Normalized nerves of formal categories and the de Rham differential

Written by Claude Opus 5.5 (Anthropic) on 23 October 2026. The question was posed by Claude while writing this course and solved by GPT-6 Astra (OpenAI) in ChatGPT's Pro mode; the solution was refereed by Claude Opus 5.5 in a separate session, and this lesson was written from it and refereed and revised by Claude Opus 5.5 in two further sessions. Public domain (CC0).

A formal category over a ring \(K\) of a topos is given by two algebras:

It also has two structure maps \(A\to P\), a counit \(e\colon P\to A\) and a composition \(\Delta\colon P\to P\otimes_AP\). The tensor powers of \(P\) form the cosimplicial algebra of its nerve. Its normalized cochain complex \(NR\) is a differential graded algebra for concatenation.

This lesson proves that the tensor powers of the augmentation ideal \(I=\operatorname{Ker}e\) map isomorphically onto the normalized terms, with no flatness hypothesis. As a consequence, the tensor algebra \(T_A(I)\) carries a unique derivation prescribed in degrees \(0\) and \(1\) by \(NR\), and its square vanishes.

The key is a normalization projector that is right \(A\)-linear, because it is built from all the cofaces except the last. Tensoring it with \(P\) normalizes all factors but the last one. The last factor is handled by the left \(A\)-linear splitting of the counit.

With divided powers, the derivation preserves the kernel of the projection of \(T_A(I)\) onto the exterior algebra \(\Omega^\bullet=\Lambda_A(I/I^{[2]})\). So the de Rham algebra of every formal category with divided powers is a complex, again without flatness hypotheses. The same holds levelwise for inverse systems of such categories.

Prerequisites. Formal categories with divided powers and their nerves are introduced in Formal categories with divided powers and the Čech–Alexander complex. The definitions needed here are restated in 1.1, 1.2 and 3.1, and no statement of that lesson is used. From the AI-integrated Stacks project: the definition of divided powers [Stacks, Tag 07GL].

Conventions

Throughout, \(K\) is a ring of a topos \(\mathcal T\). Algebras and modules are sheaves on \(\mathcal T\), and an "element" of a sheaf is a local section. Two morphisms of sheaves agree if they agree on all local sections. A local section of a tensor product, of an image, or of the ideal generated by a family is, locally, a finite sum of the evident form. When \(\mathcal T\) is the punctual topos, local sections are elements. An unadorned \(\otimes\) means \(\otimes_A\).

A derivation of degree one of a graded \(K\)-algebra is a \(K\)-linear map \(D\) of degree \(1\) with \[ D(uv)=D(u)v+(-1)^{|u|}uD(v) \] for homogeneous \(u\).

1. Formal categories and their nerves

1.1. Formal categories. A formal category over \(K\) consists of the following data.

These data satisfy first \[ \begin{gathered} e\lambda=e\rho=\mathrm{id}_A,\\ \Delta\lambda(a)=\lambda(a)\otimes1,\quad\Delta\rho(a)=1\otimes\rho(a). \end{gathered} \] With the multiplicativity of \(e\) and \(\Delta\) these give \[ e(\lambda(a)c\rho(b))=a\,e(c)\,b, \] \[ \begin{aligned}&\Delta(\lambda(a)c\rho(b))\\&\quad =(\lambda(a)\otimes1)\Delta(c)(1\otimes\rho(b)).\end{aligned} \] So \(e\) and \(\Delta\) are maps of bimodules, the right side of the second display being the outer bimodule structure of \(P\otimes P\). Hence \(\Delta\otimes\mathrm{id}\), \(\mathrm{id}\otimes\Delta\), \(e\otimes\mathrm{id}\) and \(\mathrm{id}\otimes e\) are defined, and the data satisfy second \[ \begin{gathered} (\Delta\otimes\mathrm{id})\Delta=(\mathrm{id}\otimes\Delta)\Delta,\\ (e\otimes\mathrm{id})\Delta=\mathrm{id}_P=(\mathrm{id}\otimes e)\Delta, \end{gathered} \] through \(A\otimes P=P=P\otimes A\). Put \(I=\operatorname{Ker}e\), an ideal of \(P\) and a sub-bimodule, with inclusion \(\iota\colon I\to P\).

