Introduction
Let Y Y Y be a Schubert variety in a symmetrizable Kac–Moody partial flag variety. For a divisor D D D supported on the boundary, a regular function on the open Schubert cell defines a section of O Y ( D ) \mathcal{O}_Y(D) O Y ( D ) precisely when its poles are bounded by D D D . The boundary classes freely generate Cl ( Y ) \operatorname{Cl}(Y) Cl ( Y ) , so these inequalities describe every homogeneous component of Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) , including the sections of non-Cartier classes. We encode these inequalities in a cluster seed and identify its upper cluster algebra with the full Cox ring. We then realize the geometric simple-root operators in seeds of the same mutation class.
The construction starts on the Bott–Samelson variety of a word i = ( i 1 , … , i N ) \mathbf{i}=(i_1,\ldots,i_N) i = ( i 1 , … , i N ) , which may be nonreduced. Each position k k k has a standard boundary divisor with canonical section τ k \tau_k τ k . Let x k x_k x k be the minimal homogeneous lift of the corresponding word minor. If s ( k ) s(k) s ( k ) is the
2020 Mathematics Subject Classification . Primary 13F60, 14M15; Secondary 14L30, 17B67, 14M25.
Key words and phrases. Schubert variety, Bott–Samelson variety, Cox ring, upper cluster algebra, theta basis, root operator, crystal, horospherical variety.
The author was supported in part by the National Natural Science Foundation of China (Nos. 12131015 and 12571038). next occurrence of i k i_k i k , the exchange at k k k is
x k x k ′ = τ s ( k ) M k + ( x ) + τ k M k − ( x ) , x_kx'_k=\tau_{s(k)}M_k^+(x)+\tau_kM_k^-(x), x k x k ′ = τ s ( k ) M k + ( x ) + τ k M k − ( x ) ,
where M k ± M_k^\pm M k ± are the exchange monomials on the affine chart. The same line-bundle recursion gives the two boundary coefficients and the anticanonical class. Theorem 1.2 reduces upper cluster equality to a degree calculation. Under the hypotheses of that theorem, the Cox ring equals the upper algebra if and only if the sum of the frozen degrees equals the degree of a rational top form with divisor zero. This equality also shows that the initial and adjacent charts cover every codimension-one point of the total coordinate space.
These arguments give the Cox ring of the Bott–Samelson variety (Theorem 3.3 ). Under a birational contraction preserving the affine chart, the exceptional boundary sections specialize to one. Upper cluster equality descends because the pole order along each exceptional frozen divisor is independent of the seed (Theorem 2.2 ). For Y Y Y , a reduced word for its minimal Weyl group representative therefore gives Cox ( Y ) = U p o l ( Σ Y ) \operatorname{Cox}(Y)=\mathcal{U}_{\mathrm{pol}}(\Sigma_Y) Cox ( Y ) = U pol ( Σ Y ) , and all reduced expressions give the same mutation class (Theorem 3.4 and Corollary 3.9 ).
Theorem 4.1 gives an adapted seed for each nonzero positive simple-root operator E i E_i E i and each nonzero negative simple-root operator F i F_i F i whose rank-one group preserves Y Y Y . In this seed the operator is
D = c D x v D ∂ b D , ( v D ) b D = 0 , D=c_Dx^{v_D}\partial_{b_D},\qquad(v_D)_{b_D}=0, D = c D x v D ∂ b D , ( v D ) b D = 0 ,
where b D b_D b D is frozen and x v D x^{v_D} x v D is a monomial with nonnegative exponents. For a nonzero section f f f , the largest r r r such that D r f ≠ 0 D^rf\ne0 D r f = 0 is its polynomial degree in b D b_D b D . For E i E_i E i , one mutates the successive nonterminal occurrences of i i i . For F i F_i F i , the construction depends on whether i i i is a left descent. A reduced word starting with i i i gives an adapted chart in the descent case. In the other case, reversing the word presentation identifies the terminal minor on which F i F_i F i acts.
For every divisor class D D D , the regular theta basis of H 0 ( Y , O Y ( D ) ) H^0(Y,\mathcal{O}_Y(D)) H 0 ( Y , O Y ( D )) is indexed by the lattice points of a bounded slice of one rational polyhedral cone (Theorem 5.4 ). The word seeds of Shen and Weng admit maximal green sequences [43 ] , so the theta basis is available after inverting the frozen variables [29 ] . To restore the boundary, we prove Laurent finiteness on the chiral dual by a finite monomial cover and apply theta reciprocity and valuative independence [15 ] . Theorem 5.6 gives the facets in the initial word coordinates. Each terminal minor contributes the suffix sums along its letter. One mutation sequence computes the orders at all surviving standard boundary divisors; its construction uses Qin’s interval variables for the reversed word and transposed Cartan matrix [Qin]. Each bound on the nilpotence degree of a simple-root operator then cuts out a subspace spanned by a subset of this same basis, even when several of the operators do not commute.
For a finite-type Levi subgroup acting on Y Y Y and its graded Cox ring, the leading root arrows give the normal crystal of each section space. To calculate an arrow, carry its tropical index to an adapted seed by tropical mutation, translate the index there, and transport it back (Corollary 5.16 ). Slicing the common positive-root kernel by divisor class and weight counts branching multiplicities. Section 6 also identifies adjoint sections with interior lattice points and determines the canonical modules of the associated toric degenerations.
Two extensions use the geometry of the boundary. An adapted mutation sequence lifts through an equivariant compactification when the lifted seed gives its Cox ring, since differentiation preserves the filtration of each invariant boundary divisor (Theorem 4.4). For projective horospherical varieties with free class group, the equivariant description of the Cox ring gives adapted charts and crystals from the flag variety of the open orbit (Corollary 5.17 ). The examples illustrate the Cox grading, the root charts, and the polyhedral section-space basis in both smooth and singular settings.
0.1. Relation to earlier work. The cell seeds come from the generalized minors of Berenstein, Fomin, and Zelevinsky [6 ] . Geiß, Leclerc, and Schröer lifted cluster structures to multihomogeneous coordinate rings of partial flags [26 ] . Demonet extended the categorical construction to nonsimply-laced types [16 ] , and Kadhem gave a lifting construction in arbitrary finite semisimple type [32 ] . Shen and Weng established braid compatibility and maximal green sequences for positive braid words in symmetrizable Kac–Moody type [43 ] . Bao and Ye constructed upper cluster structures on open Richardson varieties and twisted products of flag varieties [4 Theorems 5.13 and 6.5] .
Mandel realizes upper cluster algebras as Cox rings of cluster varieties and their partial compactifications [37 Theorem 3.6 and Corollary 3.7] . For a prescribed compactification, Francone’s minimal monomial lifting gives a Cox-ring inclusion and equality after the boundary sections are inverted [20 ] and [21 Theorem 3.6] . He proves equality before localization for complete flag varieties [21 Theorem 4.4] . Here the boundary recursion and the logarithmic cluster volume form [5 Appendix A] give that equality for arbitrary-word Bott–Samelson varieties and all Schubert varieties, with their full Weil-class grading. Enwright, Francone, Moraga, and Spink show that finite-type Schubert and Bott–Samelson varieties are of cluster type [17 ] . Our result identifies the upper algebra of a specified seed on each projective variety.
The author constructs root derivations from optimized frozen variables and proves perfectness when generic characters form a basis [19 ] . Here we start with the geometric simple-root operators and find a reachable chart for each of them. We use the interval mutations of Goodearl and Yakimov [28 ] , the freezing results of Qin [41 ] , and Tang’s results on monomial derivatives and opposite seeds [45 ] . Tang’s comparison uses the twist of Kimura, Qin, and Wei [34 ] . The perfect-basis theorem of Berenstein and Kazhdan identifies the resulting Levi crystals [7 ] . In the horospherical case the underlying description of the Cox ring as a polynomial extension is due to Brion and Gagliardi [10 , 24 ] .
Earlier polyhedral models include Caldero’s toric degenerations [12 ] , generalized string polytopes for Bott–Samelson varieties [22 ] , global value semigroups [38 ] , cluster valuations on Schubert varieties [23 ] , and superpotential polytopes for Grassmannian Schubert varieties [42 ] . Boundary theta cones also occur in plethysm algebras [18 ] . The cone here retains the full divisor class and the simple-root filtrations simultaneously. Its semigroup algebras give toric degenerations by the finite Khovanskii-basis construction [KM].
We work over C \mathbb{C} C and write N = Z ≥ 0 \mathbb{N} = \mathbb{Z}_{\ge0} N = Z ≥ 0 . Frozen variables are polynomial unless explicitly inverted. The rows of an exchange matrix are indexed by mutable variables and its columns by all variables. All Schubert and Bott–Samelson varieties in this paper are finite-dimensional and projective.
A criterion for upper cluster equality
Seeds with polynomial frozen variables
A seed Σ = ( x , B ) \Sigma= (\mathbf{x}, B) Σ = ( x , B ) consists of a tuple of algebraically independent coordinates x = ( x j ) j ∈ I \mathbf{x} = (x_j)_{j\in I} x = ( x j ) j ∈ I and an integer matrix B = ( B k j ) B = (B_{kj}) B = ( B k j ) , whose rows are indexed by I m u t ⊂ I I_{\mathrm{mut}} \subset I I mut ⊂ I and whose columns are indexed by I I I . Its principal part is skew-symmetrizable. The other indices form I f r = I ∖ I m u t I_{\mathrm{fr}} = I \setminus I_{\mathrm{mut}} I fr = I ∖ I mut . Mutation at a mutable index k k k replaces x k x_k x k by x k ′ x'_k x k ′ , where
x k x k ′ = ∏ j ∈ I x j [ B k j ] + + ∏ j ∈ I x j [ − B k j ] + , [ a ] + = max ( a , 0 ) . (1) x_kx'_k = \prod_{j\in I} x_j^{[B_{kj}]_+} + \prod_{j\in I} x_j^{[-B_{kj}]_+}, \qquad[a]_+ = \max(a,0). \tag*{(1)} x k x k ′ = j ∈ I ∏ x j [ B k j ] + + j ∈ I ∏ x j [ − B k j ] + , [ a ] + = max ( a , 0 ) . (1)
We use the standard matrix mutation rule in this row convention.
The mixed Laurent ring of a seed is L p o l ( Σ ) = C [ x j ± 1 : j ∈ I m u t ] [ x f : f ∈ I f r ] \mathcal{L}_{\mathrm{pol}}(\Sigma) = \mathbb{C}[x_j^{\pm1} : j \in I_{\mathrm{mut}}][x_f : f \in I_{\mathrm{fr}}] L pol ( Σ ) = C [ x j ± 1 : j ∈ I mut ] [ x f : f ∈ I fr ] . Its upper cluster algebra with polynomial frozen variables is
U p o l ( Σ ) = ⋂ Σ ′ ∼ Σ L p o l ( Σ ′ ) . \mathcal{U}_{\mathrm{pol}}(\Sigma) = \bigcap_{\Sigma'\sim\Sigma} \mathcal{L}_{\mathrm{pol}}(\Sigma'). U pol ( Σ ) = Σ ′ ∼ Σ ⋂ L pol ( Σ ′ ) .
All intersections are taken in the common rational function field. The cluster algebra A ( Σ ) \mathcal{A}(\Sigma) A ( Σ ) is the subalgebra generated by all coordinates of all reachable seeds, with the frozen variables polynomial. The Laurent phenomenon gives A ( Σ ) ⊂ U p o l ( Σ ) \mathcal{A}(\Sigma) \subset\mathcal{U}_{\mathrm{pol}}(\Sigma) A ( Σ ) ⊂ U pol ( Σ ) . We write L i n v ( Σ ) = C [ x j ± 1 : j ∈ I ] \mathcal{L}_{\mathrm{inv}}(\Sigma) = \mathbb{C}[x_j^{\pm1} : j \in I] L inv ( Σ ) = C [ x j ± 1 : j ∈ I ] and U i n v ( Σ ) \mathcal{U}_{\mathrm{inv}}(\Sigma) U inv ( Σ ) for the corresponding rings with every frozen variable inverted. We call the intersection of the initial mixed Laurent ring and those of its one-step mutations the upper bound . A grading is homogeneous if the two monomials in (1) have the same degree. Equivalently, each exchange row has degree zero. For a grading by a lattice Γ \Gamma Γ , a positive integral coarsening is a homomorphism h : Γ → Z h:\Gamma\to\mathbb{Z} h : Γ → Z for which the induced grading is nonnegative and its degree-zero part is C \mathbb{C} C .
Lemma 1.1. For each frozen index f f f , let v f v_f v f be the valuation given by the least exponent of x f x_f x f in the initial Laurent chart. It has the same coordinate description in every seed. If R = U p o l ( Σ ) R=U_{\mathrm{pol}}(\Sigma) R = U pol ( Σ ) and b = ∏ f x f b=\prod_f x_f b = ∏ f x f , then
R = U i n v ( Σ ) ∩ ⋂ f O v f , R [ b − 1 ] = U i n v ( Σ ) . R=U_{\mathrm{inv}}(\Sigma)\cap\bigcap_f\mathcal{O}_{v_f},\qquad R[b^{-1}]=U_{\mathrm{inv}}(\Sigma). R = U inv ( Σ ) ∩ f ⋂ O v f , R [ b − 1 ] = U inv ( Σ ) . Every frozen coordinate is prime in R R R , with x f R = { a ∈ R : v f ( a ) ≥ 1 } x_fR=\{a\in R:v_f(a)\geq1\} x f R = { a ∈ R : v f ( a ) ≥ 1 } .
Proof. The remaining seed coordinates have algebraically independent residues for v f v_f v f . Reducing an exchange relation modulo x f x_f x f gives a nonzero monomial if x f x_f x f occurs in the exchange, and a nonzero binomial otherwise. The mutated coordinate is a unit for v f v_f v f , and the change of residue coordinates is birational. Induction therefore proves the first assertion. In every seed, its mixed Laurent ring is its Laurent ring intersected with the finitely many valuation rings O v f \mathcal{O}_{v_f} O v f . Intersecting over all seeds gives the formula for R R R .
For a ∈ U i n v ( Σ ) a\in U_{\mathrm{inv}}(\Sigma) a ∈ U inv ( Σ ) , a sufficiently large power of b b b clears all these frozen valuations, proving the localization identity. Dividing an element of R R R by x f x_f x f decreases only its v f v_f v f -value, by one. This gives the formula for x f R x_fR x f R . Since the other values remain nonnegative and v f v_f v f is additive on products, this ideal is prime. □ \square □
The canonical degree
The logarithmic cluster volume form detects whether the initial chart and its neighbors cover all prime divisors. The product of the frozen variables cancels its poles along the frozen divisors. At a mutable divisor, regularity of the first mutation controls the remaining pole. Comparison with a generator of the canonical sheaf in codimension one yields the following degree criterion.
Theorem 1.2. Let R = ⨁ d ≥ 0 R d R=\bigoplus_{d\geq0}R_d R = ⨁ d ≥ 0 R d be a finitely generated normal C \mathbb{C} C -domain with R 0 = C R_0=\mathbb{C} R 0 = C . Suppose a homogeneous rational top form Ω \Omega Ω has divisor zero and degree κ \kappa κ . Let z 1 , … , z N z_1,\ldots,z_N z 1 , … , z N be homogeneous pairwise nonassociated prime elements that are algebraically independent and satisfy Frac R = C ( z 1 , … , z N ) \operatorname{Frac}R=\mathbb{C}(z_1,\ldots,z_N) Frac R = C ( z 1 , … , z N ) .
Equip this tuple with a seed whose first ℓ \ell ℓ variables are mutable. Suppose the extended exchange matrix has rank ℓ \ell ℓ , every row is primitive, and every first exchange quotient belongs to R R R . Then
∑ j = ℓ + 1 N deg z j ≥ κ . \sum_{j=\ell+1}^{N}\deg z_j\geq\kappa. j = ℓ + 1 ∑ N deg z j ≥ κ . The following conditions are equivalent:
(i) ∑ j = ℓ + 1 N deg z j = κ \sum_{j=\ell+1}^{N}\deg z_j=\kappa ∑ j = ℓ + 1 N deg z j = κ ;
(ii) R = L p o l ( Σ ) ∩ ⋂ i = 1 ℓ L p o l ( μ i Σ ) R=\mathcal{L}_{\mathrm{pol}}(\Sigma)\cap\bigcap_{i=1}^{\ell}\mathcal{L}_{\mathrm{pol}}(\mu_i\Sigma) R = L pol ( Σ ) ∩ ⋂ i = 1 ℓ L pol ( μ i Σ ) ;
(iii) R = U p o l ( Σ ) R=\mathcal{U}_{\mathrm{pol}}(\Sigma) R = U pol ( Σ ) .
When they hold, R R R is factorial, and the initial chart and its adjacent charts form an open subset of Spec R \operatorname{Spec}R Spec R whose complement has codimension at least two. Each chart is a product of an algebraic torus and an affine space.
Proof. Write b = ∏ j > ℓ z j b=\prod_{j>\ell}z_j b = ∏ j > ℓ z j and consider η = b d log z 1 ∧ ⋯ ∧ d log z N \eta=b\,d\log z_1\wedge\cdots\wedge d\log z_N η = b d log z 1 ∧ ⋯ ∧ d log z N . We first prove that η \eta η is regular in codimension one. It suffices to check the coordinate divisors. At a frozen divisor, the factor b b b cancels the logarithmic pole. At a mutable divisor ( z i ) (z_i) ( z i ) , write P i = M + + M − = z i z i ′ P_i=M_++M_-=z_i z_i' P i = M + + M − = z i z i ′ and z i ′ = z i a v z_i'=z_i^a v z i ′ = z i a v , where a ≥ 0 a\geq0 a ≥ 0 and v ∈ R ( z i ) × v\in R_{(z_i)}^\times v ∈ R ( z i ) × . Choose j j j with B i j ≠ 0 B_{ij}\ne0 B ij = 0 . All other initial coordinates are units at ( z i ) (z_i) ( z i ) , and M − = − M + M_-=-M_+ M − = − M + there. Consequently the coefficient of d log z j d\log z_j d log z j in d P i dP_i d P i is a unit, with residue B i j M + B_{ij}M_+ B ij M + . Wedge d P i = d ( z i a + 1 v ) dP_i=d(z_i^{a+1}v) d P i = d ( z i a + 1 v ) with d z i dz_i d z i and the d log z k d\log z_k d log z k , k ≠ i , j k\ne i,j k = i , j , to obtain
η = u z i a d z i ∧ d v ∧ ⋀ k ≠ i , j d log z k , u ∈ R ( z i ) × . (2) \eta=uz_i^a\,dz_i\wedge dv\wedge\bigwedge_{k\ne i,j}d\log z_k,\qquad u\in R_{(z_i)}^\times. \tag*{(2)} η = u z i a d z i ∧ d v ∧ k = i , j ⋀ d log z k , u ∈ R ( z i ) × . (2)
This expression is regular, since a normal variety is smooth at its codimension-one points. The ratio F = η / Ω F=\eta/\Omega F = η /Ω therefore belongs to R R R . It is nonzero and homogeneous of degree ∑ j > ℓ deg z j − κ \sum_{j>\ell}\deg z_j-\kappa ∑ j > ℓ deg z j − κ , which proves the inequality. Under condition (i), positive grading forces F F F to be constant:
η = c Ω ( c ∈ C × ) . (3) \eta=c\Omega\qquad(c\in\mathbb{C}^{\times}). \tag*{(3)} η = c Ω ( c ∈ C × ) . (3)
Set T = C [ z 1 ± 1 , … , z ℓ ± 1 ] T=\mathbb{C}[z_1^{\pm1},\ldots,z_\ell^{\pm1}] T = C [ z 1 ± 1 , … , z ℓ ± 1 ] and R ∘ = R [ ( z 1 ⋯ z N ) − 1 ] R^\circ=R[(z_1\cdots z_N)^{-1}] R ∘ = R [( z 1 ⋯ z N ) − 1 ] . The coordinate morphism from the smooth locus of Spec R ∘ \operatorname{Spec}R^\circ Spec R ∘ to Spec T \operatorname{Spec}T Spec T is birational, and (3) gives a nonvanishing Jacobian. It is therefore étale, hence an open immersion by Zariski’s Main Theorem [44 Tags 05K0 and 0AB1] . Denote its image by U U U . Normality gives R ∘ = Γ ( U , O U ) R^\circ=\Gamma(U,\mathcal{O}_U) R ∘ = Γ ( U , O U ) .
Since R × = C × R^\times=\mathbb{C}^\times R × = C × and the z j z_j z j are pairwise nonassociated primes, every unit of R ∘ R^\circ R ∘ is a scalar Laurent monomial in them. If the complement of U U U contained a prime divisor of the torus, its Laurent irreducible equation would be a further unit on U U U , a contradiction. That complement has codimension at least two, and normality of the torus gives
R ∘ = Γ ( U , O U ) = T . (4) R^\circ=\Gamma(U,\mathcal{O}_U)=T. \tag*{(4)} R ∘ = Γ ( U , O U ) = T . (4)
Nagata’s localization theorem now shows that Cl ( R ) \operatorname{Cl}(R) Cl ( R ) is generated by the principal coordinate divisors. Thus R R R is factorial [25 Theorem 1.19] .
We next recover regularity along each mutable divisor. Equations (2) and (3) imply a = 0 a=0 a = 0 , and show that the residues of z i ′ = v z_i'=v z i ′ = v and z k z_k z k , k ≠ i , j k\ne i,j k = i , j , are algebraically independent. The valuation ord z i \operatorname{ord}_{z_i} ord z i has a height-one center on T i = L inv ( μ i Σ ) T_i=\mathcal{L}_{\operatorname{inv}}(\mu_i\Sigma) T i = L inv ( μ i Σ ) , and this center contains P i P_i P i . Primitivity of the exchange row makes P i P_i P i irreducible in T i T_i T i : a unimodular change of Laurent coordinates takes it to a unit times 1 + w 1+w 1 + w . Hence its center is ( P i ) (P_i) ( P i ) . The two discrete valuation rings have the same fraction field, so their local inclusion is an equality:
R ( z i ) = ( T i ) ( P i ) . (5) R_{(z_i)}=(T_i)_{(P_i)}. \tag*{(5)} R ( z i ) = ( T i ) ( P i ) . (5)
By normality and (4) , R [ b − 1 ] = T ∩ ⋂ i = 1 ℓ R ( z i ) R[b^{-1}]=T\cap\bigcap_{i=1}^{\ell}R_{(z_i)} R [ b − 1 ] = T ∩ ⋂ i = 1 ℓ R ( z i ) . Moreover, T [ P i − 1 ] = T i [ P i − 1 ] T[P_i^{-1}]=T_i[P_i^{-1}] T [ P i − 1 ] = T i [ P i − 1 ] , so an element of T T T can have a pole on Spec T i \operatorname{Spec}T_i Spec T i only along ( P i ) (P_i) ( P i ) . Equation (5) therefore gives
R [ b − 1 ] = T ∩ ⋂ i = 1 ℓ T i = U inv ( Σ ) . R[b^{-1}]=T\cap\bigcap_{i=1}^{\ell}T_i=U_{\operatorname{inv}}(\Sigma). R [ b − 1 ] = T ∩ i = 1 ⋂ ℓ T i = U inv ( Σ ) .
The last equality is the full-rank upper-bound theorem [6 Proposition 1.8 and Corollary 1.9] ; over C \mathbb{C} C , its proof is unchanged, or follows by flat scalar extension of the finite upper bounds.
At a frozen divisor ( z j ) (z_j) ( z j ) , nonvanishing of η \eta η shows that the residues of all other coordinates are algebraically independent. Thus ord z j \operatorname{ord}_{z_j} ord z j is the usual frozen-coordinate valuation. Lemma 1.1 identifies L pol ( Σ ′ ) \mathcal{L}_{\operatorname{pol}}(\Sigma') L pol ( Σ ′ ) , for every reachable seed, with L inv ( Σ ′ ) ∩ ⋂ j > ℓ R ( z j ) \mathcal{L}_{\operatorname{inv}}(\Sigma')\cap\bigcap_{j>\ell}R_{(z_j)} L inv ( Σ ′ ) ∩ ⋂ j > ℓ R ( z j ) . Intersecting over the initial and adjacent seeds, or over all seeds, now gives R R R . This proves (ii) and (iii).
Conversely, assume (i) or (iii). For each initial or adjacent seed, localizing R R R at its mutable coordinates gives its mixed Laurent ring. These are smooth open subsets of Spec R \operatorname{Spec}R Spec R . The initial chart misses only the prime divisors ( z i ) (z_i) ( z i ) , i ≤ ℓ i\le\ell i ≤ ℓ . The adjacent i i i -chart contains the generic point of ( z i ′ ) (z_i') ( z i ′ ) : the other initial coordinates are units there, and z i ∤ z i ′ z_i\nmid z_i' z i ∤ z i ′ . Indeed, divisibility would place z i ′ / z i = ( z i ′ ) 2 / P i z_i'/z_i=(z_i')^2/P_i z i ′ / z i = ( z i ′ ) 2 / P i in R R R , although this expression does not belong to the adjacent mixed Laurent ring. The nonmonomial polynomial P i P_i P i has no coordinate factor, so its denominator cannot cancel. These charts therefore cover every codimension-one point.
Mutation preserves η \eta η up to sign. On each mixed Laurent chart it is the product of the logarithmic differentials of the mutable coordinates and the ordinary differentials of the frozen coordinates. Hence div ( η ) = 0 \operatorname{div}(\eta)=0 div ( η ) = 0 . The ratio F = η / Ω F=\eta/\Omega F = η /Ω is now a unit of R R R , so it is constant and has degree zero. This proves (i), the asserted chart cover, and all remaining claims. □ \square □
The primitivity hypothesis is essential, even when R R R is a polynomial ring. Let R = C [ x , u , y ] R=\mathbb{C}[x,u,y] R = C [ x , u , y ] , with deg x = deg u = 1 \deg x=\deg u=1 deg x = deg u = 1 and deg y = 2 \deg y=2 deg y = 2 , and set z = ( x u − y ) / i z=(xu-y)/i z = ( xu − y ) / i , where i 2 = − 1 i^2=-1 i 2 = − 1 . Use x x x as the sole mutable variable and y , z y,z y , z as frozen variables, with exchange x x ′ = y 2 + z 2 xx'=y^2+z^2 x x ′ = y 2 + z 2 . The initial coordinates are pairwise nonassociated primes and form a transcendence basis. The exchange quotient is x ′ = u ( y − i z ) ∈ R x'=u(y-iz)\in R x ′ = u ( y − i z ) ∈ R , and both the sum of the frozen degrees and the degree of d x ∧ d u ∧ d y dx\wedge du\wedge dy d x ∧ d u ∧ d y are four. Nevertheless, u = ( y + i z ) / x = x ′ / ( y − i z ) u=(y+iz)/x=x'/(y-iz) u = ( y + i z ) / x = x ′ / ( y − i z ) is not in the adjacent mixed Laurent ring. The exchange row ( 0 , 2 , − 2 ) (0,2,-2) ( 0 , 2 , − 2 ) is not primitive.
Cox rings and birational morphisms
2.1. Boundary divisors and minimal lifts. Let Y Y Y be a normal projective variety with an affine open subset C C C such that Cl ( C ) = 0 \operatorname{Cl}(C)=0 Cl ( C ) = 0 and Γ ( C , O C ) × = C × \Gamma(C,\mathcal{O}_C)^\times=\mathbb{C}^\times Γ ( C , O C ) × = C × . Let D 1 , … , D d D_1,\ldots,D_d D 1 , … , D d be the prime divisors in Y ∖ C Y\setminus C Y ∖ C . The divisor-class localization sequence gives Cl ( Y ) = ⨁ a = 1 d Z [ D a ] \operatorname{Cl}(Y)=\bigoplus_{a=1}^d\mathbb{Z}[D_a] Cl ( Y ) = ⨁ a = 1 d Z [ D a ] . They generate because Cl ( C ) = 0 \operatorname{Cl}(C)=0 Cl ( C ) = 0 , and any relation would arise from a unit on C C C . We use this basis to define the Cox ring
R ( Y ) = ⨁ m ∈ Z d H 0 ( Y , O Y ( ∑ a m a D a ) ) τ m ⊂ C [ C ] [ τ 1 ± 1 , … , τ d ± 1 ] . R(Y)=\bigoplus_{m\in\mathbb{Z}^d}H^0\left(Y,\mathcal{O}_Y\left(\sum_a m_aD_a\right)\right)\tau^m\subset\mathbb{C}[C][\tau_1^{\pm1},\ldots,\tau_d^{\pm1}]. R ( Y ) = m ∈ Z d ⨁ H 0 ( Y , O Y ( a ∑ m a D a ) ) τ m ⊂ C [ C ] [ τ 1 ± 1 , … , τ d ± 1 ] .
Here O Y ( D ) \mathcal{O}_Y(D) O Y ( D ) is the rank-one reflexive sheaf of a Weil divisor, and multiplication is multiplication of rational functions. In this realization, a homogeneous expression f τ m f\tau^m f τ m , with f ∈ C [ C ] f\in\mathbb{C}[C] f ∈ C [ C ] , belongs to R ( Y ) R(Y) R ( Y ) precisely when ord D a ( f ) + m a ≥ 0 \operatorname{ord}_{D_a}(f)+m_a\ge0 ord D a ( f ) + m a ≥ 0 for every a a a . The element τ a \tau_a τ a is the canonical section of D a D_a D a . This description agrees with the usual construction of Cox rings [2 ] . For a class λ ∈ Cl ( Y ) \lambda\in\operatorname{Cl}(Y) λ ∈ Cl ( Y ) , the notation O Y ( λ ) \mathcal{O}_Y(\lambda) O Y ( λ ) refers to the representative in this boundary basis. Thus R ( Y ) λ = H 0 ( Y , O Y ( λ ) ) R(Y)_\lambda=H^0(Y,\mathcal{O}_Y(\lambda)) R ( Y ) λ = H 0 ( Y , O Y ( λ )) , with the formal degree factor understood.
