On the dimension conjecture for global Galois deformation rings in characteristic two

Put D=D(K,n) and d_n=\lfloor n^2/2\rfloor.

If r_1(K)=0, [Boeckle2013, Proposition 1.5.1] gives

Since \dim W(k)=1, Corollary 4.3 implies

Reducing the presentation modulo 2 and applying Lemma 4.5 to the regular local ring k[[x_1,\ldots,x_{h^1}]] shows that every irreducible component of \operatorname{Spec}(R_{\bar\rho}/2R_{\bar\rho}) has dimension at least h^1-m\ge h^1-h^2=D.

Suppose now that r_1(K)>0. Let B be the completed tensor product over W(k) of the real framed deformation rings, and let A represent global deformations with a frame at each real prime. By Proposition 3.2,

At each real prime, the periodic resolution of C_2 has alternating differentials c_v-1 and 1+c_v. These coincide in characteristic two, and hence h_v^1=h_v^2. Together with Lemma 4.2, the tangent and relation counts in Proposition 3.2 give

where the last equality follows from Lemma 4.1 and h^0=1. Proposition 2.4 gives \dim B=1+r_1(K)d_n. Thus (?) yields

To prove the componentwise assertion when r_1(K)>0, put N=r_1(K)n^2-1 and reduce the relative presentation modulo 2:

By Proposition 2.4, the power series ring on the right is equidimensional of dimension r_1(K)d_n+t. Lemma 4.5 shows that every component of \operatorname{Spec}(A/2A) has dimension at least r_1(K)d_n+t-o. The minimal primes of a formal power series ring over a Noetherian ring are the extensions of the minimal primes of its coefficient ring, and adjoining N variables increases each component’s dimension by N. Thus, for every irreducible component Z of \operatorname{Spec}(R_{\bar\rho}/2R_{\bar\rho}), we have

Finally, the identity

and Corollary 4.3 give D=\delta(\bar\rho)+C(\bar\rho). If C(\bar\rho)>0, then every component of the special fiber has dimension strictly greater than \delta(\bar\rho), proving the final assertion.

The inverse image \mathcal O_0 of k under \mathcal O_F\to k_F is a finite complete local W(k)-algebra with residue field k, and \rho takes values in \operatorname{GL}_n(\mathcal O_0). Universality gives R_{\bar\rho}\to\mathcal O_0\subseteq\mathcal O_F. Let \mathfrak P be its kernel. Its image is finite over W(k) and contains W(k), so \dim(R_{\bar\rho}/\mathfrak P)=1. Thus \mathfrak P determines a closed point x of \operatorname{Spec} R_{\bar\rho}[1/2], with residue field E_x\subseteq F.

Put \mathfrak m_x=\mathfrak P R_{\bar\rho}[1/2] and

The isomorphism uses 2\notin\mathfrak P. Set

The rational representation at x becomes \rho\otimes_{\mathcal O_F}F after extending scalars to F. Real signatures and cohomology dimensions are unchanged under this finite coefficient extension. Since

by invariance of dimension under completion, Lemma 5.3 gives

For (1), Lemma 4.4 yields

Lemma 5.2 gives the stated identity and shows that its correction is positive exactly when s_v<\min(a_v,b_v) for some v\in S_{\mathbb R}.

For (2), choose a minimal prime \mathfrak q\subseteq\mathfrak P of R_{\bar\rho}. Since 2\notin\mathfrak P, we have 2\notin\mathfrak q. The complete local domain R_{\bar\rho}/\mathfrak q is catenary [Stacks, Tags 032C and 0ECF], so

Every minimal prime is associated [Stacks, Tag 05BV]. Hence

Together with Theorem 1.3, this proves the depth bound. Under the strict inequality in (2), this component has dimension smaller than \dim R_{\bar\rho}, and \operatorname{depth}R_{\bar\rho}<\dim R_{\bar\rho}. Thus R_{\bar\rho} is neither equidimensional nor Cohen–Macaulay, and hence is not a complete intersection.

Put N=|\operatorname{disc}(K)|, u=3Nm, and f(X)=X^3-4u^2X+1. Neither possible rational root \pm1 is a root, and

Thus its splitting field L/\mathbb Q is totally real with Galois group S_3. We claim that L\cap K^{\mathrm{gal}}=\mathbb Q, where K^{\mathrm{gal}} is the normal closure of K/\mathbb Q. Indeed,

The positive odd integer Q is congruent to 2 modulo 3 and is coprime to 3N. Hence some odd prime q\nmid3N occurs in Q to an odd power. It ramifies in the unique quadratic subfield \mathbb Q(\sqrt D) of L, whereas K^{\mathrm{gal}} is unramified at q. Thus this quadratic field is not contained in K^{\mathrm{gal}}, and the normal subgroups of S_3 prove the claim.

