EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP CAUCHER BIRKAR Abstract. Let (X/Z, B + A) be a Q-factorial dlt pair where B, A 0 are Q-divisors and K X +B+A Q 0/Z. We prove that any LMMP/Z on K X +B with scaling of an ample divisor terminates with a good log minimal model or a Mori fibre space assuming: either a certain semi-ampleness conjecture, or the ACC conjecture on lc thresholds. An immediate corollary of this is that log flips exist for log canonical pairs up to either of those conjectures. Contents. Introduction 2. Preliminaries 4 3. Shokurov Bss-ampleness 6 4. The klt case 0 5. Proof of Theorem.3 2 5.. The vertical case 2 5.5. The horizontal case 6 6. Proof of Theorem.5 9 6.2. The vertical case 20 6.6. The horizontal case 26 References 28. Introduction We work over C. Extending results from the klt case to the lc case is often not as easy as it may sound. For example, it took some hard work to prove the cone and contraction theorem for lc pairs as done by Ambro [3] and Fujino [3]. Another major example is the finite generation of the log canonical rings: the klt case was proved in [9] but extending this to the lc case is essentially equivalent to proving the abundance conjecture. It is well-known that log flips exist for klt pairs [9]. In this paper we study the existence of log flips for log canonical pairs. Along the way, we came across the following more general statement. Date: May 4, 20. 2000 Mathematics Subject Classification: 4E30.

2 CAUCHER BIRKAR Conjecture.. Let (X/Z, B + A) be a lc pair where B, A 0 are Q-divisors, A is Q-Cartier, and the given morphism f : X Z is surjective. Assume further that K X + B + A Q 0/Z. Then, () (X/Z, B) has a Mori fibre space or a log minimal model (Y/Z, B Y ), (2) if K Y + B Y is nef/z, then it is semi-ample/z; (3) if (X/Z, B) is Q-factorial dlt, then any LMMP/Z on K X +B with scaling of an ample/z Q-divisor terminates. Note that the surjectivity of f is obviously not necessary but we have put there for practical convenience. A statement similar to Conjecture. appeared as a conjecture in a recent paper of Kollár [20] in the context of studying singularities. However, we arrived at this statement independently and our motivation mainly came from studying log flips for log canonical pairs. The main result of this paper is that the above conjecture follows from either a certain semi-ampleness conjecture, or Shokurov s ACC conjecture on lc thresholds. Conjecture.2. Let (X/Z, B) be a Q-factorial dlt pair, and T := B where B is a Q-divisor and Z is affine. Suppose that K X + B is nef/z, (K X + B) S is semi-ample/z for each component S of T, K X + B ɛp is semi-ample/z for some Q-divisor P 0 with Supp P = T and for any sufficiently small rational number ɛ > 0. Then, K X + B is semi-ample/z. Actually, we need a very special case of this conjecture (see the proof of Theorem 5.4). If Z is a point, then the conjecture is already known as one can use Fujino-Gongyo [6, Theorem 3.] to prove the semi-ampleness of (K X +B) T and use Kollár s injectivity theorem to lift sections from T. So, one only needs to work out the latter two results in the relative situation and this seems quite within reach. Theorem.3. Conjecture.2 in dimension d implies Conjecture. in dimension d. This theorem should be considered as the conceptual approach to.. However, similar arguments also work if we replace the semi-ampleness conjecture with the ACC conjecture on lc thresholds. Conjecture.4 (ACC for lc thresholds). Suppose that Γ [0, ] and S Q are finite sets of rational numbers, and d a natural number. Then, the set {lct(m, X, B) (X, B) is lc of dimension d} satisfies the ACC where the coefficients of B belong to Γ, M is a Q-Cartier divisor with coefficients in S, and lct(m, X, B) is the lc threshold of M with respect to (X, B). The conjecture can be formulated in more general settings but we need this weaker form (actually, even weaker versions). Ein and Mustaţă [] proved the conjecture for smooth varieties (and more generally varieties with quotient singularities). Hacon, M c Kernan, and Xu recently announced that they have

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP 3 solved the conjecture in which case we could remove it from the assumptions of the next theorem hence get an unconditional proof of Conjecture.. Theorem.5. Conjecture.4 in dimension d implies Conjecture. in dimension d. As mentioned earlier, an immediate consequence of the above theorems concerns the existence of log flips for lc pairs. Corollary.6. Assume Conjecture.2 or Conjecture.4 in dimension d. Let (X/Z, B) be a lc pair of dimension d and X Y/Z an extremal K X + B- negative flipping contraction. Then, the log flip of X Y exists. In section 3, we prove the following which is similar to a result of Fujino [2]. Theorem.7. (=Theorem 3.4) Let (X/Z, B) be a Q-factorial dlt pair such that K X + B R M/Z with M 0 very exceptional/z. Then, any LMMP/Z on K X +B with scaling of an ample divisor terminates with a model Y on which K Y + B Y R M Y = 0/Z. In the same section, we prove that under finite generation (in particular in the klt case) a more general version of Conjecture. holds: Theorem.8. (=Theorem 3.7) Let (X/Z, B) be a Q-factorial dlt pair where B is a Q-divisor and f : X Z is surjective. Assume further that R(X/Z, K X + B) is a finitely generated O Z -algebra, and that (K X + B) Xη Q 0 where X η is the generic fibre of f. Then, any LMMP/Z on K X + B with scaling of an ample divisor terminates with a good log minimal model. We could generalise this by replacing (K X + B) Xη Q 0 with: existence of a good log minimal model on the generic fibre. We have a similar situation in some of the other statements below. In section 4, we study the klt case beyond the statement of.. In section 5, we give the proof of Theorem.3 by proving several inductive theorems. In section 6, we give the proof of Theorem.5. The proof goes through the following more general theorem: Theorem.9. (=Theorem 6.) Assume Conjecture.4 in dimension d. Let (X/Z, B) be a lc pair of dimension d where B is a Q-divisor and the given morphism f : X Z is surjective. Assume further that K X + B Q 0 over some non-empty open subset U Z, and if η is the generic point of a lc centre of (X/Z, B), then f(η) U. Then, () (X/Z, B) has a log minimal model (Y/Z, B Y ), (2) K Y + B Y is semi-ample/z; (3) if (X/Z, B) is Q-factorial dlt, then any LMMP/Z on K X +B with scaling of an ample/z divisor terminates. In some places we make use of the following result which is of independent interest.

