arity at most two), this approach cannot work since obstruction trees for Krom formulas can have arbitrarily many clauses.We solve this issue by adapting the separator obstructions from SAT [27] to CSP.We also exploit the fact that our base classes have bounded arity (in contrast to, e.g., the class of Horn formulas) to simplify their separator obstructions significantly.This allows us to drop the complexity for solving CSP using backdoor depth from triple to double exponential in the backdoor depth.

Separator obstructions
Obstruction trees are made up of paths, therefore, it is sufficient to separate each new path P that is added to an obstruction.Note that P can be arbitrarily long and therefore the splitter cannot simply select all variables in (constraints of) P. Instead, given such a path P that we want to separate, we will use separator obstructions to develop a splitter-algorithm (Lemma 18) that reaches in each game Game(I, ) within a bounded number of rounds either 1. a winning position, or 2. a position J such that no variable in V (J ) occurs in a constraint of P, or 3. a proof that the backdoor depth of I is at least d.
Informally, a separator obstruction is a sequence P 1 , . . ., P of paths that form a tree T together with an assignment τ of certain important variables occurring in T .The variables of τ correspond to the variables chosen by the splitter-algorithm and the assignment τ corresponds to the assignment chosen by the connector.Each path P i adds (at least one) -bad constraint b i to the separator obstruction, which is an important prerequisite to increase the backdoor depth by growing the obstruction.Moreover, by choosing the important variables and the paths carefully, we ensure that the tree T has bounded maximum degree and that every outside variable, i.e., any variable that is not an important variable assigned by τ , can occur in at most four constraints of T .Therefore assigning any outside variable can split T into only boundedly many parts.Together with the assignment τ , which we will use as a guide for the connector for the variables inside the obstruction, this will allow us to show that the connector can play in such a way that after every round at least a constant fraction of the separator obstruction remains intact.This means a large separator obstruction is a proof that the backdoor depth is larger than d.
To illustrate the growth of a separator obstruction (and motivate its definition) suppose that our splitter-algorithm is at position J of the game Game(I, ) and already has built a separator obstruction X = P 1 , . . ., P i , τ containing -bad constraints b 1 , . . ., b i ; note that τ is compatible with τ J .If J is already a winning position, then we are done.Therefore, J has to contain a -bad constraint.Note that if J does not contain a variable that occurs in a constraint of T i , then J satisfies 2) of property i and we are done.Otherwise, let Y be the set of all such variables in J and let b i+1 be a -bad constraint in J that is closest to any variable in Y .Note that it can happen that b i+1 is in T i in which case, we let P i+1 be the path that only contains b i+1 .Otherwise, let P be a shortest path from b i+1 to Y in J and let y ∈ Y be the endpoint of P in Y .Let P i+1 be the path that is equal to P if y ∈ T i and otherwise P i+1 is obtained from P i after adding an edge from y to a constraint c in T i such that y occurs in c.Then, we extend our separator obstruction X by attaching the path P i+1 to T i (and obtain the tree T i+1 ).Our next order of business is to choose a bounded number of important variables occurring on P i+1 that we will add to X (or more precisly that will be added to the set of variables assigned by τ ).Those variables need to be chosen in such a way that no outside variable can destroy too much of the separator obstruction.Apart from destroying the paths of the separator obstruction, we also need to avoid that assigning any outside variable makes too many of the -bad constraints b 1 , . . ., b i+1 -good.Therefore, a natural choice is all variables of b i+1 to X , i.e., to make those variables important.The following lemma shows that this is possible, because the number of those variables is bounded.