Consequently K(E[2])=KL has Galois group S_3=\operatorname{GL}_2(\mathbb F_2) over K. Its natural representation is absolutely irreducible: the two elementary transvections have different unique invariant lines over \overline{\mathbb F}_2. Good reduction outside S gives the stated representations of G_{K,S} [Silverman2009, Chapter VII, Proposition 5.1(a) and Theorem 7.1]. Since L is totally real, \bar\rho(c_v)=I_2 at every real prime. Thus s_v=0 and C(\bar\rho)=2r_1(K)>0. The Weil pairing identifies \det\rho with the 2-adic cyclotomic character [Silverman2009, Chapter III, Proposition 8.3]; hence \rho(c_v) has eigenvalues 1,-1, each once. Therefore s_v=0<\min(a_v,b_v)=1, and both asserted criteria apply. Theorem 1.3 gives \dim Z-\delta(\bar\rho)\ge C(\bar\rho)=2r_1(K) for every irreducible component Z of the special fiber.

To prove (2), we construct another lift \tau of \bar\rho, distinct from the Tate-module lift \rho. The permutation action of \operatorname{Gal}(KL/K)\simeq S_3 on

gives a totally even finite-image lift \tau of \bar\rho.

Now suppose that K is totally real. Put d=[K:\mathbb Q], L_K=KL, T=L_K^{A_3}, and V=\tau\otimes_{\mathbb Z_2}\mathbb Q_2. The decomposition \operatorname{ad} V\simeq\mathbf1\oplus\varepsilon\oplus V, where \varepsilon is the sign character, gives

Shapiro’s lemma and the definition of \lambda imply

The fields L_K and T are totally real of degrees 6d and 2d over \mathbb Q. By (?), \lambda(L_K)\le3d-1 and \lambda(T)\le d-1. Hence

and Theorem 1.5(2) proves the remaining assertions.

The subgroup of order \ell-1 in (\mathbb Z/\ell^{r+1}\mathbb Z)^\times contains -1, so F_r/\mathbb Q is totally real and cyclic of degree \ell^r. Since \ell\nmid[K:\mathbb Q], one has K\cap F_r=\mathbb Q and [K_r:K]=\ell^r. Every real prime of K splits completely in K_r, whence r_1(K_r)=\ell^r r_1(K). Put L=K(E[2]). Since [L:K]=6 and \ell\nmid6, one has L\cap K_r=K. Consequently \operatorname{Gal}(LK_r/K_r)\simeq S_3, so \bar\rho_r remains absolutely irreducible. Good reduction is preserved by base change, giving the stated representations of G_{K_r,S_r}.

At every real prime of K_r, the residual action is trivial and the Tate representation has signature (1,1). Thus s_v=0<\min(a_v,b_v)=1 and

Theorem 1.3 therefore gives the claimed discrepancy for every irreducible component of the special fiber. Theorem 1.5(1) independently gives the same lower bound for its total dimension.

The totally even augmentation-lattice lift restricts to K_r. If K is totally real, so is K_r, and the preceding S_3 calculation with Waldschmidt’s bound applies over K_r. This proves the remaining assertions.

Put \mathcal O=W(\mathbb F_4)=\mathbb Z_2[\zeta_3] and a=\lfloor n/2\rfloor. We first construct two induced lifts, with signatures (a,n-a) and (n,0), whose common reduction remains absolutely irreducible over K.

Step 1: the inducing character. Choose primes \ell\equiv1\pmod{2n} and q\equiv1\pmod{3\ell}, both unramified in the normal closure K^{\mathrm{gal}} of K/\mathbb Q; Dirichlet’s theorem permits these choices [Neukirch1999, Chapter VII, Theorem 5.14]. Let L be the degree-n subfield of \mathbb Q(\zeta_\ell). Then L/\mathbb Q is cyclic, totally real, and totally ramified at \ell, so L\cap K^{\mathrm{gal}}=\mathbb Q. Moreover q\mathcal O_L=\mathfrak q_1\cdots\mathfrak q_n splits completely.