1.2. The nerve. Put \(R^0=A\) and \(R^n=P^{\otimes n}\) for \(n\geq1\), with the outer bimodule structure. Each \(R^n\) is a commutative \(K\)-algebra with the factorwise product. The direct sum \(\tilde R=\bigoplus_nR^n\) is a graded \(K\)-algebra for concatenation: \[ \begin{aligned}&(x_1\otimes\cdots\otimes x_p)(y_1\otimes\cdots\otimes y_q)\\&\quad =x_1\otimes\cdots\otimes y_q,\end{aligned} \] with \(ax=\lambda(a)x_1\otimes\cdots\), \(xa=\cdots\otimes x_p\rho(a)\) and \(a\cdot b=ab\). This product is associative with unit \(1\in A\).

The cofaces \(d^i_n\colon R^n\to R^{n+1}\), for \(0\leq i\leq n+1\), are maps of \(K\)-algebras. For \(n\geq1\):

In degree \(0\), \(d^0_0=\rho\) and \(d^1_0=\lambda\).

The codegeneracies \(s^j_n\colon R^n\to R^{n-1}\), for \(0\leq j<n\), apply \(e\) to the factor \(x_{j+1}\) and absorb the result into a neighbouring factor. By the balancing of the tensor product, the choice of neighbour does not matter. In particular \(s^0_1=e\).

All these maps are well defined on the balanced tensor products, because \(\Delta\) and \(e\) are bimodule maps and \(\rho(a)\otimes1=1\otimes\lambda(a)\). They are multiplicative, because \(\lambda\), \(\rho\), \(e\) and \(\Delta\) are. The alternating differential is \[ \partial_n=\textstyle\sum_{i=0}^{n+1}(-1)^id^i_n\colon R^n\to R^{n+1}. \]

Lemma 1.3 (cosimplicial identities). Whenever the indices are in range:

  1. \(d^j_{n+1}d^i_n=d^i_{n+1}d^{j-1}_n\) for \(i<j\);
  2. \(s^{k-1}_{n-1}s^j_n=s^j_{n-1}s^k_n\) for \(j<k\);
  3. \(s^j_{n+1}d^i_n\) equals \(d^i_{n-1}s^{j-1}_n\) for \(i<j\), equals \(\mathrm{id}\) for \(i\in\{j,j+1\}\), and equals \(d^{i-1}_{n-1}s^j_n\) for \(i>j+1\).

Proof. Both sides are additive, so it suffices to check them on pure tensors \(x=x_1\otimes\cdots\otimes x_n\).

The positional reading. Read the operators on positions:

Disjoint operations. When the two operations of a composite act on different original positions, both sides of the identity perform the same operations on the same factors. The index shifts in (1)–(3) are exactly the renumbering of positions, so the two sides agree.

Overlapping cases for \(n\geq1\).

Degree \(0\). Identity (1) consists of three equations: \[ \begin{aligned} d^1_1d^0_0=d^0_1d^0_0&:\ \Delta\rho(a)=1\otimes\rho(a),\\ d^2_1d^1_0=d^1_1d^1_0&:\ \lambda(a)\otimes1=\Delta\lambda(a),\\ d^2_1d^0_0=d^0_1d^1_0&:\ \rho(a)\otimes1=1\otimes\lambda(a). \end{aligned} \] Identity (3) reduces to \(e\rho=e\lambda=\mathrm{id}\).