For 0 ≠ f ∈ C [ C ] 0\ne f\in\mathbb{C}[C] 0 = f ∈ C [ C ] , its minimal homogeneous lift is f ~ = τ ν ( f ) f \widetilde f=\tau^{\nu(f)}f f = τ ν ( f ) f , where ν ( f ) a = − ord D a ( f ) \nu(f)_a=-\operatorname{ord}_{D_a}(f) ν ( f ) a = − ord D a ( f ) . It is the unique lift with no canonical boundary section as a factor. We allow negative entries in ν ( f ) \nu(f) ν ( f ) . Minimal lifting and its application to cluster structures were studied by Francone [20 , 21 ] .
If a cell seed has coordinates f 1 , … , f N f_1,\ldots,f_N f 1 , … , f N and exchange matrix B C B_C B C , let ν \nu ν be the matrix with columns ν ( f k ) \nu(f_k) ν ( f k ) . Its homogeneous lift has coordinates x k = τ ν ∙ k f k x_k=\tau^{\nu_{\bullet k}}f_k x k = τ ν ∙ k f k and τ 1 , … , τ d \tau_1,\ldots,\tau_d τ 1 , … , τ d , with the boundary sections frozen and exchange matrix
B ~ = ( B C ∣ − B C ν T ) . \widetilde B=(B_C\mid-B_C\nu^{\mathsf T}). B = ( B C ∣ − B C ν T ) .
Homogeneity determines the added columns. To check regularity, write
v k ± = ∑ j [ ± B k j ] + ν ( f j ) , a k = max ( v k + , v k − ) , v_k^\pm=\sum_j[\pm B_{kj}]_+\nu(f_j),\qquad a_k=\max(v_k^+,v_k^-), v k ± = j ∑ [ ± B k j ] + ν ( f j ) , a k = max ( v k + , v k − ) ,
where the maximum is coordinatewise. The lifted exchange quotient is τ a k − ν ( f k ) f k ′ \tau^{a_k-\nu(f_k)}f'_k τ a k − ν ( f k ) f k ′ . If f k ′ f'_k f k ′ is regular on C C C , then
ord D a ( f k ′ ) ≥ − ( a k ) a + ν ( f k ) a \operatorname{ord}_{D_a}(f'_k)\ge-(a_k)_a+\nu(f_k)_a ord D a ( f k ′ ) ≥ − ( a k ) a + ν ( f k ) a
for every boundary divisor, by the valuation inequality for its two numerator terms. Thus the lifted quotient belongs to the Cox ring. The inequality may be strict, so the quotient may have a boundary factor. Equality with the Cox ring will show that no such factor occurs anywhere in the mutation pattern.
The divisor-class grading admits a positive integral coarsening. In positive dimension, intersection with H dim Y − 1 H^{\dim Y-1} H d i m Y − 1 , for an ample Cartier divisor H H H , is positive on every nonzero effective divisor class. The resulting degree-zero component is C \mathbb{C} C .
2.2. Contraction of boundary divisors. Under a birational contraction, the exceptional divisorial valuations no longer enter the definition of the Cox ring. In the Cox grading this amounts to setting their canonical sections equal to one. For the general contraction formula for finitely generated Cox rings, see [31 Proposition 2.2] .
Proposition 2.1. Let p : X → Y p : X \to Y p : X → Y be a proper birational morphism of normal projective varieties over C \mathbb{C} C identifying an open subvariety C ⊂ X C \subset X C ⊂ X with an open subvariety of Y Y Y . Suppose that Cl ( C ) = 0 \operatorname{Cl}(C) = 0 Cl ( C ) = 0 and Γ ( C , O C ) × = C × \Gamma(C, \mathcal{O}_C)^{\times} = \mathbb{C}^{\times} Γ ( C , O C ) × = C × . Let D 1 , … , D d D_1,\ldots,D_d D 1 , … , D d be the prime divisors in Y ∖ C Y \setminus C Y ∖ C , let D ~ a \widetilde{D}_a D a be their strict transforms, and let E 1 , … , E e E_1,\ldots,E_e E 1 , … , E e be the exceptional prime divisors. Then
Cl ( Y ) = ⨁ a = 1 d Z [ D a ] , Cl ( X ) = ⨁ a = 1 d Z [ D ~ a ] ⊕ ⨁ b = 1 e Z [ E b ] . \operatorname{Cl}(Y) = \bigoplus_{a=1}^{d} \mathbb{Z}[D_a], \qquad\operatorname{Cl}(X) = \bigoplus_{a=1}^{d} \mathbb{Z}[\widetilde{D}_a] \oplus\bigoplus_{b=1}^{e} \mathbb{Z}[E_b]. Cl ( Y ) = a = 1 ⨁ d Z [ D a ] , Cl ( X ) = a = 1 ⨁ d Z [ D a ] ⊕ b = 1 ⨁ e Z [ E b ] . Write τ a \tau_a τ a for the canonical sections of D a D_a D a and D ~ a \widetilde{D}_a D a in the respective Cox rings, and σ b \sigma_b σ b for the canonical section of E b E_b E b . The map
q p : Cox ( X ) ⟶ Cox ( Y ) , f τ m σ n ⟼ f τ m q_p : \operatorname{Cox}(X) \longrightarrow\operatorname{Cox}(Y), \qquad f\tau^{\mathbf{m}}\sigma^{\mathbf{n}} \longmapsto f\tau^{\mathbf{m}} q p : Cox ( X ) ⟶ Cox ( Y ) , f τ m σ n ⟼ f τ m is surjective, respects degrees under p ∗ : Cl ( X ) → Cl ( Y ) p_* : \operatorname{Cl}(X) \to\operatorname{Cl}(Y) p ∗ : Cl ( X ) → Cl ( Y ) , and has kernel ( σ 1 − 1 , … , σ e − 1 ) (\sigma_1 - 1,\ldots,\sigma_e - 1) ( σ 1 − 1 , … , σ e − 1 ) . These maps commute with composition of contractions over C C C . Moreover,
Cox ( X ) [ σ 1 − 1 , … , σ e − 1 ] = Cox ( Y ) [ σ 1 ± 1 , … , σ e ± 1 ] (6) \operatorname{Cox}(X)[\sigma_1^{-1},\ldots,\sigma_e^{-1}] = \operatorname{Cox}(Y)[\sigma_1^{\pm1},\ldots,\sigma_e^{\pm1}] \tag*{(6)} Cox ( X ) [ σ 1 − 1 , … , σ e − 1 ] = Cox ( Y ) [ σ 1 ± 1 , … , σ e ± 1 ] (6) inside C ( C ) ( τ , σ ) \mathbb{C}(C)(\tau,\sigma) C ( C ) ( τ , σ ) .
Proof. Since Y Y Y is normal, p p p is an isomorphism over the generic point of each D a D_a D a . Thus D a D_a D a and D ~ a \widetilde{D}_a D a define the same valuation, and the remaining boundary divisors of X X X are the E b E_b E b . The divisor-class localization sequence for C C C gives the asserted bases on X X X and Y Y Y , since a relation among boundary divisors restricts to a unit on C C C .
For f ∈ Γ ( C , O C ) f \in\Gamma(C,\mathcal{O}_C) f ∈ Γ ( C , O C ) , the homogeneous expression f τ m σ n f\tau^{\mathbf{m}}\sigma^{\mathbf{n}} f τ m σ n belongs to Cox ( X ) \operatorname{Cox}(X) Cox ( X ) precisely when
ord D a ( f ) + m a ≥ 0 ( 1 ≤ a ≤ d ) , ord E b ( f ) + n b ≥ 0 ( 1 ≤ b ≤ e ) . \operatorname{ord}_{D_a}(f) + m_a \ge0 \quad(1 \le a \le d), \qquad\operatorname{ord}_{E_b}(f) + n_b \ge0 \quad(1 \le b \le e). ord D a ( f ) + m a ≥ 0 ( 1 ≤ a ≤ d ) , ord E b ( f ) + n b ≥ 0 ( 1 ≤ b ≤ e ) . The first set of inequalities characterizes membership of f τ m f\tau^{\mathbf{m}} f τ m in Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) . For such a section we can choose n \mathbf{n} n satisfying the second set. This proves surjectivity and also gives (2.1).
It suffices to compute the kernel in each Cl ( Y ) \operatorname{Cl}(Y) Cl ( Y ) -degree. Write an element of the kernel as ∑ n f n τ m σ n \sum_{\mathbf{n}} f_{\mathbf{n}}\tau^{\mathbf{m}}\sigma^{\mathbf{n}} ∑ n f n τ m σ n , so ∑ n f n = 0 \sum_{\mathbf{n}} f_{\mathbf{n}} = 0 ∑ n f n = 0 . Choose N \mathbf{N} N dominating its finitely many exponents n \mathbf{n} n coordinatewise. Modulo ( σ b − 1 : 1 ≤ b ≤ e ) (\sigma_b - 1 : 1 \le b \le e) ( σ b − 1 : 1 ≤ b ≤ e ) , each summand equals its product with σ N − n \sigma^{\mathbf{N}-\mathbf{n}} σ N − n . The resulting sum vanishes. The reverse inclusion follows by evaluation. Under a composition the two successive evaluations set exactly the exceptional variables of the composite equal to one. □
Theorem 2.2. In the setting of Proposition 2.1, let Σ X \Sigma_X Σ X and Σ Y \Sigma_Y Σ Y be the minimal homogeneous lifts of a seed of regular functions on C C C , with all boundary sections frozen. Then
Cox ( X ) = U p o l ( Σ X ) ⟹ Cox ( Y ) = U p o l ( Σ Y ) . \operatorname{Cox}(X) = \mathcal{U}_{\mathrm{pol}}(\Sigma_X) \Longrightarrow\operatorname{Cox}(Y) = \mathcal{U}_{\mathrm{pol}}(\Sigma_Y). Cox ( X ) = U pol ( Σ X ) ⟹ Cox ( Y ) = U pol ( Σ Y ) . Equality with the initial upper bound also descends. More generally, if the mixed Laurent rings after a nonempty collection of mutation sequences intersect in Cox ( X ) \operatorname{Cox}(X) Cox ( X ) , then their counterparts on Y Y Y intersect in Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) . The target exchange matrix is obtained by deleting the exceptional frozen columns. Under any of these equalities on X X X , q p q_p q p takes each cluster variable to the corresponding target variable.
Proof. Write B B B for the exchange matrix on C C C , and let ν Y \nu_Y ν Y and ν E \nu_E ν E record the pole orders of the seed functions at the surviving and exceptional divisors. The lifted matrix on X X X is ( B ∣ − B ν Y T ∣ − B ν E T ) (B \mid-B\nu_Y^{\mathsf{T}} \mid-B\nu_E^{\mathsf{T}}) ( B ∣ − B ν Y T ∣ − B ν E T ) , whose first two blocks form the lifted matrix on Y Y Y . The initial lift on X X X of each cell coordinate differs from its lift on Y Y Y by its exceptional degree monomial. Matrix mutation commutes with deletion of frozen columns, and both terms of each source exchange have the same exceptional degree. Hence the same monomial rescaling, with the new degrees, relates the coordinates after any mutation sequence t t t . Writing L X , t L_{X,t} L X , t and L Y , t L_{Y,t} L Y , t for the two mixed Laurent rings, and S S S for the multiplicative system generated by the σ b \sigma_b σ b , we obtain
S − 1 L X , t = L Y , t [ σ 1 ± 1 , … , σ e ± 1 ] . (7) S^{-1}L_{X,t} = L_{Y,t}[\sigma_1^{\pm1},\ldots,\sigma_e^{\pm1}]. \tag*{(7)} S − 1 L X , t = L Y , t [ σ 1 ± 1 , … , σ e ± 1 ] . (7)
For any nonempty collection T T T of mutation sequences,
S − 1 ( ⋂ t ∈ T L X , t ) = ⋂ t ∈ T S − 1 L X , t . S^{-1}\left(\bigcap_{t\in T}L_{X,t}\right)=\bigcap_{t\in T}S^{-1}L_{X,t}. S − 1 ( t ∈ T ⋂ L X , t ) = t ∈ T ⋂ S − 1 L X , t .
Indeed, let h ≠ 0 h\ne0 h = 0 belong to the right side. By Lemma 1.1, the least exponent v b ( h ) v_b(h) v b ( h ) of each exceptional variable σ b \sigma_b σ b is independent of the chart. Thus ∏ b σ b max ( 0 , − v b ( h ) ) h \prod_b\sigma_b^{\max(0,-v_b(h))}h ∏ b σ b m a x ( 0 , − v b ( h )) h belongs to every L X , t L_{X,t} L X , t . This proves the reverse inclusion. The forward inclusion holds for every localization. If ⋂ t ∈ T L X , t = Cox ( X ) \bigcap_{t\in T}L_{X,t}=\operatorname{Cox}(X) ⋂ t ∈ T L X , t = Cox ( X ) , apply (6) and (7) and take the component of exceptional degree zero. The result is Cox ( Y ) = ⋂ t ∈ T L Y , t \operatorname{Cox}(Y)=\bigcap_{t\in T}L_{Y,t} Cox ( Y ) = ⋂ t ∈ T L Y , t . Take all mutation sequences for the full upper algebra, or the initial and adjacent charts for the upper bound. By the Laurent phenomenon every source cluster variable belongs to the assumed intersection. Evaluation at σ = 1 \boldsymbol{\sigma}=\mathbf{1} σ = 1 removes its exceptional degree monomial and gives the corresponding target variable. □
Bott–Samelson and Schubert Cox rings
Let A = ( a i j ) i , j ∈ I 0 A=(a_{ij})_{i,j\in I_0} A = ( a ij ) i , j ∈ I 0 be a symmetrizable generalized Cartan matrix, and let G G G be the associated simply connected Kac–Moody group in a standard realization. Fix opposite Borel subgroups B + B^+ B + , B − B^- B − , a common Cartan subgroup T T T , and a pinning. We write W W W for the Weyl group, α i \alpha_i α i and α i ∨ \alpha_i^\vee α i ∨ for the simple roots and coroots, and Λ i \Lambda_i Λ i for the fundamental weights, and set ρ = ∑ i ∈ I 0 Λ i \rho=\sum_{i\in I_0}\Lambda_i ρ = ∑ i ∈ I 0 Λ i . The standard realization includes the additional Cartan directions when A A A has nonzero corank, so the simple roots are linearly independent. We use the standard facts about the flag variety from [35 ] .
For J ⊂ I 0 J\subset I_0 J ⊂ I 0 , let P J P_J P J be the standard parabolic and let W J W^J W J be the minimal representatives of W / W J W/W_J W / W J . For w ∈ W J w\in W^J w ∈ W J , write X w P J = B + w P J / P J X^{P_J}_w=B^+wP_J/P_J X w P J = B + w P J / P J . This is a normal projective variety with open cell C w ≃ A ℓ ( w ) C_w\simeq\mathbb{A}^{\ell(w)} C w ≃ A ℓ ( w ) . The empty word gives a point with Cox ring C \mathbb{C} C .
Words and their minors
Let i = ( i 1 , … , i N ) \mathbf{i}=(i_1,\ldots,i_N) i = ( i 1 , … , i N ) be any word in the simple reflections. Here k k k denotes a position and i k i_k i k its simple-root index. The associated Bott–Samelson variety is
Z i = P i 1 × B + ⋯ × B + P i N / B + . Z_{\mathbf{i}}=P_{i_1}\times^{B^+}\cdots\times^{B^+}P_{i_N}/B^+. Z i = P i 1 × B + ⋯ × B + P i N / B + .
Let E j E_j E j be the divisor on which the j j j -th factor belongs to B + B^+ B + . The complement of their union is the affine chart C i ≃ A N C_{\mathbf{i}}\simeq\mathbb{A}^N C i ≃ A N , with coordinates given by ( z 1 , … , z N ) ↦ [ q i 1 ( z 1 ) , … , q i N ( z N ) ] (z_1,\ldots,z_N)\mapsto[q_{i_1}(z_1),\ldots,q_{i_N}(z_N)] ( z 1 , … , z N ) ↦ [ q i 1 ( z 1 ) , … , q i N ( z N )] . Here q i ( z ) = x i ( z ) s ˙ i q_i(z)=x_i(z)\dot{s}_i q i ( z ) = x i ( z ) s ˙ i , with the representative s ˙ i = x i ( − 1 ) x − i ( 1 ) x i ( − 1 ) \dot{s}_i=x_i(-1)x_{-i}(1)x_i(-1) s ˙ i = x i ( − 1 ) x − i ( 1 ) x i ( − 1 ) of [43 Equation (A.5)] . For each position k k k , define
f k = Δ Λ i k ( q i 1 ( z 1 ) ⋯ q i k ( z k ) ) . f_k=\Delta_{\Lambda_{i_k}}\left(q_{i_1}(z_1)\cdots q_{i_k}(z_k)\right). f k = Δ Λ i k ( q i 1 ( z 1 ) ⋯ q i k ( z k ) ) .
Here Δ λ \Delta_\lambda Δ λ is the highest matrix coefficient in the integrable highest-weight module of highest weight λ \lambda λ , normalized to have value one at the identity.
We write p ( k ) p(k) p ( k ) and s ( k ) s(k) s ( k ) for the preceding and following positions bearing the letter i k i_k i k , using − ∞ -\infty − ∞ and + ∞ +\infty + ∞ when the relevant position does not exist. A position is mutable when s ( k ) < + ∞ s(k)<+\infty s ( k ) < + ∞ ; the last occurrence of each letter is frozen. Define the lower unitriangular matrix L i L_{\mathbf{i}} L i and the mutable rows of the word matrix B B B by
( L i ) k = e k T + e p ( k ) T + ∑ j < k < s ( j ) a i j i k e j T , B k = ( L i ) k − ( L i ) s ( k ) , (L_{\mathbf{i}})_k=e_k^{\mathsf T}+e_{p(k)}^{\mathsf T}+\sum_{j<k<s(j)}a_{i_j i_k}e_j^{\mathsf T},\qquad B_k=(L_{\mathbf{i}})_k-(L_{\mathbf{i}})_{s(k)}, ( L i ) k = e k T + e p ( k ) T + j < k < s ( j ) ∑ a i j i k e j T , B k = ( L i ) k − ( L i ) s ( k ) ,
where a nonexistent predecessor contributes zero. The exchange at k k k is f k f k ′ = M k + ( f ) + M k − ( f ) f_kf'_k=M_k^+(f)+M_k^-(f) f k f k ′ = M k + ( f ) + M k − ( f ) , with M k ± ( f ) = ∏ j f j [ ± B k j ] + M_k^\pm(f)=\prod_j f_j^{[\pm B_{kj}]_+} M k ± ( f ) = ∏ j f j [ ± B k j ] + .
Lemma 3.1. The functions f 1 , … , f N f_1,\ldots,f_N f 1 , … , f N are pairwise nonassociated prime elements of C [ C i ] \mathbb{C}[C_{\mathbf{i}}] C [ C i ] , and they generate its fraction field. They form a seed with exchange matrix B B B , and every first exchange quotient f k ′ f'_k f k ′ belongs to C [ C i ] \mathbb{C}[C_{\mathbf{i}}] C [ C i ] .
Proof. Let Z Z Z be the Bott–Samelson variety of a prefix, let v v v be its Demazure product, and let m : Z → X v m: Z \to X_v m : Z → X v be the multiplication map. The group U ( v ) U(v) U ( v ) acts simply transitively on C v C_v C v , so equivariance gives
m − 1 ( C v ) ≃ U ( v ) × m − 1 ( v B + ) . m^{-1}(C_v) \simeq U(v) \times m^{-1}(vB^+). m − 1 ( C v ) ≃ U ( v ) × m − 1 ( v B + ) .
The left side is a nonempty open subset of Z Z Z , so it is smooth and irreducible. The same holds for the fiber, and over C v C_v C v the morphism m m m is the first projection.
Suppose that the letter i i i occurs in the prefix. The zero scheme of the highest section of L ( Λ i ) ∣ X v \mathcal{L}(\Lambda_i)|_{X_v} L ( Λ i ) ∣ X v is the Richardson divisor X v ∩ X s i X_v \cap X^{s_i} X v ∩ X s i . This divisor is reduced and irreducible, and its intersection with C v C_v C v is dense and nonempty [36 Proposition 6.6] . Its inverse image over C v C_v C v is therefore reduced and irreducible. Any other prime component W W W of the pullback would lie over the Schubert boundary. Choose C u C_u C u such that m − 1 ( C u ) ∩ W m^{-1}(C_u) \cap W m − 1 ( C u ) ∩ W is dense in W W W . Since W W W has codimension one, it is a component of the proper closed subset m − 1 ( X u ) m^{-1}(X_u) m − 1 ( X u ) . The connected group B + B^+ B + preserves each component. Thus m ( W ) m(W) m ( W ) contains C u C_u C u , and the highest section would vanish on X u X_u X u , although it takes the value one at the identity flag. This excludes every further component. The pullback is now generically reduced, and as a Cartier divisor on the smooth variety Z Z Z it has no embedded components. It is therefore reduced.
Apply this argument to the prefix ending at k k k , with i = i k i=i_k i = i k . The pullback section is nonzero at the point whose factors are all the identity. Every standard boundary divisor contains this point, so none is a component of its zero divisor. Its restriction f k f_k f k to the affine chart is consequently a nonconstant prime, and remains prime after the later coordinates are adjoined. The rank-one calculation gives
f k = f p ( k ) z k + h k ( z 1 , … , z k − 1 ) , f − ∞ = 1. f_k=f_{p(k)}z_k+h_k(z_1,\ldots,z_{k-1}), \qquad f_{-\infty}=1. f k = f p ( k ) z k + h k ( z 1 , … , z k − 1 ) , f − ∞ = 1.
The coefficient is f p ( k ) f_{p(k)} f p ( k ) because the intervening factors fix the highest-weight vector. These triangular identities recover the z k z_k z k successively from the f k f_k f k . They also show that f k f_k f k depends on z k z_k z k , whereas the preceding minors do not. Thus the primes are pairwise nonassociated and generate the fraction field.
To check regularity of the first exchanges, let C i ∘ C_i^\circ C i ∘ be the complement of the terminal minor divisors in C i C_i C i . Consider the decorated double Bott–Samelson cell for the positive braid i \mathbf{i} i and the empty braid. Normalize the initial opposite decorated flags to ( U − h , U + ) (U^-h,U^+) ( U − h , U + ) , with h ∈ T h\in T h ∈ T . Propagating the positive decorations gives q i 1 ( z 1 ) ⋯ q i k ( z k ) U + q_{i_1}(z_1)\cdots q_{i_k}(z_k)U^+ q i 1 ( z 1 ) ⋯ q i k ( z k ) U + , and the terminal opposition condition is exactly the nonvanishing of the terminal minors. Thus the decorated cell is T × C i ∘ T\times C_i^\circ T × C i ∘ . In the coordinates of [SW, Definition 3.20], the left open string on level i i i has coordinate h Λ i h^{\Lambda_i} h Λ i , and the string after position k k k has coordinate h Λ i k f k h^{\Lambda_{i_k}}f_k h Λ i k f k . Setting the left open coordinates to one gives the stated matrix B B B , directly from its string construction. For a Cartan matrix of nonzero corank, the additional Cartan characters are frozen and are also set to one. An index absent from the word contributes the constant coefficient one.
By [43 Theorem 3.45] , the first exchange quotients are regular on the decorated cell. Specializing h = 1 h=1 h = 1 gives f k ′ ∈ C [ C i ] f'_k\in\mathbb{C}[C_i] f k ′ ∈ C [ C i ] . Its exchange formula allows a pole only along f k = 0 f_k=0 f k = 0 . The generic point of this divisor lies in C i ∘ C_i^\circ C i ∘ , since the mutable prime f k f_k f k is distinct from every terminal prime. Thus f k ′ f'_k f k ′ is regular on C i C_i C i . ◻
For k ≤ N k\leq N k ≤ N , let m k : Z i → G / B + m_k: Z_i\to G/B^+ m k : Z i → G / B + multiply the first k k k factors, and write L k ( λ ) = m k ∗ L ( λ ) \mathcal{L}_k(\lambda)=m_k^*\mathcal{L}(\lambda) L k ( λ ) = m k ∗ L ( λ ) , with L 0 ( λ ) = O \mathcal{L}_0(\lambda)=\mathcal{O} L 0 ( λ ) = O . The same recursion on k k k determines the boundary coefficients of the lifted exchanges and the class appearing in Theorem 1.2 .
Lemma 3.2. The minimal lift x k x_k x k of f k f_k f k has degree d k = [ L k ( Λ i k ) ] d_k=[\mathcal{L}_k(\Lambda_{i_k})] d k = [ L k ( Λ i k )] . In the standard boundary basis, its coefficients are
ν j k = { ⟨ Λ i k , s i k ⋯ s i j + 1 ( α i j ∨ ) ⟩ , j ≤ k , 0 , j > k . (8) \nu_{jk}=
\begin{cases}
\left\langle\Lambda_{i_k},s_{i_k}\cdots s_{i_{j+1}}(\alpha_{i_j}^{\vee})\right\rangle, & j\leq k,\\
0, & j>k.
\end{cases} \tag*{(8)} ν j k = { ⟨ Λ i k , s i k ⋯ s i j + 1 ( α i j ∨ ) ⟩ , 0 , j ≤ k , j > k . (8) Furthermore,
L i V T = I N , ( B V T ) k = e k T − e s ( k ) T . (9) L_iV^{\mathsf T}=I_N,\qquad(BV^{\mathsf T})_k=e_k^{\mathsf T}-e_{s(k)}^{\mathsf T}. \tag*{(9)} L i V T = I N , ( B V T ) k = e k T − e s ( k ) T . (9)
The terminal degrees, together with the standard boundary classes, sum to − K Z i -K_{Z_i} − K Z i .
Proof. The highest section of L k ( Λ i k ) \mathcal{L}_k(\Lambda_{i_k}) L k ( Λ i k ) restricts to f k f_k f k in the standard frame of the affine chart. It has no standard boundary component, as in the previous proof, so it is the minimal lift. For a general weight λ \lambda λ , the last P 1 \mathbb{P}^1 P 1 -bundle in a prefix gives
[ L k ( λ ) ] = ⟨ λ , α i k ∨ ⟩ [ E k ] + [ L k − 1 ( s i k λ ) ] . (10) [\mathcal{L}_k(\lambda)] = \langle\lambda,\alpha_{i_k}^{\vee}\rangle[E_k] + [\mathcal{L}_{k-1}(s_{i_k}\lambda)]. \tag*{(10)} [ L k ( λ )] = ⟨ λ , α i k ∨ ⟩ [ E k ] + [ L k − 1 ( s i k λ )] . (10)
Indeed, L k ( λ ) \mathcal{L}_k(\lambda) L k ( λ ) has fiber degree m = ⟨ λ , α i k ∨ ⟩ m=\langle\lambda,\alpha_{i_k}^{\vee}\rangle m = ⟨ λ , α i k ∨ ⟩ , so L k ( λ ) ( − m E k ) \mathcal{L}_k(\lambda)(-mE_k) L k ( λ ) ( − m E k ) descends to the preceding prefix. Its restriction to the standard section E k E_k E k is L k − 1 ( λ − m α i k ) \mathcal{L}_{k-1}(\lambda-m\alpha_{i_k}) L k − 1 ( λ − m α i k ) , since the normal bundle of that section is L k − 1 ( α i k ) \mathcal{L}_{k-1}(\alpha_{i_k}) L k − 1 ( α i k ) . Iterating (10) proves (8) .
On a prefix, the fundamental line bundle indexed by a simple root is determined by the last occurrence of that index. Since a weight that pairs trivially with all simple coroots gives a trivial line bundle on a Bott–Samelson variety, the simple-coroot pairings give
[ L k − 1 ( α i k ) ] = 2 d p ( k ) + ∑ j < k < s ( j ) a i j i k d j . [\mathcal{L}_{k-1}(\alpha_{i_k})] = 2d_{p(k)}+\sum_{j<k<s(j)}a_{i_j i_k}d_j. [ L k − 1 ( α i k )] = 2 d p ( k ) + j < k < s ( j ) ∑ a i j i k d j .
with d − ∞ = 0 d_{-\infty}=0 d − ∞ = 0 . Substituting s i k Λ i k = Λ i k − α i k s_{i_k}\Lambda_{i_k}=\Lambda_{i_k}-\alpha_{i_k} s i k Λ i k = Λ i k − α i k into the recursion therefore gives
d k + d p ( k ) + ∑ j < k < s ( j ) a i j i k d j = [ E k ] . d_k+d_{p(k)}+\sum_{j<k<s(j)}a_{i_j i_k}d_j=[E_k]. d k + d p ( k ) + j < k < s ( j ) ∑ a i j i k d j = [ E k ] .
Writing this relation in the boundary basis gives the first identity in (9) . Subtracting the relation at s ( k ) s(k) s ( k ) gives the second. In particular, ν \nu ν is upper unitriangular.
The relative canonical bundle of the last P 1 \mathbb{P}^1 P 1 -bundle has class − 2 [ E k ] + [ L k − 1 ( α i k ) ] -2[E_k]+[\mathcal{L}_{k-1}(\alpha_{i_k})] − 2 [ E k ] + [ L k − 1 ( α i k )] . Together with (10) , induction therefore gives
− K Z i = ∑ j = 1 N [ E j ] + [ L N ( ρ ) ] . -K_{Z_i}=\sum_{j=1}^{N}[E_j]+[\mathcal{L}_N(\rho)]. − K Z i = j = 1 ∑ N [ E j ] + [ L N ( ρ )] .