We need a cubic character of G_L ramified at \mathfrak q_1 and unramified at the other primes above q. Use Shanks’s polynomial [Shanks1974, p. 1138, equations (3)–(4)]

Choose an integer r whose reduction is a nontrivial cube root of unity modulo q. Since f_{3r}(X)\equiv(X-r)^3\pmod q and the coefficient of t in f_t(r) is -(r^2+r)\equiv1\pmod q, there is an integer t_0\equiv3r\pmod q with f_{t_0}(r)\equiv q\pmod{q^2}. By the Chinese remainder theorem, choose t\in\mathcal O_L with

Then f_t(X+r) is Eisenstein at \mathfrak q_1, whereas its discriminant is a unit at every \mathfrak q_i with i>1. Thus f_t is irreducible with square discriminant, so a root generates a cyclic cubic extension M/L. A faithful character \psi:G_L\to\operatorname{Gal}(M/L)\simeq\mu_3\subset\mathcal O^\times has the required ramification, as does its reduction \bar\psi.

Step 2: the two lifts and absolute irreducibility. Choose \alpha\in L^\times positive at exactly a real embeddings by weak approximation, and let \eta be the quadratic character of L(\sqrt\alpha)/L. Using one inducing basis, put

Both have finite image, consisting of monomial matrices with nonzero entries in \mu_6, and have the same reduction \bar\rho_{K,n}. Since L is totally real and \psi has odd order, their signatures at every real prime of K are (a,n-a) and (n,0), respectively; in particular \bar\rho_{K,n}(c_v)=I_n.

It remains to check absolute irreducibility. Since L\cap K^{\mathrm{gal}}=\mathbb Q, Mackey’s formula gives

The completions of KL above q are unramified over \mathbb Q_q. Hence the totally ramified cubic extension at \mathfrak q_1 remains totally ramified of degree three after base change. The n conjugates of \bar\psi|_{G_{KL}} therefore remain ramified above mutually distinct primes \mathfrak q_i and unramified above the others, so they are pairwise distinct. Their joint image has odd order. Over \overline{\mathbb F}_2, every invariant subspace for this image is consequently a sum of its distinct character lines, and G_K permutes these lines transitively. Thus \bar\rho_{K,n} is absolutely irreducible, also for even n.

Let S_{K,n} be the finite set consisting of the primes above 2, all archimedean primes, and all primes ramified in the splitting fields of these two finite-image representations. Both lifts factor through G_{K,S_{K,n}}, and

Theorem 1.3 proves the first assertion.

Step 3: dimension comparison over a totally real field. Suppose that K is totally real, and put d=[K:\mathbb Q], b=d\lfloor n^2/2\rfloor, and R=R_{\bar\rho_{K,n}}. Set F=\mathbb Q_2(\zeta_3) and V=(\operatorname{Ind}_{G_L}^{G_\mathbb Q}\psi)\otimes_{\mathcal O}F. Write \operatorname{ad}^0V=D^0\oplus W, where D^0 consists of the traceless diagonal matrices and W of the matrices with zero diagonal. Both summands are stable under conjugation by the monomial representation V. Since L/\mathbb Q is cyclic and totally real, after a finite extension of F the representation D^0 splits into one-dimensional totally even characters. Proposition 6.1 gives h_\mathbb Q^2(\chi)\le1/2, hence h_\mathbb Q^2(\chi)=0, for each such character. The same proposition gives h_\mathbb Q^2(W)\le n(n-1)/2. Thus (6.1) yields

Since D_{\mathrm{lift}}(\tau)=1 and h_K^2(\operatorname{ad} V|_{G_K})=\lambda(K)+e_0, (?) gives a minimal prime \mathfrak q of R not containing 2 such that \dim(R/\mathfrak q)\le2+\lambda(K)+e_0.

For the lower bound, let x be the characteristic-zero closed point defined by \rho_{K,n}, with E_x, V_x, and \mathfrak m_x as in Lemma 5.3. Since n is invertible in E_x, taking the unique nth root of the determinant ratio congruent to one splits the deformation functor into fixed-determinant deformations and character deformations. By logarithms and (?), the latter factor is formally smooth of dimension 1+\lambda(K), so

where R_{V_x}^{\det} is the universal deformation ring of V_x over E_x with determinant fixed to \det V_x. The fixed-determinant presentation, as in the proof of Lemma 5.3, and Lemma 5.1 give

using absolute irreducibility and the real signatures (a,n-a). Catenarity, as in the proof of Theorem 1.5, gives

Consequently,

Thus R is not equidimensional, not Cohen–Macaulay, and not a complete intersection.