Lemma 12
Let I be a CSP instance and be a finite constraint language.If I has -backdoor depth at most some integer d, then every constraint of I has arity at most d + arity( ).
Proof As stated in the preliminaries, we can assume that I does not contain any tautological constraints.Given a non-tautological constraint c and a variable v ∈ V (c), then there always exists a value x ∈ D(I ) such that c[v = x] also is non-tautological.This means, the connector has a strategy which guarantees that with each round of the game Game(I, ), c remains in the game and the arity of c decreases at most by one.
Suppose that I contains a constraint c having arity larger than d + arity( ).Then the connector can play such that after d rounds, c still has arity larger than arity( ) and therefore still is -bad.By Observation 4, I has backdoor depth larger than d.
The next thing that we need to ensure is that any outside variable can not destroy too many paths.Note that by choosing a shortest path P i+1 , we have already ensured that no variable occurs on more than two constraints of P i+1 (such a variable would be a shortcut, meaning P i+1 was not a shortest path).Moreover, because P i+1 is a shortest path from b i+1 to T i , we know that every variable that occurs on T i and on P i+1 must occur in the constraint c in P i+1 that is closest to T i , but not in T i itself.Similarly, to how we dealt with the -bad constraints, we will now add all variables that occur in c to X .This ensures that no outside variable can occur in both T i and P i+1 , which (by induction over i) implies that every outside variable occurs in at most two constraints (either from T i or from P i+1 ).Finally, if the endpoint of P i+1 in T i is a variable, we also add this variable to the separator obstruction to ensure that no variable has degree larger than three in T i+1 and therefore the deletion of any variable does not split T into too many parts.Note that we use Lemma 12 to bound the degree of constraints for the same reason.This leads us to the following definition of separator obstructions (see also Fig. 2 for an illustration).
Definition 13 Let X be a pair P 1 , . . ., P , τ , where τ is a partial assignment to variables of I and each P i is a path in I such that T i = j≤i P j is a tree for every i ∈ [1, ].For every i ∈ [2, ], let e i be the constraint in P i that is closest to T i−1 ; if such a constraint exists.Then, X is a -separator obstruction for I if there are constraints b 0 , . . ., b and assignments τ 0 , . . ., τ such that: • P 1 is a shortest path between the two -bad constraints b 0 and b 1 in I.
• For every i ∈ [1, ], τ i is the restriction of τ to the set of variables V i occurring in any constraint of {b 0 , b 1 , b 2 , e 2 , . . ., b i , e i } and τ = τ .

• For every i ∈ [1, ], if G i denotes the subgraph of I induced by the vertex set of I[τ i ] together with C(T i ), then for every
We define the size of X to be the number of leaves of T = T .
We start by showing some simple but important properties of separator obstructions.

Lemma 14
Let be a finite constraint language that is closed under assignments and let X = P 1 , . . ., P , τ be a -separator obstruction in I, then for every i ∈ [ ]:

(P3) Every variable v / ∈ V i is contained in at most two constraints of T i and moreover those constraints are consecutive in T i . (P4) Every variable v ∈ V i \ V i−1 is contained in at most 4 constraints of T i . (P5) If β is any assignment compatible with τ that does not assign any variable in