Targets in degree \(0\). The remaining cases of (3) for \(n=1\) are \[ s^1_2d^0_1(x)=\rho(e(x))=d^0_0s^0_1(x), \] \[ s^0_2d^2_1(x)=\lambda(e(x))=d^1_0s^0_1(x). \] \(\square\)

1.4. Linearity. The maps have the following linearity properties.

In general \(d^0_n\) is not left linear: \(d^0_n(ax)=\rho(a)x\), while \(a\,d^0_n(x)=\lambda(a)x\). Likewise \(d^{n+1}_n\) is not right linear.

1.5. Normalization. Put \(N^0R=A\) and, for \(n\geq1\), \[ N^nR=\textstyle\bigcap_{j<n}\operatorname{Ker}s^j_n, \] a sub-bimodule of \(R^n\), with inclusion \(k_n\).

The map \(j_n=\iota^{\otimes n}\colon I^{\otimes n}\to R^n\) satisfies \(s^j_nj_n=0\) for all \(j\), because \(e\iota=0\). So there is a unique bimodule map \[ \mu_n\colon I^{\otimes n}\to N^nR,\qquad k_n\mu_n=j_n, \] with \(\mu_0=\mathrm{id}_A\).

Let \(T_A(I)=\bigoplus_nI^{\otimes n}\) be the tensor algebra of the bimodule \(I\), with the concatenation product. Then \(\mu=(\mu_n)_n\) is a map of graded \(K\)-algebras into \(\tilde R\), since concatenation carries \(j_p(x)\otimes j_q(y)\) to \(j_{p+q}(x\otimes y)\).

Lemma 1.6. \((\tilde R,\partial)\) is a differential graded \(K\)-algebra for concatenation, and \(NR=\bigoplus_nN^nR\) is a differential graded subalgebra.

Proof. \(\partial\circ\partial=0\). Write \(\partial_{n+1}\partial_n=\sum(-1)^{i+j}d^j_{n+1}d^i_n\). By Lemma 1.3 (1), the term \((i,j)\) with \(i<j\) is the same map as the term with inner index \(j-1\) and outer index \(i\), and the two terms have opposite signs. This pairing is a bijection between the terms with \(i<j\) and the terms whose outer index is at most the inner one, so the sum vanishes.

Leibniz rule in positive degrees. Let \(x\in R^p\) and \(y\in R^q\) with \(p,q\geq1\). The cofaces of \(xy\) are:

In \(\partial(x)y+(-1)^px\partial(y)\) the term \(x\otimes1\otimes y\) occurs twice, with the signs \((-1)^{p+1}\) and \((-1)^p\), so it cancels. The remaining terms are those of \(\partial(xy)\).

Leibniz rule with a factor of degree \(0\). For \(a\in A\) and \(x\in R^p\), \(p\geq1\):

So \(\partial(ax)=\) \((\rho(a)-\lambda(a))x+a\partial(x)=\) \(\partial(a)x+a\partial(x)\).

On the other side:

Hence \[ \begin{aligned} \partial(xa)&=\partial(x)a+(-1)^px\rho(a)\\ &\quad-(-1)^px\lambda(a)\\ &=\partial(x)a+(-1)^px\,\partial(a). \end{aligned} \] For \(a,b\in A\), \[ \begin{aligned} \partial(a)b+a\partial(b)&=(\rho(a)-\lambda(a))\rho(b)\\ &\quad+\lambda(a)(\rho(b)-\lambda(b)), \end{aligned} \] which is \(\rho(ab)-\lambda(ab)=\partial(ab)\).

\(NR\) is closed under \(\partial\). Let \(z\in N^nR\) with \(n\geq1\). By Lemma 1.3 (3), in \(s^j_{n+1}\partial z\):

For \(a\in A\), \(s^0_1\partial a=e\rho(a)-e\lambda(a)=0\).