If k k k is the last occurrence of i i i , every later simple reflection fixes Λ i \Lambda_i Λ i ; consequently L N ( Λ i ) = L k ( Λ i ) \mathcal{L}_N(\Lambda_i)=\mathcal{L}_k(\Lambda_i) L N ( Λ i ) = L k ( Λ i ) . The fundamental weights indexed by simple roots absent from the word give trivial bundles. Since ρ \rho ρ is the sum of the fundamental weights, the terminal degrees sum to [ L N ( ρ ) ] [\mathcal{L}_N(\rho)] [ L N ( ρ )] , as required. □
Theorem 3.3. For every word in the simple reflections of a symmetrizable Kac–Moody group, minimally lift the seed of Lemma 3.1 to Cox ( Z i ) \operatorname{Cox}(Z_i) Cox ( Z i ) and adjoin the standard boundary sections τ 1 , … , τ N \tau_1,\ldots,\tau_N τ 1 , … , τ N as frozen variables. The resulting seed Σ i \Sigma_i Σ i satisfies
Cox ( Z i ) = L p o l ( Σ i ) ∩ ⋂ k m u t a b l e L p o l ( μ k Σ i ) = U p o l ( Σ i ) . \operatorname{Cox}(Z_i)=\mathcal{L}_{\mathrm{pol}}(\Sigma_i)\cap\bigcap_{k\ \mathrm{mutable}}\mathcal{L}_{\mathrm{pol}}(\mu_k\Sigma_i)=\mathcal{U}_{\mathrm{pol}}(\Sigma_i). Cox ( Z i ) = L pol ( Σ i ) ∩ k mutable ⋂ L pol ( μ k Σ i ) = U pol ( Σ i ) . If the word contains s s s distinct letters, the seed has N − s N-s N − s mutable and N + s N+s N + s frozen variables. Its extended exchange matrix B ~ = ( B ∣ − B ν T ) \widetilde{B}=(B\mid-B\nu^{\mathsf{T}}) B = ( B ∣ − B ν T ) has full row rank and primitive rows. The exchanges are
x k x k ′ = τ s ( k ) M k + ( x ) + τ k M k − ( x ) . (11) x_kx_k'= \tau_{s(k)}M_k^+(x)+\tau_kM_k^-(x). \tag*{(11)} x k x k ′ = τ s ( k ) M k + ( x ) + τ k M k − ( x ) . (11) Its frozen degrees sum to − K Z i -K_{Z_i} − K Z i . The initial and adjacent mixed Laurent charts are smooth and cover the Cox space outside a subset of codimension at least two.
Proof. The boundary classes freely generate Pic ( Z i ) \operatorname{Pic}(Z_i) Pic ( Z i ) , and the Cox ring is factorial [3 Theorem 1.1] . Bott–Samelson varieties for arbitrary words are log Fano [1 Theorem A.2] , so their Cox rings are finitely generated [27 Corollary 5.3] . Choose an ample divisor H H H on Z i Z_i Z i . For N > 0 N>0 N > 0 , intersection with H N − 1 H^{N-1} H N − 1 makes the Cox ring positively graded with degree-zero part C \mathbb{C} C , since every nonzero effective divisor meets H N − 1 H^{N-1} H N − 1 positively. The case N = 0 N=0 N = 0 is immediate. Choose the coordinate volume form ω C = d z 1 ∧ ⋯ ∧ d z N \omega_C = dz_1 \wedge\cdots\wedge dz_N ω C = d z 1 ∧ ⋯ ∧ d z N . Its divisor is supported on the standard boundary. Write − K Z i = ∑ j κ j [ E j ] -K_{Z_i} = \sum_j \kappa_j [E_j] − K Z i = ∑ j κ j [ E j ] , using Lemma 3.2 . Independence of the boundary classes gives div ( ω C ) = − ∑ j κ j E j \operatorname{div}(\omega_C) = -\sum_j \kappa_j E_j div ( ω C ) = − ∑ j κ j E j . Consequently
Ω = τ κ ω C ∧ d log τ 1 ∧ ⋯ ∧ d log τ N \Omega= \tau^\kappa\omega_C \wedge d\log\tau_1 \wedge\cdots\wedge d\log\tau_N Ω = τ κ ω C ∧ d log τ 1 ∧ ⋯ ∧ d log τ N
has degree − K Z i -K_{Z_i} − K Z i . To compute its divisor, recall that the characteristic space is a torus torsor over the smooth variety Z i Z_i Z i , and is an open subset of Spec Cox ( Z i ) \operatorname{Spec} \operatorname{Cox}(Z_i) Spec Cox ( Z i ) with complement of codimension at least two [2 Construction 1.6.3.1] . Near the generic point of E j E_j E j , choose a local equation t t t and torus coordinates u a u_a u a such that τ j = t u j \tau_j = t u_j τ j = t u j and τ a = u a \tau_a = u_a τ a = u a for a ≠ j a \ne j a = j . Choose local parameters t , t 2 , … , t N t,t_2,\ldots,t_N t , t 2 , … , t N on the base and write ω C = t − κ j v d t ∧ d t 2 ∧ ⋯ ∧ d t N \omega_C = t^{-\kappa_j}v\,dt \wedge dt_2 \wedge\cdots\wedge dt_N ω C = t − κ j v d t ∧ d t 2 ∧ ⋯ ∧ d t N , where v v v is a unit. In the expression for Ω \Omega Ω , the powers of t t t cancel, and the term d t / t dt/t d t / t from d log τ j d\log\tau_j d log τ j vanishes when wedged with d t dt d t . The remaining coefficient v u κ v u^\kappa v u κ is a unit for every integral κ \kappa κ . Over the affine chart the form is also nowhere vanishing, so its divisor on the Cox space is zero.
The lifted minors and the boundary sections are the canonical sections of distinct prime divisors and are therefore pairwise nonassociated primes in the Cox ring. Together they generate its fraction field and form a transcendence basis. Lemma 3.1 gives regular first exchange quotients on the affine chart. At each boundary prime the initial minimal lift in the denominator has order zero, and both terms of the homogenized numerator have nonnegative order. Thus the lifted quotients are regular on the Cox space by normality.
The second identity in (9) proves (11) and makes every exchange row primitive. For each fixed letter, the vectors e k − e s ( k ) e_k-e_{s(k)} e k − e s ( k ) are the successive edge vectors of a path, and different letters have disjoint supports. Hence B ~ \widetilde{B} B has full row rank. Finally, the frozen-degree identity in Lemma 3.2 is precisely the equality required by the canonical-degree criterion. That criterion proves all the asserted equalities and the statement about mixed Laurent charts. □
For explicit canonical coefficients, define γ j = s i N ⋯ s i j + 1 α i j \gamma_j = s_{i_N}\cdots s_{i_{j+1}}\alpha_{i_j} γ j = s i N ⋯ s i j + 1 α i j . These roots need not be positive. Formula (8) gives
− K Z i = ∑ j = 1 N ( 1 + ⟨ ρ , γ j ∨ ⟩ ) [ E j ] . -K_{Z_i} = \sum_{j=1}^{N}\left(1+\langle\rho,\gamma_j^\vee\rangle\right)[E_j]. − K Z i = j = 1 ∑ N ( 1 + ⟨ ρ , γ j ∨ ⟩ ) [ E j ] .
For example, in type C 2 C_2 C 2 take a 12 = − 2 a_{12}=-2 a 12 = − 2 , a 21 = − 1 a_{21}=-1 a 21 = − 1 and the word ( 2 , 1 , 2 ) (2,1,2) ( 2 , 1 , 2 ) . Its minors are z 1 z_1 z 1 , z 2 z_2 z 2 , z 1 z 3 − z 2 2 z_1z_3-z_2^2 z 1 z 3 − z 2 2 , so the first mutant on the cell is z 3 z_3 z 3 . The lifted exchange x ~ 1 x 1 ′ = τ 3 x 2 2 + τ 1 x 3 \widetilde{x}_1x'_1=\tau_3x_2^2+\tau_1x_3 x 1 x 1 ′ = τ 3 x 2 2 + τ 1 x 3 records both the Cartan entry and the two boundary positions indexed by the simple root 2.
For a nonreduced example, take the word i = ( 1 , … , 1 ) \mathbf{i}=(1,\ldots,1) i = ( 1 , … , 1 ) of length n n n in type A 1 A_1 A 1 . Then Z i = ( P 1 ) n Z_{\mathbf{i}}=(\mathbb{P}^1)^n Z i = ( P 1 ) n and its Cox ring is polynomial in the homogeneous coordinates [ a j : b j ] [a_j:b_j] [ a j : b j ] of the factors. With a 0 = 1 a_0=1 a 0 = 1 , the boundary sections are τ 1 = b 1 \tau_1=b_1 τ 1 = b 1 and τ j = a j − 1 b j − b j − 1 a j \tau_j=a_{j-1}b_j-b_{j-1}a_j τ j = a j − 1 b j − b j − 1 a j for j > 1 j>1 j > 1 ; the lifted exchanges are
a k a k ′ = τ k + 1 a k − 1 + τ k a k + 1 ( 1 ≤ k < n ) . a_k a'_k=\tau_{k+1}a_{k-1}+\tau_k a_{k+1}\qquad(1\le k<n). a k a k ′ = τ k + 1 a k − 1 + τ k a k + 1 ( 1 ≤ k < n ) .
Successively mutating at 1 , … , n − 1 1,\ldots,n-1 1 , … , n − 1 gives the chart adapted to the diagonal positive-root action. This illustrates the arbitrary-word construction without any reducedness assumption.
Contractions and partial flag varieties.
Theorem 3.4. Let p : Z i → Y p:Z_i\to Y p : Z i → Y be a proper birational morphism to a normal projective variety, and suppose that p p p identifies C i C_i C i with an open subset of Y Y Y . Delete from Σ i \Sigma_i Σ i the frozen columns of the standard boundary divisors contracted by p p p , and set their canonical sections equal to one. The resulting seed Σ Y \Sigma_Y Σ Y satisfies
Cox ( Y ) = L pol ( Σ Y ) ∩ ⋂ k m u t a b l e L pol ( μ k Σ Y ) = U pol ( Σ Y ) . \operatorname{Cox}(Y)=\mathcal{L}_{\operatorname{pol}}(\Sigma_Y)\cap\bigcap_{k\ \mathrm{mutable}}\mathcal{L}_{\operatorname{pol}}(\mu_k\Sigma_Y)=\mathcal{U}_{\operatorname{pol}}(\Sigma_Y). Cox ( Y ) = L pol ( Σ Y ) ∩ k mutable ⋂ L pol ( μ k Σ Y ) = U pol ( Σ Y ) .
Every nonboundary seed coordinate is its minimal homogeneous lift from C i C_i C i . The frozen degrees sum to − K Y -K_Y − K Y , and the initial and adjacent mixed Laurent charts cover the Cox space in codimension one.
In particular, this construction applies to every Schubert variety Y = X w P J Y = X^{P_J}_w Y = X w P J and every reduced expression i \mathbf{i} i of w w w . If w j ^ \widehat{w^j} w j is the Weyl-group product obtained by deleting position j j j , the surviving standard boundary sections are exactly those satisfying ℓ ( w j ^ ) = ℓ ( w ) − 1 \ell(\widehat{w^j}) = \ell(w) - 1 ℓ ( w j ) = ℓ ( w ) − 1 and w j ^ ∈ W J \widehat{w^j} \in W^J w j ∈ W J .
Proof. Apply Theorem 2.2 to the seed in Theorem 3.3 . This gives the two ring equalities and identifies the surviving boundary columns. A rational top form has the same order along corresponding nonexceptional prime divisors. Thus p ∗ K Z i = K Y p_*K_{Z_{\mathbf{i}}} = K_Y p ∗ K Z i = K Y , and pushing forward the frozen degrees gives their sum on Y Y Y .
The quotient description makes Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) finitely generated. Since Cl ( Y ) \operatorname{Cl}(Y) Cl ( Y ) is free, its Cox ring is factorial [3 ] . A minimal lift of a cell prime is the canonical homogeneous section of the closure of its prime zero divisor and is therefore prime in the Cox ring. The same holds for the boundary sections. Each initial or adjacent mixed Laurent ring is a localization of Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) at that seed’s mutable coordinates. As in the last paragraph of the proof of Theorem 1.2 , these charts cover every prime divisor. Proposition 3.7 below gives minimality throughout the mutation pattern.
For a reduced word of w ∈ W J w \in W^J w ∈ W J , multiplication gives a proper birational morphism to X w P J X^{P_J}_w X w P J and identifies the open cells. The image of the standard divisor at position j j j is indexed by the minimal representative modulo W J W_J W J of the Demazure product of the deleted word. It has dimension ℓ ( w ) − 1 \ell(w) - 1 ℓ ( w ) − 1 exactly when the deleted word is reduced and its product belongs to W J W^J W J . This proves the survival criterion. □
For a reduced word the cell functions above are the usual unipotent minors . Under C w ≃ U ( w ) = U + ∩ w U − w − 1 C_w \simeq U(w) = U^+ \cap wU^-w^{-1} C w ≃ U ( w ) = U + ∩ w U − w − 1 , they are f k = Δ Λ i k , w ≤ k Λ i k ∣ U ( w ) f_k = \Delta_{\Lambda_{i_k},w\leq k\Lambda_{i_k}}|_{U(w)} f k = Δ Λ i k , w ≤ k Λ i k ∣ U ( w ) . The valuation matrix for the partial flag variety consists of the surviving rows of ν \nu ν .
Proposition 3.5. Let J ⊂ K J \subset K J ⊂ K be subsets of the simple-root index set, let w ∈ W K w \in W^K w ∈ W K , and let p : X w P J ⟶ X w P K p : X^{P_J}_w \longrightarrow X^{P_K}_w p : X w P J ⟶ X w P K be the standard projection. Construct both lifted seeds from the same reduced expression of w w w . The map q p q_p q p of Proposition 2.1 sets equal to one the boundary frozen variables indexed by
v ∈ W J , v < w , ℓ ( v ) = ℓ ( w ) − 1 , v ∉ W K , v \in W^J,\qquad v < w,\qquad\ell(v) = \ell(w) - 1,\qquad v \notin W^K, v ∈ W J , v < w , ℓ ( v ) = ℓ ( w ) − 1 , v ∈ / W K , and deletes their columns from the extended exchange matrix. It sends every initial minimal lift to the corresponding minimal lift on X w P K X^{P_K}_w X w P K , and it sends corresponding cluster variables to one another after any fixed sequence of mutations. Consequently,
U p o l ( Σ K ) ≃ U p o l ( Σ J ) / ( τ v − 1 : v ∈ W J ∖ W K , v < w , ℓ ( v ) = ℓ ( w ) − 1 ) . \mathcal{U}_{\mathrm{pol}}(\Sigma_K) \simeq\mathcal{U}_{\mathrm{pol}}(\Sigma_J)/(\tau_v - 1 : v \in W^J \setminus W^K,\ v < w,\ \ell(v) = \ell(w) - 1). U pol ( Σ K ) ≃ U pol ( Σ J ) / ( τ v − 1 : v ∈ W J ∖ W K , v < w , ℓ ( v ) = ℓ ( w ) − 1 ) . For J ⊂ K ⊂ L J \subset K \subset L J ⊂ K ⊂ L and w ∈ W L w \in W^L w ∈ W L , these specializations commute with composition.
Proof. Since w ∈ W K ⊂ W J w \in W^K \subset W^J w ∈ W K ⊂ W J , the projection identifies the open cells of the two Schubert varieties. Both cells have dimension ℓ ( w ) \ell(w) ℓ ( w ) , so the projection is birational. The boundary prime divisors of X w P J X^{P_J}_w X w P J are X v P J X^{P_J}_v X v P J , where v ∈ W J v \in W^J v ∈ W J , v < w v < w v < w , and ℓ ( v ) = ℓ ( w ) − 1 \ell(v) = \ell(w) - 1 ℓ ( v ) = ℓ ( w ) − 1 . The image of this divisor is X v K P K X^{P_K}_{v^K} X v K P K , where v K v^K v K is the minimal representative of v W K vW_K v W K . Its dimension is ℓ ( v K ) \ell(v^K) ℓ ( v K ) , so it is a divisor precisely when v = v K v = v^K v = v K . This identifies the exceptional columns.
Theorem 2.2 now gives the specialization of both the initial lifts and all their mutations. By Theorem 3.4, each upper algebra equals the corresponding Cox ring. The quotient formula and the composition law follow from Proposition 2.1 . □
The universal grading
Let S S S be the set of simple-root indices occurring in i \mathbf{i} i , and let Q S = ⨁ i ∈ S Z α i Q_S = \bigoplus_{i \in S} \mathbb{Z}\alpha_i Q S = ⨁ i ∈ S Z α i be the support root lattice. The word seed carries the Q S Q_S Q S -grading
λ k = deg Q S f k = Λ i k − w ≤ k Λ i k . \lambda_k = \deg_{Q_S} f_k = \Lambda_{i_k} - w_{\leq k}\Lambda_{i_k}. λ k = deg Q S f k = Λ i k − w ≤ k Λ i k .
Here w ≤ k = s i 1 ⋯ s i k w_{\le k}=s_{i_1}\cdots s_{i_k} w ≤ k = s i 1 ⋯ s i k is the ordinary Weyl-group product, whether or not the prefix is reduced. Giving the canonical boundary sections root degree zero makes deg Q S x k = λ k \deg_{Q_S}x_k=\lambda_k deg Q S x k = λ k in every compactification. On cell functions the inverse-pullback Cartan weight is − λ k -\lambda_k − λ k . In a divisorial section space the chosen linearization adds a character depending on the divisor class. The representation-theoretic statements use these shifted Cartan weights.
Proposition 3.6. Let Y Y Y be a Schubert variety, or a normal projective contraction of Z i Z_i Z i that is an isomorphism on the open cell. Write d d d for the number of boundary prime divisors of Y Y Y , and let B ~ = ( B ∣ − B ν T ) \widetilde{B}=(B\mid-B\nu^{\mathsf T}) B = ( B ∣ − B ν T ) be its lifted exchange matrix. The homomorphism
D : Z N + d ⟶ Cl ( Y ) ⊕ Q S , e k ⟼ ( ν ∙ k , λ k ) , e N + a ⟼ ( [ D a ] , 0 ) , D:\mathbb{Z}^{N+d}\longrightarrow\operatorname{Cl}(Y)\oplus Q_S,\qquad e_k\longmapsto(\nu_{\bullet k},\lambda_k),\quad e_{N+a}\longmapsto([D_a],0), D : Z N + d ⟶ Cl ( Y ) ⊕ Q S , e k ⟼ ( ν ∙ k , λ k ) , e N + a ⟼ ([ D a ] , 0 ) , fits into an exact sequence
0 ⟶ Z N − ∣ S ∣ → B ~ T Z N + d → D Cl ( Y ) ⊕ Q S ⟶ 0. (12) 0\longrightarrow\mathbb{Z}^{N-|S|}\xrightarrow{\widetilde{B}^{\mathsf T}}\mathbb{Z}^{N+d}\xrightarrow{D}\operatorname{Cl}(Y)\oplus Q_S\longrightarrow0. \tag*{(12)} 0 ⟶ Z N − ∣ S ∣ B T Z N + d D Cl ( Y ) ⊕ Q S ⟶ 0. (12) Consequently, the divisor-class and root-lattice gradings together give the universal grading of the lifted seed: every homogeneous grading of the seed with values in an abelian group factors uniquely through D D D .
Proof. The cell exchanges are homogeneous for the root grading, so B ( λ 1 , … , λ N ) T = 0 B(\lambda_1,\ldots,\lambda_N)^{\mathsf T}=0 B ( λ 1 , … , λ N ) T = 0 . The formula for the added frozen columns also gives D B ~ T = 0 D\widetilde{B}^{\mathsf T}=0 D B T = 0 .
We first show that D D D is surjective. The boundary columns generate Cl ( Y ) \operatorname{Cl}(Y) Cl ( Y ) . If k k k is the first occurrence of i k i_k i k , then λ k = w ≤ k − 1 α i k \lambda_k=w_{\le k-1}\alpha_{i_k} λ k = w ≤ k − 1 α i k . The preceding product belongs to the parabolic subgroup generated by the earlier letters, none of which equals i k i_k i k . Consequently this root is positive, its α i k \alpha_{i_k} α i k -coefficient is one, and every other nonzero simple-root coefficient is indexed by a letter occurring earlier in the word. Ordering these first occurrences from left to right therefore gives a unitriangular matrix in the simple-root basis of Q S Q_S Q S . The λ k \lambda_k λ k generate Q S Q_S Q S , proving surjectivity.
The successor columns give a maximal minor of B B B with determinant ± 1 \pm1 ± 1 . Indeed, the entry in column s ( k ) s(k) s ( k ) of mutable row k k k is − 1 -1 − 1 , and that row has no nonzero entry in a column greater than s ( k ) s(k) s ( k ) . Order the mutable rows by increasing successor and use the corresponding successor columns in the same order. The resulting square matrix is triangular with diagonal entries − 1 -1 − 1 . Hence B ~ T \widetilde{B}^{\mathsf T} B T is injective and its image is a saturated sublattice of Z N + d \mathbb{Z}^{N+d} Z N + d . Its rank is N − ∣ S ∣ N-|S| N − ∣ S ∣ , which is also the rank of ker D \ker D ker D . The inclusion already proved must therefore be an equality. This gives (12) and the universal property. □
In particular, the torus of diagonal rescalings compatible with the exchange relations is the product of the Cox grading torus and the effective Cartan torus Hom ( Q S , C × ) \operatorname{Hom}(Q_S,\mathbb{C}^{\times}) Hom ( Q S , C × ) . A birational projection acts on the universal grading by pushforward on Cl ( Y ) \operatorname{Cl}(Y) Cl ( Y ) and the identity on Q S Q_S Q S .
Minimal lifting and changes of reduced expression
Different root operators may require different reduced expressions of w w w . The corresponding lifted seeds lie in one mutation class: boundary minimality is preserved by mutation, and braid moves connect the unipotent-minor seeds.
Proposition 3.7. Let Y Y Y be a normal projective variety with an affine open subset C C C , and suppose that the prime divisors D 1 , … , D d D_1,\ldots,D_d D 1 , … , D d in Y ∖ C Y\setminus C Y ∖ C freely generate Cl ( Y ) \operatorname{Cl}(Y) Cl ( Y ) . Let τ a \tau_a τ a be the canonical section of D a D_a D a , and realize the Cox ring as R ⊂ C [ C ] [ τ 1 ± 1 , … , τ d ± 1 ] R\subset\mathbb{C}[C][\tau_1^{\pm1},\ldots,\tau_d^{\pm1}] R ⊂ C [ C ] [ τ 1 ± 1 , … , τ d ± 1 ] . Suppose that the minimal homogeneous lift Σ ~ \widetilde{\Sigma} Σ of a seed Σ \Sigma Σ on C C C , with the τ a \tau_a τ a adjoined as polynomial frozen variables, satisfies R = U p o l ( Σ ~ ) R=\mathcal{U}_{\mathrm{pol}}(\widetilde{\Sigma}) R = U pol ( Σ ) . Then every seed obtained from Σ ~ \widetilde{\Sigma} Σ by mutation is the minimal homogeneous lift of the corresponding seed obtained from Σ \Sigma Σ .
In particular, minimal lifting commutes with mutation and relabeling, and it identifies the labeled exchange graphs of these two cluster patterns, with the boundary variables fixed.
Proof. The lifted exchanges are homogeneous and specialize, after setting τ a = 1 \tau_a=1 τ a = 1 , to the cell exchanges at every seed. Hence every lifted cluster coordinate has the form X = τ m f X=\tau^m f X = τ m f , where f f f is the corresponding cell coordinate. The Laurent phenomenon with polynomial frozen variables gives X ∈ R X \in R X ∈ R .
In its own mixed Laurent chart a mutable coordinate X X X is invertible, while each τ a \tau_a τ a cuts out a divisor. Thus no τ a \tau_a τ a divides X X X in R R R . A homogeneous section τ m f \tau^m f τ m f is divisible by τ a \tau_a τ a precisely when m a + ord D a ( f ) > 0 m_a+\operatorname{ord}_{D_a}(f)>0 m a + ord D a ( f ) > 0 . Therefore
m a = − ord D a ( f ) ( 1 ≤ a ≤ d ) . m_a=-\operatorname{ord}_{D_a}(f) \qquad(1\leq a\leq d). m a = − ord D a ( f ) ( 1 ≤ a ≤ d ) .
This is exactly the minimal lift. The original frozen coordinates do not change under mutation, so their lifts remain minimal as well.
The boundary part of the exchange matrix is also determined by these degrees. If B ′ B' B ′ is the cell exchange matrix and ν ′ \nu' ν ′ records the degrees of its minimal lifts, homogeneity forces the lifted matrix to be ( B ′ ∣ − B ′ ( ν ′ ) T ) (B'\mid-B'(\nu')^{\mathsf T}) ( B ′ ∣ − B ′ ( ν ′ ) T ) . Consequently two mutation paths with the same terminal labeled seed on C C C have the same lifted terminal seed. Specialization gives the inverse correspondence.
Monomial lifting and its compatibility with mutation were developed by Francone [20 Proposition 1.0.3 and Lemma 3.0.8] . The point of Proposition 3.7 is that the equality with the Cox ring forces every coordinate in the lifted pattern to remain minimal.
Lemma 3.8. Let w w w belong to the Weyl group of a symmetrizable Kac–Moody group. The normalized unipotent-minor seeds of C [ U ( w ) ] \mathbb{C}[U(w)] C [ U ( w )] associated with all reduced expressions of w w w belong to one mutation class.
Proof. We compare these seeds with the double Bruhat cell G e , w − 1 G^{e,w^{-1}} G e , w − 1 . For a reduced expression i = ( i 1 , … , i N ) \mathbf{i}=(i_1,\ldots,i_N) i = ( i 1 , … , i N ) of w w w , use the reversed positive double word i o p = ( i N , … , i 1 ) \mathbf{i}^{\mathrm{op}}=(i_N,\ldots,i_1) i op = ( i N , … , i 1 ) . The chamber minors of this double word consist of the functions
Δ Λ i k , w ≤ k Λ i k ( 1 ≤ k ≤ N ) , h i = Δ Λ i , Λ i ( i ∈ I 0 ) . \Delta_{\Lambda_{i_k},w_{\leq k}\Lambda_{i_k}}\quad(1\leq k\leq N),\qquad h_i=\Delta_{\Lambda_i,\Lambda_i}\quad(i\in I_0). Δ Λ i k , w ≤ k Λ i k ( 1 ≤ k ≤ N ) , h i = Δ Λ i , Λ i ( i ∈ I 0 ) . The chamber minor at the position corresponding to k k k is the prefix minor at the preceding occurrence of i k i_k i k , or h i k h_{i_k} h i k when there is no preceding occurrence. The initial frozen minor indexed by i i i is the terminal prefix minor for that index. This follows directly from the conventions for double-word minors in [6 Equations (2.9)–(2.12)] .
The torus factor in G e , w − 1 ⊂ B + G^{e,w^{-1}}\subset B^+ G e , w − 1 ⊂ B + is specified by the h i h_i h i . For a realization with additional torus directions, we also adjoin the corresponding character functions as frozen coordinates and specialize them to one. Write U = U + U=U^+ U = U + . After setting the torus coordinates to one, the chamber minors restrict to U ∩ B − w − 1 B − U\cap B^-w^{-1}B^- U ∩ B − w − 1 B − . To compare them with functions on U ( w ) U(w) U ( w ) , write U = U ( w ) U ′ U=U(w)U' U = U ( w ) U ′ , where U ′ = U ∩ w U w − 1 U'=U\cap wUw^{-1} U ′ = U ∩ w U w − 1 , and let q q q be the projection to U ( w ) U(w) U ( w ) . If u u u is a prefix of a reduced expression of w w w , then U ′ ⊂ u U u − 1 U'\subset uUu^{-1} U ′ ⊂ uU u − 1 . Hence U ′ U' U ′ fixes the extremal vector of weight u Λ i u\Lambda_i u Λ i , and the minor Δ Λ i , u Λ i \Delta_{\Lambda_i,u\Lambda_i} Δ Λ i , u Λ i is right U ′ U' U ′ -invariant. Its restriction is therefore the pullback under q q q of the same unipotent minor on U ( w ) U(w) U ( w ) .
These restricted chamber minors are algebraically independent: the chamber minors and the h i h_i h i form a torus chart on the double Bruhat cell, and the torus factor splits off. It follows that the restriction of q q q to U ∩ B − w − 1 B − U\cap B^-w^{-1}B^- U ∩ B − w − 1 B − is dominant. Thus identities among these minors can be checked on that double Bruhat cell.
Shen and Weng prove that braid moves act by the corresponding sequences of mutations, both on the exchange matrices and on the chamber-minor functions, for every symmetrizable Cartan matrix [43 Propositions 3.7 and 3.25] . Their sequences leave the principal minors h i h_i h i frozen. Specializing them to 1 1 1 therefore gives the same mutation sequences for our unipotent minors. The exchange matrices specialize to the reduced-word matrices; this also follows directly by reversing the word and replacing each position by its predecessor. The equality of functions follows from the dominance just proved. Finally, any two reduced expressions are joined by braid moves.
Corollary 3.9. For every Schubert variety Y = X w P Y=X^P_w Y = X w P , the minimally lifted seeds associated with all reduced expressions of w w w belong to one mutation class. Every seed in this class is the
minimal lift of its cell seed, and every sequence of cell mutations lifts uniquely to a sequence of Cox-ring mutations fixing the boundary sections.
Proof. Apply Proposition 3.7 to the Cox-ring equality and then use Lemma 3.8 .
Lemma 3.10. Let p : X → Y p:X \to Y p : X → Y be a proper birational morphism between normal projective varieties. Every action of a connected algebraic group on X X X descends uniquely to Y Y Y . In particular, the canonical B + B^{+} B + -action on a Bott–Samelson variety descends along every such morphism.
Proof. Normality of Y Y Y gives p ∗ O X = O Y p_{*}\mathcal{O}_{X}=\mathcal{O}_{Y} p ∗ O X = O Y . Blanchard’s lemma therefore gives the first assertion [Bri2, Theorem 7.2.1]. More generally, it gives a homomorphism Aut 0 ( X ) → Aut 0 ( Y ) \operatorname{Aut}^{0}(X) \to\operatorname{Aut}^{0}(Y) Aut 0 ( X ) → Aut 0 ( Y ) making p p p equivariant [Bri2, Corollary 7.2.2]. The canonical Borel action has image in Aut 0 ( X ) \operatorname{Aut}^{0}(X) Aut 0 ( X ) , so this also proves the assertion for a Kac–Moody Borel.
Simple-root operators
A seed is adapted to a derivation D D D when D = c x v ∂ b D=cx^{v}\partial_{b} D = c x v ∂ b for a frozen coordinate b b b , a scalar c ∈ C × c \in\mathbb{C}^{\times} c ∈ C × , and a vector v ∈ N I v \in\mathbb{N}^{I} v ∈ N I with v b = 0 v_{b}=0 v b = 0 . Thus D D D kills every seed coordinate except the frozen variable b b b , which we call the active coordinate . Choose positive rational numbers t i t_{i} t i satisfying a i j / t i = a j i / t j a_{ij}/t_{i}=a_{ji}/t_{j} a ij / t i = a j i / t j , and assign δ k = t i k \delta_{k}=t_{i_{k}} δ k = t i k to each mutable position of a word seed. The symmetrizer weights δ k \delta_{k} δ k remain attached to the corresponding mutable indices under mutation. Equivalently, t i = 2 / ( α i , α i ) t_{i}=2/(\alpha_{i},\alpha_{i}) t i = 2/ ( α i , α i ) for a positive normalization of the invariant form on each connected component.