Proof (P1) holds trivially by induction because attaching a path to a tree results in a tree.We show (P2) by induction on i.Again, we see that (P2) holds for i = 1.Let i > 1 and v be the endpoint of P i contained in T i−1 .Because of the induction hypothesis and because P i is a path, we obtain that all vertices in T i (except possibly v) have degree at most 3.Moreover, if v is a variable, then v / . Therefore, v is not the endpoint of any path P j for j < i (since otherwise v would be contained in e j and therefore also in V i−1 ), which shows that v has degree two in T i−1 and hence v has degree at most 3 in T i .
We also establish (P3) by induction on i.For i = 1, assume for contradiction that v is adjacent to more than just two consecutive constraints of T 1 = P 1 .Then v would be a "shortcut" and P 1 would not have been a shortest path, contradicting our choice of P 1 .Now suppose that the claim holds for i − 1.Then, v is contained in at most two consecutive constraints of T i−1 and because P i is a shortest -good path, v is also contained in at most two consecutive constraints of P i .Moreover, because v / ∈ V i , it holds that if v is in some constraint of P i , it is not contained in e i .Therefore, if v were in some constraint of P i and of T i−1 then v would be a "shortcut" from b i to T i−1 , skipping e i .This would be a contradiction to our assumption that P i is a shortest -good path from b i to any vertex in T i−1 .Hence, v is either contained in at most 2 consecutive constraints of T i−1 but not contained in any constraint of P i or vice versa, which shows (P3).
Towards showing (P4), if i = 1, then because P 1 is a shortest -good path from b 0 to b 1 , v can occur in at most 2 consecutive constraints of P 1 .Moreover, if i > 1, then by (P3), it holds that v occurs in at most 2 consecutive constraints of T i−1 .Moreover, because P i is a shortest -good path from b i to T i−1 , v can also occur in at most 2 consecutive constraints of P i .Therefore, v can occur in at most 4 constraints of T i in total, as required.
Towards showing (P5), note that by the definition of a -separator obstruction, it holds that b i is a -bad constraint in Our next aim is to show that separator obstructions-just like obstruction trees-can be employed to obtain a lower bound on the backdoor depth of a CSP instance.For this it is important to show that assigning a single variable cannot sufficiently destroy a separator obstruction.
Note that Lemma 14 already provides a first step in this direction.In particular, (P3) limits the influence of variables outside of V to only two constraints and (P4) limits the influence of variables inside V , at least towards the part of the separator obstruction that was constructed before the variable was added.To limit the influence of variables in V also on the remaining part of the separator obstruction, we show that even though these variables can appear in arbitrary many constraints of the remaining part, their influence is still limited as long as we only consider CSP instances obtained by assigning those variables according to τ .Definition 15 Let X = P 1 , . . ., P , τ be a -separator obstruction for I and let β be an assignment that is compatible with τ .Moreover, let c be a constraint contained in T and let i be minimal such that c is contained in T i .We say that c is tainted by Otherwise we say that c is untainted by β.Similarly, we say that a subtree T of T is untainted by β if so is every constraint of T and moreover V (β) does not contain a variable of T .