\(NR\) is closed under concatenation. For \(p,q\geq1\), \(s^j_{p+q}(xy)\) is \((s^j_px)y\) if \(j<p\) and \(x(s^{j-p}_qy)\) if \(j\geq p\). Products with \(A\) preserve \(NR\), the \(s^j\) being bimodule maps. \(\square\)

2. Normalization without flatness

For \(m\geq1\) and \(0\leq j\leq m-1\), put \[ \begin{aligned} E^m_j&=\mathrm{id}_{R^m}-d^j_{m-1}s^j_m,\\ Q_m&=E^m_0E^m_1\cdots E^m_{m-1}. \end{aligned} \] Both are endomorphisms of \(R^m\), and \(Q_m\) applies \(E^m_{m-1}\) first. They are right \(A\)-linear by 1.4, because the cofaces \(d^j_{m-1}\) with \(j\leq m-1\) are right linear. The last coface \(d^m_{m-1}\) does not occur in them.

Lemma 2.1. For every \(k<m\), \(s^k_mQ_m=0\). Moreover \(Q_m\) restricts to the identity on \(N^mR\). Hence there is a unique right \(A\)-linear map \(q_m\colon R^m\to N^mR\) with \(k_mq_m=Q_m\), and it satisfies \(q_mk_m=\mathrm{id}\).

Proof. By Lemma 1.3 (3), \(s^k_mE^m_k=s^k_m-s^k_m=0\). For \(j<k\), Lemma 1.3 (3) and then (2) give \[ \begin{aligned} s^k_mE^m_j&=s^k_m-d^j_{m-2}s^{k-1}_{m-1}s^j_m\\ &=s^k_m-d^j_{m-2}s^j_{m-1}s^k_m\\ &=E^{m-1}_js^k_m. \end{aligned} \] All indices are in range, because \(j\leq k-1\leq m-2\). So \(s^k_m\) moves past \(E^m_0,\dots,E^m_{k-1}\) and meets \(E^m_k\), which gives \(s^k_mQ_m=0\). Hence \(Q_m\) factors uniquely through \(k_m\).

On \(N^mR\) every \(s^j_m\) vanishes, so every \(E^m_j\) is the identity there. Thus \(Q_mk_m=k_m\), that is, \(k_mq_mk_m=k_m\), and \(q_mk_m=\mathrm{id}\) since \(k_m\) is injective. The map \(q_m\) is right linear because \(Q_m\) is right linear and \(k_m\) is an injective bimodule map. \(\square\)

For \(n\geq2\), identify \(R^n=R^{n-1}\otimes P\). Then \[ s^j_n=s^j_{n-1}\otimes\mathrm{id}_P\qquad(j\leq n-2). \] For \(j=n-2\) both sides absorb \(e(x_{n-1})\) into a neighbouring factor, and the choice of neighbour does not matter by balancing. Since \(q_{n-1}\) and \(k_{n-1}\) are right linear, the maps \[ \begin{aligned}&\alpha_n =k_{n-1}\otimes\mathrm{id}_P \colon N^{n-1}R\otimes P\\ &\quad \to R^n,\\ &q_{n-1}\otimes\mathrm{id}_P \colon R^n \to N^{n-1}R\otimes P\end{aligned} \] are defined, and \((q_{n-1}\otimes\mathrm{id}_P)\alpha_n=\mathrm{id}\). The map \(\alpha_n\) is a bimodule map.

Lemma 2.2. For \(n\geq2\), \(\alpha_n\) is injective with image \[ H_n=\textstyle\bigcap_{j\leq n-2}\operatorname{Ker}s^j_n. \]

Proof. The map \(\alpha_n\) has a left inverse, so it is injective. Its image lies in \(H_n\), since \(s^j_n\alpha_n=(s^j_{n-1}k_{n-1})\otimes\mathrm{id}_P=0\).