We use inverse pullback for the action on functions and normalize the infinitesimal operators by [ E i , F i ] = H i [E_{i},F_{i}]=H_{i} [ E i , F i ] = H i and [ H i , E i ] = 2 E i [H_{i},E_{i}]=2E_{i} [ H i , E i ] = 2 E i . Scalar factors from the pinning are written explicitly in the monomial formulas.
Let Y = X w P J Y=X^{P_{J}}_{w} Y = X w P J , R = Cox ( Y ) R=\operatorname{Cox}(Y) R = Cox ( Y ) , and let b i b_{i} b i be the minimal lift of the terminal minor indexed by i i i . The positive simple-root derivations E i E_{i} E i act on R R R . The negative derivation F i F_{i} F i also acts whenever the rank-one group G i G_{i} G i preserves Y Y Y , that is, whenever i ∈ I Y = { i : s i w W J ≤ w W J } i \in I_{Y}=\{i:s_{i}wW_{J}\leq wW_{J}\} i ∈ I Y = { i : s i w W J ≤ w W J } . These actions preserve the divisor-class grading. We use the common mutation class of Corollary 3.9 .
Theorem 4.1. Every nonzero E i E_{i} E i , and every nonzero F i F_{i} F i with i ∈ I Y i \in I_{Y} i ∈ I Y , has an adapted seed in this mutation class. In such a seed, with active frozen variable b ( D ) b(D) b ( D ) ,
D = c D ∏ x ≠ b ( D ) x m x ( D ) ∂ b ( D ) , c D ∈ C × , m x ( D ) ∈ Z ≥ 0 . D=c_{D}\prod_{x\ne b(D)}x^{m_{x}(D)}\partial_{b(D)},\qquad c_{D}\in\mathbb{C}^{\times},\qquad m_{x}(D)\in\mathbb{Z}_{\geq0}. D = c D x = b ( D ) ∏ x m x ( D ) ∂ b ( D ) , c D ∈ C × , m x ( D ) ∈ Z ≥ 0 . For D = E i D=E_{i} D = E i or D = F i D=F_{i} D = F i , its mutable exponents are
m x k ( D ) = δ k t i ∣ B k , b ( D ) D ∣ . m_{x_{k}}(D)=\frac{\delta_{k}}{t_{i}}\left|B^{D}_{k,b(D)}\right|. m x k ( D ) = t i δ k B k , b ( D ) D . Here the nonzero entries in the active column are positive for E i E_{i} E i and negative for F i F_{i} F i . The following constructions give the adapted seeds.
For E i E_{i} E i , start from any reduced-word seed and mutate the nonterminal occurrences of i i i in increasing order. The active variable is b i b_{i} b i .
For F i F_{i} F i with s i w < w s_{i}w<w s i w < w , choose a reduced expression beginning with i i i . In its initial lift F i = c i x 1 ∂ τ a F_{i}=c_{i}x_{1}\partial_{\tau_{a}} F i = c i x 1 ∂ τ a , where τ a \tau_{a} τ a is the canonical section of X s i w P J X^{P_{J}}_{s_{i}w} X s i w P J .
For F i F_{i} F i with s i w > w s_{i}w>w s i w > w and i ∈ I Y i\in I_{Y} i ∈ I Y , one has w − 1 α i = α j w^{-1}\alpha_{i}=\alpha_{j} w − 1 α i = α j for a unique j ∈ J j\in J j ∈ J . Pass to the reverse interval order of the symmetric CGL presentation, then mutate the nonterminal occurrences of j j j in that order. The active variable is b j b_{j} b j .
The operator E i E_{i} E i is nonzero exactly when i ∈ supp ( w ) i\in\operatorname{supp}(w) i ∈ supp ( w ) . For i ∈ I Y i\in I_{Y} i ∈ I Y , the same criterion holds for F i F_{i} F i .
A mutation criterion
Let R R R be a complex domain with A ( Σ ) ⊆ R ⊆ U p o l ( Σ ) \mathcal{A}(\Sigma) \subseteq R \subseteq\mathcal{U}_{\mathrm{pol}}(\Sigma) A ( Σ ) ⊆ R ⊆ U pol ( Σ ) . Assume that the extended exchange matrix B B B has full row rank. For mutable positions, choose a positive skew-symmetrizer ( δ j ) (\delta_j) ( δ j ) , so that δ j B j k = − δ k B k j \delta_j B_{jk}=-\delta_k B_{kj} δ j B j k = − δ k B k j .
Lemma 4.2. A Laurent monomial in any reachable seed belongs to R R R only if all its exponents are nonnegative. If R R R is factorial, every coordinate of every reachable seed is prime in R R R .
Proof. Frozen exponents are nonnegative by membership in the mixed Laurent ring. A negative mutable exponent introduces a power of the exchange binomial in the denominator after mutation. Its two terms have no common coordinate factor, and the exchange row is nonzero by full rank. Thus this binomial cannot divide the monomial numerator, which contradicts regularity in the adjacent chart.
If a mutable coordinate factors in R R R , both factors are units in its mixed Laurent ring, hence scalar Laurent monomials in the mutable coordinates. The first assertion makes them ordinary monomials, so one is scalar. If a frozen coordinate factors, its primality in the mixed Laurent ring makes one factor a unit c x a cx^{a} c x a . Both c x a cx^{a} c x a and the complementary monomial are regular, so the first assertion gives a = 0 a=0 a = 0 . Every coordinate is therefore irreducible in R R R , and factoriality makes it prime. □
A simple oriented exchange chain of sign ε ∈ { 1 , − 1 } \varepsilon\in\{1,-1\} ε ∈ { 1 , − 1 } is a sequence of distinct seed positions a 1 , … , a r a_1,\ldots,a_r a 1 , … , a r whose exchange entries satisfy
B a p , a q = { ε , q = p − 1 , − ε , q = p + 1 , 0 , ∣ p − q ∣ ≠ 1 , p < r . B_{a_p,a_q} =
\begin{cases}
\varepsilon, & q=p-1,\\
-\varepsilon, & q=p+1,\\
0, & |p-q|\ne1,
\end{cases}
\qquad p<r. B a p , a q = ⎩ ⎨ ⎧ ε , − ε , 0 , q = p − 1 , q = p + 1 , ∣ p − q ∣ = 1 , p < r .
Proposition 4.3. Let D D D be a locally nilpotent derivation of R R R , and let a 1 , … , a r = b a_1,\ldots,a_r=b a 1 , … , a r = b be a simple oriented exchange chain with a 1 , … , a r − 1 a_1,\ldots,a_{r-1} a 1 , … , a r − 1 mutable and b b b frozen. Assume D 2 x a p = 0 ≠ D x a p D^2x_{a_p}=0\ne Dx_{a_p} D 2 x a p = 0 = D x a p for all p p p , that D D D kills every other initial coordinate, and that D x a 1 Dx_{a_1} D x a 1 is a Laurent monomial in invariant coordinates. Mutation at a 1 , … , a r − 1 a_1,\ldots,a_{r-1} a 1 , … , a r − 1 in that order gives an adapted seed.
Let h h h be the indicator vector of the chain, and let m ( 0 ) m^{(0)} m ( 0 ) be the exponent vector of D x a 1 Dx_{a_1} D x a 1 . If δ a p = t > 0 \delta_{a_p}=t>0 δ a p = t > 0 for every mutable chain position and
m j ( 0 ) = ε δ j t ( B h ) j + 1 j = a 1 ( j mutable ) , (13) m_j^{(0)}=\varepsilon\frac{\delta_j}{t}(Bh)_j+\mathbf{1}_{j=a_1}
\qquad(j\ \text{mutable}), \tag*{(13)} m j ( 0 ) = ε t δ j ( B h ) j + 1 j = a 1 ( j mutable ) , (13) then the final mutable exponents satisfy
m j = ε δ j t B j , b f i n ( j mutable ) . (14) m_j=\varepsilon\frac{\delta_j}{t}B^{\mathrm{fin}}_{j,b}
\qquad(j\ \text{mutable}). \tag*{(14)} m j = ε t δ j B j , b fin ( j mutable ) . (14) In particular, ε B j , b f i n ≥ 0 \varepsilon B^{\mathrm{fin}}_{j,b}\ge0 ε B j , b fin ≥ 0 .
Proof. The degree of a nonzero element under a locally nilpotent derivation is additive on products. This follows from the binomial formula for the highest nonzero derivative, using characteristic zero.
At a stage whose first remaining chain position is a a a , let s s s be the next one. Among the remaining chain positions, a a a is joined only to s s s , with entry − ε -\varepsilon − ε . Mutation at a a a leaves the chain on the remaining positions unchanged. Its exchange therefore reads
x a x a ′ = M + x s C . x_a x'_a=M+x_sC. x a x a ′ = M + x s C .
where M M M and C C C are monomials in invariant coordinates. The numerator has derivation degree one. Since the old coordinate has degree one and the new coordinate belongs to R R R , additivity gives D x a ′ = 0 Dx'_a=0 D x a ′ = 0 . Differentiation of this exchange yields
D x s = x a ′ C D x a . (15) Dx_s=\frac{x'_a}{C}Dx_a. \tag*{(15)} D x s = C x a ′ D x a . (15)
Iterating this recurrence leaves b b b as the only moving coordinate and expresses D x b Dx_b D x b as a Laurent monomial in the others. Since D x b ∈ R Dx_b \in R D x b ∈ R , Lemma 4.2 makes its exponents nonnegative.
It remains to track the mutable exponents. At each stage, let h h h indicate the remaining chain positions, and let m m m be the exponent vector of D x a Dx_a D x a . We maintain
m j = ε δ j / t ( B h ) j + 1 j = a ( j mutable ) . (16) m_j=\varepsilon^{\delta_j/t}(Bh)_j+1_{j=a}\qquad(j\text{ mutable}). \tag*{(16)} m j = ε δ j / t ( B h ) j + 1 j = a ( j mutable ) . (16)
Its initial instance is (13) . After mutation at a a a , we have h ′ = h − e a h'=h-e_a h ′ = h − e a . Since B a , s = − ε B_{a,s}=-\varepsilon B a , s = − ε is the only nonzero entry at a a a remaining chain position, matrix mutation gives
( B ′ h ′ ) j = ( B h ) j − ε [ ε B j , a ] + ( j ≠ a ) , ( B ′ h ′ ) a = ε . (B'h')_j=(Bh)_j-\varepsilon[\varepsilon B_{j,a}]_+\quad(j\ne a),\qquad(B'h')_a=\varepsilon. ( B ′ h ′ ) j = ( B h ) j − ε [ ε B j , a ] + ( j = a ) , ( B ′ h ′ ) a = ε .
The recurrence (15) gives
m j ′ = m j − [ − ε B a , j ] + + 1 j = s ( j ≠ a ) , m a ′ = 1. m'_j=m_j-[-\varepsilon B_{a,j}]_++1_{j=s}\quad(j\ne a),\qquad m'_a=1. m j ′ = m j − [ − ε B a , j ] + + 1 j = s ( j = a ) , m a ′ = 1.
The identity δ j B j , a = − t B a , j \delta_jB_{j,a}=-tB_{a,j} δ j B j , a = − t B a , j shows that these formulas preserve (16) . At the last stage, h = e b h=e_b h = e b , and the indicator term vanishes at all mutable positions because b b b is frozen. This proves (14) .
Equivariant compactifications
An adapted mutation sequence on an affine open set lifts to an equivariant compactification when the boundary valuations are invariant and the minimally lifted seed describes its Cox ring.
Let Y Y Y be a normal projective G a \mathbb{G}_a G a -variety, and let C ⊂ Y C\subset Y C ⊂ Y be an invariant factorial affine open set with C [ C ] × = C × \mathbb{C}[C]^\times=\mathbb{C}^\times C [ C ] × = C × . Write D 1 , … , D d D_1,\ldots,D_d D 1 , … , D d for its boundary prime divisors, τ a \tau_a τ a for their Cox sections, and ν a ( f ) = − ord D a ( f ) \nu_a(f)=-\operatorname{ord}_{D_a}(f) ν a ( f ) = − ord D a ( f ) . Assume that the minimally lifted seed Σ ~ \widetilde{\Sigma} Σ of a seed Σ \Sigma Σ on C C C satisfies Cox ( Y ) = U p o l ( Σ ~ ) \operatorname{Cox}(Y)=\mathcal{U}_{\mathrm{pol}}(\widetilde{\Sigma}) Cox ( Y ) = U pol ( Σ ) , with full-rank extended exchange matrix and polynomial frozen variables. Let D D D be the infinitesimal action on C [ C ] \mathbb{C}[C] C [ C ] .
Proposition 4.4. The action on Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) fixes the τ a \tau_a τ a and satisfies, for D f ≠ 0 Df\ne0 D f = 0 ,
D ~ ( f ~ ) = τ ν ( f ) − ν ( D f ) D f , ν ( f ) − ν ( D f ) ∈ Z ≥ 0 d . (17) \widetilde{D}(\widetilde{f})=\tau^{\nu(f)-\nu(Df)}Df,\qquad\nu(f)-\nu(Df)\in\mathbb{Z}_{\geq0}^{d}. \tag*{(17)} D ( f ) = τ ν ( f ) − ν ( D f ) D f , ν ( f ) − ν ( D f ) ∈ Z ≥ 0 d . (17) It preserves the nilpotence degree of every minimal lift.
If a mutation sequence on C C C reaches a seed Σ ′ \Sigma' Σ ′ with D = c ∏ j ≠ b ( f j ′ ) m j ∂ f b D=c\prod_{j\ne b}(f'_j)^{m_j}\partial_{f_b} D = c ∏ j = b ( f j ′ ) m j ∂ f b , where b b b is frozen and m j ∈ Z m_j\in\mathbb{Z} m j ∈ Z , the lifted sequence gives an adapted Cox seed. In that seed,
D ~ = c τ ν ( f b ) − ∑ j ≠ b m j ν ( f j ′ ) ∏ j ≠ b ( x j ′ ) m j ∂ x b , x j ′ = f j ′ ~ , (18) \widetilde{D}=c\tau^{\nu(f_b)-\sum_{j\ne b}m_j\nu(f'_j)}\prod_{j\ne b}(x'_j)^{m_j}\partial_{x_b},
\qquad x'_j=\widetilde{f'_j}, \tag*{(18)} D = c τ ν ( f b ) − ∑ j = b m j ν ( f j ′ ) j = b ∏ ( x j ′ ) m j ∂ x b , x j ′ = f j ′ , (18) and all displayed exponents are nonnegative. The construction commutes with equivariant projective birational contractions that preserve C C C and satisfy the same Cox-ring hypotheses.
Proof. Because the action is connected, each boundary prime divisor is invariant. At its generic point the infinitesimal action preserves the local discrete valuation ring and its maximal ideal. It therefore preserves every integral power of that ideal, including negative powers in the fraction field. Hence
ord D a ( D f ) ≥ ord D a ( f ) . \operatorname{ord}_{D_a}(Df)\geq\operatorname{ord}_{D_a}(f). ord D a ( D f ) ≥ ord D a ( f ) .
The valuation realization of the Cox ring now shows that the rule D ~ ( f τ μ ) = ( D f ) τ μ \widetilde{D}(f\tau^\mu)=(Df)\tau^\mu D ( f τ μ ) = ( D f ) τ μ is a regular homogeneous derivation. It is locally nilpotent because D D D is locally nilpotent on C [ C ] \mathbb{C}[C] C [ C ] , and its exponential is the lifted action. This proves (17) . Iterating the rule gives equality of the two degrees.
By Proposition 3.7 , every coordinate reached by the lifted sequence is the minimal lift of the corresponding cell coordinate. The boundary sections and every x j ′ x'_j x j ′ with j ≠ b j\ne b j = b are invariant. Applying (17) to f b f_b f b , and using additivity of every boundary valuation on products, gives (18) and its nonnegative boundary exponents. Its coefficient is a Laurent monomial regular in Cox ( Y ) = U p o l ( Σ ~ ) \operatorname{Cox}(Y)=\mathcal{U}_{\mathrm{pol}}(\widetilde{\Sigma}) Cox ( Y ) = U pol ( Σ ) ; Lemma 4.2 therefore gives m j ≥ 0 m_j\geq0 m j ≥ 0 for every j j j . Under a contraction, Proposition 2.1 sets the exceptional canonical sections equal to one. They are invariant, so the quotient commutes with the derivation and with the displayed mutation sequence. □
In particular, the chain construction of Proposition 4.3 lifts from C C C to Y Y Y : minimal lifting preserves invariance, derivation degree one, the first monomial derivative, and the exchange entries along the chain. The mutable exponents are unchanged. If the adapted cell column is sign-coherent and contains every mutable coordinate appearing in the coefficient, the lifted active column retains both properties.
The chain conditions have a geometric interpretation. The degree of a regular function with respect to D D D is its degree on a general orbit, with the orbit parametrized by G a \mathbb{G}_a G a . Thus degree one means that its zero divisor meets a general orbit once, with multiplicity one. A prime principal divisor stable under G a \mathbb{G}_a G a has invariant defining equation, since the only units on C C C are constants and G a \mathbb{G}_a G a has no nontrivial characters. Finally, if the invariant coordinate divisors are prime and the zero divisor of the first derivative is supported on them, factoriality and the unit hypothesis make that derivative a monomial in their equations.
Remark 4.5. An additive action need not have an adapted seed in a given mutation class. On Cox ( P 2 ) = C [ x , y , z ] \operatorname{Cox}(\mathbb{P}^2)=\mathbb{C}[x,y,z] Cox ( P 2 ) = C [ x , y , z ] , take the rank-zero seed and the homogeneous locally nilpotent derivation D = x ∂ y + y ∂ z D=x\partial_y+y\partial_z D = x ∂ y + y ∂ z . Since D 2 z = x ≠ 0 D^2z=x\ne0 D 2 z = x = 0 , this seed is not adapted, and no mutation is available. The hypothesis of an adapted cell seed in Proposition 4.4 is therefore substantive.
Positive simple roots
Let Y Y Y be a Bott–Samelson variety for an arbitrary word i = ( i 1 , … , i N ) \mathbf{i}=(i_1,\ldots,i_N) i = ( i 1 , … , i N ) , or a normal projective birational contraction that is an isomorphism on C i C_i C i . Equip R = Cox ( Y ) R=\operatorname{Cox}(Y) R = Cox ( Y ) with its minimally lifted word seed. Write E i E_i E i for the derivation induced by the positive simple root subgroup, and b i b_i b i for the lifted terminal minor indexed by i i i . We set τ j = 1 \tau_j=1 τ j = 1 whenever the corresponding standard boundary divisor is contracted.
Proposition 4.6. Each E i E_i E i is locally nilpotent and preserves the Cox grading. It is nonzero precisely when i i i occurs in i \mathbf{i} i . For such an i i i , mutate its nonterminal occurrences from left to right. The resulting seed is adapted to E i E_i E i , with active frozen variable b i b_i b i , and
E i = c i ∏ z ≠ b i z m z ( i ) ∂ b i , c i ∈ C × , m z ( i ) ∈ Z ≥ 0 . E_i=c_i\prod_{z\ne b_i}z^{m_z(i)}\partial_{b_i},\qquad c_i\in\mathbb{C}^{\times},\qquad m_z(i)\in\mathbb{Z}_{\ge0}. E i = c i z = b i ∏ z m z ( i ) ∂ b i , c i ∈ C × , m z ( i ) ∈ Z ≥ 0 .
The mutable exponents are
m x k ( i ) = δ k t i p k , b i . m_{x_k}(i)=\frac{\delta_k}{t_i}p_{k,b_i}. m x k ( i ) = t i δ k p k , b i .
The operators and adapted seeds commute with every birational contraction preserving the affine chart.
Proof. Every positive root subgroup preserves the boundary divisors. It therefore acts on the valuation realization of the Cox ring through the coefficient functions and fixes all canonical boundary sections. The resulting derivations are regular, locally nilpotent and homogeneous for the Cox grading.
In the fundamental highest-weight module V ( Λ j ) V(\Lambda_j) V ( Λ j ) , the highest coordinate functional has e i e_i e i -string length one when i = j i=j i = j , and zero otherwise. Thus the initial minors indexed by i i i have E i E_i E i -degree one; the others are invariant. The derivative is nonzero whenever i i i occurs: its section is nonzero at s i B s_iB s i B , which belongs to the prefix Schubert image. Invariant boundary sections make the same assertions true for the minimal lifts.
Let a a a be the first occurrence of i i i . The preceding factors lie in the derived Levi subgroup generated by indices other than i i i . The vector of weight Λ i − α i \Lambda_i-\alpha_i Λ i − α i in the fundamental simple-root string is highest for this Levi, with pairings − a j i -a_{ji} − a j i , for j ≠ i j\ne i j = i . Normalized highest matrix coefficients multiply under addition of weights, as their restrictions to the Gauss cell are torus characters. Extra Cartan characters are trivial on the derived Levi. It follows, up to a nonzero pinning scalar, that E i f a = ∏ j < a < s ( j ) f j − a i j E_i f_a=\prod_{j<a<s(j)} f_j^{-a_{ij}} E i f a = ∏ j < a < s ( j ) f j − a ij . The prefix line-bundle recursion then gives
E i x a = c i τ a ∏ j < a < s ( j ) x j − a i j (19) E_i x_a=c_i\tau_a\prod_{j<a<s(j)}x_j^{-a_{ij}} \tag*{(19)} E i x a = c i τ a j < a < s ( j ) ∏ x j − a ij (19)
This proves nonvanishing when i i i occurs. If i i i is absent, every initial coordinate is invariant, so E i = 0 E_i=0 E i = 0 .
The occurrences of i i i form a simple oriented chain ending at the frozen terminal minor. Formula (19) supplies the monomial hypothesis of Proposition 4.3 . Its column hypothesis follows from the word matrix. If h h h is the indicator of all i i i -positions, then at a mutable position with i j = i i_j=i i j = i , ( B h ) j (Bh)_j ( B h ) j equals − 1 -1 − 1 at the first occurrence and zero elsewhere. At a position of another letter, it equals − a i , i j -a_{i,i_j} − a i , i j when j < a < s ( j ) j<a<s(j) j < a < s ( j ) , and zero otherwise. The identity t i j a i , i j = t i a a i j , i t_{ij}a_{i,i_j}=t_{i a}a_{i_j,i} t ij a i , i j = t ia a i j , i converts these values into the exponents in (19) . The proposition gives the adapted seed and its exponent formula.
A birational contraction sets the exceptional boundary sections equal to one. This commutes with each E i E_i E i , which kills those sections, and with every exchange in the same mutation sequence. The final assertion follows. □
Negative roots on Bott–Samelson varieties
Left multiplication on the first factor of a Bott–Samelson variety gives a negative root operator in the initial Cox seed.
Proposition 4.7. Let i = ( i 1 , … , i N ) \mathbf{i}=(i_1,\ldots,i_N) i = ( i 1 , … , i N ) be any nonempty word, and let p : Z i → Y p:Z_{\mathbf{i}}\to Y p : Z i → Y be a proper birational morphism to a normal projective variety that is an isomorphism on C i C_{\mathbf{i}} C i . The first standard boundary divisor E 1 E_1 E 1 is not contracted. The rank-one group G i 1 G_{i_1} G i 1 acting on the first factor descends to Y Y Y and its Cox ring. Its negative simple-root derivation satisfies
F i 1 = c x 1 ∂ τ 1 , c ∈ C × , F_{i_1}=c x_1\partial_{\tau_1},\qquad c\in\mathbb{C}^{\times}, F i 1 = c x 1 ∂ τ 1 , c ∈ C × , in the initial lifted word seed. Here x 1 x_1 x 1 is the first lifted minor, and τ 1 \tau_1 τ 1 is the canonical section of the image of E 1 E_1 E 1 . Its active column is zero except for the entry − 1 -1 − 1 at position one, when that position is mutable. Thus its mutable coefficient exponents are
m x k = − δ k t i 1 B k , τ 1 . m_{x_k}=-\frac{\delta_k}{t_{i_1}}B_{k,\tau_1}. m x k = − t i 1 δ k B k , τ 1 . These formulas commute with further projective birational contractions preserving the affine chart.
Proof. The degree of the first minor is [ E 1 ] [E_1] [ E 1 ] , so the nonconstant rational function f 1 = x 1 / τ 1 f_1=x_1/\tau_1 f 1 = x 1 / τ 1 has its only pole along E 1 E_1 E 1 . If E 1 E_1 E 1 were contracted, this function would have nonnegative order along every prime divisor of Y Y Y . Normality would make it regular on Y Y Y , contradicting the fact that every regular function on an irreducible projective variety is constant. Hence E 1 E_1 E 1 survives.
On Z i Z_{\mathbf{i}} Z i , the group G i 1 G_{i_1} G i 1 acts by left multiplication in the first factor. The line bundle O ( E 1 ) = m 1 ∗ O P 1 ( 1 ) \mathcal{O}(E_1)=m_1^*\mathcal{O}_{\mathbb{P}^1}(1) O ( E 1 ) = m 1 ∗ O P 1 ( 1 ) has its natural rank-one linearization. Every standard boundary divisor E j E_j E j , for j > 1 j>1 j > 1 , is stable under this action, so its line bundle has the canonical divisor linearization. These line bundles form a basis of the Picard group and define an action on the Cox ring. The canonical section of E j E_j E j , for j > 1 j>1 j > 1 , is fixed by G i 1 G_{i_1} G i 1 , which has no nontrivial characters. The negative root subgroup also fixes every prefix highest section x k x_k x k : a highest coordinate functional is annihilated by negative simple-root multiplication on the left. The first prefix is P 1 \mathbb{P}^1 P 1 , and its two fundamental sections give F i 1 τ 1 = c x 1 F_{i_1}\tau_1=cx_1 F i 1 τ 1 = c x 1 , with c ≠ 0 c\ne0 c = 0 . The resulting linearization of each prefix line bundle agrees with its natural rank-one linearization, since a connected semisimple group has no nontrivial characters. Since the initial coordinates generate the fraction field, this proves the formula on Cox ( Z i ) \operatorname{Cox}(Z_{\mathbf{i}}) Cox ( Z i ) .
Every exceptional prime divisor of p p p is a standard boundary divisor, and the first one is not exceptional. Thus all its exceptional canonical sections are G i 1 G_{i_1} G i 1 -fixed. The Cox quotient ideal generated by these sections minus one is stable under G i 1 G_{i_1} G i 1 . The action consequently descends to Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) , preserving its divisor-class grading. Taking the section ring of an ample Cartier divisor recovers an action on Y Y Y . It agrees with the induced rational action on the common function field, so p p p is equivariant. The operator formula descends at the same time.
Finally, the boundary incidence identity gives entry − 1 -1 − 1 in the τ 1 \tau_1 τ 1 -column only at the first mutable position; there is no earlier position whose successor is one. Since δ 1 = t i 1 \delta_1=t_{i_1} δ 1 = t i 1 , the exponent formula follows. The same quotient argument proves compatibility with further contractions. □
The following criterion extends the corresponding finite-type result for reduced words [14 Proposition 6.3] to arbitrary words in a symmetrizable Kac–Moody group.
Proposition 4.8. Let i i i occur in i \mathbf{i} i , and let a a a be its first occurrence. The canonical action of the rank-one Borel B i = G i ∩ B + B_i=G_i\cap B^+ B i = G i ∩ B + on Z i Z_{\mathbf{i}} Z i extends to G i G_i G i if and only if a i , i k = 0 a_{i,i_k}=0 a i , i k = 0 for every k < a k<a k < a . Equivalently, commuting adjacent orthogonal letters can move this occurrence of i i i to the beginning of the word.
When these conditions hold, the action descends to every contraction in Theorem 3.4, the divisor E a E_a E a survives, and in the initial lifted seed its negative root derivation is F i = c i x a ∂ τ a F_i=c_i x_a\partial_{\tau_a} F i = c i x a ∂ τ a , with c i ∈ C × c_i\in\mathbb{C}^{\times} c i ∈ C × .
Proof. Suppose first that the action extends. Each prefix projection π k : Z i → Z ( i 1 , … , i k ) \pi_k:Z_{\mathbf{i}}\to Z_{(i_1,\ldots,i_k)} π k : Z i → Z ( i 1 , … , i k ) is an iterated P 1 \mathbb{P}^1 P 1 -bundle, so ( π k ) ∗ O Z i = O Z ( i 1 , … , i k ) (\pi_k)_*\mathcal{O}_{Z_{\mathbf{i}}}=\mathcal{O}_{Z_{(i_1,\ldots,i_k)}} ( π k ) ∗ O Z i = O Z ( i 1 , … , i k ) . Blanchard’s lemma makes π k \pi_k π k equivariant for a unique G i G_i G i -action on the prefix [Bri2, Theorem 7.2.1].
For k < a k<a k < a , the positive simple-root subgroup indexed by i i i acts trivially on the prefix, since the word has no occurrence of i i i . Hence the relation [ E i , F i ] = H i [E_i,F_i]=H_i [ E i , F i ] = H i forces its coroot torus to act trivially there as well. The standard curve obtained by varying only the factor at position k k k is P i k / B + P_{i_k}/B^+ P i k / B + ; the coroot torus acts on its affine coordinate with weight a i , i k a_{i,i_k} a i , i k . Thus a i , i k = 0 a_{i,i_k}=0 a i , i k = 0 for every preceding position.
Conversely, under these orthogonality conditions the commuting rank-two factor identifications move the occurrence at position a a a to the first factor. They are B + B^+ B + -equivariant and identify the corresponding lifted minors and boundary sections. Left multiplication on that factor gives the G i G_i G i -action. Proposition 4.7 then gives its descent, the survival of E a E_a E a , and the formula for F i F_i F i . □
Corollary 4.9. Let w ∈ W J w\in W^J w ∈ W J , and suppose that i i i is a left descent of w w w . Choose a reduced word beginning with i i i . The initial lifted seed of Cox ( X w P J ) \operatorname{Cox}(X^{P_J}_w) Cox ( X w P J ) is adapted to F i F_i F i , with active frozen variable the canonical section of X s i w P J X^{P_J}_{s_iw} X s i w P J , and coefficient a nonzero scalar multiple of the first lifted minor.