Lemma 16
Let be a finite constraint language that is closed under assignments, let X = P 1 , . . ., P , τ be a -separator obstruction in I, let β be an assignment that is compatible with τ , and let T be a subtree of T untainted by β.Then, I[β] contains T .
Proof Consider a constraint c of T .Because c is in T , c must appear on some path P i and moreover by the definition of a -separator obstruction c appears in I[τ i−1 ] and has some scope S in I[τ i−1 ].Moreover, because c is untainted, it holds that S ∩ V (β) = ∅, which implies that c appears with some scope containing S in I[β] and since c is not tautological in I[τ i−1 ] and β is compatible with τ , c is not tautological in I [β].Therefore, all constraints of T appear in I[β] and since V (β) does not contain any variable of T , also all variables of T appear in I[β].
We are now ready to show our main result of this subsection, namely, that separator obstructions can be used to obtain lower bounds on the backdoor depth of a CSP instance.Consider the following strategy S for the connector in the game Game(I, ).Suppose that we have reached position J in the game and suppose that the splitter chooses a variable v as his next move.We distinguish the following two cases: Let J be a position reached in the game Game(I, ) against S at round i.We show by induction on i that J contains a subtree of T untainted by τ J containing at least n i = n/(3 + 4(d + arity( ))) i − 1 elements from B.
The claim clearly holds for i = 0 since the connector chooses the component of I containing T .Moreover, for i > 0 let J be the predecessor (position) of J in Game(I, ).By the induction hypothesis J contains a subtree T of T untainted by τ J containing at least Let v be the variable chosen by the splitter at position J and let α be the assignment of v chosen by the connector.
If v / ∈ V , then it follows from Lemma 14 (P3) with i = that v is contained in at most 2 constraints of T and therefore α can taint at most 2 constraints of T .Otherwise let 1 ≤ i ≤ be minimal such that v ∈ V i .Assume for contradiction α taints a constraint c in T \ T i .Then let j be minimal such that c is contained in T j .Obviously, j > i.But then v / ∈ V j−1 , a contradiction to our choice of i.This means α cannot taint any constraints in T \ T i .Since 1 ≤ i ≤ is minimal with v ∈ V i , we have v ∈ V i \ V i−1 and by (P4) v is contained in at most 4 constraints of T i .This means α can taint at most 4 constraints of T i .In total, α can taint at most 4 constraints of T and therefore also of T .Further, since T is untainted by τ J and τ J = τ J ∪ α, the assignment τ J taints at most 4 constraints of T .
Moreover, because of Lemma 16 and the fact that τ J ∪ α is compatible with τ , it follows that every subtree of T untainted by α is contained in some connected component of J [α].Because of Lemma 14 (P2), we obtain that the variable v has degree at most 3 in T .Moreover, because of Lemma 12, we can assume that every constraint in T has degree at most d + arity( ), since otherwise I has -backdoor depth at least d.Therefore, after removing the at most 4 constraints together with the variable v from J , there is a component of J [α] containing a subtree of T untainted by τ J with at least (n i−1 − 5)/( 3 Therefore, we obtain that if J is a position reached after i rounds in the game Game(I, ) against S, then J contains a subtree of T untainted by τ J containing at least n i = n/(3 + 4(d + arity( ))) i − 1 constraints from B. In particular, this implies that if J is a position reached after d rounds against S, then J contains a subtree of T untainted by τ J containing at least n/(3 + 4(d + arity( ))) d − 1 = d + 1 constraints from B. Finally, because of Lemma 14 (P5) at least one of these constraints is -bad in J , which concludes the proof of the lemma.
splitter-algorithm that takes a shortest path P and separates it from all future obstructions.By reaching a position J such that no variable in V (J ) occurs in a constraint of P, we are guaranteed that all future obstructions are separated from P, as future obstructions will only contain constraints and variables from J .Lemma 18 Let be a finite constraint language that is closed under assignments.There exists a splitter-algorithm that implements a strategy to reach for each game Game(I, ), non-negative integer d, and shortest path P between two -bad constraints in I within at most Proof Let X = P 1 , . . ., P , τ be a -separator obstruction for I and let τ be a subassignment of τ assigning at least all variables in V −1 .Then, we call X = P 1 , . . ., P , τ a partial -separator obstruction for I.
Consider the following splitter-algorithm, where for each position J of the game Game(I, ), we additionally associate a partial -separator obstruction denoted by X (J ) = P 1 , . . ., P , τ J with P 1 = P to every position J .We set X (S) = P , ∅ for the starting position S of the game.
Then, the splitter-algorithm does the following for a position J in Game(I, ).If X (J ) = P 1 , . . ., P , τ J and there is at least one variable in V \ V −1 (assuming that V 0 = ∅) that has not yet been assigned by τ J , then the splitter chooses any such variable.Otherwise, X (J ) is a -separator obstruction and we distinguish the following cases: 1.If there is a -bad constraint in J that has a path to some vertex of T in G , then let b +1 be a -bad constraint that is closest to any vertex of T in G and let P +1 be a shortest path from b +1 to some vertex of T in G .Note that P 1 , . . ., P , P +1 , τ J is a partial -separator obstruction for I.The splitter now chooses any variable in V +1 \ V and assigns X (J ) = P 1 , . . ., P , P +1 , τ J for the position J resulting from this move.2. Otherwise, X (J ) can no longer be extended and either: (a) there is no -bad constraint in J , in which case we reached a winning position, i.e., we achieved case 1), or (b) every -bad constraint of J has no path to T in G , which implies that no variable of J is contained in a constraint of T and therefore also of P, i.e., we achieved case 2).This completes the description of the splitter-algorithm.Moreover, if every play against the splitter-algorithm ends after at most (2 • arity( ) + d)(d + 2)(3 + 4(d + arity( ))) d rounds, every position is either of type i) or type ii) and we are done.
Otherwise, there is a position J that is reached after playing at least Finally, the splitter-algorithm takes time at most O(|I|) per round since a -bad constraint that is closest to the current -separator obstruction and the associated shortest path can be found using a simple breadth-first search.
Since selecting more variables can only help the splitter in achieving their goal, we immediately also get the following statement.