The matrix N_v=\bar\rho_v(c_v)-I_n satisfies N_v^2=0. As a k[T]-module, with T acting by N_v, one has

The space of homomorphisms from k[T]/(T^i) to k[T]/(T^j) has dimension \min(i,j). The endomorphism algebra is the centralizer of \bar\rho_v(c_v), and hence its dimension is

Modulo 2, the substitution X=I_n+N changes X^2=I_n into N^2=0. Over an algebraic closure of k, the square-zero matrices of rank q, for 0\le q\le m:=\lfloor n/2\rfloor, form one conjugacy class, of Jordan type (2^q,1^{n-2q}). Its stabilizer is the open set of units in its centralizer algebra. The calculation in Lemma 2.1 therefore gives its dimension as

Every such class lies in the closure of the rank-m class: multiply the upper-right entries of m-q size-two blocks by a parameter and specialize it to zero, then conjugate. Thus the variety of square-zero matrices is irreducible of dimension 2m(n-m)=d_n. Since k is perfect, the reduced coordinate ring

is geometrically integral of dimension d_n.

Since T is a finite-type domain over k, its localization at the maximal ideal induced by J_v has dimension d_n [Stacks, Tag 02IJ]. Reduction modulo 2 commutes with Noetherian local completion. Removing nilpotents and completing preserve Krull dimension [Stacks, Tag 07NU], which proves the assertion.

In a Jordan basis, each size-two block lifts to

The matrix H^{\oplus s_v}\oplus I_{a-s_v}\oplus(-I_{b-s_v}) has the required reduction and eigenvalue multiplicities. Lift the change of basis to \operatorname{GL}_n(W(k)) and conjugate back; the resulting matrix defines \rho_v(c_v).

We first prove the assertion for R_v^{\square}. Let a=\lfloor n/2\rfloor and b=\lceil n/2\rceil. Over F, the conjugacy class of \operatorname{diag}(I_a,-I_b) is closed and geometrically integral, of dimension

Indeed, every involution is semisimple, and its trace determines the two multiplicities. Let \mathfrak p\subset A_n be the contraction of the prime ideal defining this conjugacy class in A_n[1/2]. Then A_n/\mathfrak p is a domain containing W(k), and its fraction field has transcendence degree d_n over F. Evaluation at the lift in Lemma 2.3 kills \mathfrak p, so \mathfrak p\subset J_v. The dimension formula over the universally catenary discrete valuation ring W(k) gives

[Stacks, Tag 02IJ]. The surjection

therefore gives \dim R_v^{\square}\ge1+d_n. Conversely, the principal ideal dimension inequality and Lemma 2.2 give

We now prove the assertions for B. They are immediate if S_{\mathbb R}=\varnothing. Otherwise, form the ordinary tensor product

Here each factor is evaluated at its own \bar\rho_v(c_v). The completed tensor product satisfies

Indeed, both rings represent tuples of framed lifts of the \bar\rho_v. In the notation of the proof of Lemma 2.2,

This is a geometrically integral k-algebra of dimension r_1(K)d_n. Localizing at the maximal ideal induced by J and completing, as in Lemma 2.2, gives \dim(B/2B)=r_1(K)d_n. The same calculation also proves equidimensionality after completion. Indeed, the localization of (A_{S_{\mathbb R}}/2A_{S_{\mathbb R}})_{\mathrm{red}} at the maximal ideal induced by J is a local domain essentially of finite type over k, hence is universally catenary. By Ratliff’s theorem, its completion is equidimensional [Stacks, Tags 0AW2 and 0AW6]. The natural surjection from B/2B to this completion has nilpotent kernel, since completion is exact on finite modules over a Noetherian local ring. It therefore induces a homeomorphism on spectra.

For the integral dimension, take the product over F of the conjugacy classes used in the single-prime argument. It is geometrically integral of dimension r_1(K)d_n. Let \mathfrak P\subset A_{S_{\mathbb R}} be the contraction of its defining prime ideal in A_{S_{\mathbb R}}[1/2]. The lifts from Lemma 2.3, chosen for every v, give an evaluation map A_{S_{\mathbb R}}\to W(k) that kills \mathfrak P and reduces to the map defining J. Thus \mathfrak P\subset J. The same dimension formula gives

Its completion is a quotient of B, so \dim B\ge1+r_1(K)d_n. The reverse inequality follows from \dim B\le1+\dim(B/2B).