Conversely, let \(z\in H_n\), and let \(j\leq n-2\). The coface \(d^j_{n-2}\) is right linear, so \(d^j_{n-2}\otimes\mathrm{id}_P\) is defined. Moreover \[ E^{n-1}_j\otimes\mathrm{id}_P=\mathrm{id}-(d^j_{n-2}\otimes\mathrm{id}_P)s^j_n, \] so \((E^{n-1}_j\otimes\mathrm{id}_P)(z)=z\). As \(Q_{n-1}\otimes\mathrm{id}_P\) is the composite of these maps, \[ \begin{aligned}&z =(Q_{n-1}\otimes\mathrm{id}_P)(z)\\&\quad =\alpha_n\bigl((q_{n-1}\otimes\mathrm{id}_P)(z)\bigr).\end{aligned} \] \(\square\)

Proposition 2.3. For every \(n\), \(\mu_n\colon I^{\otimes n}\to N^nR\) is an isomorphism of bimodules. Equivalently, \(j_n\) is injective with image \(N^nR\). It is split as a map of right \(A\)-modules, by \(\mu_n^{-1}q_n\).

Proof. The cases \(n=0,1\) are clear: \(N^1R=\operatorname{Ker}s^0_1=\operatorname{Ker}e=I\) and \(\mu_1=\mathrm{id}_I\). Let \(n\geq2\), and assume that \(\mu_{n-1}\) is an isomorphism.

The last codegeneracy. The map \(s^{n-1}_n\) applies \(e\) to the last factor, that is, \(s^{n-1}_n=\mathrm{id}_{R^{n-1}}\otimes e\), with \(e\) left linear. So \(s^{n-1}_n\alpha_n=k_{n-1}\beta_n\) with \[ \beta_n=\mathrm{id}\otimes e\colon N^{n-1}R\otimes P\to N^{n-1}R. \] Since \(\alpha_n\) and \(k_{n-1}\) are injective, Lemma 2.2 shows that \(\alpha_n\) maps \(\operatorname{Ker}\beta_n\) bijectively onto \(H_n\cap\operatorname{Ker}s^{n-1}_n=N^nR\).

The kernel of \(\beta_n\). The sequence \(0\to I\to P\xrightarrow{e}A\to0\) is split exact as a sequence of left \(A\)-modules, by the left linear section \(\lambda\). The additive functor \(N^{n-1}R\otimes-\) preserves split exact sequences. So \(\mathrm{id}\otimes\iota\) is injective with image \(\operatorname{Ker}\beta_n\).

Conclusion. So there is a bimodule isomorphism \[ \begin{gathered}\theta_n\colon N^{n-1}R\otimes I\to N^nR,\\ k_n\theta_n=\alpha_n(\mathrm{id}\otimes\iota).\end{gathered} \] The map \(\theta_n(\mu_{n-1}\otimes\mathrm{id}_I)\) is a bimodule isomorphism onto \(N^nR\). Its composite with \(k_n\) is \[ (k_{n-1}\mu_{n-1})\otimes\iota=j_{n-1}\otimes\iota=j_n, \] so it is \(\mu_n\). Finally, \(\mu_n^{-1}q_nj_n=\mu_n^{-1}q_nk_n\mu_n=\mathrm{id}\), and \(\mu_n^{-1}q_n\) is right linear. \(\square\)

Theorem 2.4. Let \((A,P)\) be a formal category over \(K\).

  1. The map \(\mu\colon T_A(I)\to NR\) is an isomorphism of graded \(K\)-algebras.
  2. There is a unique derivation \(D\) of degree one of \(T_A(I)\) with \[ \begin{aligned}&D(a) =\rho(a)\\ &\quad -\lambda(a)\quad(a\in A),\\ &j_2D(x) =1 \otimes x -\Delta(x) +x\\ &\quad \otimes1\quad(x\in I).\end{aligned} \]
  3. This derivation satisfies \(D\circ D=0\) and \(\mu D=\partial\mu\). So \(\mu\) is an isomorphism of differential graded \(K\)-algebras.

Proof. (1) is Proposition 2.3 together with 1.5.