Corollary 19
Consider a finite constraint language that is closed under assignments, a game Game(I, ) and a position J in this game, a non-negative integer d and shortest path P between two -bad constraints in I.There exists a splitter-algorithm that implements a strategy that continues the game from position J and reaches within at most (2 • arity( ) + d)(d + 2)(3 + 4(d + arity( ))) d rounds either: 1. a winning position, or 2. a position J such that no variable in V (J ) is contained in a constraint of P, or 3. a proof that the -backdoor depth of I is larger than d.

This algorithm takes at most O(|I|) time per move.
As described at the end of Section 4, we can now construct in the following lemma obstruction trees of growing size, using the previous corollary to separate them from potential future obstruction trees.
Lemma 20 Let be a finite constraint language that is closed under assignments.There is a splitter-algorithm that implements a strategy to reach for a game Game(I, ) and nonnegative integers i and d with 1. a winning position, or 2. a position J and a -obstruction tree T of depth i in I such that no variable in V (J ) is contained in a constraint of T , or 3. a proof that the -backdoor depth of I is larger than d.

This algorithm takes at most O(|I|) time per move.
Proof We will prove this lemma by induction over i.Our splitter-algorithm will try construct an obstruction tree of depth i by first using the induction hypothesis to build two obstruction trees T 1 and T 2 of depth i − 1 and then joining them together.After the construction of the first tree T 1 , we reach a position J 1 and by our induction hypothesis no variable in V (J 1 ) is contained in a constraint of T 1 .This encapsulates the core idea behind our approach, as it means that T 1 is separated from all potential future obstruction trees T 2 .Therefore, we can compute the next tree T 2 and join them together in accordance with Definition 7. At last, we connect them via a path and use Corollary 19 to also separate this path from all future obstructions.If at any point of this process we reach a winning position or a proof that the -backdoor depth of I is larger than d, we can stop.Let us now describe this approach in detail.
For convenience, let x = (2 • arity( ) + d)(d + 2)(3 + 4(d + arity( ))) d .We start our induction with i = 0.If there is no -bad constraint in I, then it is a winning position and we can stop.Otherwise, there is a -bad constraint c containing variables v 1 , . . ., v t with t ≤ d + arity( ) by Lemma 12.We define T = {c} to be our -obstruction tree of depth i = 0. To reach a position J such that no variable v ∈ V (J ) is contained in c, we let the splitter select all variables in c.This takes at most d + arity( ) ≤ (2 i+1 − 1)x rounds.
We now assume the statement of this lemma to hold for i − 1 and we show it also holds for i.To this end, we start playing the game Game(I, ) according to the existing splitteralgorithm for i − 1.If we reach (within at most (2 i − 1)x rounds) a winning position or a proof that the -backdoor depth of I is larger than d then we are done.Assuming this is not the case, we reach a position J 1 and a -obstruction tree T 1 of depth i − 1 in I such that no variable v ∈ V (J 1 ) is contained in a constraint of T 1 .
We continue playing the game at position J 1 according to the existing splitter-algorithm for Game(I, )J One and i − 1.The -backdoor depth of I is larger or equal to the -backdoor depth of J 1 .Thus again (after at most (2 i − 1)x rounds) we either are done (because we reach a winning position or can conclude that the -backdoor depth of J 1 is larger than d) or we reach a position J 2 and a -obstruction tree T 2 of depth i − 1 in J 1 such that no variable v ∈ V (J 2 ) is contained in a constraint of T 2 .
Let β = τ J 1 be the assignment that assigns all the variables the splitter chose until reaching position J 1 to the value given by the connector.Note that J 1 is a connected component of I [β].By Lemma 10, T 2 is a -obstruction tree of depth i − 1 not only in J 1 , but also in I [β].
Let v ∈ V (I [β]).We show that v is not contained both in some constraint of T 1 and of T 2 .To this end, assume v is contained in a constraint of T 2 .Since all constraints of T 2 are in J 1 and J 1 is a connected component of I [β], we further have v ∈ V (J 1 ).On the other hand (as discussed earlier), no variable v ∈ V (J 1 ) is contained in a constraint of T 1 .
We pick two constraints c 1 ∈ T 1 and c 2 ∈ T 2 that are -bad in I and compute a shortest path P between c 1 and c 2 in I.By Definition 7, T = V (P)∪C(P)∪T 1 ∪T 2 is a -obstruction tree of depth i in I.
We use Corollary 19 to continue playing the game at position J 2 .Again, if we reach a winning position or a proof that the -backdoor depth of I is larger than d we are done.So we focus on the third case that we reach (within at most x rounds) a position J such that no variable v ∈ V (J ) is contained in a constraint of P. We know already that no variable v ∈ V (J 1 ) is contained in a constraint of T 1 and no variable v ∈ V (J 2 ) is contained in a constraint of T 2 .Since V (J 1 ), V (J 2 ) ⊆ V (J ), we can conclude that no variable v ∈ V (J ) is contained in a constraint of T .
In total, we played for (2 i −1)x +(2 i −1)x +x = (2 i+1 −1)x rounds.The splitter-algorithm in Corollary 19 takes at most O(|I|) time per move.The same holds for the splitter-algorithm for i − 1 that we use as a subroutine.Thus, the whole algorithm takes at most O(|I|) time per move.
Given Lemma 20, the remaining results now follow easily.
Theorem 21 Let be a finite constraint language that is closed under assignments.We can, for a given CSP instance I and a non-negative integer d, in time at most |D(I )| 2 (d) |I | either: 1. compute a component -backdoor tree of I of depth at most 2 O(d) , or 2. conclude that the -backdoor depth of I is larger than d.
Proof An obstruction tree of depth d is a proof that the backdoor depth is higher than d, thus for the case i = d the output of the splitter-algorithm in Lemma 20 after 2 O(d) rounds reduces to either a winning position, or a proof that the -backdoor depth of I is larger than d.The algorithm takes at most O(|I|) time per move.The statement then follows from Lemma 6.