Choose elements of \mathfrak m_A lifting a basis of the relative cotangent space. Sending the X_i to these elements defines a map B[[X_1,\ldots,X_t]]\to A, which is surjective by Nakayama’s lemma and completeness.

Absolute irreducibility implies that the matrices \bar\rho(g) span M_n(k), by Burnside’s theorem after extension to an algebraic closure. Choose n^2 such matrices forming a basis. For every lift over C, their lifts form a C-basis of M_n(C) by Nakayama’s lemma; hence its centralizer consists of scalars. Fix a representative of the universal deformation over R_{\bar\rho}. The choices of frames are then r_1(K) matrices in I_n+M_n(\mathfrak m_C) modulo simultaneous multiplication by 1+\mathfrak m_C. Normalizing the (1,1)-entry of one frame to 1 removes this scalar uniquely and leaves r_1(K)n^2-1 free parameters. This proves (3.2) and gives B\to A.

The induced tangent map is

Varying the frames supplies all local coboundaries, so

Since H^0(G_{K,S},\operatorname{ad}\bar\rho)=k and S_{\mathbb R}\ne\varnothing, \operatorname{res}^0 is injective. Hence \dim_k\ker\theta=t, and Lemma 3.1 gives a minimal presentation

We bound its relations by following the proof of [Boeckle2013, Key Lemma 5.2.2, Section 5.6], using \operatorname{ad}\bar\rho in place of the trace-zero adjoint module. Put J=\ker\varphi, let \mathfrak m_P be the maximal ideal of P, and set E=J/\mathfrak m_PJ. Nakayama’s lemma gives \mu(J)=\dim_k E. Consider the square-zero extension

For u\in E^\vee=\operatorname{Hom}_k(E,k), pushing out along u gives an extension of complete local W(k)-algebras

Choose a representative (\rho_A,(\gamma_{v,A})_v) of the universal S_{\mathbb R}-framed deformation over A and a continuous set-theoretic lift \widetilde\rho_u:G_{K,S}\to\operatorname{GL}_n(P_u). Such a lift exists: the surjection \operatorname{GL}_n(P_u)\to\operatorname{GL}_n(A) has finite kernel and admits a continuous set-theoretic section, obtained by choosing coset representatives for an open subgroup on which it is injective. Its multiplicative defect defines a continuous cocycle

The class [z_u] is independent of the lift and vanishes precisely when \rho_A lifts to a representation over P_u. These classes are contractions of the obstruction class with coefficients in E, so u\mapsto[z_u] is k-linear. The map B\to P_u supplies local framed lifts; conjugating by lifts of the \gamma_{v,A} gives lifts of \rho_A|_{G_v}. Thus every local restriction of [z_u] vanishes, and we obtain

It remains to bound \ker\alpha. Put

There is a natural surjection \beta:E\twoheadrightarrow K_0, induced by J\subset P. Since the images of the X_i form a basis of the relative cotangent space, the decomposition \operatorname{ct}(P)=\operatorname{ct}(B)\oplus k^t identifies

For 0\ne u\in\ker\alpha, lift \rho_A and all its frames to P_u. Universality gives a W(k)-algebra section s:A\to P_u of \varphi_u. The decomposition P_u=s(A)\oplus I_u, together with \mathfrak m_{P_u}I_u=0, implies

Hence I_u injects into \operatorname{ct}(P_u). If e\in\ker\beta, its image in \operatorname{ct}(P_u) is zero, while its image in I_u\simeq k is u(e). Thus u(e)=0. Consequently u factors uniquely through K_0, giving a linear injection \ker\alpha\hookrightarrow K_0^\vee. It follows that

Finally, \dim P=\dim B+t, and quotienting by at most o elements gives \dim A\ge\dim B+t-o.

It follows from Tate’s global Euler characteristic formula in [Milne2006, Chapter I, Theorem 5.1].

It follows from [Milne2006, Chapter I, Corollary 4.16].

Absolute irreducibility gives h^0=1. Lemma 2.1 gives

Substituting these values into Lemma 4.1 proves the formula.

Every chain of prime ideals of C avoiding t can be extended by appending \mathfrak m, which proves the first inequality in the display. For the other inequality, put d=\dim C/tC and lift a system of d parameters of C/tC to y_1,\ldots,y_d\in C. The ideal (t,y_1,\ldots,y_d) is \mathfrak m-primary. Krull’s height theorem gives \dim C\le d+1.