Proof. Apply Proposition 4.7 to the Bott–Samelson map. The image of E 1 E_1 E 1 is X s i w P J X^{P_J}_{s_iw} X s i w P J . It is a divisor: if s i w s_iw s i w had a right descent in J J J , then w w w would also have that right descent, contrary to w ∈ W J w\in W^J w ∈ W J . Thus s i w ∈ W J s_iw\in W^J s i w ∈ W J and its length is ℓ ( w ) − 1 \ell(w)-1 ℓ ( w ) − 1 . □
Stabilizing negative simple roots
Return to Y = X w P J Y=X^{P_J}_w Y = X w P J . For i ∈ I Y i\in I_Y i ∈ I Y , the action of G i G_i G i lifts uniquely to the graded Cox ring because the simply connected rank-one group has trivial character group and Picard group [Vez, Section 2.2 and Remark 2.3.4]. Corollary 4.9 treats the left descents. For the remaining i i i , a change of coordinates on the open cell turns F i F_i F i into a positive root operator on U ( w − 1 ) U(w^{-1}) U ( w − 1 ) .
Lemma 4.10. Let i = ( i 1 , … , i N ) \mathbf{i}=(i_1,\ldots,i_N) i = ( i 1 , … , i N ) be a reduced expression of w w w , and let i o p = ( i N , … , i 1 ) \mathbf{i}^{\mathrm{op}}=(i_N,\ldots,i_1) i op = ( i N , … , i 1 ) . The reverse interval-order seed of C [ U ( w ) ] \mathbb{C}[U(w)] C [ U ( w )] is reachable from the word seed. In reverse order, its coordinates are, up to nonzero pinning scalars,
g r = Δ w < k Λ i k , w Λ i k ∣ U ( w ) , k = N + 1 − r . g_r=\Delta_{w_{<k}\Lambda_{i_k},\,w\Lambda_{i_k}}|_{U(w)},\qquad k=N+1-r. g r = Δ w < k Λ i k , w Λ i k ∣ U ( w ) , k = N + 1 − r . Its full cell exchange matrix and mutable symmetrizer are
B r e v = − B i o p , δ r = t i N + 1 − r . (20) B^{\mathrm{rev}}=-B_{i^{\mathrm{op}}},\qquad\delta_r=t_{i_{N+1-r}}. \tag*{(20)} B rev = − B i op , δ r = t i N + 1 − r . (20)
Chevalley transposition gives an isomorphism of varieties
ι w : U ( w ) ⟶ U ( w − 1 ) , n ⟼ w ‾ − 1 n T w ‾ , \iota_w: U(w) \longrightarrow U(w^{-1}), \qquad n \longmapsto\overline{w}^{-1}n^{T}\overline{w}, ι w : U ( w ) ⟶ U ( w − 1 ) , n ⟼ w − 1 n T w ,
which identifies these coordinates, up to the same scalars, with the inverse-word minors. Each terminal minor is the original terminal minor indexed by its letter.
Proof. The interval-order mutations of [28 Theorems 3.11(c), 4.8(b), and 7.3] reach the reverse order. If a ( k ) a(k) a ( k ) is the last occurrence of i k i_k i k , the corresponding prime is the interval prime for [ k , a ( k ) ] [k,a(k)] [ k , a ( k )] . By [28 Equation (3.33) and Theorem 7.1(c)] , this prime specializes at q = 1 q=1 q = 1 , up to a nonzero scalar, to Δ w < k Λ i k , w ≤ a ( k ) Λ i k \Delta_{w_{<k}\Lambda_{i_k},w_{\leq a(k)}\Lambda_{i_k}} Δ w < k Λ i k , w ≤ a ( k ) Λ i k . Since the later reflections fix Λ i k \Lambda_{i_k} Λ i k , this is g r g_r g r , with k = N + 1 − r k=N+1-r k = N + 1 − r . It is terminal in the reverse order precisely when k k k is the first occurrence of i k i_k i k . Then w < k Λ i k = Λ i k w_{<k}\Lambda_{i_k}=\Lambda_{i_k} w < k Λ i k = Λ i k , so it is the original terminal minor indexed by that simple root.
We check the exchange matrix before specializing q q q . Write β k = w < k α i k \beta_k=w_{<k}\alpha_{i_k} β k = w < k α i k and k r = N + 1 − r k_r=N+1-r k r = N + 1 − r . The inverse-word root sequence satisfies β r o p = − w − 1 β k r \beta_r^{\mathrm{op}}=-w^{-1}\beta_{k_r} β r op = − w − 1 β k r . The elementary commutation parameters are ν k l = q ( β k , β l ) / 2 \nu_{kl}=q^{(\beta_k,\beta_l)/2} ν k l = q ( β k , β l ) /2 for k > l k>l k > l , with ν l k = ν k l − 1 \nu_{lk}=\nu_{kl}^{-1} ν l k = ν k l − 1 . Reversal changes the order of indices, while Weyl invariance preserves the pairings; hence ν k r , k s = ( ν r s o p ) − 1 \nu_{k_r,k_s}=(\nu_{rs}^{\mathrm{op}})^{-1} ν k r , k s = ( ν r s op ) − 1 . The prime-coordinate parameters are products of these elementary parameters over intervals joining equal letters [28 Equation (3.32)] . The reverse toric matrix is therefore the entrywise inverse of the inverse-word toric matrix. Their weight maps have the same kernel: at corresponding positions,
w < k Λ i k − w Λ i k = − w ( Λ i k − w − 1 w < k Λ i k ) . w_{<k}\Lambda_{i_k}-w\Lambda_{i_k}=-w\left(\Lambda_{i_k}-w^{-1}w_{<k}\Lambda_{i_k}\right). w < k Λ i k − w Λ i k = − w ( Λ i k − w − 1 w < k Λ i k ) .
In the column convention of [28 ] , an exchange vector is uniquely determined by zero character weight and its commutation pairings with every seed coordinate [28 Theorem 3.11(a)] . The normalization for the simple-root index i k i_k i k is λ ∗ = q i k − 2 \lambda^*=q_{i_k}^{-2} λ ∗ = q i k − 2 in both presentations [28 Equation (7.4)] . Negating an inverse-word exchange vector compensates for inversion of the toric matrix and satisfies exactly these conditions for the reverse presentation. Uniqueness proves (20) for the full extended matrix, including its frozen part. The same interval-order mutations transport the symmetrizer weights with the mutable indices.
Chevalley transposition reverses multiplication and sends the root subgroup of β \beta β to that of − β -\beta − β . Conjugation by w ‾ − 1 \overline{w}^{-1} w − 1 then gives an anti-isomorphism from U ( w ) U(w) U ( w ) to U ( w − 1 ) U(w^{-1}) U ( w − 1 ) . The inverse-word prefix at position r r r is w − 1 w < k w^{-1}w_{<k} w − 1 w < k . Applying the contravariant form on the highest-weight module shows that its highest minor pulls back to g r g_r g r ; the extremal-vector normalizations account for the nonzero scalars.
Proposition 4.11. Suppose that i ∈ I Y i\in I_Y i ∈ I Y is not a left descent of w w w . Then w − 1 α i = α j w^{-1}\alpha_i=\alpha_j w − 1 α i = α j for a unique j ∈ J j\in J j ∈ J . The derivation F i F_i F i is nonzero exactly when j ∈ supp ( w ) j\in\operatorname{supp}(w) j ∈ supp ( w ) . In that case, every minimally lifted word mutation class contains a seed adapted to F i F_i F i , with active frozen variable b j b_j b j . Its mutable exponents satisfy
m x k ( F i ) = − δ k t i B k , b j i . m_{x_k}(F_i)=-\frac{\delta_k}{t_i}B^i_{k,b_j}. m x k ( F i ) = − t i δ k B k , b j i .
Proof. Since w < w s i w w<w s_iw w < w s i w and i ∈ I Y i\in I_Y i ∈ I Y , the two cosets coincide. Thus w − 1 s i w ∈ W J w^{-1}s_iw\in W_J w − 1 s i w ∈ W J , and w − 1 α i w^{-1}\alpha_i w − 1 α i is a positive root of its root system. Write
w − 1 α i = ∑ j ∈ J c j α j , c j ∈ Z ≥ 0 . w^{-1}\alpha_i=\sum_{j\in J}c_j\alpha_j,\qquad c_j\in\mathbb{Z}_{\geq0}. w − 1 α i = j ∈ J ∑ c j α j , c j ∈ Z ≥ 0 .
All the w α j w\alpha_j w α j are positive by minimality of w ∈ W J w\in W^J w ∈ W J . Their sum with coefficients c j c_j c j is the simple root α i \alpha_i α i , so precisely one coefficient is nonzero and it equals one. Thus w − 1 α i = α j w^{-1}\alpha_i=\alpha_j w − 1 α i = α j , and Weyl invariance gives
t i = t j . (21) t_i=t_j. \tag*{(21)} t i = t j . (21)
To identify the action on the cell, take an inversion root β ∈ Φ + ∩ w Φ − \beta\in\Phi^+\cap w\Phi^- β ∈ Φ + ∩ w Φ − . Since β ≠ α i \beta\ne\alpha_i β = α i , some simple-root coefficient other than the i i i th is positive throughout its α i \alpha_i α i -string. Likewise w − 1 β ≠ − α j w^{-1}\beta\ne-\alpha_j w − 1 β = − α j , so its α j \alpha_j α j -string retains a negative coefficient away from j j j . Thus the entire α i \alpha_i α i -string of β \beta β lies in the inversion set. The inversion Lie algebra is stable under s l 2 i \mathfrak{sl}_{2i} sl 2 i , and G i G_i G i normalizes U ( w ) U(w) U ( w ) . Since w ‾ − 1 G i w ‾ = G j ⊂ P J \overline{w}^{-1}G_i\overline{w}=G_j\subset P_J w − 1 G i w = G j ⊂ P J , G i G_i G i fixes w P J wP_J w P J . It therefore acts by conjugation on the open cell U ( w ) w P J U(w)wP_J U ( w ) w P J . Connectedness fixes each boundary prime, and its canonical linearization fixes τ a \tau_a τ a . In the Cox-ring realization by boundary valuations, F i F_i F i thus acts on the cell functions and kills every τ a \tau_a τ a .
For a nonzero pinning constant c w c_w c w , transposition gives
ι w ( x − i ( u ) n x − i ( − u ) ) = x j ( − c w u ) ι w ( n ) x j ( c w u ) . \iota_w\bigl(x_{-i}(u) n x_{-i}(-u)\bigr)=x_j(-c_wu)\iota_w(n)x_j(c_wu). ι w ( x − i ( u ) n x − i ( − u ) ) = x j ( − c w u ) ι w ( n ) x j ( c w u ) .
Since w α j = α i > 0 w\alpha_j=\alpha_i>0 w α j = α i > 0 , the subgroup x j x_j x j belongs to the complementary positive subgroup for U ( w − 1 ) U(w^{-1}) U ( w − 1 ) . The same root-string argument shows that it normalizes U ( w − 1 ) U(w^{-1}) U ( w − 1 ) . Its projected left action on the inverse cell is therefore conjugation. With our inverse-pullback convention, the displayed identity identifies F i F_i F i with a nonzero scalar multiple of E j E_j E j on cell functions.
Now lift the mutation sequence of Lemma 4.10 to R R R . By Proposition 3.7 , the resulting Cox coordinates y r y_r y r are the minimal lifts of the reversed minors g r g_r g r . Their F i F_i F i -degrees are the E j E_j E j -degrees of the inverse-word minors, since F i F_i F i fixes the boundary sections. Thus the moving coordinates are precisely the positions bearing the letter j j j , and each has degree one. If j j j is absent, F i = 0 F_i=0 F i = 0 . Otherwise let a a a be its first occurrence in i o p \mathbf{i}^{\mathrm{op}} i op . The first-occurrence calculation of Proposition 4.6 on the inverse cell gives
F i y a = c τ μ ∏ r < a < s o p ( r ) y r − a r , j o p , c ∈ C × , μ ∈ Z { boundary primes } . F_i y_a=c\tau^\mu\prod_{r<a<s^{\mathrm{op}}(r)}y_r^{-a^{\mathrm{op}}_{r,j}},\qquad c\in\mathbb{C}^{\times},\quad\mu\in\mathbb{Z}^{\{\text{boundary primes}\}}. F i y a = c τ μ r < a < s op ( r ) ∏ y r − a r , j op , c ∈ C × , μ ∈ Z { boundary primes } .
Every factor is invariant. The boundary exponent μ \mu μ records the difference of the lift degrees on Y Y Y .
Let h h h indicate all positions bearing the letter j j j . By (20) , their chain has orientation ε = − 1 \varepsilon=-1 ε = − 1 . Negating the inverse-word column identity gives
m r ( 0 ) = − δ r ( 0 ) t j ( B ~ r e v h ) r + 1 r = a ( r mutable ) . m_r^{(0)}=-\frac{\delta_r^{(0)}}{t_j}(\widetilde{B}^{\mathrm{rev}}h)_r+\mathbf{1}_{r=a}\qquad(r\text{ mutable}). m r ( 0 ) = − t j δ r ( 0 ) ( B rev h ) r + 1 r = a ( r mutable ) .
Here the boundary entries of h h h vanish. Proposition 4.3 now applies: mutating the nonterminal occurrences of j j j gives
F i = c D ∏ z ≠ b j z m z ∂ b j , c D ≠ 0 , m z ∈ Z ≥ 0 , m x k = − δ k t j B k , b j f i n = − δ k t i B k , b j f i n . F_i=c_D\prod_{z\ne b_j}z^{m_z}\partial_{b_j},\qquad c_D\ne0,\qquad m_z\in\mathbb{Z}_{\ge0},\qquad m_{x_k}=-\frac{\delta_k}{t_j}B^{\mathrm{fin}}_{k,b_j}=-\frac{\delta_k}{t_i}B^{\mathrm{fin}}_{k,b_j}. F i = c D z = b j ∏ z m z ∂ b j , c D = 0 , m z ∈ Z ≥ 0 , m x k = − t j δ k B k , b j fin = − t i δ k B k , b j fin .
The final equality uses (21) . By Lemma 4.10 , the terminal minor in this chain lifts to the original frozen coordinate b j b_j b j . Its degree-one string gives F i ≠ 0 F_i\ne0 F i = 0 .
Proof of Theorem 4.1. For E i E_i E i apply Proposition 4.6. For F i F_i F i , Corollary 4.9 treats the left descents and Proposition 4.11 treats the remaining stabilizing simple-root indices. Corollary 3.9 allows us to choose a word beginning with any prescribed left descent. Lemma 4.10 realizes reversal by mutations as well. All three constructions therefore take place in the same Cox mutation class.
For the nonvanishing statement, only the third case requires an additional observation. Write K = supp ( w ) K=\operatorname{supp}(w) K = supp ( w ) . The action of w ∈ W K w\in W_K w ∈ W K is the identity modulo the span of { α k : k ∈ K } \{\alpha_k:k\in K\} { α k : k ∈ K } . From w α j = α i w\alpha_j=\alpha_i w α j = α i , we therefore obtain j ∈ K j\in K j ∈ K if and only if i ∈ K i\in K i ∈ K . The nonvanishing criterion in Proposition 4.11 is consequently equivalent to the one asserted here.
Proposition 4.12. Let R R R be a complex upper cluster algebra with polynomial frozen variables and a full-rank extended exchange matrix. Suppose that a reachable seed is adapted to a locally nilpotent derivation D D D , so that D = M ∂ b D=M\partial_b D = M ∂ b with b b b frozen. Assume that the nonzero entries in the b b b -column have one sign and that every mutable variable dividing M M M belongs to the support of this column. If z z z is a cluster monomial and d = max { j : D j z ≠ 0 } d=\max\{j:D^jz\ne0\} d = max { j : D j z = 0 } , then D d z / d ! D^dz/d! D d z / d ! is a nonzero scalar multiple of a cluster monomial in a reachable seed.
Proof. Replacing the exchange matrix by its negative leaves every exchange relation unchanged. We may therefore assume that B k , b ≤ 0 B_{k,b} \le0 B k , b ≤ 0 for every mutable index k k k , and write J = { k : B k , b < 0 } J=\{k:B_{k,b}<0\} J = { k : B k , b < 0 } . The separation formula gives z = x g F ( y ) z=x^{g}F(y) z = x g F ( y ) , where F F F is a polynomial with constant term one and y k = ∏ j x j B k , j y_k=\prod_j x_j^{B_{k,j}} y k = ∏ j x j B k , j ; see [41 Theorem 2.5] . Every y k y_k y k with k ∈ J k\in J k ∈ J lowers the exponent of b b b , and the remaining y k y_k y k have b b b -exponent zero. Since the exchange matrix has full rank, the constant monomial of F F F cannot cancel after substitution. Thus the largest exponent of b b b in z z z is g b = d g_b=d g b = d , and its terms of that degree form
z t o p = x g F ( y ) ∣ y k = 0 ( k ∈ J ) . z^{\mathrm{top}}=x^gF(y)\big|_{y_k=0\ (k\in J)}. z top = x g F ( y ) y k = 0 ( k ∈ J ) .
By Qin’s freezing theorem [41 Theorem 5.11] , z t o p z^{\mathrm{top}} z top is a localized cluster monomial in a seed of the pattern obtained by freezing J J J . Its mutation path avoids J J J and is also a path in the original pattern. Every variable dividing M M M , and the frozen variable b b b , is unchanged along this path. Hence
D d z d ! = M d b − d z t o p \frac{D^d z}{d!}=M^d b^{-d}z^{\mathrm{top}} d ! D d z = M d b − d z top
is a Laurent monomial in that reachable seed. The derivative belongs to R R R , so Lemma 4.2 makes this an ordinary monomial in the reachable seed, up to a nonzero scalar. □
Corollary 4.13. Let D D D be any nonzero operator in Theorem 4.1 , and let z z z be a cluster monomial. If d = max { q : D q z ≠ 0 } d=\max\{q:D^qz\ne0\} d = max { q : D q z = 0 } , then D d z / d ! D^dz/d! D d z / d ! is a nonzero scalar multiple of a cluster monomial in the same mutation class.
Proof. The exponent formula shows that the active column has one sign and that its support is exactly the set of mutable variables occurring in the coefficient of D D D . Apply Proposition 4.12. □
Corollary 4.14. Let L L L be a finite-type Levi subgroup generated by rank-one groups preserving Y Y Y , and let V ⊂ R V\subset R V ⊂ R be a simple finite-dimensional L L L -module. If one extremal vector of V V V is a scalar multiple of a cluster monomial, then every extremal vector has this property. If the original vector is a cluster variable, every extremal vector is a scalar multiple of a cluster variable.
Proof. Successive simple reflections connect the extremal weights of V V V . On an extremal vector, a nontrivial reflection is realized, up to a nonzero scalar, by the maximal power of the corresponding raising or lowering operator. Apply Corollary 4.13 along such a sequence. For the last assertion, cluster variables are prime in R R R by Lemma 4.2 . A Weyl-group representative in L L L acts by a ring automorphism, so each resulting extremal vector is prime as well. A nonconstant cluster monomial can be prime only when it is one cluster variable to the first power. This proves the assertion. □
Theta bases and root filtrations
A theta function extends across the frozen boundary when its order along each boundary divisor is nonnegative. Theta reciprocity turns these orders into inequalities on its tropical index. In a seed adapted to a simple root, the same basis records both the length of the root string and its leading arrow.
Let Σ \Sigma Σ be a seed with mutable index set I m u t I_{\mathrm{mut}} I mut , frozen index set I f r I_{\mathrm{fr}} I fr , and character lattice M = Z I M=\mathbb{Z}^I M = Z I . We index theta functions by tropical points m ∈ M m\in M m ∈ M , distinct from the simple-root indices i i i and the word positions k k k . We write U i n ( Σ ) \mathcal{U}_{\mathrm{in}}(\Sigma) U in ( Σ ) for the upper cluster algebra in which every frozen variable is inverted. For n ∈ N = Hom ( M , Z ) n\in\mathbb{N}=\operatorname{Hom}(M,\mathbb{Z}) n ∈ N = Hom ( M , Z ) , the monomial valuation in this seed is
ν n ( ∑ m c m x m ) = min c m ≠ 0 ⟨ m , n ⟩ . \nu_n\left(\sum_m c_mx^m\right)=\min_{c_m\ne0}\langle m,n\rangle. ν n ( m ∑ c m x m ) = c m = 0 min ⟨ m , n ⟩ .
These valuations are the integral tropical points of the cluster variety. Coordinates of a tropical point are transported when the seed changes. We normalize ϑ 0 = 1 \vartheta_0=1 ϑ 0 = 1 .
A reddening sequence is a mutation sequence whose terminal c c c -vectors, calculated with principal coefficients, are all nonpositive. Its terminal c c c -matrix is a negative permutation matrix. The corresponding mutable g g g -vectors are the negatives of the coordinate vectors after relabeling; in this situation the seed is also called injective-reachable . A maximal green sequence is a reddening sequence in which each mutation is at a green vertex. The word seeds used here satisfy this stronger condition.
Lemma 5.1. The principal exchange matrix of every word seed has a maximal green sequence. The same holds for its negative.
Proof. Shen–Weng construct a maximal green sequence for the one-sided string seed of every positive braid [43 Theorem 4.1] . Its principal matrix is the word matrix above, up to the global sign determined by the side carrying the word. Reversing the word sends the closed string ( k , s ( k ) ) (k,s(k)) ( k , s ( k )) to ( N + 1 − s ( k ) , N + 1 − k ) (N+1-s(k),N+1-k) ( N + 1 − s ( k ) , N + 1 − k ) and negates the principal matrix. Apply their theorem to the word and its reversal. □
To apply theta reciprocity over the integral theta-indexing lattice, we use the chiral dual ( Q ∙ , P ∙ ) (Q^\bullet,P^\bullet) ( Q ∙ , P ∙ ) of the seed datum ( P , Q ) (P,Q) ( P , Q ) in Cheung–Magee–Mandel–Muller [15 Section 4.1] .
Lemma 5.2. Let Σ \Sigma Σ have full-rank extended exchange matrix and a reddening sequence. In the conventions of [15 Section 4.1] , let ( P , Q ) (P,Q) ( P , Q ) be its seed datum, with character lattice M M M , and let ( Q ∙ , P ∙ ) (Q^\bullet,P^\bullet) ( Q ∙ , P ∙ ) be its chiral dual, with character lattice N ∙ N^\bullet N ∙ . Every theta function of the chiral dual indexed by an integral point is a Laurent polynomial in its positive seed chart.
For every n ∈ N Q n\in N_{\mathbb{Q}} n ∈ N Q , there is a finite A n ⊂ N Q A_n\subset N_{\mathbb{Q}} A n ⊂ N Q such that
v n ( ϑ m ) = min a ∈ A n ⟨ m , a ⟩ ( m ∈ M ) . v_n(\vartheta_m)=\min_{a\in A_n}\langle m,a\rangle\qquad(m\in M). v n ( ϑ m ) = a ∈ A n min ⟨ m , a ⟩ ( m ∈ M ) . Specifically, choose q ∈ Z > 0 q\in\mathbb{Z}_{>0} q ∈ Z > 0 with q n ∈ N ∙ qn\in N^\bullet q n ∈ N ∙ , and let S q n S_{qn} S q n be the Laurent support of the dual theta function ϑ q n , + ∨ \vartheta^{\vee}_{qn,+} ϑ q n , + ∨ . One may take A n = q − 1 S q n A_n=q^{-1}S_{qn} A n = q − 1 S q n . If n ∈ N n\in N n ∈ N , every vertex of conv ( A n ) \operatorname{conv}(A_n) conv ( A n ) belongs to N N N , so the minimum can be taken over finitely many integral linear forms.
Proof. The maps P P P and Q Q Q are injective over Q \mathbb{Q} Q : this is full rank for P P P , whereas the columns of Q Q Q are part of the seed basis. The two maps of the chiral-dual datum are therefore injective as well. Write them as integral d × r d\times r d × r matrices ( P ′ , Q ′ ) (P',Q') ( P ′ , Q ′ ) in mutually dual bases of the chiral-dual lattices. Complete the columns of Q ′ Q' Q ′ to a nonsingular integral matrix A = ( Q ′ ∣ K ) A=(Q'\mid K) A = ( Q ′ ∣ K ) , and define
P 0 = A T P ′ , Q 0 = ( I 0 ) . P_0=A^{\mathsf T}P', \qquad Q_0=\begin{pmatrix}I\\0\end{pmatrix}. P 0 = A T P ′ , Q 0 = ( I 0 ) . The identities A T P ′ = P 0 A^{\mathsf T}P'=P_0 A T P ′ = P 0 and A Q 0 = Q ′ AQ_0=Q' A Q 0 = Q ′ make A T A^{\mathsf T} A T an injective linear morphism of seed data from ( P ′ , Q ′ ) (P',Q') ( P ′ , Q ′ ) to ( P 0 , Q 0 ) (P_0,Q_0) ( P 0 , Q 0 ) . The latter is an ordinary A \mathcal{A} A -seed with full-rank extended matrix. To check its principal matrix in the nonsimply-laced case, put C = Q T P C=Q^{\mathsf T}P C = Q T P and use the symmetrizing matrix D D D of [15 Section 4.1] . Since P ∙ = P D P^\bullet=PD P ∙ = P D , Q ∙ = Q D − 1 Q^\bullet=QD^{-1} Q ∙ = Q D − 1 , and D − 1 C D^{-1}C D − 1 C is skew-symmetric, one obtains
Q ′ T P ′ = ( P ∙ ) T Q ∙ = D C T D − 1 = − C . Q'^{\mathsf T}P'=(P^\bullet)^{\mathsf T}Q^\bullet=DC^{\mathsf T}D^{-1}=-C. Q ′ T P ′ = ( P ∙ ) T Q ∙ = D C T D − 1 = − C . Thus it has the opposite principal matrix. The map on tori is finite; the integral matrix A A A need not be unimodular. Moreover, Q 0 T A T p = Q ′ T p Q_0^{\mathsf T}A^{\mathsf T}p=Q'^{\mathsf T}p Q 0 T A T p = Q ′ T p , so the map of exponent spaces takes the positive chamber to the positive chamber.
The negative of a matrix admitting a reddening sequence also admits one: reverse the path between its framed and coframed seeds and negate the exchange matrices, using the terminal negative permutation matrix for the relabeling [43 Appendix A, equation (A.25)] . The negative chamber therefore belongs to the cluster complex [29 Lemma 5.12 and Lemma 8.29] . Full rank also implies the strict convexity hypothesis of [29 Theorem 0.3(7)] . Theorem 0.3 and Proposition 8.28 of [29 ] give Laurent finiteness for all target theta
functions. By [15 Theorem 4.9] , the injective monomial map x m ↦ x A T m x^m \mapsto x^{A^{\mathsf{T}}m} x m ↦ x A T m sends each source theta function to the target theta function with the transported index. Its injectivity prevents distinct exponents from merging, so the source support is finite too. Full rank also makes each exchange-direction exponent determine its auxiliary coefficient exponent. Consequently no infinite auxiliary-coefficient support is concealed by specializing those coefficients to one.
For rational tropical points, use the rational broken-line exponent sets of [15 Section 3.1] . Their bending rule is homogeneous: scaling attached exponents by q q q and reparametrizing the path gives Θ q n , + ∨ = q Θ n , + ∨ \Theta^{\vee}_{qn,+}=q\Theta^{\vee}_{n,+} Θ q n , + ∨ = q Θ n , + ∨ . Corollary 3.2 of [15 ] therefore identifies their valuation with the minimum over q − 1 S q n q^{-1}S_{qn} q − 1 S q n . The original theta functions are Laurent finite by the same GHKK criterion. Both sides of theta reciprocity are consequently finite, and [15 Claim 4.16 and Theorem 5.19] gives the stated equality.
Finally, suppose n ∈ N n\in N n ∈ N , and write F ( m ) = min a ∈ A n ⟨ m , a ⟩ F(m)=\min_{a\in A_n}\langle m,a\rangle F ( m ) = min a ∈ A n ⟨ m , a ⟩ . The equality already proved makes F F F integral on M M M . If a a a is a vertex of conv ( A n ) \operatorname{conv}(A_n) conv ( A n ) , the region where it is the unique minimizer has nonempty interior and is rational. Choose an integral point u u u in this region. For a lattice basis m 1 , … , m d m_1,\ldots,m_d m 1 , … , m d and all sufficiently large integers k k k , both k u ku k u and k u + m j ku+m_j k u + m j remain in this region. Hence
⟨ m j , a ⟩ = F ( k u + m j ) − F ( k u ) ∈ Z . \langle m_j,a\rangle=F(ku+m_j)-F(ku)\in\mathbb{Z}. ⟨ m j , a ⟩ = F ( k u + m j ) − F ( k u ) ∈ Z .
It follows that a ∈ N a\in N a ∈ N . Taking just these vertices proves the integral assertion. □
Proposition 5.3. Let R = U p o l ( Σ ) R=\mathcal{U}_{\mathrm{pol}}(\Sigma) R = U pol ( Σ ) , where the extended exchange matrix has full rank and its principal part admits a reddening sequence. Then there is a rational polyhedral cone Ξ Σ ⊂ M R \Xi_\Sigma\subset M_{\mathbb{R}} Ξ Σ ⊂ M R such that
R = ⨁ m ∈ Ξ Σ ∩ M C ϑ m . R=\bigoplus_{m\in\Xi_\Sigma\cap M}\mathbb{C}\vartheta_m. R = m ∈ Ξ Σ ∩ M ⨁ C ϑ m . The basis is homogeneous for every grading for which the seed and all exchange relations are homogeneous, and its multiplication has nonnegative integral structure constants.