Corollary 22
Let be a tractable constraint language that is finite and semi-conservative.The CSP can be solved in time δ Proof According to Lemma 1, the closure * of is also tractable.Furthermore, * is more permissive than and therefore depth * (I) ≤ depth (I) = d.We use Theorem 21 to compute a component * -backdoor tree of depth 2 O(d) in I and then use Lemma 2 to solve I in time δ 2 O(d) (|I |) (1) .
We would like to mention a corollary of Theorem 21 that we can derive very similarly to Corollary 22. Consider the #CSP problem, which asks for the number of satisfying assignments.A constraint language is #tractable if #CSP is solvable in polynomial time for instances from C [34].The Proof of Lemma 2 can easily be adapted to #CSP, as at a variable node, we have to add, and at a component node we have to multiply.Hence, we can substitute in the statement of Corollary 22 CSP with #CSP and tractable with #tractable.

Conclusion
In this work, we compute backdoors of bounded depth for the CSP to base classes defined via finite semi-conservative constraint languages.Our approach via obstruction trees seems to be fundamentally limited to semi-conservative languages.However, we are optimistic that our techniques can be extended to base classes of unbounded arity.A first step in this direction has already been obtained in the context of SAT for the base class of Horn formulas [27].In this setting, it is particularly interesting to consider tractable classes (of unbounded arity) of CSPs based on restrictions on the graphical structure [7][8][9], as well as hybrid restrictions [10][11][12].
Another interesting direction for future research, which has also been mentioned in the context of SAT [27], are the so-called scattered and heterogeneous extensions of (strong) backdoor sets [28,29].These extensions can be readily lifted to backdoor depth by allowing each component to be in any of a given set of (heterogeneous) tractable base classes.Interestingly, while those two notions lead to orthogonal tractable classes in the context of backdoor size, they lead to the same notion for backdoor depth.Therefore, lifting these two extensions to backdoor depth, would result in a unified and significantly more general approach.Moreover, we think that obtaining a heterogeneous version of backdoor depth seems to be particularly promising within the context of CSP.This is because, in contrast to SAT, there is a wide range of tractable classes (even of bounded arity) that can be characterized in a unified manner via algebraic properties.

Fig. 2 A 3 (
Fig.2A separator obstruction containing three paths P 1 , P 2 , and P 3 .The figure shows the vertices and edges of the incidence graph.Variables are represented by circles and constraints are represented by rectangles.Filled variables are contained in V 3 (all other variables are not) and filled rectangles are bad constraints (all other constraints are good).Only the red, blue, and green variables and edges are part of the tree of the separator obstruction, grey variables and edges are not part of the tree but are part of V 3

Lemma 17
Let be a finite constraint language that is closed under assignments and let I be a CSP instance.If I has a -separator obstruction of size at least n = (d + 2)(3 + 4(d + arity( ))) d , then I has -backdoor depth at least d.Proof Let X = P 1 , . . ., P , τ be a -separator obstruction for I of size at least n = (d + 2)(3 + 4(d + arity( ))) d and let B = {b 0 , . . ., b }.

1 .
If v / ∈ V , then the connector plays an arbitrary assignment α for v and chooses a component of J [α] containing a subtree untainted by τ J ∪ α of T containing the largest subset of B among all components of J [α]. 2. If v ∈ V , then the connector plays the assignment α(v) = τ (v) for v and chooses the component of J [α] containing a subtree of T untainted by τ J ∪ α containing the largest subset of B among all components of J [α].

( 2 •
arity( ) + d)(d + 2)(3 + 4(d + arity( ))) d rounds either: 1. a winning position, or 2. a position J such that no variable in V (J ) is contained in a constraint of P, or 3. a proof that the -backdoor depth of I is larger than d.This algorithm takes at most O(|I|) time per move.

( 2 •
arity( ) + d)(d + 2)(3 + 4(d + arity( ))) d rounds.Then, X (J ) has size at least (d + 2)(3 + 4(d + arity( ))) d because the size of the -separator obstruction increases by at least 1 after at most 2 • arity( ) + d steps.This is because every time the -separator obstruction increases by 1, we only add the at most arity( ) + d + 1 variables of at most one -bad constraint b i (because of Lemma 12) and the at most arity( ) variables of the -good constraint e i .Therefore, it follows from Lemma 17 that I has -backdoor depth at least d.