4 CAUCHER BIRKAR Theorem.0. (=Theorem 5.6) Let f : X Y/Z be a surjective morphism of normal varieties, projective over an affine variety Z = Spec R, and L a Cartier divisor on Y. If R(X/Z, f L) is a finitely generated R-algebra, then R(Y/Z, L) is also a finitely generated R-algebra. Proof. (of Corollary.6) First assume that B has rational coefficients. Since (K X + B) is ample/y, we can find 0 A Q (K X + B)/Y such that (X/Z, B + A) is lc and K X + B + A Q 0/Y. Now by Theorem.3 or Theorem.5, (X/Y, B) has a good log minimal model and its lc model gives the flip of X Y. When B is not rational, we can find rational lc divisors K X + B i and real numbers r i > 0 such that K X + B = r i (K X + B i ) and r i =. Moreover, we can assume that (K X + B i ) is ample/y for every i. For each i, j there is a rational number b i,j such that K X +B i b i,j (K X +B j )/Y. By the cone theorem for lc pairs proved by Ambro [3] and Fujino [3], K X + B i Q b i,j (K X + B j )/Y. Therefore, there is a morphism X + Y which gives the flip of K X + B i for every i. The morphism also gives the K X + B-flip. Remark. Assume Conjecture.2 or Conjecture.4 in dimension d. Let (X/Z, B) be a lc pair of dimension d. In view of Ambro [3] and Fujino [3], and Corollary.6, we can run the LMMP/Z on K X + B. Termination of such an LMMP can be reduced to the Q-factorial dlt case. We remark that it should not be difficult to generalise Theorem.3 and Theorem.5 to the case when B, C are R-divisors. One would have to use appropriate rational approximations to reduce to the rational case. Finally, we briefly mention some previous works on flips. Mori proved the existence of flips for 3-folds with terminal singularities [23]. Shokurov proved it in full generality in dimension 3 [27][26], in dimension 4 in the klt case [25], and also worked out a significant proportion of what we know about flips in every dimension. Hacon-M c Kernan [7] filled in the missing parts of Shokurov s program on pl flips. Birkar-Cascini-Hacon-M c Kernan [9] proved the problem in the klt case in all dimensions (see also [0]). Fujino [4] proved it for lc pairs in dimension 4. Alexeev-Hacon-Kawamata [] and Birkar [8] proved it in dimension 5 in the klt case. 2. Preliminaries We work over k = C. A pair (X/Z, B) consists of normal quasi-projective varieties X, Z over k, an R-divisor B on X with coefficients in [0, ] such that K X + B is R-Cartier, and a projective morphism X Z. For a prime divisor D on some birational model of X with a nonempty centre on X, a(d, X, B) denotes the log discrepancy. Let (X/Z, B) be a lc pair. By a K X +B-flip/Z we mean the flip of a K X +Bnegative extremal flipping contraction/z. A sequence of log flips/z starting with (X/Z, B) is a sequence X i X i+ /Z i in which X i Z i X i+ is a K Xi +B i - flip/z, B i is the birational transform of B on X, and (X /Z, B ) = (X/Z, B). Special termination means termination near B.

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP 5 Definition 2. (Weak lc and log minimal models) A pair (Y/Z, B Y ) is a log birational model of (X/Z, B) if we are given a birational map φ: X Y/Z and B Y = B + E where B is the birational transform of B and E is the reduced exceptional divisor of φ, that is, E = E j where E j are the exceptional/x prime divisors on Y. A log birational model (Y/Z, B Y ) is a weak lc model of (X/Z, B) if K Y + B Y is nef/z, and for any prime divisor D on X which is exceptional/y, we have a(d, X, B) a(d, Y, B Y ) A weak lc model (Y/Z, B Y ) is a log minimal model of (X/Z, B) if (Y/Z, B Y ) is Q-factorial dlt, the above inequality on log discrepancies is strict. Definition 2.2 (Mori fibre space) A log birational model (Y/Z, B Y ) of a lc pair (X/Z, B) is called a Mori fibre space if (Y/Z, B Y ) is Q-factorial dlt, there is a K Y + B Y -negative extremal contraction Y T/Z with dim Y > dim T, and a(d, X, B) a(d, Y, B Y ) for any prime divisor D (on birational models of X) and strict inequality holds if D is on X and contracted/y. Definition 2.3 (Log smooth model) We need to define various kinds of log smooth models. Let (X/Z, B) be a lc pair, and let g : W X be a log resolution. Let B W 0 be a boundary on W so that K W + B W = g (K X + B) + E where E 0 is exceptional/x and the support of E contains each prime exceptional/x divisor D on W if a(d, X, B) > 0. We call (W/Z, B W ) a log smooth model of (X/Z, B). However, in practice we usually need further assumptions. We list the ones we will need: Type (): We take B W to be the birational transform of B plus the reduced exceptional divisor of g, that is, we assume that a(d, W, B W ) = 0 for each prime exceptional/x divisor D on W. Type (2): We assume that a(d, W, B W ) > 0 if a(d, X, B) > 0, for each prime exceptional/x divisor D on W. Type (3): here (X/Z, B) is klt, and we assume that (W/Z, B W ) is also klt. Remark 2.4 A log smooth model (W/Z, B W ) of (X/Z, B) has the property that any log minimal model of (W/Z, B W ) is also a log minimal model of (X/Z, B). Indeed let (Y/Z, B Y ) be a log minimal model of (W/Z, B W ). Let e: V W and h: V Y be a common resolution. Then, e (K W + B W ) = h (K Y + B Y ) + G where G 0 is exceptional/y [8, Remark 2.6]. Thus, e g (K X + B) = h (K Y + B Y ) + G e E. Since G e E is antinef/x, and g e (G e E) 0, by the negativity lemma, G e E 0. Now if D is a prime divisor on X which is exceptional/y, then a(d, X, B) = a(d, W, B W ) < a(d, Y, B Y ). On the other hand, if D is a prime divisor on Y which is exceptional/x, then a(d, Y, B Y ) = 0 otherwise D is not exceptional/w and a(d, X, B) > a(d, W, B W ) = a(d, Y, B Y ) which contradicts G e E 0. Moreover, if

6 CAUCHER BIRKAR D is a prime divisor on Y which is not exceptional/x, then a(d, X, B) = a(d, W, B W ) = a(d, Y, B Y ). So, B Y is the birational transform of B plus the reduced exceptional divisor of Y X hence (Y/Z, B Y ) is a log minimal model of (X/Z, B). Definition 2.5 (LMMP with scaling) Let (X /Z, B + C ) be a lc pair such that K X + B + C is nef/z, B 0, and C 0 is R-Cartier. Suppose that either K X + B is nef/z or there is an extremal ray R /Z such that (K X + B ) R < 0 and (K X + B + λ C ) R = 0 where λ := inf{t 0 K X + B + tc is nef/z} If R defines a Mori fibre structure, we stop. Otherwise assume that R gives a divisorial contraction or a log flip X X 2. We can now consider (X 2 /Z, B 2 + λ C 2 ) where B 2 + λ C 2 is the birational transform of B + λ C and continue. That is, suppose that either K X2 +B 2 is nef/z or there is an extremal ray R 2 /Z such that (K X2 + B 2 ) R 2 < 0 and (K X2 + B 2 + λ 2 C 2 ) R 2 = 0 where λ 2 := inf{t 0 K X2 + B 2 + tc 2 is nef/z} By continuing this process, we obtain a sequence of numbers λ i and a special kind of LMMP/Z which is called the LMMP/Z on K X + B with scaling of C. When we refer to termination with scaling we mean termination of such an LMMP. We usually put λ = lim λ i. By a contraction f : X Z we mean a projective morphism of quasiprojective varieties with f O X = O Z. Definition 2.6 Let f : X Z be a contraction of normal varieties, and D an R-divisor on X. Define the divisorial sheaf algebra of D as R(X/Z, D) = m 0 f O X ( md ) where m Z. If Z is affine, we take R(X/Z, D) to be the global sections of R(X/Z, D). 3. Shokurov Bss-ampleness Most of the ideas in this section are explicit or implicit in Shokurov [25]. However, we would like to give full details here. Definition 3. (cf. Shokurov [25, Definition 3.2]) Let f : X Y be a contraction of normal varieties, D an R-divisor on X, and V X a closed subset. We say that V is vertical over Y if f(v ) is a proper subset of Y. We say that D is very exceptional/y if D is vertical/y and for any prime divisor P on Y there is a prime divisor Q on X which is not a component of D but f(q) = P, i.e. over the generic point of P we have: Supp f P Supp D.