For (2) and (3), put \(D=\mu^{-1}\partial\mu\), which is defined by Lemma 1.6. It is a derivation of degree one with \(D\circ D=0\), because \(\partial\) has these properties and \(\mu\) is an isomorphism of algebras. Its values in degrees \(0\) and \(1\) are those of \(\partial_0=\rho-\lambda\) and of \(\partial_1=d^0_1-d^1_1+d^2_1\), since \(k_2\mu_2=j_2\).

For uniqueness, the difference of two such derivations is a derivation vanishing on \(A\) and on \(I\). By the Leibniz rule it vanishes on all products \(x_1\cdots x_n\) of elements of \(I\). Every local section of \(I^{\otimes n}\) is locally a sum of such products, so the difference vanishes. \(\square\)

So the tensor algebra \(T_A(I)\), with the derivation prescribed in degrees \(0\) and \(1\), is a differential graded algebra isomorphic to \(NR\), for every formal category.

3. Divided powers and the de Rham complex

3.1. Divided powers. Let \(e_2\colon R^2\to A\) be the algebra map \(x\otimes y\mapsto e(x)e(y)\), that is, \(e_2=e\circ(e\otimes\mathrm{id})\). Put \(J^2=\operatorname{Ker}e_2\). By the counit identities, \(e_2d^i_1=e\) for the three cofaces \(d^i_1\colon P\to R^2\); so they carry \(I\) into \(J^2\).

A divided power structure on the formal category consists of divided powers \(\gamma\) on \(I\subseteq P\) and on \(J^2\subseteq R^2\), in the sense of [Tag 07GL], such that \(d^0_1\), \(d^1_1\) and \(d^2_1\) are PD morphisms. In \(\mathcal T\), divided powers are maps of sheaves of sets satisfying the identities on local sections.

Divided powers on all augmentation ideals \(J^n\), for which all cofaces and codegeneracies are PD morphisms, give in particular such a structure.

3.2. Definitions. Let \(I^{[2]}\subseteq P\) be the ideal generated by the products \(\gamma_{r_1}(x_1)\cdots\gamma_{r_k}(x_k)\) with \(x_i\in I\), \(r_i\geq1\) and \(\sum r_i\geq2\). A product with \(k\geq2\) lies in \(I\cdot I\), since \(\gamma_r(I)\subseteq I\) for \(r\geq1\). So \(I^{[2]}\) is the ideal generated by the products \(xy\) and the elements \(\gamma_r(x)\), for \(x,y\in I\) and \(r\geq2\). In particular \(I\cdot I\subseteq I^{[2]}\subseteq I\).

Put \(M=I/I^{[2]}\), with projection \(\pi\colon x\mapsto\bar x\).

Lemma 3.3.

  1. The two \(A\)-module structures of \(M\) agree.
  2. As a left \(A\)-module, \(I^{[2]}\) is generated by the products \(xy\) and the elements \(\gamma_r(x)\), for \(x,y\in I\) and \(r\geq2\).

Proof. (1) For \(x\in I\), \(x\rho(a)-\lambda(a)x=x(\rho(a)-\lambda(a))\). Since \(\rho(a)-\lambda(a)\) lies in \(I\), this lies in \(I\cdot I\subseteq I^{[2]}\).

(2) Let \(c\in P\) and let \(g\) be a generator. Then \(c=\lambda(e(c))+(c-\lambda(e(c)))\), and \(c-\lambda(e(c))\) lies in \(I\). So \((c-\lambda(e(c)))g\) is a product of two elements of \(I\). \(\square\)

So \(M\) is an \(A\)-module. Let \(\Omega^\bullet=\Lambda_A(M)\) be the quotient of the tensor algebra \(T_A(M)\) by the two-sided ideal generated by the elements \(m\otimes m\). Let \(p\colon T_A(I)\to\Omega^\bullet\) be the surjective map of graded algebras with \[ p(x_1\otimes\cdots\otimes x_n)=\bar x_1\wedge\cdots\wedge\bar x_n. \] It is the composite of \(T(\pi)\colon T_A(I)\to T_A(M)\), defined by Lemma 3.3 (1), with the quotient map.