For any prime \mathfrak p minimal over I, choose a minimal prime \mathfrak q of C contained in \mathfrak p. Then \mathfrak p/\mathfrak q is minimal over (I+\mathfrak q)/\mathfrak q, so its height is at most o. Since C/\mathfrak q is a catenary local domain of dimension e,

It follows from [Kisin2003, Lemma 9.7].

The matrices \rho(c_v)-I_n and \rho(c_v)+I_n have ranks b_v and a_v over F, respectively. Their reductions over k_F are equal and have rank s_v, since rank is preserved by k\subseteq k_F. Specialization gives both inequalities. Lemma 2.1 and diagonalization over F give

which is (5.2). Combining this identity with Corollary 4.3 gives the asserted formula for D_{\mathrm{lift}}(\rho). Finally, n=a_v+b_v implies

and summation proves (5.2).

The assertions about E_x, \rho_x, and R_{\bar\rho}/\mathfrak P follow from [deJong1995, Lemma 7.1.9 and its proof, pp. 78–79]. The representation V_x is absolutely irreducible, since \rho_x has absolutely irreducible reduction. By [Kisin2003, Proposition 9.5], \widehat{(R_{\bar\rho}[1/2])_{\mathfrak m_x}} is its universal deformation ring over E_x. Write h^i=\dim_{E_x}H^i(G_{K,S},\operatorname{ad} V_x). Lemma 5.1 gives h^1-h^2=D_{\mathrm{lift}}(\rho_x). The presentation in the proof of [Kisin2003, Corollary 9.8] has h^1 variables and at most h^2=e_x relations, so

Let V=F^n with its given Galois action, let N/K be the extension cut out by \rho, and put \Delta=\operatorname{Gal}(N/K). Let S_N be the set of primes above S, and let \mathcal D_N be the logarithmic kernel in (?), with N in place of K. Write (-)^\vee for the linear dual. Poitou–Tate duality [Milne2006, Chapter I, Theorem 4.10], after passing to rational coefficients, and Kummer theory give a \Delta-equivariant isomorphism

Indeed, the local H^2 groups with coefficients in \mathbb Q_2 vanish, while the localization kernel on S_N-units is \mathcal D_N: local valuations force its elements to be ordinary units, and the remaining conditions are precisely the vanishing of their 2-adic logarithms. Hochschild–Serre therefore yields

After extending scalars to \overline{\mathbb Q}_2, the dimension of the right-hand side is the Leopoldt defect \delta_K^L(V\otimes_F\overline{\mathbb Q}_2) of [Maksoud2023, (3.10)]. Since K is totally real and V is totally even,

Thus [Maksoud2023, Corollary 4.2(1)] gives

The displayed description by \operatorname{Hom} also proves independence of S and invariance under finite extension of F.

For (6.1), we write d=[K:\mathbb Q] and choose an auxiliary finite set of rational primes containing 2, \infty, and the primes where K/\mathbb Q or \eta ramifies. Write \operatorname{Ind}_{G_K}^{G_\mathbb Q}\mathbf1=\mathbf1\oplus P_0. The representation P_0 is totally even of dimension d-1. The projection formula and Shapiro’s lemma give

The bound just proved, applied over \mathbb Q to \eta\otimes_F P_0, gives the assertion for d>1; the case d=1 is immediate. Independence of S permits the auxiliary enlargement of the ramification sets.

By Theorem 1.3 and Lemma 4.4, we have 4\le\dim R_{\bar\rho}\le1+\dim(R_{\bar\rho}/2R_{\bar\rho})=4. Choose a minimal prime \mathfrak q with \dim(R_{\bar\rho}/\mathfrak q)=4. Since \dim(R_{\bar\rho}/2R_{\bar\rho})=3, we have 2\notin\mathfrak q and hence (R_{\bar\rho}/\mathfrak q)[1/2]\ne0. Choose a closed point of \operatorname{Spec} R_{\bar\rho}[1/2] on this component, corresponding to a prime \mathfrak P\supseteq\mathfrak q of R_{\bar\rho}. By Lemma 5.3, we have \dim(R_{\bar\rho}/\mathfrak P)=1. By catenarity of the complete local domain R_{\bar\rho}/\mathfrak q [Stacks, Tags 032C and 0ECF],

Thus \dim R_{\bar\rho}[1/2]\ge3, while Lemma 4.4 gives the reverse inequality.