If e f ∈ N e_f\in N e f ∈ N is the coordinate covector of a frozen variable x f x_f x f , then
Ξ Σ = { m ∈ M R : ⟨ m , a ⟩ ≥ 0 for all a ∈ A e f , f ∈ I f r } . \Xi_\Sigma=\{m\in M_{\mathbb{R}}:\langle m,a\rangle\ge0\text{ for all }a\in A_{e_f},\ f\in I_{\mathrm{fr}}\}. Ξ Σ = { m ∈ M R : ⟨ m , a ⟩ ≥ 0 for all a ∈ A e f , f ∈ I fr } . If a homogeneous lattice grading γ : M → Γ \gamma:M\to\Gamma γ : M → Γ has a positive integral coarsening, then every degree slice Ξ Σ ∩ γ − 1 ( λ ) \Xi_\Sigma\cap\gamma^{-1}(\lambda) Ξ Σ ∩ γ − 1 ( λ ) is a bounded rational polyhedron, possibly empty.
Proof. The reddening sequence and full rank give a theta basis { ϑ m : m ∈ M } \{\vartheta_m:m\in M\} { ϑ m : m ∈ M } of U i n v ( Σ ) \mathcal{U}_{\mathrm{inv}}(\Sigma) U inv ( Σ ) , by [29 Proposition 8.28] , as in the preceding lemma. Lemma 1.1 gives
R = { g ∈ U i n v ( Σ ) : v e f ( g ) ≥ 0 for every f ∈ I f r } . R=\{g\in\mathcal{U}_{\mathrm{inv}}(\Sigma):v_{e_f}(g)\ge0\text{ for every }f\in I_{\mathrm{fr}}\}. R = { g ∈ U inv ( Σ ) : v e f ( g ) ≥ 0 for every f ∈ I fr } . The cluster scattering functions satisfy [15 Assumption 3.9 and Lemma 4.3] ; specialization of the auxiliary coefficients to one over C \mathbb{C} C sends no nonzero integer monomial to zero. Their valuative independence theorem [15 Theorems 3.8 and 3.10] therefore gives
v e f ( ∑ m c m ϑ m ) = min c m ≠ 0 v e f ( ϑ m ) . v_{e_f}\left(\sum_m c_m\vartheta_m\right)=\min_{c_m\ne0}v_{e_f}(\vartheta_m). v e f ( m ∑ c m ϑ m ) = c m = 0 min v e f ( ϑ m ) . Thus a theta expansion belongs to R R R exactly when each of its nonzero terms belongs to R R R . Lemma 5.2 turns the finitely many frozen-valuation conditions into finitely many integral linear inequalities, proving the asserted formula.
Let γ : M → Γ \gamma:M\to\Gamma γ : M → Γ describe a homogeneous grading. Every exponent in the seed expansion of ϑ m \vartheta_m ϑ m differs from m m m by an integral combination of exchange vectors, all of degree zero. Hence deg ϑ m = γ ( m ) \deg\vartheta_m=\gamma(m) deg ϑ m = γ ( m ) . Positivity and integrality of the structure constants follow from the theta multiplication in [29 ] . Since R R R is a subalgebra and its theta expansion is unique, multiplication cannot introduce basis elements outside Ξ Σ \Xi_\Sigma Ξ Σ .
For the final assertion, write h : Γ → Z h:\Gamma\to\mathbb{Z} h : Γ → Z for the positive coarsening. Nonnegativity of the grading on the theta functions gives h γ ≥ 0 h\gamma\ge0 hγ ≥ 0 on Ξ Σ \Xi_\Sigma Ξ Σ , by rationality. The rational cone
Ξ Σ ∩ ker ( h γ ) \Xi_\Sigma\cap\ker(h\gamma) Ξ Σ ∩ ker ( hγ ) is zero: otherwise it would contain a nonzero lattice point, and the theta function with that index would be linearly independent of ϑ 0 = 1 \vartheta_0 = 1 ϑ 0 = 1 in degree zero. The recession cone of any degree slice is Ξ Σ ∩ ker γ \Xi_\Sigma\cap\ker\gamma Ξ Σ ∩ ker γ , so it too is zero. Every slice is therefore bounded. Its lattice points are finite, which also proves that each homogeneous component of R R R is finite-dimensional. □
Theorem 5.4. Let Y Y Y be a Bott–Samelson variety for any word, or a normal projective birational contraction preserving its standard affine chart. Let Σ \Sigma Σ be its minimally lifted word seed. The Cox ring of Y Y Y has a theta basis indexed by the integral points of a rational polyhedral cone Ξ Σ \Xi_\Sigma Ξ Σ . The basis is homogeneous for the divisor-class and Cartan-weight gradings.
Let γ : M → Cl ( Y ) \gamma: M \to\operatorname{Cl}(Y) γ : M → Cl ( Y ) be the degree map. For every divisor class λ \lambda λ , the set P λ = Ξ Σ ∩ γ − 1 ( λ ) P_\lambda= \Xi_\Sigma\cap\gamma^{-1}(\lambda) P λ = Ξ Σ ∩ γ − 1 ( λ ) is a bounded rational polyhedron, possibly empty, and
H 0 ( Y , O Y ( λ ) ) = ⨁ m ∈ P λ ∩ M C ϑ m . H^0(Y,\mathcal{O}_Y(\lambda)) = \bigoplus_{m\in P_\lambda\cap M} \mathbb{C}\vartheta_m. H 0 ( Y , O Y ( λ )) = m ∈ P λ ∩ M ⨁ C ϑ m . The analogous formula for a Cartan-weight space is obtained by adding the corresponding linear weight equation.
Proof. Theorem 3.4 gives the upper-algebra equality and full row rank. Minimal lifting and contraction preserve the principal matrix, so Lemma 5.1 supplies a maximal green sequence. The Cox grading has a positive integral coarsening. Proposition 5.3 now gives the basis and the bounded degree slices. □
In a seed adapted to a locally nilpotent derivation D D D , the valuation at − e b -e_b − e b records the D D D -degree of each nonzero element of R R R , where b b b is the active coordinate.
Proposition 5.5. Assume the hypotheses of Proposition 5.3 . Let D 1 , … , D t D_1,\ldots,D_t D 1 , … , D t be locally nilpotent derivations of R R R . Suppose that each D j D_j D j admits a mutation-equivalent seed in which D j = M j ∂ b j D_j = M_j\partial_{b_j} D j = M j ∂ b j , where M j ≠ 0 M_j \ne0 M j = 0 is a Laurent monomial independent of b j b_j b j and all other seed variables are killed by D j D_j D j . Let n j n_j n j be the cluster tropical point represented by − e b j -e_{b_j} − e b j in that seed. Then the same theta basis is compatible with every intersection of the kernel filtrations:
⋂ j = 1 t ker D j r j + 1 = ⨁ m ∈ Ξ Σ ∩ M v n j ( ϑ m ) ≥ − r j ( 1 ≤ j ≤ t ) C ϑ m ( r j ∈ N ) . \bigcap_{j=1}^{t}\ker D_j^{r_j+1}
=
\bigoplus_{\substack{m\in\Xi_\Sigma\cap M\\ v_{n_j}(\vartheta_m)\ge-r_j\ (1\le j\le t)}}\mathbb{C}\vartheta_m
\qquad(r_j\in\mathbb{N}). j = 1 ⋂ t ker D j r j + 1 = m ∈ Ξ Σ ∩ M v n j ( ϑ m ) ≥ − r j ( 1 ≤ j ≤ t ) ⨁ C ϑ m ( r j ∈ N ) . Each condition on the right is a finite collection of integral linear inequalities. If the grading has a positive integral coarsening, each homogeneous component of the left side is indexed by a bounded rational polyhedron.
Proof. Fix D = D j D=D_j D = D j , and let K K K be the field generated by the other coordinates of its adapted seed. The seed expansion embeds R R R in K [ b , b − 1 ] K[b,b^{-1}] K [ b , b − 1 ] , while D = M ∂ b D=M\partial_b D = M ∂ b with M ∈ K × M\in K^\times M ∈ K × . Local nilpotence on R R R rules out negative powers of b b b , since a nonzero negative-power term survives every iterated derivative in characteristic zero. Thus, for every nonzero f ∈ R f\in R f ∈ R ,
max { q : D q f ≠ 0 } = deg b f = − v n j ( f ) . \max\{q : D^qf\ne0\}=\deg_b f=-v_{n_j}(f). max { q : D q f = 0 } = deg b f = − v n j ( f ) .
Valuative independence shows that this degree of a finite theta expansion is the maximum of the degrees of its nonzero terms. Hence ker D j r j + 1 \ker D_j^{r_j+1} ker D j r j + 1 is spanned by precisely the indicated subset of the theta basis. Intersecting subspaces spanned by subsets of one basis intersects those subsets, whether or not the derivations commute. Lemma 5.2 gives the finite linear inequalities, and the boundedness assertion of Proposition 5.3 applies to each homogeneous slice under the stated grading hypothesis. □
By Theorem 4.1 , this description applies on every Schubert divisorial section space to all positive simple-root operators and all stabilizing negative simple-root operators.
The regular cone in word coordinates
We give inequalities for the cone of Theorem 5.4 in the initial seed. Use the sign of the word matrix in § 4 , so B k , s ( k ) = − 1 B_{k,s(k)}=-1 B k , s ( k ) = − 1 .
Write a theta index as m = ( ( m k ) k = 1 N , ( u j ) j ∈ S ) m=((m_k)_{k=1}^{N},(u_j)_{j\in S}) m = (( m k ) k = 1 N , ( u j ) j ∈ S ) , where S S S indexes the standard divisors that survive the contraction and u j u_j u j is the coordinate at τ j \tau_j τ j . The row matrix of the lifted seed is B ~ = ( B ∣ C ) \widetilde{B}=(B\mid C) B = ( B ∣ C ) , where
C k j = 1 s ( k ) = j − 1 k = j . C_{kj}=\mathbf{1}_{s(k)=j}-\mathbf{1}_{k=j}. C k j = 1 s ( k ) = j − 1 k = j .
For a Schubert variety in a partial flag variety, one obtains S S S from the reduced word for the minimal coset representative: deletion of j j j must give a reduced word for a minimal representative of length N − 1 N-1 N − 1 .
The following sweep depends only on repetitions of the word letters. Start with the lifted seed and a theta index m m m . Read j = N , … , 1 j=N,\ldots,1 j = N , … , 1 . Just before processing j j j , record the current τ j \tau_j τ j -coordinate as U j ( m ) U_j(m) U j ( m ) if j ∈ S j\in S j ∈ S . To process j j j , write the occurrences of its letter as a 1 < ⋯ < a d a_1<\cdots<a_d a 1 < ⋯ < a d , with j = a h j=a_h j = a h , and mutate at
a d − 1 , a d − 2 , … , a d − h + 1 in this order . (22) a_{d-1},a_{d-2},\ldots,a_{d-h+1}\qquad\text{in this order}. \tag*{(22)} a d − 1 , a d − 2 , … , a d − h + 1 in this order . (22)
The list is empty for h = 1 h=1 h = 1 . At every mutation transport the theta index as well. Explicitly, for the current row matrix D D D , mutation at k k k sends m k m_k m k to − m k -m_k − m k and sends each m l m_l m l , l ≠ k l\ne k l = k , to
m l + [ D k l ] + m k − D k l min ( m k , 0 ) . (23) m_l+[D_{kl}]_+m_k-D_{kl}\min(m_k,0). \tag*{(23)} m l + [ D k l ] + m k − D k l min ( m k , 0 ) . (23)
Here [ v ] + = max ( v , 0 ) [v]_+=\max(v,0) [ v ] + = max ( v , 0 ) . The sweep includes every position, even when its boundary section has been set equal to one.
Theorem 5.6. For each simple-root index i i i with occurrences a 1 < ⋯ < a d a_1<\cdots<a_d a 1 < ⋯ < a d , the order of a theta function along its terminal minor is
v e b i ( ϑ m ) = min 1 ≤ h ≤ d ∑ r = h d m a r . v_{e_{b_i}}(\vartheta_m)=\min_{1\le h\le d}\sum_{r=h}^{d}m_{a_r}. v e b i ( ϑ m ) = 1 ≤ h ≤ d min r = h ∑ d m a r . For every j ∈ S j\in S j ∈ S we have v e τ j ( ϑ m ) = U j ( m ) v_{e_{\tau_j}}(\vartheta_m)=U_j(m) v e τ j ( ϑ m ) = U j ( m ) . Thus Ξ Σ \Xi_\Sigma Ξ Σ is cut out by the nonnegativity of all the displayed suffix sums and of the functions U j U_j U j . A single sweep computes all U j ( m ) U_j(m) U j ( m ) using ∑ i ( d i 2 ) \sum_i\binom{d_i}{2} ∑ i ( 2 d i ) mutations, where d i d_i d i is the number of occurrences of i i i . The distinct linear forms attained by each frozen order on full-dimensional regions give precisely the primitive facet inequalities of Ξ Σ \Xi_\Sigma Ξ Σ .
Proof. Suppose the column of a frozen variable f f f is nonnegative in a seed with theta index m ′ m' m ′ . In that seed every monomial in the pointed expansion of ϑ m \vartheta_m ϑ m has exponent m ′ + n D m'+nD m ′ + n D , n ≥ 0 n\ge0 n ≥ 0 , and the term of exponent m ′ m' m ′ has coefficient one. Thus the least f f f -exponent is m f ′ m'_f m f ′ . By Lemma 1.1 this computes the same valuation in every seed.
For f = b i f=b_i f = b i , Proposition 4.6 gives a nonnegative column after mutations at a 1 , … , a d − 1 a_1,\ldots,a_{d-1} a 1 , … , a d − 1 . The submatrix on the i i i -positions is a path with successive entries B a r , a r + 1 = − 1 B_{a_r,a_{r+1}}=-1 B a r , a r + 1 = − 1 . Mutation at the first remaining vertex leaves the subpath on later vertices unchanged, since that vertex meets only its successor there. Hence, if t r t_r t r is the coordinate at a r a_r a r just before its mutation and t d t_d t d the final terminal coordinate, Eq. (23) gives
t 1 = m a 1 , t r + 1 = m a r + 1 + min ( 0 , t r ) . t_1=m_{a_1},\qquad t_{r+1}=m_{a_{r+1}}+\min(0,t_r). t 1 = m a 1 , t r + 1 = m a r + 1 + min ( 0 , t r ) .
Induction gives the asserted minimum, including the case d = 1 d=1 d = 1 . On the cell seed these suffix inequalities also follow from Qin’s description of the dominant degrees [40 Lemma 8.11] .
For the boundary sections, consider the auxiliary word obtained by reversing i i i and using the transposed Cartan matrix. Write j ∗ = N + 1 − j j^*=N+1-j j ∗ = N + 1 − j . A mutable position k k k represents the interval ( k , s ( k ) ) (k,s(k)) ( k , s ( k )) , whose reflection starts at ρ ( k ) = s ( k ) ∗ \rho(k)=s(k)^* ρ ( k ) = s ( k ) ∗ . Under this identification the auxiliary principal row matrix is H = B p r i n T H=B_{\mathrm{prin}}^{\mathsf{T}} H = B prin T . Indeed, reversal reverses each arrow between adjacent intervals of one letter and interchanges the two crossing orders for intervals of distinct letters. In the latter case the magnitude of B k l B_{kl} B k l is − a i l i k -a_{i_li_k} − a i l i k . Transposing the Cartan matrix exchanges the two magnitudes. Nested and disjoint intervals of distinct letters contribute zero. This gives the identity on every entry, and matrix mutation preserves transposition.
Fix j ∈ S j \in S j ∈ S . On mutable rows its initial column is c = e p ( j ) − e j c=e_{p(j)}-e_j c = e p ( j ) − e j , with absent or frozen indices omitted. After identification by ρ \rho ρ , this is e j ∗ − e p r e v ( j ∗ ) e_{j^*}-e_{p_{\mathrm{rev}}(j^*)} e j ∗ − e p rev ( j ∗ ) , again restricted to mutable indices, where p r e v p_{\mathrm{rev}} p rev is the predecessor in the reversed word. This difference is the mutable part of the initial degree of Qin’s singleton interval variable W [ j ∗ , j ∗ ] W_{[j^*,j^*]} W [ j ∗ , j ∗ ] for the auxiliary word [40 Definition 8.3 and Lemma 8.4] . Here Qin uses mutable columns, so his matrix is the transpose of our auxiliary row matrix; the signs and Cartan entries agree with [40 Equation (6.1)] .
This identification persists under mutation. Indeed, the update of a frozen column at k k k , in the original row convention, is
c k ′ = − c k , c l ′ = c l + [ D l k ] + [ c k ] + − [ − D l k ] + [ − c k ] + ( l ≠ k ) . c'_k=-c_k,\qquad c'_l=c_l+[D_{lk}]_+[c_k]_+-[-D_{lk}]_+[-c_k]_+ \quad(l\ne k). c k ′ = − c k , c l ′ = c l + [ D l k ] + [ c k ] + − [ − D l k ] + [ − c k ] + ( l = k ) .
This agrees with (23) for the mutable coordinate in the transposed principal matrix. Thus it suffices to reach the seed containing the singleton. Qin’s sequence processes the reversed word from the left. At the v v v -th occurrence of a letter of multiplicity d d d , it mutates at the first d − v d-v d − v occurrences of that letter in increasing order. Just before the position j ∗ j^* j ∗ is processed, W [ j ∗ , j ∗ ] W_{[j^*,j^*]} W [ j ∗ , j ∗ ] occupies the first slot of its letter [40 Sections 8.1–8.2] . Its mutable degree there is a positive unit vector if the letter is repeated, and zero otherwise. Translating back by ρ \rho ρ gives exactly (22) : the v v v -th reversed occurrence is the h h h -th original one with v = d + 1 − h v=d+1-h v = d + 1 − h . Consequently the column of τ j \tau_j τ j is nonnegative when its coordinate is recorded. Deleting the other contracted columns has no effect on this calculation. The frozen-valuation criterion in Proposition 5.3 now gives the cone. For a letter occurring d d d times, the block has length 0 + 1 + ⋯ + ( d − 1 ) 0+1+\cdots+(d-1) 0 + 1 + ⋯ + ( d − 1 ) .
We finally identify the facets. Write z z z for the mutable part of m m m and v f v_f v f for its initial coordinate at a frozen variable f f f . By (23) , the computed order at f f f has the form v f + ψ f ( z ) v_f+\psi_f(z) v f + ψ f ( z ) : each increment uses only the current mutable coordinates. By Lemma 5.2, ψ f \psi_f ψ f is concave, first on rational tropical points by homogeneity and then on real tropical points by continuity. Subdivide the real space of mutable theta indices by the signs of the coordinates at the successive mutations. On each full-dimensional region ψ f \psi_f ψ f agrees with an integral linear form L L L . Concavity gives L ≥ ψ f L\ge\psi_f L ≥ ψ f everywhere. The regions cover a dense set, so the minimum of their distinct forms is ψ f \psi_f ψ f everywhere.
Each inequality v f + L ( z ) ≥ 0 v_f+L(z)\ge0 v f + L ( z ) ≥ 0 defines a facet. Choose z z z in an open region attaining L L L , away from the equality hyperplanes of the other distinct forms for f f f . Set v f = − L ( z ) v_f=-L(z) v f = − L ( z ) and make all other frozen coordinates sufficiently large. Only the chosen inequality is tight; all are strict if v f v_f v f is then increased. Its normal is primitive because its v f v_f v f -coefficient is one. For a terminal minor every suffix sum occurs on an open region: consecutive sums differ by independent mutable coordinates. This proves the facet assertion. To test a single theta index, one needs only (23) .
Example 5.7. In type A 4 A_4 A 4 take the reduced word 2123243 2123243 2123243 . In one-line notation this word represents 35421 35421 35421 ; deleting the third position gives 34512 34512 34512 . Thus τ 3 \tau_3 τ 3 survives on the full-flag Schubert variety. The mutable positions are 1 , 3 , 4 1,3,4 1 , 3 , 4 ; in that order the principal matrix and the column of τ 3 \tau_3 τ 3 are
( 0 − 1 0 1 0 1 0 − 1 0 ) , ( 1 , − 1 , 0 ) T . \begin{pmatrix}
0 & -1 & 0 \\
1 & 0 & 1 \\
0 & -1 & 0
\end{pmatrix},\qquad(1,-1,0)^{\mathsf T}. 0 1 0 − 1 0 − 1 0 1 0 , ( 1 , − 1 , 0 ) T . For this single divisor there is a shorter calculation. Mutations at 3 3 3 and 4 4 4 send its column first to ( 0 , 1 , − 1 ) T (0,1,-1)^{\mathsf T} ( 0 , 1 , − 1 ) T and then to ( 0 , 0 , 1 ) T (0,0,1)^{\mathsf T} ( 0 , 0 , 1 ) T . The resulting coordinate is
u 3 + min ( 0 , m 3 ) + min ( 0 , m 4 + max ( m 3 , 0 ) ) = u 3 + min ( 0 , m 3 , m 3 + m 4 ) . u_3+\min(0,m_3)+\min(0,m_4+\max(m_3,0))=u_3+\min(0,m_3,m_3+m_4). u 3 + min ( 0 , m 3 ) + min ( 0 , m 4 + max ( m 3 , 0 )) = u 3 + min ( 0 , m 3 , m 3 + m 4 ) . The facet u 3 + m 3 + m 4 ≥ 0 u_3+m_3+m_4\ge0 u 3 + m 3 + m 4 ≥ 0 involves two different letters.
Factoriality and two cluster charts
We use two results on upper cluster algebras in their polynomial-frozen form: a class-group formula and a two-chart criterion.
Proposition 5.8. Let R = U p o l ( Σ ) R = U_{\mathrm{pol}}(\Sigma) R = U pol ( Σ ) be finitely generated, and suppose that the exchange matrix has full row rank. Then
Cl ( R ) ≃ Z ∑ k ∈ I m u t ( c k − 1 ) , c k = gcd { ∣ B k j ∣ : j ∈ I } . \operatorname{Cl}(R) \simeq\mathbb{Z}^{\sum_{k\in I_{\mathrm{mut}}}(c_k-1)}, \qquad c_k = \gcd\{|B_{kj}| : j \in I\}. Cl ( R ) ≃ Z ∑ k ∈ I mut ( c k − 1 ) , c k = g cd{ ∣ B k j ∣ : j ∈ I } . Thus R R R is factorial precisely when every exchange row is primitive.
Proof. The ring R R R is normal. By Lemma 1.1, inverting the frozen coordinates gives U i n v ( Σ ) U_{\mathrm{inv}}(\Sigma) U inv ( Σ ) and removes only principal prime divisors. Nagata’s exact sequence gives Cl ( R ) ≃ Cl ( U i n v ( Σ ) ) \operatorname{Cl}(R) \simeq\operatorname{Cl}(U_{\mathrm{inv}}(\Sigma)) Cl ( R ) ≃ Cl ( U inv ( Σ )) . Apply the full-rank class-group formula of [39 Theorem 4.8] . Over C \mathbb{C} C , the exchange binomial at k k k has exactly c k c_k c k irreducible factors: up to a Laurent monomial, it is 1 + x c k a 1 + x^{c_k a} 1 + x c k a with a a a primitive. This gives the formula and the polynomial-frozen form of the factoriality criterion in [13 Theorems 4.9 and 4.13] . □
Corollary 5.9. Suppose that the exchange matrix has full row rank and primitive rows, and that U p o l ( Σ ) U_{\mathrm{pol}}(\Sigma) U pol ( Σ ) is finitely generated. If two seeds Σ 1 , Σ 2 \Sigma_1,\Sigma_2 Σ 1 , Σ 2 have no common mutable cluster variable, then
U p o l ( Σ ) = L p o l ( Σ 1 ) ∩ L p o l ( Σ 2 ) . U_{\mathrm{pol}}(\Sigma) = L_{\mathrm{pol}}(\Sigma_1) \cap L_{\mathrm{pol}}(\Sigma_2). U pol ( Σ ) = L pol ( Σ 1 ) ∩ L pol ( Σ 2 ) . Their mixed Laurent charts cover Spec U p o l ( Σ ) \operatorname{Spec}U_{\mathrm{pol}}(\Sigma) Spec U pol ( Σ ) in codimension one. For every Cox ring in Theorem 3.4, one may take a word seed and the terminal seed of a reddening sequence.
Proof. The ring R = U p o l ( Σ ) R = U_{\mathrm{pol}}(\Sigma) R = U pol ( Σ ) is factorial by Proposition 5.8 . Every mutable cluster variable x x x is prime in U i n v ( Σ ) U_{\mathrm{inv}}(\Sigma) U inv ( Σ ) [13 Theorem 4.9] . Its frozen valuations are zero. If a = x h a = xh a = x h with a ∈ R a \in R a ∈ R and h ∈ U i n v ( Σ ) h \in U_{\mathrm{inv}}(\Sigma) h ∈ U inv ( Σ ) , then v f ( h ) = v f ( a ) ≥ 0 v_f(h) = v_f(a) \ge0 v f ( h ) = v f ( a ) ≥ 0 for every frozen index f f f , so Lemma 1.1 gives h ∈ R h \in R h ∈ R . Primality thus descends to R R R . Distinct cluster variables are nonassociated [13 Lemma 4.22] . Each mixed Laurent ring is the localization of R R R at its mutable coordinates. A reduced denominator in their intersection could therefore involve only prime coordinates common to the two mutable clusters. There are none. The same argument shows that every height-one prime lies in at least one chart. This is the polynomial-frozen form of [13 Theorem 4.23] .
For the Cox rings, the word matrix has full row rank and primitive rows by the successor minor. A reddening sequence has terminal mutable g g g -vectors equal, after relabeling, to the negative coordinate vectors. None of its mutable variables is initial, so the two clusters are disjoint. □
Theta functions and root strings
We use theta functions for a full-rank classical seed, with rows of the extended exchange matrix B B B indexed by mutable vertices. For g ∈ Z I g \in\mathbb{Z}^{I} g ∈ Z I , the corresponding theta function has Laurent expansion
ϑ g = x g ∑ n ∈ N I m u t c g , n x n B , c g , 0 = 1. (24) \vartheta_g = x^g \sum_{n\in\mathbb{N}^{I_{\mathrm{mut}}}} c_{g,n}x^{nB}, \qquad c_{g,0}=1. \tag*{(24)} ϑ g = x g n ∈ N I mut ∑ c g , n x n B , c g , 0 = 1. (24)
For a full-rank injective-reachable seed, the results of Gross, Hacking, Keel, and Kontsevich [29 ] give a basis of the upper cluster algebra with all frozen variables inverted, consisting of these Laurent polynomials. A regular theta basis of a subalgebra R R R is a basis consisting of theta functions regular on Spec R \operatorname{Spec}R Spec R .
Proposition 5.10 (Tang). Let Σ \Sigma Σ be a full-rank injective-reachable classical seed. Write Σ o p \Sigma^{\mathrm{op}} Σ op for the seed with the same coordinates and exchange matrix − B -B − B . The two seeds define the same theta functions in their common rational function field. If a subset of the theta functions is a basis of a subalgebra R R R , it consists of all theta functions in R R R for either choice of sign.
Proof. The first assertion is [45 Proposition 4.2] , which includes the frozen coordinates. The second follows from linear independence in the common field [45 Lemma 4.3] . □
The tropical index of a theta function depends on the choice of sign for the exchange matrix. In the next proposition we use the index for the chosen orientation. The leading derivative formula follows Tang’s freezing argument [45 Proposition 3.2] . Retaining the scalar in the derivation fixes the normalization of the maximal divided derivative.
Proposition 5.11. Let R R R be a complex subalgebra of the rational function field of a full-rank injective-reachable seed, and suppose that R R R has a regular theta basis B \mathcal{B} B . Let D D D be a locally nilpotent derivation of R R R which, in this seed, has the form D = c x v ∂ / ∂ x b D=cx^{v}\partial/\partial x_b D = c x v ∂ / ∂ x b , where c ∈ C × c \in\mathbb{C}^{\times} c ∈ C × , b b b is frozen, and v b = 0 v_b=0 v b = 0 . Suppose that the b b b -column of B B B is sign-coherent and that v k = 0 v_k=0 v k = 0 for each mutable k k k with B k b = 0 B_{kb}=0 B k b = 0 .
Choose the sign of B B B so that B k b ≤ 0 B_{kb}\leq0 B k b ≤ 0 for all mutable k k k , and write B = { ϑ g : g ∈ S } \mathcal{B}=\{\vartheta_g:g\in S\} B = { ϑ g : g ∈ S } in this orientation. For every g ∈ S g\in S g ∈ S , the integer d = g b d=g_b d = g b is nonnegative and is the largest integer such that D d ϑ g ≠ 0 D^d\vartheta_g\ne0 D d ϑ g = 0 . If d > 0 d>0 d > 0 , then g + v − e b ∈ S g+v-e_b\in S g + v − e b ∈ S and
D ϑ g − c d ϑ g + v − e b ∈ ker D d − 1 . (25) D\vartheta_g-cd\vartheta_{g+v-e_b}\in\ker D^{d-1}. \tag*{(25)} D ϑ g − c d ϑ g + v − e b ∈ ker D d − 1 . (25) In particular, for every g ∈ S g\in S g ∈ S ,
D d ϑ g d ! = c d ϑ g + d ( v − e b ) . (26) \frac{D^d\vartheta_g}{d!}=c^d\vartheta_{g+d(v-e_b)}. \tag*{(26)} d ! D d ϑ g = c d ϑ g + d ( v − e b ) . (26) For each integer r ≥ 0 r\geq0 r ≥ 0 , the kernel of D r D^r D r has basis { ϑ g : g ∈ S , g b < r } \{\vartheta_g:g\in S,\ g_b<r\} { ϑ g : g ∈ S , g b < r } . The assignment ϑ g ↦ ϑ g + v − e b \vartheta_g\mapsto\vartheta_{g+v-e_b} ϑ g ↦ ϑ g + v − e b , defined when g b > 0 g_b>0 g b > 0 , is injective.
Proof. Proposition 5.10 permits the choice of orientation in the statement without changing B \mathcal{B} B . Let K 0 K_0 K 0 be the rational function field in the variables other than x b x_b x b . Every theta function is in K 0 [ x b ± 1 ] K_0[x_b^{\pm1}] K 0 [ x b ± 1 ] , and hence so is every element of R R R . Since the coefficient c x v cx^v c x v is in K 0 × K_0^{\times} K 0 × , a nonzero term of negative x b x_b x b -degree would survive every power of D D D . Local nilpotence therefore implies R ⊂ K 0 [ x b ] R\subset K_0[x_b] R ⊂ K 0 [ x b ] . For nonzero f ∈ R f\in R f ∈ R , its D D D -degree is its degree as a polynomial in x b x_b x b .