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP 7 Lemma 3.2 (cf. Shokurov [25, Lemma 3.9]). Let f : X Y be a contraction of normal varieties, projective over a normal variety Z. Let A be an ample/z divisor on Y and F = f A. If E 0 is a divisor on X which is vertical/y and such that me = Fix(mF + me) for every integer m 0, then E is very exceptional/y. Proof. We can assume that A is very ample/z. Since E is effective, for each integer l > 0, we have the natural exact sequence 0 O X O X (le) given by the inclusion O X O X (le). Since mf = Mov(mF + le) for each m 0, the induced homomorphism H 0 (X, le + mf ) m Z H 0 (X, mf ) m Z of R(X/Z, F )-modules is an isomorphism in large degrees. This in turn induces a homomorphism H 0 (Y, (f O X )(ma)) m Z H 0 (Y, ma) = m Z m Z H 0 (Y, (f O X (le))(ma)) of R(Y/Z, A)-modules which is an isomorphism in large degrees. Therefore, by [8, II, Proposition 5.5 and Exercise 5.9], the morphism O Y = f O X f O X (le) is surjective hence we get the equality O Y = f O X = f O X (le). Now let P be a prime divisor on Y and let U be a smooth open subset of Y so that O U O U (P U ), P U is Cartier, and each component of f P U maps onto P U. Let V = f U. Suppose that Supp f P U Supp E V. Then, for some l > 0, f P U le V. But then, O U O U (P U ) f O V (le V ) = O U, a contradiction. Therefore, E is very exceptional/y. We need the following general negativity lemma of Shokurov [25, Lemma 3.22] (see also Prokhorov [24, Lemma.7]). Let S Z be a projective morphism of varieties and M an R-Cartier divisor on S. We say that M is nef on the very general curves of S/Z if there is a countable union Λ of proper closed subsets of S such that M C 0 for any curve C on S contracted over Z satisfying C Λ. Lemma 3.3 (Negativity). Let f : X Z be a contraction of normal varieties. Let D be an R-Cartier divisor on X written as D = D + D with D +, D 0 having no common components. Assume that D is very exceptional/z, and that for each component S of D, D is nef on the very general curves of S/Z. Then, D 0. Proof. Assume that D 0 otherwise there is nothing to prove, and let P = f(supp D ). By shrinking Z if necessary we can assume that P is irreducible and that every component of D maps onto P. By assumptions, there is a countable union Λ of codim 2 proper closed subsets of X such that Λ Supp D and such that for any component S of D and any curve C of S/Z satisfying C Λ, we have D C 0.

8 CAUCHER BIRKAR Assume that dim P > 0. Let Z be a general hyperplane section of Z, X = f Z, and let f be the induced contraction X Z. Since Z is general, it intersects P, and X does not contain any component of D nor any component of Λ. So, Λ X is again a countable union of codim 2 proper closed subsets of X. Since X is general, it does not contain any component of the singular locus of X hence D X is determined on the smooth locus of X and the negative part of D X is given by D X defined on the smooth locus. Moreover, the choice of X ensures that it does not contain any component of S S 2 if S, S 2 are prime divisors mapping onto P. Now, if T is any component of D X with negative coefficient, then T is a component of S X for some component S of D hence D X is nef on the curves of T/Z outside Λ X (note that by construction T Λ). Moreover, the negative part of D X, say D X, is very exceptional/z : if codim P 2, the claim is trivial; if codim P =, we can assume that Z is smooth at every point of P ; now since D is very exceptional/z, there is a component S of f P mapping onto P such that S is not a component of D ; let T be a component of S X having codimension one in X ; then, by our choice of X, T is not a component of D X which means that D X is very exceptional/z as claimed. Now by induction on dimension applied to X Z, D X = 0 which is a contradiction. So, from now on we can assume that dim P = 0. If dim Z =, then f P is a divisor which is numerically zero over Z. Let t be the smallest real number such that D + tf P 0. Then, there is a component S of D which has coefficient zero in D +tf P 0 but such that S intersects Supp(D +tf P ). If C is any curve on S which is not inside S Supp(D + tf P ) but such that C intersects Supp(D + tf P ), then D C = (D + tf P ) C > 0. This contradicts our assumptions. On the other hand, if dim X = dim Z = 2, the lemma is quite well-known and elementary and it can be proved similarly. So, from now on we assume that dim Z 2, and dim X 3. Let H be a general hyperplane section of X and let g : H G be the contraction given by the Stein factorisation of H Z. Since f(d ) is just a point and dim Z 2 and dim X 3, it is trivial that D H the negative part of D H is very exceptional/g. Moreover, arguments similar to the above show that if T is any component of D H, then D H is nef on the very general curves of T/G. So, we can apply induction to deduce that D H = 0 which gives a contradiction. The negativity lemma implies the following. Theorem 3.4. Let (X/Z, B) be a Q-factorial dlt pair such that K X + B R M/Z with M 0 very exceptional/z. Then, any LMMP/Z on K X + B with scaling of an ample divisor terminates with a model Y on which K Y + B Y R M Y = 0/Z. Proof. Assume that C 0 is an ample/z R-divisor such that K X + B + C is lc and nef/z. Run the LMMP/Z on K X + B with scaling of C. The only divisors that can be contracted are the components of M hence M remains very