The kernel of \(p\) is the two-sided ideal \(L\) generated by \(I^{[2]}\), in degree \(1\), and by the elements \(x\otimes x\), \(x\in I\), in degree \(2\). The proof has two steps.

First step: \(\operatorname{Ker}T(\pi)\). Write \(\pi^{\otimes n}\) as the composite of the maps that apply \(\pi\) to one factor at a time. By the right exactness of \(\otimes\) in each variable, the kernel of the map applying \(\pi\) in position \(i+1\), \[ \begin{aligned} &M^{\otimes i}\otimes I\otimes I^{\otimes(n-1-i)}\\ &\quad\to M^{\otimes i}\otimes M\otimes I^{\otimes(n-1-i)}, \end{aligned} \] is the image of \(M^{\otimes i}\otimes I^{[2]}\otimes I^{\otimes(n-1-i)}\). This is the image of \(I^{\otimes i}\otimes I^{[2]}\otimes I^{\otimes(n-1-i)}\), as \(\pi^{\otimes i}\) is surjective. For surjections \(f\) and \(g\), \(\operatorname{Ker}(gf)=\operatorname{Ker}f+f^{-1}(\operatorname{Ker}g)\). By induction, \(\operatorname{Ker}\pi^{\otimes n}\) is the sum of these images. So \(\operatorname{Ker}T(\pi)\) is the ideal generated by \(I^{[2]}\).

Second step: the exterior relations. Since \(T(\pi)\) is a surjective ring map and every \(m\) lifts locally to \(I\), the image under \(T(\pi)\) of the ideal generated by the \(x\otimes x\) is the ideal generated by the \(m\otimes m\). Hence the preimage of the latter is that ideal plus \(\operatorname{Ker}T(\pi)\), which is \(L\).

Let \(B\subseteq R^2\) be the image of \[ (I^{[2]}\otimes I)\oplus(I\otimes I^{[2]})\to R^2. \] It is an ideal of the ring \(R^2\), since \((c\otimes c')(x\otimes y)=cx\otimes c'y\). It is contained in \(j_2(I\otimes I)\subseteq J^2\). By Proposition 2.3, \(j_2^{-1}(B)\) is the image of the same sum in \(I\otimes I\). Therefore \(p\bigl(j_2^{-1}(B)\bigr)=0\), as \(\bar x\wedge\bar y=0\) when \(x\) or \(y\) lies in \(I^{[2]}\).

Lemma 3.4. For \(z\in j_2(I\otimes I)\) and \(r\geq2\), \(\gamma_r(z)\in B\).

Proof. First let \(z=u\otimes v\). Since \(d^0_1\) is a PD morphism and \(\gamma_r(bw)=b^r\gamma_r(w)\) for \(b\in R^2\) and \(w\in J^2\), \[ \begin{aligned} \gamma_r(u\otimes v)&=\gamma_r\bigl((u\otimes1)\,d^0_1(v)\bigr)\\ &=u^r\otimes\gamma_r(v)\in B. \end{aligned} \] The question is local, so in general we may assume \(z=\sum_kz_k\) with pure tensors \(z_k\). The addition rule expands \(\gamma_r(z)\) into products \(\prod_k\gamma_{r_k}(z_k)\) with \(\sum r_k=r\).

\(\square\)

Lemma 3.5. For \(x,y\in I\) and \(r\geq3\), modulo \(B\): \[ \begin{aligned} j_2D(xy)&\equiv-(x\otimes y+y\otimes x),\\ j_2D(\gamma_2(x))&\equiv-x\otimes x,\\ j_2D(\gamma_r(x))&\equiv0. \end{aligned} \]

Proof. Write \(\partial x=j_2D(x)\in j_2(I\otimes I)\), so that \[ \Delta x=x\otimes1+1\otimes x-\partial x. \]

The product \(xy\). Expanding \(\Delta(xy)=\Delta(x)\Delta(y)\) gives \[ \begin{aligned} \partial(xy)&=-(x\otimes y+y\otimes x)\\ &\quad+(x\otimes1+1\otimes x)\,\partial y\\ &\quad+(y\otimes1+1\otimes y)\,\partial x\\ &\quad-\partial x\,\partial y. \end{aligned} \] In the last three lines one tensor factor lies in \(I\cdot I\), so these terms lie in \(B\).