Let J = { k ∈ I m u t : B k b < 0 } J=\{k\in I_{\mathrm{mut}}:B_{kb}<0\} J = { k ∈ I mut : B k b < 0 } and freeze the vertices in J J J . Denote theta functions for the resulting seed by ϑ g ′ \vartheta'_g ϑ g ′ . The freezing theorem [41 Theorem 5.2] says that ϑ g ′ \vartheta'_g ϑ g ′ is obtained from ϑ g \vartheta_g ϑ g by selecting the terms with n k = 0 n_k=0 n k = 0 for all k ∈ J k\in J k ∈ J . These are exactly the terms of largest x b x_b x b -degree. Thus the highest homogeneous part of ϑ g \vartheta_g ϑ g is ϑ g ′ \vartheta'_g ϑ g ′ and has degree g b g_b g b .
The remaining rows of B B B are linearly independent. Choose a real linear functional on the exponent lattice that is strictly positive on each of them. In any finite linear relation among distinct ϑ g ′ \vartheta'_g ϑ g ′ , an index of least value under this functional contributes its leading monomial with coefficient one, and no other term can cancel it. Hence the functions ϑ g ′ \vartheta'_g ϑ g ′ , for all integral indices g g g , are linearly independent. It follows that the degree of a nonzero finite sum ∑ g a g ϑ g \sum_g a_g\vartheta_g ∑ g a g ϑ g is the largest g b g_b g b for which a g ≠ 0 a_g\ne0 a g = 0 . This proves the assertion about ker D r \ker D^r ker D r .
Suppose d = g b > 0 d=g_b>0 d = g b > 0 . All variables occurring in x v − e b x^{v-e_b} x v − e b are frozen after freezing J J J . Multiplication by this Laurent monomial therefore translates theta indices, so the highest homogeneous part of D ϑ g D\vartheta_g D ϑ g is c d x v − e b ϑ g ′ = c d ϑ g + v − e b ′ cd\,x^{v-e_b}\vartheta'_g=cd\,\vartheta'_{g+v-e_b} c d x v − e b ϑ g ′ = c d ϑ g + v − e b ′ . Expand D ϑ g D\vartheta_g D ϑ g in the basis B \mathcal{B} B . The independence just proved forces the index g + v − e b g+v-e_b g + v − e b to occur with coefficient c d cd c d and excludes every other term of degree d − 1 d-1 d − 1 . Every remaining term has degree at most d − 2 d-2 d − 2 . This proves (25) .
The target index has b b b -coordinate d − 1 d-1 d − 1 . Repeatedly applying (25) , the lower-degree terms vanish after d d d derivatives, and the leading coefficients multiply to c d d ! c^d d! c d d ! . This proves (26) . Injectivity follows because translation by v − e b v-e_b v − e b is injective.
Corollary 5.12. Let Y Y Y be as in Theorem 3.4. Its regular theta basis satisfies Proposition 5.11 for every nonzero positive simple-root derivation, using the adapted seeds of Proposition 4.6. For a Schubert variety, the same assertion holds for every nonzero stabilizing negative simple-root derivation by Theorem 4.1 . In particular, every maximal divided root derivative of a theta function is a nonzero scalar multiple of a theta function.
For any finite family of these derivations D a D_a D a and nonnegative integers r a r_a r a , the intersection ⋂ a ker D a r a \bigcap_a\ker D_a^{r_a} ⋂ a ker D a r a is spanned by the theta functions that it contains.
Proof. The adapted seeds are full-rank and injective-reachable. In each, the active column has one sign, and its nonzero mutable entries are precisely the mutable positions occurring in the coefficient of the derivation. Thus Proposition 5.11 applies in each chart. Its description of the kernels gives the assertion about their intersections. □
Perfect bases and Levi crystals.
Lemma 5.13. Let V V V be a finite-dimensional s l 2 \mathfrak{sl}_2 sl 2 -module, with operators E , F , H E,F,H E , F , H , and let B \mathcal{B} B be a weight basis compatible with the kernels of all powers of E E E . Suppose that c ∈ C × c \in\mathbb{C}^{\times} c ∈ C × and an injective partial map e ~ : B ⇢ B \widetilde{e}:\mathcal{B} \dashrightarrow\mathcal{B} e : B ⇢ B , defined exactly on { b ∈ B : E b ≠ 0 } \{b \in\mathcal{B}: Eb \ne0\} { b ∈ B : E b = 0 } , satisfy
E b − c ε E ( b ) e ~ ( b ) ∈ ker E ε E ( b ) − 1 Eb-c\varepsilon_E(b)\widetilde{e}(b)\in\ker E^{\varepsilon_E(b)-1} E b − c ε E ( b ) e ( b ) ∈ ker E ε E ( b ) − 1 for every b b b with E b ≠ 0 Eb\ne0 E b = 0 , where ε D ( b ) = max { r : D r b ≠ 0 } \varepsilon_D(b)=\max\{r:D^rb\ne0\} ε D ( b ) = max { r : D r b = 0 } . Then B \mathcal{B} B is also compatible with the kernels of all powers of F F F . The inverse partial map f ~ = e ~ − 1 \widetilde{f}=\widetilde{e}^{-1} f = e − 1 is defined precisely when F b ≠ 0 Fb\ne0 F b = 0 , and
F b − c − 1 ε F ( b ) f ~ ( b ) ∈ ker F ε F ( b ) − 1 . Fb-c^{-1}\varepsilon_F(b)\widetilde{f}(b)\in\ker F^{\varepsilon_F(b)-1}. F b − c − 1 ε F ( b ) f ( b ) ∈ ker F ε F ( b ) − 1 . For a basis vector of weight μ \mu μ , we have ε F ( b ) = ε E ( b ) + μ \varepsilon_F(b)=\varepsilon_E(b)+\mu ε F ( b ) = ε E ( b ) + μ .
Proof. The leading relation forces e ~ ( b ) \widetilde{e}(b) e ( b ) to have weight μ + 2 \mu+2 μ + 2 and E E E -degree ε E ( b ) − 1 \varepsilon_E(b)-1 ε E ( b ) − 1 when b b b has weight μ \mu μ . Write V ≤ n V_{\le n} V ≤ n for the sum of the isotypic components of highest weight at most n n n . If a nonzero weight-μ \mu μ vector v v v has a nonzero component of highest weight n n n and none of larger highest weight, the rank-one representation formulas give
ε E ( v ) = ( n − μ ) / 2 , ε F ( v ) = ( n + μ ) / 2. \varepsilon_E(v)=(n-\mu)/2,\qquad\varepsilon_F(v)=(n+\mu)/2. ε E ( v ) = ( n − μ ) /2 , ε F ( v ) = ( n + μ ) /2. Thus, in each weight space, the filtrations by E E E -kernels, F F F -kernels, and V ≤ n V_{\le n} V ≤ n coincide after reindexing. The basis is consequently compatible with all three.
Fix b b b of weight μ \mu μ , and write p = ε E ( b ) p=\varepsilon_E(b) p = ε E ( b ) , q = p + μ q=p+\mu q = p + μ , and n = μ + 2 p n=\mu+2p n = μ + 2 p . Its image belongs to the induced basis of gr n V = V ≤ n / V ≤ n − 1 \operatorname{gr}_n V=V_{\le n}/V_{\le n-1} gr n V = V ≤ n / V ≤ n − 1 , a direct sum of copies of the irreducible module of highest weight n n n . The leading E E E -relation preserves n n n , and its remainder belongs to V ≤ n − 2 V_{\le n-2} V ≤ n − 2 . If q > 0 q>0 q > 0 , the map E : ( gr n V ) μ − 2 → ( gr n V ) μ E:(\operatorname{gr}_n V)_{\mu-2}\to(\operatorname{gr}_n V)_\mu E : ( gr n V ) μ − 2 → ( gr n V ) μ is an isomorphism. Its matrix in the induced bases has exactly one nonzero entry in each column, so the leading arrows give a bijection between those bases. There is therefore a unique b ′ b' b ′ with e ~ ( b ′ ) = b \widetilde{e}(b')=b e ( b ′ ) = b , and E b ′ = c ( p + 1 ) b Eb'=c(p+1)b E b ′ = c ( p + 1 ) b modulo V ≤ n − 2 V_{\le n-2} V ≤ n − 2 .
On ( gr n V ) μ (\operatorname{gr}_n V)_\mu ( gr n V ) μ , the operator E F EF E F is multiplication by ( p + 1 ) q (p+1)q ( p + 1 ) q . Applying the preceding isomorphism gives F b = c − 1 q b ′ Fb=c^{-1}qb' F b = c − 1 q b ′ modulo V ≤ n − 2 V_{\le n-2} V ≤ n − 2 . At weight μ − 2 \mu-2 μ − 2 , every vector in the remainder has F F F -degree at most q − 2 q-2 q − 2 , which proves the asserted leading relation. If q = 1 q=1 q = 1 , this remainder is zero. If q = 0 q=0 q = 0 , then F b = 0 Fb=0 F b = 0 , and the weight-μ − 2 \mu-2 μ − 2 space of gr n V \operatorname{gr}_n V gr n V is zero, so no leading E E E -arrow ends at b b b . Thus the two partial maps are mutually inverse on their stated domains. □
Proposition 5.14. Let Y Y Y be as in Theorem 3.4, and suppose that a connected reductive group L L L of finite type acts on Y Y Y and on Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) , preserving Cox \operatorname{Cox} Cox degrees. Assume that the regular theta basis is homogeneous for a maximal torus T L T_L T L of L L L and satisfies Proposition 5.11 for every nonzero positive simple-root operator E i E_i E i , with leading constant c i c_i c i . Set c i = 1 c_i=1 c i = 1 when E i = 0 E_i=0 E i = 0 . Let P L = X ∗ ( T L ) P_L=X^*(T_L) P L = X ∗ ( T L ) be its character lattice, and choose a character χ : P L → C × \chi:P_L\to\mathbb{C}^{\times} χ : P L → C × satisfying χ ( α i ) = c i \chi(\alpha_i)=c_i χ ( α i ) = c i . Then the basis ϑ ~ m = χ ( wt ϑ m ) ϑ m \widetilde{\vartheta}_m=\chi(\operatorname{wt}\vartheta_m)\vartheta_m ϑ m = χ ( wt ϑ m ) ϑ m is perfect for both simple-root directions, with leading coefficient ε D ( ϑ m ) \varepsilon_D(\vartheta_m) ε D ( ϑ m ) for each D = E i , F i D=E_i,F_i D = E i , F i . In particular,
D ε D ( ϑ m ) ϑ ~ m ε D ( ϑ m ) ! = ϑ ~ m max D . \frac{D^{\varepsilon_D(\vartheta_m)}\widetilde{\vartheta}_m}{\varepsilon_D(\vartheta_m)!}
=\widetilde{\vartheta}_{m_{\max}^D}. ε D ( ϑ m )! D ε D ( ϑ m ) ϑ m = ϑ m m a x D . Its multiplication constants agree with those of the original theta basis. The required character exists because the simple roots of L L L are linearly independent in ( P L ) Q (P_L)_{\mathbb{Q}} ( P L ) Q .
Proof. Each Cox-degree piece is finite-dimensional. Apply Lemma 5.13 to the pinned s l 2 \mathfrak{sl}_2 sl 2 action in each piece. It supplies the negative leading arrows and their constants c i − 1 c_i^{-1} c i − 1 . For an operator of weight β = α i \beta=\alpha_i β = α i or − α i -\alpha_i − α i , the leading target has weight wt ϑ m + β \operatorname{wt}\vartheta_m+\beta wt ϑ m + β . Rescaling therefore changes the leading coefficient to
χ ( β ) ε D ( ϑ m ) χ ( wt ϑ m ) χ ( wt ϑ m + β ) = ε D ( ϑ m ) . \chi(\beta)\varepsilon_D(\vartheta_m)\frac{\chi(\operatorname{wt}\vartheta_m)}{\chi(\operatorname{wt}\vartheta_m+\beta)}
=\varepsilon_D(\vartheta_m). χ ( β ) ε D ( ϑ m ) χ ( wt ϑ m + β ) χ ( wt ϑ m ) = ε D ( ϑ m ) .
Iteration gives the maximal divided-derivative formula. Every term of a theta product has the sum of the input weights, so rescaling leaves all multiplication constants unchanged. Finally, the prescribed values define a character on the root lattice, and this character extends to P L P_L P L because C × \mathbb{C}^{\times} C × is divisible.
Corollary 5.15. Under the hypotheses of Proposition 5.14, the leading root arrows on the regular theta basis of each section space H 0 ( Y , O Y ( λ ) ) H^0(Y,\mathcal{O}_Y(\lambda)) H 0 ( Y , O Y ( λ )) form the normal Kashiwara crystal of that L L L -module. More precisely, write B λ = { ϑ m : m ∈ Ξ Σ ∩ M , γ ( m ) = λ } B_\lambda=\{\vartheta_m:m\in\Xi_\Sigma\cap M,\ \gamma(m)=\lambda\} B λ = { ϑ m : m ∈ Ξ Σ ∩ M , γ ( m ) = λ } . If H 0 ( Y , O Y ( λ ) ) ≃ ⨁ μ V L ( μ ) ⊕ m λ , μ H^0(Y,\mathcal{O}_Y(\lambda))\simeq\bigoplus_\mu V_L(\mu)^{\oplus m_{\lambda,\mu}} H 0 ( Y , O Y ( λ )) ≃ ⨁ μ V L ( μ ) ⊕ m λ , μ , then
B λ ≃ ⨆ μ B L ( μ ) ⊔ m λ , μ . B_\lambda\simeq\bigsqcup_\mu B_L(\mu)^{\sqcup m_{\lambda,\mu}}. B λ ≃ μ ⨆ B L ( μ ) ⊔ m λ , μ . The weight of m m m is η ( m ) \eta(m) η ( m ) , and its string lengths are ε i ( ϑ m ) = deg E i ϑ m \varepsilon_i(\vartheta_m)=\deg_{E_i}\vartheta_m ε i ( ϑ m ) = deg E i ϑ m and φ i ( ϑ m ) = deg F i ϑ m \varphi_i(\vartheta_m)=\deg_{F_i}\vartheta_m φ i ( ϑ m ) = deg F i ϑ m . The highest vertices are precisely the theta functions in the common kernel of the E i E_i E i . Their multiplicities are given by Corollary 6.2. The arrows are unchanged by the weight-character rescaling and are independent of the adapted seed used to calculate them.
Proof. The preceding proposition gives a perfect basis in the sense of Berenstein–Kazhdan [7 Definition 5.49] . Lemma 5.13 identifies the two arrows on each rank-one string and gives φ i − ε i = ⟨ η ( m ) , α i ∨ ⟩ \varphi_i-\varepsilon_i=\langle\eta(m),\alpha_i^\vee\rangle φ i − ε i = ⟨ η ( m ) , α i ∨ ⟩ . The perfect-basis theorem [7 Lemma 5.51 and Theorem 5.55] identifies this crystal with that of any other perfect basis. Comparing with the union of global crystal bases in an irreducible decomposition gives the stated disjoint union [7 Lemma 5.50] . Kernel compatibility identifies the highest vertices with a basis of H 0 ( Y , O Y ( λ ) ) U L H^0(Y,\mathcal{O}_Y(\lambda))^{U_L} H 0 ( Y , O Y ( λ ) ) U L , whose weight-μ \mu μ part is the corresponding multiplicity space. Finally, the leading target is uniquely determined modulo the smaller kernel filtration, so it does not depend on the chart.
For the canonical root actions, the positive hypothesis follows from Proposition 4.6. Hence the corollary applies to every finite-type Levi subgroup whose action extends to Y Y Y and its graded Cox ring, including Bott–Samelson varieties and their contractions. Only the positive adapted charts are needed for this conclusion.
For the explicit crystal arrows we negate the word exchange matrix and use this orientation throughout its mutation class. Every positive-root adapted column is then nonpositive. By Proposition 5.10 , the indices in this orientation parametrize the same theta basis.
Corollary 5.16. Let Y Y Y be as in Theorem 3.4, and let L L L be a finite-type Levi whose action extends to Y Y Y and its graded Cox ring. For each nonzero E i E_i E i , choose the adapted seed Σ i \Sigma_i Σ i of Proposition 4.6, with E i = c i x v i ∂ b i E_i=c_i x^{v_i}\partial_{b_i} E i = c i x v i ∂ b i . Write T i T_i T i for the tropical coordinate change from the reference seed to Σ i \Sigma_i Σ i , and put ζ i = v i − e b i \zeta_i=v_i-e_{b_i} ζ i = v i − e b i . Use the opposite exchange orientation fixed above. For m ∈ Ξ Σ ∩ M m\in\Xi_\Sigma\cap M m ∈ Ξ Σ ∩ M , the crystal string lengths are
ε i ( m ) = ( T i ( m ) ) b i , φ i ( m ) = ε i ( m ) + ⟨ η ( m ) , α i ∨ ⟩ . \varepsilon_i(m)=(T_i(m))_{b_i},\qquad\varphi_i(m)=\varepsilon_i(m)+\langle\eta(m),\alpha_i^\vee\rangle. ε i ( m ) = ( T i ( m ) ) b i , φ i ( m ) = ε i ( m ) + ⟨ η ( m ) , α i ∨ ⟩ . The arrows are given by
e ~ i ( m ) = T i − 1 ( T i ( m ) + ζ i ) , f ~ i ( m ) = T i − 1 ( T i ( m ) − ζ i ) , \widetilde e_i(m)=T_i^{-1}\bigl(T_i(m)+\zeta_i\bigr),\qquad\widetilde f_i(m)=T_i^{-1}\bigl(T_i(m)-\zeta_i\bigr), e i ( m ) = T i − 1 ( T i ( m ) + ζ i ) , f i ( m ) = T i − 1 ( T i ( m ) − ζ i ) , on the respective domains ε i ( m ) > 0 \varepsilon_i(m)>0 ε i ( m ) > 0 and φ i ( m ) > 0 \varphi_i(m)>0 φ i ( m ) > 0 . The iterated arrows are obtained by replacing ζ i \zeta_i ζ i by r ζ i r\zeta_i r ζ i or s ζ i s\zeta_i s ζ i , as long as the corresponding string lengths permit the iteration. These formulas are integral piecewise-linear, preserve the divisor-class slices, and do not depend on the adapted chart.
Proof. Theta indices are tropical points, so a mutation sequence induces an integral piecewise-linear bijection between their seed coordinates [GHKK, Section 3], [Qin2, Section 2.4]. For clarity, mutation at a mutable index k k k has the formula
( T k ( m ) ) k = − m k , (T_k(m))_k=-m_k, ( T k ( m ) ) k = − m k , ( T k ( m ) ) j = m j + [ B k j ] + m k − B k j min ( m k , 0 ) ( j ≠ k ) , (T_k(m))_j=m_j+[B_{kj}]_+m_k-B_{kj}\min(m_k,0)\qquad(j\ne k), ( T k ( m ) ) j = m j + [ B k j ] + m k − B k j min ( m k , 0 ) ( j = k ) , where [ a ] + = max ( a , 0 ) [a]_+=\max(a,0) [ a ] + = max ( a , 0 ) and we use the exchange matrix before the mutation. In the adapted chart, Proposition 5.11 identifies the E i E_i E i -degree with the b i b_i b i -coordinate and the leading
arrow with translation by ζ i \zeta_i ζ i . Conjugating this translation by T i T_i T i gives the formula for e ~ i \widetilde e_i e i . Lemma 5.13 identifies f ~ i \widetilde f_i f i with its inverse and gives its domain and string length. Iteration gives the remaining formulas. Their degree and independence assertions follow because they calculate the intrinsic leading root arrows on the theta basis. □ \square □
Horospherical varieties
The adapted operators also extend to a class of equivariant compactifications through a known description of their Cox rings.
Corollary 5.17. Let G G G be a semisimple simply connected complex group, and let X X X be a normal projective horospherical G G G -variety with free divisor class group. Write G / H G/H G / H for its open orbit, P = N G ( H ) P=N_G(H) P = N G ( H ) , taken standard after conjugation, and D 1 , … , D s D_1,\ldots,D_s D 1 , … , D s for its G G G -stable prime divisors. Adjoin s s s isolated polynomial frozen variables to a Cox seed for G / P G/P G / P , assigning them the degrees [ D 1 ] , … , [ D s ] [D_1],\ldots,[D_s] [ D 1 ] , … , [ D s ] . Assign each color degree of G / P G/P G / P the class of the corresponding color of X X X . The resulting graded upper cluster algebra is Cox ( X ) \operatorname{Cox}(X) Cox ( X ) . Every nonzero geometric simple-root operator E i E_i E i or F i F_i F i has an adapted seed obtained by the same mutation sequence as for G / P G/P G / P , with the same monomial coefficient.
Proof. The Cox-ring description of Brion and Gagliardi gives an isomorphism of graded G G G -algebras
Cox ( X ) ≃ Cox ( G / P ) [ z 1 , … , z s ] , \operatorname{Cox}(X) \simeq\operatorname{Cox}(G/P)[z_1,\ldots,z_s], Cox ( X ) ≃ Cox ( G / P ) [ z 1 , … , z s ] ,
with the indicated degrees, where z j z_j z j is the canonical section of D j D_j D j and is G G G -invariant [10 Theorem 4.3.2 and Lemma 4.1.2] [24 Theorems 3.6 and 3.8] . For the semisimple simply connected group, the equivariant divisor class group is the ordinary divisor class group, so this is the lift of the given geometric action. Polynomial extension by isolated frozen variables commutes with the intersection defining the upper cluster algebra, by comparing coefficients of the z z z -monomials. Theorem 3.4 identifies the flag factor. Every root operator kills the z j z_j z j , and hence the adapted seeds of Theorem 4.1 remain adapted after this extension. □ \square □
The corresponding basis is { ϑ m z a : a ∈ N s } \{\vartheta_m z^a : a \in\mathbb{N}^s\} { ϑ m z a : a ∈ N s } , where ϑ m \vartheta_m ϑ m runs through the regular theta basis of Cox ( G / P ) \operatorname{Cox}(G/P) Cox ( G / P ) . Isolated frozen variables contribute no scattering walls. The regular theta cone is Ξ G / P × R ≥ 0 s \Xi_{G/P} \times\mathbb{R}_{\geq0}^s Ξ G / P × R ≥ 0 s . If ι : Pic ( G / P ) → Cl ( X ) \iota: \operatorname{Pic}(G/P) \to\operatorname{Cl}(X) ι : Pic ( G / P ) → Cl ( X ) sends flag color classes to their counterparts on X X X , the Cox degree of the basis element indexed by ( m , a ) (m,a) ( m , a ) is
γ X ( m , a ) = ι ( γ G / P ( m ) ) + ∑ j = 1 s a j [ D j ] . \gamma_X(m,a)=\iota(\gamma_{G/P}(m))+\sum_{j=1}^{s}a_j[D_j]. γ X ( m , a ) = ι ( γ G / P ( m )) + j = 1 ∑ s a j [ D j ] .
Slicing by γ X ( m , a ) = λ \gamma_X(m,a)=\lambda γ X ( m , a ) = λ gives the theta basis of H 0 ( X , O X ( λ ) ) H^0(X,\mathcal{O}_X(\lambda)) H 0 ( X , O X ( λ )) . Root operators preserve the flag degree and a a a , so the crystal is the disjoint union of the flag crystals occurring in this slice. Only finitely many may occur, because the section space is finite-dimensional.
For a general spherical variety, Brion’s description involves monomial base change of invariant boundary parameters [10 Theorem 4.3.2] . Such substitutions can introduce common boundary factors into exchanges. Transferring adapted seeds through that base change requires an additional argument; the polynomial-extension proof above treats the horospherical case.
Root invariants and polyhedral applications
For an adapted derivation D D D , the D D D -degree of a section is its polynomial degree in the active coordinate. Imposing bounds on several such degrees cuts the regular theta cone by linear inequalities. We use the resulting semigroups to describe the common invariant ring and the Rees algebra of the joint filtration. The corresponding toric degenerations come from finite Khovanskii bases [KM].
For the next theorem, let R = U p o l ( Σ ) R=\mathcal{U}_{\mathrm{pol}}(\Sigma) R = U pol ( Σ ) have full-rank extended exchange matrix, a reddening sequence for its principal part, and a positive integral coarsening of its grading. Let $D_1,\ldots,D_t be nonzero homogeneous locally nilpotent derivations, each adapted to a frozen coordinate b j b_j b j . Write B = ⋂ j ker D j B = \bigcap_j \ker D_j B = ⋂ j ker D j , and set
d j ( f ) = max { q : D j q f ≠ 0 } ( f ≠ 0 ) , F r B = ⋂ j ker D j r j + 1 , ϕ j ( m ) = v − e b j ( ϑ m ) . d_j(f) = \max\{q : D_j^q f \ne0\}\quad(f \ne0),\qquad F_rB = \bigcap_j \ker D_j^{r_j+1},\qquad\phi_j(m) = v_{-e_{b_j}}(\vartheta_m). d j ( f ) = max { q : D j q f = 0 } ( f = 0 ) , F r B = j ⋂ ker D j r j + 1 , ϕ j ( m ) = v − e b j ( ϑ m ) .
Here d j ( 0 ) = − ∞ d_j(0)=-\infty d j ( 0 ) = − ∞ , and the tropical point − e b j -e_{b_j} − e b j is expressed in the j j j -th adapted seed.
Theorem 6.1. The cones
C d = { m ∈ Ξ Σ : ϕ j ( m ) ≥ 0 for all j } , C_d=\{m\in\Xi_\Sigma:\phi_j(m)\ge0\text{ for all }j\}, C d = { m ∈ Ξ Σ : ϕ j ( m ) ≥ 0 for all j } , C ~ d = { ( m , r ) : m ∈ Ξ Σ , ϕ j ( m ) + r j ≥ 0 for all j } \widetilde C_d=\{(m,r):m\in\Xi_\Sigma,\ \phi_j(m)+r_j\ge0\text{ for all }j\} C d = {( m , r ) : m ∈ Ξ Σ , ϕ j ( m ) + r j ≥ 0 for all j } are rational polyhedral. Their lattice points index theta bases of B B B and of R d ( R ) = ⨁ r ∈ N t F r R t r \mathcal{R}_d(R)=\bigoplus_{r\in\mathbb{N}^t}F_rR\,t^r R d ( R ) = ⨁ r ∈ N t F r R t r , respectively. The rings R R R , B B B , and R d ( R ) \mathcal{R}_d(R) R d ( R ) are finitely generated normal domains with flat graded degenerations to the normal semigroup algebras of their cones. In each case, the theta functions indexed by a Hilbert basis generate the ring. The operators D j D_j D j need not commute.
Proof. Proposition 5.5 identifies the displayed theta bases. Lemma 5.2 expresses each ϕ j \phi_j ϕ j as the minimum of finitely many integral linear forms. Moreover, ϕ j = − d j ≤ 0 \phi_j=-d_j\le0 ϕ j = − d j ≤ 0 on the lattice points of Ξ Σ \Xi_\Sigma Ξ Σ , hence on the entire cone. The inequalities for C ~ d \widetilde C_d C d therefore imply r j ≥ 0 r_j\ge0 r j ≥ 0 .
For each of the three algebras, write C C C for its cone and b m b_m b m for the basis element indexed by m m m . Pointed theta multiplication gives
b m b n = b m + n + ∑ q ∈ N r ∖ { 0 } c q b m + n + ∑ i q i p i . b_m b_n=b_{m+n}+\sum_{q\in\mathbb{N}^r\setminus\{0\}}c_qb_{m+n+\sum_iq_ip_i}. b m b n = b m + n + q ∈ N r ∖ { 0 } ∑ c q b m + n + ∑ i q i p i .
where the exchange directions p i p_i p i are linearly independent and have degree zero. In the Rees algebra the directions are ( p i , 0 ) (p_i,0) ( p i , 0 ) . Choose an integral linear form ℓ \ell ℓ positive on each direction. Adding a multiple of the positive grading makes ℓ \ell ℓ positive on the Hilbert basis H H H of the relevant lattice semigroup. For the Rees algebra use the grading h ( γ ( m ) ) + ∑ j r j h(\gamma(m))+\sum_jr_j h ( γ ( m )) + ∑ j r j .
Each homogeneous component is finite-dimensional by Proposition 5.3. If m = ∑ i h i m=\sum_i h_i m = ∑ i h i with h i ∈ H h_i\in H h i ∈ H , the product ∏ i b h i \prod_i b_{h_i} ∏ i b h i has leading term b m b_m b m . Every correction term has the same degree and strictly larger ℓ \ell ℓ -value. Descending induction on ℓ \ell ℓ within this component therefore expresses b m b_m b m as a polynomial in the b h b_h b h . The filtration by ℓ \ell ℓ has associated graded algebra C [ C ∩ M ] \mathbb{C}[C\cap M] C [ C ∩ M ] , using M ⊕ Z t M\oplus\mathbb{Z}^t M ⊕ Z t for the Rees cone. The same induction generates its extended Rees algebra by u u u and b h u − ℓ ( h ) b_hu^{-\ell(h)} b h u − ℓ ( h ) for h ∈ H h\in H h ∈ H . This algebra is torsion-free over C [ u ] \mathbb{C}[u] C [ u ] , hence flat. Saturation of the lattice semigroup gives normality of the toric fiber.