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP 9 exceptional/z during the LMMP. We may assume that the LMMP consists of only a sequence X i X i+ /Z i of K X + B-flips/Z with X = X. If λ i are the numbers appearing in the LMMP, and λ := lim λ i, then by [9], we may assume that λ = 0. Since λ = 0, K X + B is a limit of movable/z R-divisors hence for any prime divisor S on X, (K X + B) Γ = M Γ 0 for the very general curves Γ of S/Z. Now by assumptions M is very exceptional/z hence by the Negativity Lemma 3.3, M 0 which implies that M = 0. The previous theorem implies the existence of Q-factorial dlt blowups (cf. Fujino [2, Theorem 4.]). Corollary 3.5. Let (X/Z, B) be a lc pair. blowup of (X/Z, B). Then, there is a Q-factorial dlt Proof. Let (W/Z, B W ) be a log smooth model of (X/Z, B) of type () as in Definition 2.3 constructed via a log resolution f : W X. Then, K W + B W = f (K X + B) + E where E 0 is exceptional/x. Now, by Theorem 3.4 some LMMP/X on K W + B W with scaling of an ample divisor terminates with a model Y on which K Y + B Y R 0/X. If g : Y X is the induced morphism, then K Y + B Y = g (K X + B) because E is contracted over Y. Now (Y/Z, B Y ) is the desired Q-factorial dlt blowup. Corollary 3.6. Let (X/Z, B) be a lc pair. If it has a weak lc model, then it has a log minimal model. Proof. Assume that (Y /Z, B Y ) is a weak lc model of (X/Z, B). By replacing (X/Z, B) with a log smooth model over Y, we can assume that the given map φ: X Y is a morphism, and that B Y = φ B. Now, K X + B = φ (K Y + B Y ) + E where E 0 is exceptional/y. By Theorem 3.4 some LMMP/Y on K X + B X with scaling of an ample divisor terminates with a model Y on which K Y + B Y R 0/Y which gives the desired log minimal model. The next theorem shows that a more general version of Conjecture. holds when (X/Z, B) has a finitely generated lc algebra. The result is an easy consequence of Lemma 3.2 and Lemma 3.3 and it should be considered as a very special case of Shokurov s attempt in relating bss-ampleness and finite generation [25, Theorem 3.8](see also [22][2] for more recent adaptations). Though we are not using Shokurov s terminology of bss-ample divisors but the next theorem is saying that finite generation implies bss-ampleness in the specific situation we are concerned with. Shokurov proves that in general finite generation together with the so-called global almost generation property implies bss-ampleness. Theorem 3.7. Let (X/Z, B) be a Q-factorial dlt pair where B is a Q-divisor and f : X Z is surjective. Assume further that R(X/Z, K X + B) is a finitely generated O Z -algebra, and that (K X + B) Xη Q 0 where X η is the generic fibre of f. Then, any LMMP/Z on K X + B with scaling of an ample divisor terminates with a good log minimal model.

0 CAUCHER BIRKAR Proof. We can assume that f is a contraction. Run an LMMP/Z on K X + B with scaling of some ample/z divisor. Since termination and semi-ampleness/z are both local on Z, we can assume that Z is affine, say Spec R. By [7], the claim follows if we prove that (X/Z, B) has a good log minimal model. Let I be a positive integer so that I(K X + B) is Cartier. Thus, perhaps after replacing I with a multiple, there is a resolution g : W X, a divisor E 0 and a free divisor F on W such that Fix g mi(k X + B) = me and Mov g mi(k X + B) = mf for every m > 0 (cf [25] or [4, Theorem 4.3]). Let h: W T/Z be the contraction defined by F. Then, since (K X + B) Xη Q 0, the map T Z is birational and E is vertical/t. Choose a boundary B W so that (W/Z, B W ) is a log smooth model of (X/Z, B) of type () as in Definition 2.3. We can write where E 0 is exceptional/x. So, I(K W + B W ) = g I(K X + B) + E Fix mi(k W + B W ) = me + me and Mov mi(k W + B W ) = mf Run the LMMP/T on K W +B W with scaling of some ample/t divisor. Over the generic point of T, a suitable Q-factorial dlt blowup of (X/Z, B) is a log minimal model of (W/Z, B W ). Thus, by [7], the LMMP terminates over the generic point of T hence in particular, we reach a model Y on which E Y + E Y is vertical/t. By Lemma 3.2, E Y +E Y is very exceptional/t. Now, by Theorem 3.4, there is an LMMP/T on K Y + B Y which ends up with a log minimal model (Y/T, B Y ) on which K Y +B Y Q 0/T. In particular, E Y +E Y = 0 and I(K Y +B Y ) Q F Y. Thus, K Y + B Y is semi-ample/z and (Y/Z, B Y ) is a good log minimal model of (W/Z, B W ) hence of (X/Z, B). 4. The klt case Theorem 4.. Conjecture. holds when (X/Z, B) is klt. Proof. Suppose that (X/Z, B) is klt under the assumptions of Conjecture.. If A is not vertical/z, then K X + B is not pseudo-effective/z hence (X/Z, B) has a Mori fibre space by [9]. So, we can assume that A is vertical/z. By [9], R(X/Z, K X + B) is a finitely generated O Z -algebra hence we can apply Theorem 3.7 in this case. Note that if (X/Z, B) is not Q-factorial we can reduce the problem to the Q-factorial situation. Though the last theorem settles the klt case of Conjecture. but we prove further results in this direction as we will need them to deal with the lc case. Theorem 4.2. Let (X/Z, B) be a klt pair where f : X Z is surjective. Assume further that there is a contraction g : X S/Z such that K X +B R 0/S and that S Z is generically finite. Then, (X/Z, B) has a good log minimal model. (Note that B is not necessarily a Q-divisor.)

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP Proof. We can assume that f is a contraction hence S Z is birational. Moreover, by taking a Q-factorialisation we can assume that X is Q-factorial. Since K X + B R 0/S, by Ambro [2] and Fujino-Gongyo [6, Theorem 3.], there is a boundary B S on S such that K X + B R g (K S + B S ) and such that (S/Z, B S ) is klt and of general type. So, by [9], there is a log minimal model (T/Z, B T ) for (S/Z, B S ) obtained by running an LMMP with scaling. Moreover, since T S does not contract divisors, there are open subsets U S and V T such that the induced map U V is an isomorphism and such that codim(t \ V ) 2. Take a log smooth model (W/Z, B W ) of (X/Z, B) of type (3) as in Definition 2.3 using a log resolution e: W X. Then, K W +B W = e (K X +B)+E where E 0 is exceptional/x such that any log minimal model of (W/Z, B W ) is also a log minimal model of (X/Z, B). Moreover, we can assume that W dominates S and T. Run an LMMP/T on K W + B W with scaling of some ample divisor. The LMMP terminates over the generic point of T. Actually, it terminates over V on some model Y, i.e. we reach a model Y such the remaining steps of the LMMP would all be over T \ V. Let φ: W S, ψ : W T, and π : W Y be the induced maps. Let M X, N W, and O Y be the inverse images of V. Then, (M/V, B M ) is a log minimal model of (N/V, B W N ) hence M and O are isomorphic in codimension one. Thus, E N is contracted over O. So, in particular, E Y = π E is a divisor on Y which is mapped into T \ V hence it is very exceptional/t. On the other hand, φ (K S + B S ) = ψ (K T + B T ) + G where G 0 is very exceptional/t (G is contracted to a closed subset of T of codimension 2). Thus, K Y + B Y = π (K W + B W ) = π (e (K X + B X ) + E) = π e (K X + B X ) + π E R π φ (K S + B S ) + π E = π ψ (K T + B T ) + π G + π E By construction, π G + π E 0 is very exceptional/t. So, by Theorem 3.4, there is an LMMP/T on K Y +B Y with scaling of some ample divisor which ends up with a model Y on which K Y +B Y R 0/T hence K Y +B Y R ν (K T +B T ) where ν : Y T is the induced morphism. In particular, K Y + B Y is semiample/z. So, (Y/Z, B Y ) is a good log minimal model of (W/Z, B W ) hence of (X/Z, B). Theorem 4.3. Let (X/Z, B) be a klt pair where B is a Q-divisor and f : X Z is surjective. Assume further that (K X +B) F Q 0 for the generic fibre F of f. If K X + B + H is klt and nef/z for some Q-divisor H 0, then the LMMP/Z on K X + B with scaling of H terminates if either () B is big/z or H is big/z, or (2) H is vertical/z and λ λ j for any j where λ j are the numbers appearing in the LMMP with scaling (as in Definition 2.5) and λ = lim λ i.

2 CAUCHER BIRKAR Proof. We can assume that f is a contraction. By Theorem 3.7, (X/Z, B) has a good log minimal model. Assume that condition () holds. If B is big/z, the termination follows from [9]. If H is big/z, and if λ > 0, then again the termination follows from [9] as we can replace B with B + λh. If λ = 0, use [7]. Now assume that condition (2) holds. So, from now on we assume that H is vertical/z and λ λ j for any j. We may assume that the LMMP consists of only a sequence X i X i+ /Z i of log flips and that X = X. Now K X +B+λ H is nef/z and actually semi-ample/z by Theorem 3.7 (note that λ i are rational numbers by [8, Lemma 3.]). Let g : X T be the contraction associated to it. Since H is vertical/z, T Z is birational so the generic fibre of g and f are the same. By (), there is an LMMP/T on K X + B which ends up with a weak lc model (Y/T, B Y ) of (X/T, B) such that K Y + B Y is semiample/t. Let h: Y S/T be the contraction associated to K Y + B Y. Since, K X + B + λ H Q 0/T, K Y + B Y + λ H Y Q 0/S. This combined with K Y +B Y Q 0/S implies that H Y Q 0/S. Therefore, K Y +B Y +λh Y R 0/S. So, by Theorem 4.2, (Y/Z, B Y + λh Y ) has a log minimal model which is also a log minimal model of (X/Z, B + λh). Now the termination follows from [7]. 5. Proof of Theorem.3 5.. The vertical case. In this subsection we deal with Theorem.3 in the vertical case, i.e. when every lc centre of (X/Z, B) is vertical/z, in particular, when X Z is birational. Lemma 5.2. Let (X/Z, B + A) be a lc pair of dimension d such that B, A 0 are Q-divisors, f : X Z is surjective, and K X + B + A Q 0/Z. Assume further that (X/Z, B) is Q-factorial dlt, K X + B + C is lc and nef/z for some Q-divisor C 0, C Q P/Z for some P 0 with Supp P = Supp B, A and C are vertical/z. Then, assuming Conjecture. in dimension d, there is an LMMP/Z on K X + B with scaling of C which terminates. Proof. Step. Since C Q P/Z, for each rational number t (0, ] we can write K X + B + tc Q K X + B ɛp + (t ɛ)c/z for some sufficiently small rational number ɛ > 0. Under our assumptions, (X, B ɛp + (t ɛ)c) is klt. In particular, by Theorem 3.7, K X + B + tc is semi-ample/z if K X + B + tc is nef/z. We will use this observation on X and on the birational models that will be constructed. Step 2. Put Y := X, B := B, and C := C. Let λ 0 be the smallest number such that K Y + B + λ C is nef/z (note that λ is rational by [8,

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP 3 Lemma 3.]). We may assume that λ > 0. By Step, K Y +B +λ C is semiample/z; let Y V /Z be the associated contraction. Run the LMMP/V on K Y + B + tc with scaling of some ample divisor, for some rational number t (0, λ ). This terminates by Step and by Theorem 4.3. The LMMP is also an LMMP/V on K Y + B with scaling of λ C because K Y + B + tc Q (λ t)c Q (λ t) λ (K Y + B )/V So, we get a model Y 2 such that K Y2 + B 2 is semi-ample/v where B 2 is the birational transform of B (similar notation will be used for other divisors and models). Now since K Y2 + B 2 + λ C 2 is the pullback of some ample/z divisor on V, K Y2 + B 2 + λ C 2 + δ(k Y2 + B 2 ) is semi-ample/z for some sufficiently small δ > 0. Put it in another way, K Y2 + B 2 + τc 2 is semi-ample/z for some rational number τ < λ. We can consider Y Y 2 as a partial LMMP/Z on K Y + B with scaling of C. We can continue as before. That is, let λ 2 0 be the smallest number such that K Y2 + B 2 + λ 2 C 2 is nef/z, and so on. This process is an LMMP/Z on K X + B with scaling of C. Step 3. In Step 2, we constructed an LMMP/Z on K X + B with scaling of C such that the numbers λ i that appear in the LMMP satisfy λ := lim λ i λ j for any j. In this case, actually λ = 0 otherwise we can consider the LMMP as an LMMP/Z on K X + B + λ C with scaling of ( λ )C for some rational number λ (0, λ), and Step and Theorem 4.3 give a contradiction. Perhaps after modifying the notation, we may assume that the LMMP consists of only a sequence X i X i+ /Z i of log flips and that λ i corresponds to the contraction X i Z i as in Definition 2.5; that is, K Xi + B i + λ i C i is nef/z but numerically trivial/z i, and X = X. Step 4. Since C Q P/Z for some P 0 with Supp P = Supp B, for each i, some component of B i intersects R i negatively where R i is the extremal ray defining X i Z i. Using standard special termination arguments the termination is reduced to termination in lower dimensions. Indeed, Let S be a component of B and S i its birational transform on X i. It is well-known that the induced map S i S i+ /T i is an isomorphism in codimension one if i 0 where T i is the normalisation of the image of S i in Z i. So, we may assume that these maps are all isomorphisms in codimension one for any i. In general, S i S i+ /T i is not a K Si +B Si -flip where K Si +B Si := (K Xi +B i ) Si. However, (S, B S ) is dlt and we can take a Q-factorial dlt blowup (S, B S ) of (S, B S ) such that S S is small (this follows from the proof of Corollary 3.5 and definition of dlt pairs). Now, K S + B S + λ i C S Q 0/T where C S := C S and C S is the birational transform of C S. By Conjecture. in dimension d, there is an LMMP/T on K S + B S with scaling of some ample divisor which terminates. The outcome is a model S 2 on which K S 2 +B S 2 is semi-ample/t. Since K S2 +B S2 is ample/t, S 2 maps to S 2 and K S 2 + B S 2 is the pullback of K S2 + B S2. Continuing this process, we lift the sequence to an LMMP/T on K S + B S with scaling of C S where T is the normaliation of the image of S in Z.