Finally, R R R is normal as an intersection of mixed Laurent rings. If K j K_j K j is the field generated by the inactive coordinates of the adapted seed, then B = R ∩ ⋂ j K j B=R\cap\bigcap_jK_j B = R ∩ ⋂ j K j , proving its normality. Since − d j -d_j − d j is a valuation, extend it by the Gauss rule v ~ j ( t k ) = δ j k \widetilde v_j(t_k)=\delta_{jk} v j ( t k ) = δ j k . Then
R d ( R ) = R [ t 1 , … , t t ] ∩ ⋂ j O v ~ j , \mathcal{R}_d(R)=R[t_1,\ldots,t_t]\cap\bigcap_j\mathcal{O}_{\widetilde v_j}, R d ( R ) = R [ t 1 , … , t t ] ∩ j ⋂ O v j ,
which proves normality of the Rees algebra. □
Corollary 6.2. Let a connected reductive group L L L of finite type act on a Cox ring R R R from Theorem 3.4, preserving Cox degrees. Assume that the theta basis is homogeneous for a maximal torus and that the positive simple-root operators are adapted. Write γ \gamma γ for the Cox-degree map and η \eta η for the weight map. For D = ( deg E i ) i ≠ 0 D=(\deg_{E_i})_{i\ne0} D = ( deg E i ) i = 0 , the ring R L U R^U_L R L U has theta basis indexed by C d ∩ M C_d\cap M C d ∩ M and degenerates to C [ C d ∩ M ] \mathbb{C}[C_d\cap M] C [ C d ∩ M ] . The multiplicity of the simple module of highest weight μ \mu μ in H 0 ( Y , O Y ( λ ) ) H^0(Y,\mathcal{O}_Y(\lambda)) H 0 ( Y , O Y ( λ )) is
# { m ∈ C d ∩ M : γ ( m ) = λ , η ( m ) = μ } . \#\{m\in C_d\cap M:\gamma(m)=\lambda,\ \eta(m)=\mu\}. # { m ∈ C d ∩ M : γ ( m ) = λ , η ( m ) = μ } .
Proof. The U L U_L U L -invariants are the common kernels of the positive simple-root operators. Complete reducibility identifies their weight-μ \mu μ part with the corresponding multiplicity space. Apply Theorem 6.1. □
Finite generation of these invariants is classical [30 ] . The point here is the common system of inequalities for branching spaces and simultaneous root filtrations. For a finite-type Levi, the negative-root inequalities also follow from the positive ones: Lemma 5.13 gives
− deg F i ϑ m = − deg E i ϑ m − ⟨ η ( m ) , α i ∨ ⟩ . -\deg_{F_i}\vartheta_m=-\deg_{E_i}\vartheta_m-\langle\eta(m),\alpha_i^\vee\rangle. − deg F i ϑ m = − deg E i ϑ m − ⟨ η ( m ) , α i ∨ ⟩ .
Proposition 6.3. Under the hypotheses of Theorem 6.1 , let g g g be the theta index of b = ∏ f ∈ I f r x f b=\prod_{f\in I_{\mathrm{fr}}}x_f b = ∏ f ∈ I fr x f . Then
int ( Ξ Σ ) ∩ M = g + ( Ξ Σ ∩ M ) , int ( C ~ d ) ∩ ( M ⊕ Z t ) = ( g , 2 , … , 2 ) + ( C ~ d ∩ ( M ⊕ Z t ) ) . \operatorname{int}(\Xi_\Sigma)\cap M=g+(\Xi_\Sigma\cap M),\qquad\operatorname{int}(\widetilde C_d)\cap(M\oplus\mathbb{Z}^t)=(g,2,\ldots,2)+(\widetilde C_d\cap(M\oplus\mathbb{Z}^t)). int ( Ξ Σ ) ∩ M = g + ( Ξ Σ ∩ M ) , int ( C d ) ∩ ( M ⊕ Z t ) = ( g , 2 , … , 2 ) + ( C d ∩ ( M ⊕ Z t )) . Consequently the two normal semigroup algebras are Gorenstein, with canonical-module generators in degrees γ ( g ) \gamma(g) γ ( g ) and ( γ ( g ) , 2 , … , 2 ) (\gamma(g),2,\ldots,2) ( γ ( g ) , 2 , … , 2 ) , respectively. For the Cox rings of Theorem 3.4, γ ( g ) = − K Y \gamma(g)=-K_Y γ ( g ) = − K Y .
Proof. Multiplication by frozen monomials translates theta indices, so ϑ m + g = b ϑ m \vartheta_{m+g}=b\vartheta_m ϑ m + g = b ϑ m . Thus the frozen valuation functions ψ f ( m ) = v e f ( ϑ m ) \psi_f(m)=v_{e_f}(\vartheta_m) ψ f ( m ) = v e f ( ϑ m ) satisfy ψ f ( m + g ) = ψ f ( m ) + 1 \psi_f(m+g)=\psi_f(m)+1 ψ f ( m + g ) = ψ f ( m ) + 1 . Express each as a minimum of finitely many integral linear forms, keeping only vertices of their convex hull. Each retained form is uniquely minimal on a full-dimensional chamber. Comparing sufficiently deep points in that chamber with their translates by g g g shows that every retained form takes value one at g g g . These forms define Ξ Σ \Xi_\Sigma Ξ Σ ; hence g g g is interior, and strict positivity at lattice points is equivalent to positivity after subtracting g g g .
Since the active variable is frozen, d j ( b ) = 1 d_j(b)=1 d j ( b ) = 1 , and φ j ( m + g ) = φ j ( m ) − 1 \varphi_j(m+g)=\varphi_j(m)-1 φ j ( m + g ) = φ j ( m ) − 1 . The same argument shows that every retained defining form a a a of φ j \varphi_j φ j satisfies a ( g ) = − 1 a(g)=-1 a ( g ) = − 1 . The forms defining the Rees cone are therefore ( a f , 0 ) (a_f,0) ( a f , 0 ) and ( a , e j ) (a,e_j) ( a , e j ) , all of which take value one at ( g , 2 , … , 2 ) (g,2,\ldots,2) ( g , 2 , … , 2 ) . This proves the second translation formula. The canonical-module statement follows from the Danilov–Stanley description by interior lattice points [11 Theorem 6.3.5] . For Cox rings the frozen-degree identity gives γ ( g ) = − K Y \gamma(g)=-K_Y γ ( g ) = − K Y .
Corollary 6.4. Let Y Y Y be as in Theorem 3.4, and put P λ = Ξ Σ ∩ γ − 1 ( λ ) P_\lambda=\Xi_\Sigma\cap\gamma^{-1}(\lambda) P λ = Ξ Σ ∩ γ − 1 ( λ ) . For every integral class λ \lambda λ in the interior of the effective cone,
relint ( P λ ) ∩ M = g + ( P λ + K Y ∩ M ) , # ( relint ( P λ ) ∩ M ) = h 0 ( Y , O Y ( λ + K Y ) ) . \operatorname{relint}(P_\lambda)\cap M=g+(P_{\lambda+K_Y}\cap M),\qquad\#(\operatorname{relint}(P_\lambda)\cap M)=h^0(Y,\mathcal{O}_Y(\lambda+K_Y)). relint ( P λ ) ∩ M = g + ( P λ + K Y ∩ M ) , # ( relint ( P λ ) ∩ M ) = h 0 ( Y , O Y ( λ + K Y )) . In particular, − K Y -K_Y − K Y is interior to the effective cone and g g g is the unique relative-interior lattice point of P − K Y P_{-K_Y} P − K Y . If K = ker γ K=\ker\gamma K = ker γ , the polar of P − K Y − g ⊂ K R P_{-K_Y}-g\subset K_{\mathbb{R}} P − K Y − g ⊂ K R is a lattice polytope in K ∗ K^* K ∗ .
Proof. The effective cone is γ ( Ξ Σ ) \gamma(\Xi_\Sigma) γ ( Ξ Σ ) . For an interior class its affine fiber meets int ( Ξ Σ ) \operatorname{int}(\Xi_\Sigma) int ( Ξ Σ ) , so relint ( P λ ) = int ( Ξ Σ ) ∩ γ − 1 ( λ ) \operatorname{relint}(P_\lambda)=\operatorname{int}(\Xi_\Sigma)\cap\gamma^{-1}(\lambda) relint ( P λ ) = int ( Ξ Σ ) ∩ γ − 1 ( λ ) . Slice Proposition 6.3 by degree. The anticanonical assertions follow from γ ( g ) = − K Y \gamma(g)=-K_Y γ ( g ) = − K Y and P 0 = { 0 } P_0=\{0\} P 0 = { 0 } . Finally, if the integral defining forms of Ξ Σ \Xi_\Sigma Ξ Σ are a j a_j a j , then
P − K Y − g = { v ∈ K R : a j ( v ) ≥ − 1 for all j } . P_{-K_Y}-g=\{v\in K_{\mathbb{R}}:a_j(v)\geq-1\text{ for all }j\}. P − K Y − g = { v ∈ K R : a j ( v ) ≥ − 1 for all j } . Its polar is conv { − a j ∣ K } \operatorname{conv}\{-a_j|_K\} conv { − a j ∣ K } , whose vertices are integral.
Examples The full flag variety of S L 3 \mathrm{SL}_3 SL 3 The Cox ring of the incidence variety in P 2 × ( P 2 ) ∗ \mathbb{P}^2\times(\mathbb{P}^2)^* P 2 × ( P 2 ) ∗ has the presentation
R = C [ x , x ′ , a , b , c , d ] / ( x x ′ − a b − c d ) , deg ( x , a , c ) = ( 1 , 0 ) , deg ( x ′ , b , d ) = ( 0 , 1 ) . R=\mathbb{C}[x,x',a,b,c,d]/(xx'-ab-cd),\qquad\deg(x,a,c)=(1,0),\qquad\deg(x',b,d)=(0,1). R = C [ x , x ′ , a , b , c , d ] / ( x x ′ − ab − c d ) , deg ( x , a , c ) = ( 1 , 0 ) , deg ( x ′ , b , d ) = ( 0 , 1 ) . Use the seed with mutable coordinate x x x and frozen coordinates a , b , c , d a,b,c,d a , b , c , d . To index the theta basis, choose the exchange orientation p = ( 0 , 1 , 1 , − 1 , − 1 ) p=(0,1,1,-1,-1) p = ( 0 , 1 , 1 , − 1 , − 1 ) , opposite to the positive-root convention of the word seed. The theta functions themselves are unchanged by this choice.
For m = ( q , A , B , C , D ) ∈ Z 5 m=(q,A,B,C,D)\in\mathbb{Z}^5 m = ( q , A , B , C , D ) ∈ Z 5 , the regular theta cone is
A ≥ 0 , B ≥ 0 , C ≥ 0 , D ≥ 0 , C + q ≥ 0 , D + q ≥ 0. (27) A\geq0,\quad B\geq0,\quad C\geq0,\quad D\geq0,\quad C+q\geq0,\quad D+q\geq0. \tag*{(27)} A ≥ 0 , B ≥ 0 , C ≥ 0 , D ≥ 0 , C + q ≥ 0 , D + q ≥ 0. (27)
The corresponding basis element is
ϑ m = { x q a A b B c C d D , q ≥ 0 , ( x ′ ) − q a A b B c C + q d D + q , q < 0. \vartheta_m =
\begin{cases}
x^q a^A b^B c^C d^D, & q \ge0,\\
(x')^{-q}a^A b^B c^{C+q}d^{D+q}, & q < 0.
\end{cases} ϑ m = { x q a A b B c C d D , ( x ′ ) − q a A b B c C + q d D + q , q ≥ 0 , q < 0.
These are exactly the standard monomials containing at most one of x , x ′ x,x' x , x ′ . The relation proves that they span. Their leading monomials in the seed ( x , a , b , c , d ) (x,a,b,c,d) ( x , a , b , c , d ) are distinct, which proves independence. The Cox degree is γ ( m ) = ( q + A + C , B + D ) \gamma(m)=(q+A+C,B+D) γ ( m ) = ( q + A + C , B + D ) . The slice of degree ( r , s ) (r,s) ( r , s ) contains ( r + 1 ) ( s + 1 ) ( r + s + 2 ) / 2 (r+1)(s+1)(r+s+2)/2 ( r + 1 ) ( s + 1 ) ( r + s + 2 ) /2 points, as expected from the dimension formula for the corresponding irreducible S L 3 \mathrm{SL}_3 SL 3 -module.
Two positive simple-root operators are E 1 = a ∂ x + x ′ ∂ b E_1=a\partial_x+x'\partial_b E 1 = a ∂ x + x ′ ∂ b and E 2 = x ∂ c + d ∂ x ′ E_2=x\partial_c+d\partial_{x'} E 2 = x ∂ c + d ∂ x ′ . Both preserve the incidence relation. The initial seed is adapted to E 2 = x ∂ c E_2=x\partial_c E 2 = x ∂ c , and mutation at x x x gives the seed adapted to E 1 = x ′ ∂ b E_1=x'\partial_b E 1 = x ′ ∂ b . Consequently
deg E 1 ϑ m = B + max ( q , 0 ) , deg E 2 ϑ m = C . \deg_{E_1}\vartheta_m=B+\max(q,0),\qquad\deg_{E_2}\vartheta_m=C. deg E 1 ϑ m = B + max ( q , 0 ) , deg E 2 ϑ m = C .
The common invariant cone is the face given by q = B = C = 0 q=B=C=0 q = B = C = 0 , so R U + = C [ a , d ] R^{U^+}=\mathbb{C}[a,d] R U + = C [ a , d ] . In degree ( r , s ) (r,s) ( r , s ) its unique basis element is a r d s a^r d^s a r d s . For instance, in the adjoint degree ( 1 , 1 ) (1,1) ( 1 , 1 ) , six of the eight standard monomials belong to ker E 1 2 ∩ ker E 2 2 \ker E_1^2\cap\ker E_2^2 ker E 1 2 ∩ ker E 2 2 . The two excluded monomials are x b xb x b , of E 1 E_1 E 1 -degree two, and c x ′ cx' c x ′ , of E 2 E_2 E 2 -degree two.
The frozen product a b c d abcd ab c d is the theta function of index g = ( 0 , 1 , 1 , 1 , 1 ) g=(0,1,1,1,1) g = ( 0 , 1 , 1 , 1 , 1 ) , and all primitive defining forms of (27) take value one at g g g . Thus the interior translation is visible in this elementary example.
A singular Schubert variety
Let Y = X 3412 ⊂ S L 4 / B + Y=X_{3412}\subset\mathrm{SL}_4/B^+ Y = X 3412 ⊂ SL 4 / B + , with reduced word i = ( 2 , 1 , 3 , 2 ) \mathbf{i}=(2,1,3,2) i = ( 2 , 1 , 3 , 2 ) , and let F ∙ F_\bullet F ∙ be the standard flag in C 4 \mathbb{C}^4 C 4 . A flag L ⊂ P ⊂ H L\subset P\subset H L ⊂ P ⊂ H belongs to Y Y Y precisely when L ⊂ F 3 L\subset F_3 L ⊂ F 3 and F 1 ⊂ H F_1\subset H F 1 ⊂ H . Forgetting P P P exhibits Y Y Y as a P 1 \mathbb{P}^1 P 1 -bundle over the incidence hypersurface
Q = { ( [ ℓ 1 : ℓ 2 : ℓ 3 ] , [ h 2 : h 3 : h 4 ] ) : ℓ 2 h 2 + ℓ 3 h 3 = 0 } ⊂ P 2 × P 2 . Q=\{([\ell_1:\ell_2:\ell_3],[h_2:h_3:h_4]):\ell_2h_2+\ell_3h_3=0\}\subset\mathbb{P}^2\times\mathbb{P}^2. Q = {([ ℓ 1 : ℓ 2 : ℓ 3 ] , [ h 2 : h 3 : h 4 ]) : ℓ 2 h 2 + ℓ 3 h 3 = 0 } ⊂ P 2 × P 2 .
The singular locus of Q Q Q is the point ( F 1 , F 3 ) (F_1,F_3) ( F 1 , F 3 ) , where its local equation is u v + z t = 0 uv+zt=0 uv + z t = 0 . The singular locus of Y Y Y is the P 1 \mathbb{P}^1 P 1 above this point.
Deleting any position from i \mathbf{i} i leaves a reduced expression of length three, and the four deletions give distinct elements. The standard boundary divisors of Z i Z_{\mathbf{i}} Z i therefore map to distinct Schubert divisors D j D_j D j on Y Y Y , and e j = [ D j ] e_j=[D_j] e j = [ D j ] form a basis of Cl ( Y ) \operatorname{Cl}(Y) Cl ( Y ) . On the standard affine chart the four prefix minors and the first mutant are
( f 1 , f 2 , f 3 , f 4 ) = ( z 1 , z 2 , z 3 , z 1 z 4 − z 2 z 3 ) , f 1 ′ = z 4 . (f_1,f_2,f_3,f_4)=(z_1,z_2,z_3,z_1z_4-z_2z_3),\qquad f'_1=z_4. ( f 1 , f 2 , f 3 , f 4 ) = ( z 1 , z 2 , z 3 , z 1 z 4 − z 2 z 3 ) , f 1 ′ = z 4 .
Write x , a , b , c x,a,b,c x , a , b , c for their four initial lifts. Only x x x is mutable. Writing y y y for its mutant, we obtain
Cox ( Y ) = C [ x , y , a , b , c , τ 1 , τ 2 , τ 3 , τ 4 ] / ( x y − τ 4 a b − τ 1 c ) . \operatorname{Cox}(Y)=\mathbb{C}[x,y,a,b,c,\tau_1,\tau_2,\tau_3,\tau_4]/(xy-\tau_4ab-\tau_1c). Cox ( Y ) = C [ x , y , a , b , c , τ 1 , τ 2 , τ 3 , τ 4 ] / ( x y − τ 4 ab − τ 1 c ) .
Indeed, an initial Laurent expansion belongs to the y y y -chart exactly when its coefficient of x − k x^{-k} x − k is divisible by ( τ 4 a b + τ 1 c ) k (\tau_4ab+\tau_1c)^k ( τ 4 ab + τ 1 c ) k for every k > 0 k>0 k > 0 . Hence the ordinary and upper cluster algebras both have the indicated presentation.
The degrees are
deg x = e 1 , deg a = e 1 + e 2 , deg b = e 1 + e 3 , deg c = deg y = e 1 + e 2 + e 3 + e 4 , deg τ j = e j . \deg x=e_1,\qquad\deg a=e_1+e_2,\qquad\deg b=e_1+e_3,\qquad\deg c=\deg y=e_1+e_2+e_3+e_4,\qquad\deg\tau_j=e_j. deg x = e 1 , deg a = e 1 + e 2 , deg b = e 1 + e 3 , deg c = deg y = e 1 + e 2 + e 3 + e 4 , deg τ j = e j .
The restrictions of the three fundamental line bundles, of degrees deg a , deg c , deg b \deg a,\deg c,\deg b deg a , deg c , deg b respectively, form a basis of Pic ( Y ) \operatorname{Pic}(Y) Pic ( Y ) [Bril, Proposition 2.2.8]. Thus
Pic ( Y ) = { d ∈ Z 4 : d 1 − d 2 − d 3 + d 4 = 0 } , − K Y = 4 e 1 + 3 e 2 + 3 e 3 + 2 e 4 , \operatorname{Pic}(Y)=\{d\in\mathbb{Z}^4:d_1-d_2-d_3+d_4=0\},\qquad-K_Y=4e_1+3e_2+3e_3+2e_4, Pic ( Y ) = { d ∈ Z 4 : d 1 − d 2 − d 3 + d 4 = 0 } , − K Y = 4 e 1 + 3 e 2 + 3 e 3 + 2 e 4 ,
where the anticanonical class is the sum of the frozen degrees.
Choose B = ( 0 , 1 , 1 , − 1 , − 1 , 0 , 0 , 1 ) B=(0,1,1,-1,-1,0,0,1) B = ( 0 , 1 , 1 , − 1 , − 1 , 0 , 0 , 1 ) in the coordinate order ( x , a , b , c , τ 1 , … , τ 4 ) (x,a,b,c,\tau_1,\ldots,\tau_4) ( x , a , b , c , τ 1 , … , τ 4 ) . For m = ( p , q , r , s , u 1 , u 2 , u 3 , u 4 ) m=(p,q,r,s,u_1,u_2,u_3,u_4) m = ( p , q , r , s , u 1 , u 2 , u 3 , u 4 ) , the regular theta cone is
q , r , s , u 1 , u 2 , u 3 , u 4 ≥ 0 , p + s ≥ 0 , p + u 1 ≥ 0. (28) q,r,s,u_1,u_2,u_3,u_4\ge0,\qquad p+s\ge0,\qquad p+u_1\ge0. \tag*{(28)} q , r , s , u 1 , u 2 , u 3 , u 4 ≥ 0 , p + s ≥ 0 , p + u 1 ≥ 0. (28)
If p ≥ 0 p \ge0 p ≥ 0 , the theta function is x p a q b r c s τ u x^p a^q b^r c^s \tau^u x p a q b r c s τ u . If p = − k < 0 p=-k<0 p = − k < 0 , it is y k a q b r c s − k τ 1 u 1 − k τ 2 u 2 τ 3 u 3 τ 4 u 4 y^k a^q b^r c^{s-k}\tau_1^{u_1-k}\tau_2^{u_2}\tau_3^{u_3}\tau_4^{u_4} y k a q b r c s − k τ 1 u 1 − k τ 2 u 2 τ 3 u 3 τ 4 u 4 . These are all cluster monomials. The exchange relation eliminates each product x y xy x y , so they span. Their distinct leading Laurent monomials prove independence, and nonnegativity of their exponents gives (28) . Theorem 5.6 gives the same inequalities directly: the suffix sums at c c c are s , p + s s,p+s s , p + s , and the order at τ 1 \tau_1 τ 1 is u 1 + min ( 0 , p ) u_1+\min(0,p) u 1 + min ( 0 , p ) . To obtain the section space of degree d d d , impose
d 1 = p + q + r + s + u 1 , d 2 = q + s + u 2 , d 3 = r + s + u 3 , d 4 = s + u 4 . d_1=p+q+r+s+u_1,\qquad d_2=q+s+u_2,\qquad d_3=r+s+u_3,\qquad d_4=s+u_4. d 1 = p + q + r + s + u 1 , d 2 = q + s + u 2 , d 3 = r + s + u 3 , d 4 = s + u 4 .
All three positive simple-root operators act nontrivially. Only the rank-one subgroup G 2 G_2 G 2 preserves Y Y Y . With suitable pinning, the four operators and their nilpotence degrees are
For example, differentiating the exchange gives E 2 x = τ 1 E_2x=\tau_1 E 2 x = τ 1 in the y y y -chart and F 2 y = c F_2y=c F 2 y = c in the x x x -chart. The table also gives the simultaneous filtrations. The bound ε E 1 ( m ) ≤ r 1 \varepsilon_{E_1}(m)\le r_1 ε E 1 ( m ) ≤ r 1 imposes q ≤ r 1 q\le r_1 q ≤ r 1 and q − p ≤ r 1 q-p\le r_1 q − p ≤ r 1 , and the common kernel of all three positive operators is C [ τ 1 , … , τ 4 ] \mathbb{C}[\tau_1,\ldots,\tau_4] C [ τ 1 , … , τ 4 ] .
The crystal for the simple root 2 has coroot weight u 1 − p − s u_1-p-s u 1 − p − s and arrows
e ~ 2 ( m ) = m − e x + e τ 1 ( p + s > 0 ) , f ~ 2 ( m ) = m + e x − e τ 1 ( u 1 > 0 ) . \widetilde e_2(m)=m-e_x+e_{\tau_1}\quad(p+s>0),\qquad\widetilde f_2(m)=m+e_x-e_{\tau_1}\quad(u_1>0). e 2 ( m ) = m − e x + e τ 1 ( p + s > 0 ) , f 2 ( m ) = m + e x − e τ 1 ( u 1 > 0 ) .
Here e x , e τ 1 e_x,e_{\tau_1} e x , e τ 1 are exponent vectors. Both arrows stay in the same Cox-degree slice of the cone. For instance, D = e 1 + e 2 + e 3 D=e_1+e_2+e_3 D = e 1 + e 2 + e 3 is non-Cartier, and its slice consists of the four sections x τ 2 τ 3 x\tau_2\tau_3 x τ 2 τ 3 , τ 1 τ 2 τ 3 \tau_1\tau_2\tau_3 τ 1 τ 2 τ 3 , a τ 3 a\tau_3 a τ 3 , b τ 2 b\tau_2 b τ 2 . The first two are a single string of length one. Both root operators kill the last two. Thus H 0 ( Y , O Y ( D ) ) ≃ V ( 1 ) ⊕ 2 V ( 0 ) H^0(Y,\mathcal{O}_Y(D))\simeq V(1)\oplus2V(0) H 0 ( Y , O Y ( D )) ≃ V ( 1 ) ⊕ 2 V ( 0 ) for this S L 2 \mathrm{SL}_2 SL 2 , where V ( n ) V(n) V ( n ) has highest weight n n n . The frozen product has theta index g = ( 0 , 1 , 1 , 1 , 1 , 1 , 1 , 1 ) g=(0,1,1,1,1,1,1,1) g = ( 0 , 1 , 1 , 1 , 1 , 1 , 1 , 1 ) , at which each primitive facet form in (28) has value one. This realizes the interior translation of Proposition 6.3.
Projecting Y Y Y to the plane P P P gives Gr ( 2 , 4 ) \operatorname{Gr}(2,4) Gr ( 2 , 4 ) . Only D 1 D_1 D 1 survives as a divisor, so τ 2 , τ 3 , τ 4 \tau_2,\tau_3,\tau_4 τ 2 , τ 3 , τ 4 specialize to one. In Plücker coordinates the resulting tuple ( x , a , b , c , τ 1 ; y ) (x,a,b,c,\tau_1;y) ( x , a , b , c , τ 1 ; y ) is ( p 24 , p 14 , p 23 , p 12 , p 34 ; p 13 ) (p_{24},p_{14},p_{23},p_{12},p_{34};p_{13}) ( p 24 , p 14 , p 23 , p 12 , p 34 ; p 13 ) , and the exchange is p 24 p 13 = p 14 p 23 + p 12 p 34 p_{24}p_{13}=p_{14}p_{23}+p_{12}p_{34} p 24 p 13 = p 14 p 23 + p 12 p 34 . Here w − 1 α 1 = α 3 w^{-1}\alpha_1=\alpha_3 w − 1 α 1 = α 3 , as in the non-left-descent case of Theorem 4.1. The negative simple root subgroup adds a multiple of row one to row two. Its derivation sends p 24 , p 23 p_{24},p_{23} p 24 , p 23 to p 14 , p 13 p_{14},p_{13} p 14 , p 13 , respectively, up to a common pinning scalar. After mutating p 24 p_{24} p 24 to p 13 p_{13} p 13 , the only moving coordinate is the terminal minor p 23 p_{23} p 23 indexed by 3, and F 1 F_1 F 1 is a scalar multiple of p 13 ∂ p 23 p_{13}\partial_{p_{23}} p 13 ∂ p 23 .
An equivariant compactification of the affine plane
Let Y Y Y be the blowup of P 2 \mathbb{P}^2 P 2 at [ 1 : 0 : 0 ] [1:0:0] [ 1 : 0 : 0 ] and [ 0 : 1 : 0 ] [0:1:0] [ 0 : 1 : 0 ] . Its toric Cox ring is R = C [ a , b , c , e , f ] [ 1 ADH ] R=\mathbb{C}[a,b,c,e,f]\left[\frac{1}{\operatorname{ADH}}\right] R = C [ a , b , c , e , f ] [ ADH 1 ] . On the invariant affine plane C = { [ Z ≠ 0 ] } C=\{[Z\ne0]\} C = {[ Z = 0 ]} , set x = X / Z x=X/Z x = X / Z and y = Y / Z y=Y/Z y = Y / Z ; then x = a c e x=\frac{a}{ce} x = ce a and y = b c f y=\frac{b}{cf} y = c f b . The action [ X : Y : Z ] ↦ [ X : Y + t Z : Z ] [X:Y:Z]\mapsto[X:Y+tZ:Z] [ X : Y : Z ] ↦ [ X : Y + tZ : Z ] fixes the line at infinity pointwise, so it lifts through both blowups. Its infinitesimal action on C C C , after choosing the sign of the parameter, is D = ∂ y D=\partial_y D = ∂ y .
Give C [ C ] \mathbb{C}[C] C [ C ] the rank-one seed with mutable coordinate y y y , frozen coordinate z 0 = x y − 1 z_0=xy-1 z 0 = x y − 1 , and exchange y x = z 0 + 1 yx=z_0+1 y x = z 0 + 1 . The boundary valuations, in the order ( c , e , f ) (c,e,f) ( c , e , f ) , give
ν ( y ) = ( 1 , 0 , 1 ) , ν ( x ) = ( 1 , 1 , 0 ) , ν ( z 0 ) = ( 2 , 1 , 1 ) . \nu(y)=(1,0,1),\qquad\nu(x)=(1,1,0),\qquad\nu(z_0)=(2,1,1). ν ( y ) = ( 1 , 0 , 1 ) , ν ( x ) = ( 1 , 1 , 0 ) , ν ( z 0 ) = ( 2 , 1 , 1 ) .
Thus the lifted seed is ( b , z , c , e , f ) (b,z,c,e,f) ( b , z , c , e , f ) , where b b b is mutable and z = a b − c 2 e f z=ab-c^2ef z = ab − c 2 e f is frozen. Its exchange is b a = z + c 2 e f ba=z+c^2ef ba = z + c 2 e f . The two mixed Laurent charts are R [ b − 1 ] R[b^{-1}] R [ b − 1 ] and R [ a − 1 ] R[a^{-1}] R [ a − 1 ] . Since a a a and b b b are relatively prime in the polynomial ring R R R , their intersection is R R R . Consequently this lifted seed has polynomial upper cluster algebra exactly Cox ( Y ) \operatorname{Cox}(Y) Cox ( Y ) .
The lifted action fixes a , c , e , f a,c,e,f a , c , e , f and satisfies D ~ b = c f \widetilde D b=cf D b = c f and D ~ z = a c f \widetilde D z=acf D z = a c f . The initial moving coordinates ( b , z ) (b,z) ( b , z ) form a chain. Mutation at b b b reaches the seed ( a , z , c , e , f ) (a,z,c,e,f) ( a , z , c , e , f ) , in which D ~ = a c f ∂ z \widetilde D=acf\,\partial_z D = a c f ∂ z . This is an instance of Theorem 4.4 with a nontrivial mutation. Since Y Y Y has dimension two and Picard rank three, it is not a Bott–Samelson variety or a contraction of one of the same dimension. The calculation is therefore an instance of equivariant adaptation beyond the Schubert setting.
Declaration of generative AI and AI-assisted technologies
During the development of the paper, the author used ChatGPT to assist in exploring and formulating most proof arguments. The author subsequently checked, revised, and independently verified all arguments and takes full responsibility for the mathematical content of the paper.
References
School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai, China
Email address: jiarui@sjtu.edu.cn