4 CAUCHER BIRKAR Step 5. On the other hand, K S + B S + A S Q 0/T where A Si := A i Si and A S i is the birational transform of A Si. So, Conjecture. in dimension d implies that (S /T, B S ) has a log minimal model because K S + B S is pseudoeffective/t. Now termination of the LMMP/T on K S + B S with scaling of C S in Step 3 follows from [7]. Step 6. The LMMP of Step 3 terminates near B hence it terminates near C which implies that it terminates everywhere. Theorem 5.3. Conjecture. in dimension d implies Conjecture. ()(3) in dimension d when every lc centre of (X/Z, B) is vertical/z. Proof. Step. After taking a Q-factorial dlt blowup using Corollary 3.5 we may assume that (X/Z, B) is Q-factorial dlt, in particular, B is vertical/z. Run an LMMP/Z on K X + B with scaling of an ample/z divisor. If K X + B is not pseudo-effective/z, then the LMMP ends up with a Mori fibre space by [9]. So, from now on we assume that K X + B is pseudo-effective/z. Since K X +B Q A/Z, A F = 0 where F is the generic fibre of f : X Z. Thus, A is vertical/z and (K X + B) F Q 0. By [7], it is enough to prove that (X/Z, B) has a log minimal model, locally over Z. Step 2. Let t > t 2 > be a sequence of sufficiently small rational numbers with lim i + t i = 0. Each (X/Z, B t i B ) is klt and (K X +B t i B ) F Q 0 hence by Theorem 4.3 each (X/Z, B t i B ) has a good log minimal model (Y i /Z, B Yi t i B Yi ) obtained by running some LMMP/Z with scaling on K X + B t i B. Thus, Y i X does not contract divisors. Moreover, Supp B Yi contains all the lc centres of (Y i /Z, B Yi ) because (Y i /Z, B Yi t i B Yi ) is klt hence (Y i /Z, B Yi ) is klt outside Supp B Yi. Now, since A and B are vertical/z, there are vertical/z Q-divisors M, N 0 such that K X + B Q A Q M/Z and B Q N/Z. So, any prime divisor contracted by X Y i is a component of M +N hence after replacing the sequence with a subsequence we can assume that the morphisms X Y i contract the same divisors, i.e. Y i are isomorphic in codimension one. Step 3. First assume that B Y = 0. Then, since K Y + B Y t B Y is semi-ample/z, and since K Y + B Y + A Y is lc, K Y + B Y t B Y Q D Y /Z for some D Y 0 such that K Y + B Y + A Y + D Y which implies that C Y := A Y + D Y Q t B Y /Z is lc. Note that K Y + B Y + t i t C Y Q K Y + B Y t i B Y /Z So, (Y i /Z, B Yi + t i t C Yi ) is a weak lc model of (Y /Z, B Y + t i t C Y ). Let (Y /Z, B Y ) be a Q-factorial dlt blowup of (Y /Z, B Y ), and let C Y on Y be the birational transform of C Y which is the same as the pullback of C Y since C Y does not contain any lc centre of (Y /Z, B Y ). Then, K Y + B Y + C Y is the pullback of K Y + B Y + C Y hence it is nef/z. Since C Y Q t B Y /Z

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP 5 and since B Y contains all the lc centres of (Y /Z, B Y ), C Y Q P /Z for some P 0 with Supp P = Supp B Y. Moreover, by construction B Y is vertical/z hence C Y is vertical/z which in turn implies that C Y is vertical/z. Similarly, A Y, the pullback of A Y is vertical/z. Now by Lemma 5.2, there is an LMMP/Z on K Y + B Y with scaling of C Y which terminates on a model Y on which K Y + B Y + δc Y is nef/z for any sufficiently small δ 0. In particular, (Y/Z, B Y + t i t C Y ) is a log minimal model of (Y /Z, B Y + t i t C Y ) and (Y i /Z, B Yi + t i t C Yi ) for i 0. Step 4. Let g : W X, h: W Y, and e: W Y i be a common resolution. Then, g (K X + B t i B ) = e (K Yi + B Yi t i B Yi ) + E i and e (K Yi + B Yi + t i C Yi ) = h (K Y + B Y + t i C Y ) t t where E i 0 [5, Remark 2.4] (note that we are using the fact that (Y/Z, B Y + t i t C Y ) and (Y i /Z, B Yi + t i t C Yi ) are both weak lc models of (Y /Z, B Y + t i t C Y )). If we put L i := E i + g t i B + h t i C Y e t i B Yi e t i C Yi t t then L i = g (K X + B) h (K Y + B Y ) which in particular means that g L i is independent of i. Since C Y Q t B Y /Z, we can write C Y + t B Y = a(α) + Q where (α) is the divisor of some rational function α, a Q, and Q is the pullback of a Q-Cartier divisor R on Z. In particular, g e (C Yi + t B Yi ) = g e (a(α) + Q i ) is independent of i where Q i is the birational transform of Q and again (α) is the divisor (on Y i ) of the rational function α. Step 5. Now, since lim i + t i = 0, we have g L i = lim i + g L i = lim g (E i + g t i B + h t i C Y ) lim i + t g (e t i B Yi + e t i C Yi ) i + t t i lim g e (a(α) + Q i ) = 0 i + t On the other hand, L i is nef/x. Therefore, by the negativity lemma, L i 0. We compare B Y and B: let D be a component of B Y ; if D is not exceptional/y, then D is the birational transform of a component of B Y hence the birational transform of a component of B having the same coefficient; if D is not exceptional/x but exceptional/y, then D appears in B Y with coefficient one hence the property L i 0 ensures that D also appears with coefficients one in B; on the other hand, if D is any prime divisor on Y which is exceptional/x then it is also exceptional/y hence D appears in B Y with coefficient one.

6 CAUCHER BIRKAR Thus, B Y is the birational transform of B plus the reduced exceptional divisor of Y X. Therefore, (Y/Z, B Y ) is a weak lc model of (X/Z, B) and by Corollary 3.6 we can construct a log minimal model of (X/Z, B). Step 6. If B Y = 0, then by taking Y = Y, similar but simpler calculations show that (Y/Z, B Y ) is already a good log minimal model of (X/Z, B). Theorem 5.4. Conjecture. in dimension d and Conjecture.2 in dimension d imply Conjecture. in dimension d when every lc centre of (X/Z, B) is vertical/z. Proof. By Theorem 5.3, the claims () and (3) of Conjecture. hold for (X/Z, B), in particular, (X/Z, B) has a Mori fibre space or a log minimal model (Y/Z, B Y ). Assume that (Y/Z, B Y ) is a log minimal model. We only need to prove that K Y + B Y is semi-ample/z. So, by replacing (X/Z, B) with (Y/Z, B Y ) we may assume that (X/Z, B) is Q-factorial dlt and that K X + B is nef/z. Moreover, we could assume that Z is affine. Run an LMMP/Z on K X + B ɛ B with scaling of some ample divisor, for some sufficiently small rational number ɛ > 0. By [6, Proposition 3.2], K X + B is trivial on each step of the LMMP. By Theorem 4., the LMMP terminates on a model X on which K X +B X and K X +B X ɛ B X are both nef/z and the latter is semi-ample/z. Moreover, X X is an isomorphism over the generic point of Z. The pair (X /Z, B X ) is Q-factorial and lc and (X /Z, B X ɛ B X ) is klt. If B X = 0 we are done so we can assume that B X = 0. Let (Y/Z, B Y ) be a Q-factorial dlt blowup of (X /Z, B X ). Let S be a component of T := B Y. Then, by induction, K S + B S := (K Y +B Y ) S is semi-ample/z. On the other hand, if P := h B X and if δ > 0 is any sufficiently small rational number, where h is the constructed morphism Y X, then K Y + B Y δp is semi-ample/z because K X + B X δ B X is is klt and nef/z hence semi-ample/z. Now, by Conjecture.2 in dimension d, K Y + B Y is semi-ample/z hence K X + B is also semi-ample/z. 5.5. The horizontal case. In this subsection, we deal with the horizontal case of Conjecture., that is, when some lc centre of (X/Z, B) is horizontal/z. We need a result from [4] concerning finite generation. For the convenience of the reader we include its proof. Theorem 5.6. Let f : X Y/Z be a surjective morphism of normal varieties, projective over an affine variety Z = Spec R, and L a Cartier divisor on Y. If R(X/Z, f L) is a finitely generated R-algebra, then R(Y/Z, L) is also a finitely generated R-algebra. Proof. If f is a contraction, the claim is trivial by the projection formula. So, replacing f by the finite part of the Stein factorisation we may assume that f is finite. We have a natural injective O Y -morphism φ: O Y f O X. Moreover, since we work over C, we also have a O Y -morphism ψ : f O X O Y given by Trace deg f X/Y such that ψφ is the identity morphism on O Y (cf. [2, Proposition

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP 7 5.7]). We actually have a splitting f O X O Y F such that φ corresponds to the natural injection and ψ to the first projection. So, for each m > 0, we have induced maps H 0 (Y, O Y (ml)) H 0 (Y, (f O X )(ml)) H 0 (X, O X (mf L)) H 0 (Y, (f O X )(ml)) H 0 (X, (O X (mf L)) H 0 (Y, O Y (ml)) where the upper map is injective and the lower map is surjective. Thus, we get an injection π : R(Y/Z, L) R(X/Z, f L) and a surjection µ: R(X/Z, f L) R(Y/Z, L) whose composition gives the identity map on R(Y/Z, L). Here, π is an R(Y/Z, L)-algebra homomorphism but µ is only an R(Y/Z, L)-module homomorphism. Let I be an ideal of R(Y/Z, L) and let I be the ideal of R(X/Z, f L) generated by π(i). We show that π I = I. Let c π I. We can write π(c) = π(a i )α i where a i I and α i R(X/Z, f L). Now, c = µ(π(c)) = µ( π(a i )α i ) = µ(π(a i ))µ(α i ) = a i µ(α i ) hence c is in the ideal I. Assume that R(X/Z, f L) is a finitely generated R-algebra. Then, R(X/Z, f L) is Noetherian. Let I I 2 be a chain of ideals of R(Y/Z, L) and let I I 2 be the corresponding chain of ideals in R(X/Z, f L). Since R(X/Z, f L) is Noetherian, the latter chain stabilises which implies that the former sequence also stabilises as π I j = I j. Therefore, R(Y/Z, L) is Noetherian which is equivalent to saying that it is a finitely generated R-algebra because R(Y/Z, L) is a graded R-algebra and the zero-degree piece of R(Y/Z, L) is a finitely generated R-algebra. Lemma 5.7. Assume Conjecture. in dimension d. Let (X/Z, B + A) be of dimension d as in Conjecture. such that (K X + B) F Q 0 for the generic fibre F of f. Moreover, assume that there is a contraction g : X T/Z such that () K X + B Q 0/T, (2) some lc centre of (X/Z, B) is horizontal over T. Then, (X/Z, B) has a good log minimal model. Proof. By replacing (X/Z, B) with a suitable Q-factorial dlt blowup, we can assume that (X/Z, B) is Q-factorial dlt and that there is a component S of B which is horizontal/t. Run an LMMP/Z on K X + B with scaling of some ample divisor. Since termination and semi-ampleness/z are local on Z, we can assume that Z is affine, say Spec R. By [7], the LMMP terminates with a good log minimal model if we prove that (X/Z, B) has a good log minimal model. By adjunction define K S + B S := (K X + B) S and let A S := A S. Then, K S + B S + A S = (K X + B + A) S Q 0/Z and K S + B S Q 0/T. Moreover, (K S + B S ) H Q 0 where H is the generic fibre of the induced morphism h: S Z. By assumptions, (S/Z, B S ) has a good log minimal model. Therefore, if I(K S + B S ) is Cartier for some I N, then R(S/Z, I(K S + B S )) is a finitely

8 CAUCHER BIRKAR generated R-algebra (cf. [4]). We can choose I such that I(K X + B) is Cartier and such that I(K X + B) g L for some Cartier divisor L on T. So, I(K S + B S ) e L where e: S T is the induced morphism. Now, by Theorem 5.6, R(T/Z, L) is a finitely generated R-algebra which in turn implies that R(X/Z, I(K X + B)) is a finitely generated R-algebra since X T is a contraction. Therefore, according to Theorem 3.7, (X/Z, B) has a good log minimal model. Lemma 5.8. Assume Conjecture. in dimension d in the vertical case. Let (X/Z, B +A) be of dimension d as in Conjecture. such that (K X +B) F Q 0 for the generic fibre F of f. Moreover, assume that (X/Z, B) is Q-factorial dlt, there exist contractions e: X Y and g : Y Z, e is birational, A is exceptional/y, and Y is Q-factorial, D := K Y + B Y ɛ B Y is klt and g is a D-negative extremal contraction for some ɛ > 0 where B Y := e B. Then, any LMMP/Z on K X + B with scaling of an ample/z divisor terminates with a good log minimal model of (X/Z, B). Proof. By [7], it is enough to prove that (X/Z, B) has a good log minimal model, locally over Z. By adding a small multiple of A to B we can assume that Supp A Supp B. Since e is birational, by Conjecture. in dimension d in the vertical case, we can assume that K X + B is actually semi-ample/y. By replacing X with the lc model of (X/Y, B) we can assume that K X + B is ample/y. Since A is ample/y, Supp A contains all the prime exceptional/y divisors on X. In particular, Supp e B Y Supp B. Replacing (X/Z, B) with a Q-factorial dlt blowup, we can again assume that (X/Z, B) is Q-factorial dlt; note that this preserves the property Supp e B Y Supp B. Since A is exceptional/y, K Y + B Y Q 0/Z. So, B Y is ample/z hence e B Y is semi-ample/z. Thus, for a small rational number τ > 0 we can write K X + B = K X + B τe B Y + τe B Y Q K X + /Z where is some rational boundary such that (X/Z, ) is klt. Note that since (Y/Z, B Y ɛ B Y ) is klt, B Y contains all the lc centres of (Y/Z, B Y ), in particular, the image of all the lc centres of (X/Z, B). So, Supp e B Y contains all the components of B hence (X/Z, B τe B Y ) is klt and we can indeed find with the required properties. By Theorem 4.3, we can run an LMMP/Z on K X + which ends up with a good log minimal model of (X/Z, ) hence a good log minimal model of (X/Z, B). Theorem 5.9. Assume Conjecture. in dimension d, and assume Conjecture. in dimension d in the vertical case. Then, Conjecture. holds in dimension d in the horizontal case. i.e. when some lc centre of (X/Z, B) is horizontal over Z. Proof. We can assume that (X/Z, B) is Q-factorial dlt and that f : X Z is a contraction. Run an LMMP/Z on K X + B with scaling of some ample/z

EXISTENCE OF LOG CANONICAL FLIPS AND A SPECIAL LMMP 9 divisor. By [9], this terminates with a Mori fibre space if A is horizontal/z. So, we can assume that A is vertical/z hence (K X + B) F Q 0 where F is the generic fibre of f : X Z. By [7], it is enough to prove that (X/Z, B) has a good log minimal model, locally over Z. By assumptions, some component of B is horizontal/z. If ɛ > 0 is a sufficiently small rational number, then (X/Z, B ɛ B ) is klt and K X + B ɛ B is not pseudo-effective/z hence by [9] there is a Mori fibre space (Y/Z, B Y ɛ B Y ) for (X/Z, B ɛ B ) obtained by running an LMMP. Let g : Y T/Z be the K Y +B Y ɛ B Y -negative extremal contraction which defines the Mori fibre space structure. Since K Y + B Y + A Y Q 0/Z and since A is vertical/z, A Y is vertical over T and K Y + B Y is numerically trivial on the generic fibre of g. But g is an extremal contraction so K Y + B Y Q 0/T. Since K Y + B Y ɛ B Y is numerically negative/t, B Y = 0. Let (W/Z, B W ) be a log smooth model of (X/Z, B) of type () as in Definition 2.3 such that W dominates Y, say by a morphism h: W Y. We can write K W + B W = h (K Y + B Y ) + G where G is exceptional/y. Run the LMMP/Y on K W + B W with scaling of some ample divisor. We reach a model X on which K X + B X is a limit of movable/y Q-divisors. Now by the Negativity Lemma 3.3, G X 0. So, K X + B X G X Q 0/T as it is the pullback of K Y + B Y. Since K X +B X is pseudo-effective/t, G X is vertical/t hence K X +B X Q 0 over the generic point of T. By construction, B Y is ample/t hence in particular horizontal/t. This implies that B X is horizontal/t because B Y is the pushdown of B X. By Lemma 5.8, we can run an LMMP/T on K X +B X which ends up with a model X on which K X + B X is semi-ample/t. Since K X + B X Q 0 over the generic point of T, X X is an isomorphism over the generic point of T. So, if X T /T is the contraction associated to K X + B X, then T T is birational hence in particular B X is horizontal/t. On the other hand, if we denote X Y by e, then we can write K X + B X + A X = e (K Y + B Y + A Y ) (which means that A X = G X + e A Y ), then K X + B X + A X Q 0/Z hence K X + B X + A X Q 0/Z. Now we can apply Lemma 5.7. Note that since K X + B X is pseudo-effective/z, A X is vertical/z. Proof. (of Theorem.3) We argue by induction so in particular we may assume that Conjecture. holds in dimension d. By Theorems 5.4, Conjecture. holds in dimension d in the vertical case. On the other hand, by Theorem 5.9, Conjecture. holds in dimension d in the horizontal case. 6. Proof of Theorem.5 We will derive Theorem.5 from the following: Theorem 6.. Assume Conjecture.4 in dimension d. Let (X/Z, B) be a lc pair of dimension d where B is a Q-divisor and the given morphism f : X Z

20 CAUCHER BIRKAR is surjective. Assume further that K X + B Q 0 over some non-empty open subset U Z, and if η is the generic point of a lc centre of (X/Z, B), then f(η) U. Then, () (X/Z, B) has a log minimal model (Y/Z, B Y ), (2) K Y + B Y is semi-ample/z; (3) if (X/Z, B) is Q-factorial dlt, then any LMMP/Z on K X +B with scaling of an ample/z divisor terminates. Proof. (of Theorem.5) We use induction on dimension so we can assume that Conjecture. holds in dimension d. Let (X/Z, B+A) be of dimension d as in Conjecture.. We may assume that f : X Z is a contraction. First assume that every lc centre of (X/Z, B) is vertical/z. By Theorem 5.3, Conjecture. ()(3) holds for (X/Z, B + A). We need to verify part (2) of the Conjecture. In particular, we may assume that K X + B is nef/z hence A is nef/z. Let U = Z \ f(supp A). Then, K X + B Q 0 over U. Now let η be the generic point of a lc centre of (X/Z, B). Since (X/Z, B + A) is lc, η does not belong to Supp A. On the other hand, since A is nef/z, Supp A = f f(supp A); indeed let z be a closed point of f(supp A) and let X z = f {z}; if X z is not a subset of Supp A, then there is a projective curve C contracted to z such that C intersects Supp A but C is not inside Supp A; hence A C > 0 which is a contradiction. Thus, η is mapped into U. Therefore, by Theorem 6., K X + B is semi-ample/z. So, Conjecture. in dimension d holds in the vertical case. Now assume that some lc centre of (X/Z, B) is horizontal/z. Then by Theorem 5.9, Conjecture. holds for (X/Z, B + A). The rest of this section is devoted to the proof of Theorem 6. which is parallel to the arguments of Section 5 with some small changes. 6.2. The vertical case. In this subsection we deal with Theorem 6. in the vertical case. Lemma 6.3. Let (X/Z, B) be a lc pair of dimension d as in Theorem 6.. Assume further that (X/Z, B) is Q-factorial dlt, K X + B + C is lc and nef/z for some Q-divisor C 0, K X + B Q P + cc/z for some Q-divisor P 0 with Supp P = Supp B, and some rational number c > 0. Then, assuming Conjecture.4 and Theorem 6. in dimension d, there is an LMMP/Z on K X + B with scaling of C which terminates. Proof. Step. Note that the assumptions imply that P and C are both vertical/z. Since K X + B Q P + cc/z, for each rational number t (0, ] we can write K X + B + tc Q ( + ɛ)(k X + B) ɛp ɛcc + tc Q ( + ɛ)(k X + B ɛ + ɛ P + t ɛc + ɛ C)/Z