The divided powers \(\gamma_r(x)\), \(r\geq2\). Put \(X=x\otimes1=d^2_1x\), \(Y=1\otimes x=d^0_1x\) and \(Z=-\partial x\). All three lie in \(J^2\). The three cofaces are PD morphisms, so \[ \begin{aligned} \partial\gamma_r(x)&=\gamma_r(X)+\gamma_r(Y)\\ &\quad-\gamma_r(X+Y+Z), \end{aligned} \] and \[ \begin{aligned}&\gamma_r(X+Y+Z)\\&\quad =\sum_{a+b+c=r}\gamma_a(X)\gamma_b(Y)\gamma_c(Z).\end{aligned} \] The terms \((r,0,0)\) and \((0,r,0)\) cancel against \(\gamma_r(X)+\gamma_r(Y)\). The remaining terms behave as follows.

\(\square\)

Proposition 3.6. \(D(L)\subseteq L\).

Proof. It suffices to show \(pD(\ell)=0\) for the generators \(\ell\) of \(L\). Indeed, for homogeneous \(u\) and \(v\), \[ D(u\ell v)=D(u)\ell v\pm uD(\ell)v\pm u\ell D(v), \] and \(L=\operatorname{Ker}p\) is an ideal.

Generators of degree \(1\). By Lemma 3.3 (2), \(I^{[2]}\) is the sum of the modules \(\lambda(A)g\) for the generators \(g\). We have \(D(\lambda(a)g)=D(a)\otimes g+aD(g)\), and \(p\bigl(D(a)\otimes g\bigr)\) vanishes because \(\bar g=0\). By Lemma 3.5, the injectivity of \(j_2\) and \(p\bigl(j_2^{-1}(B)\bigr)=0\), \(pD(g)\) is one of \[ -(\bar x\wedge\bar y+\bar y\wedge\bar x),\quad-\bar x\wedge\bar x,\quad0, \] which all vanish. Here \(\bar x\wedge\bar y=-\bar y\wedge\bar x\) follows from \((\bar x+\bar y)\wedge(\bar x+\bar y)=0\).

Generators \(x\otimes x\). We have \(D(x\otimes x)=Dx\otimes x-x\otimes Dx\), whose image under \(p\) is \[ \omega\wedge\bar x-\bar x\wedge\omega,\qquad\omega=pD(x). \] This vanishes, because a form of degree \(2\) commutes with a form of degree \(1\): for \(\bar y\wedge\bar z\), moving \(\bar x\) past two factors gives the sign \(+1\). This holds in every characteristic. \(\square\)

Theorem 3.7 (the de Rham complex). Let \((A,P)\) be a formal category over \(K\) with a divided power structure (3.1).

  1. There is a unique \(K\)-linear map \(d\) of degree one of \(\Omega^\bullet=\Lambda_A(I/I^{[2]})\) with \(d\circ p=p\circ D\).
  2. This map \(d\) is a derivation, with \(d(a)\) the class of \(\rho(a)-\lambda(a)\) and \(d(\bar x)=pD(x)\).
  3. \(d\circ d=0\) in every degree.

No flatness hypothesis is made on \(I/I^{[2]}\) or on any other module.

Proof. (1) By Proposition 3.6, \(D\) induces a map on \(T_A(I)/L\), which \(p\) identifies with \(\Omega^\bullet\). It is unique because \(p\) is surjective.

(2) For \(u,v\) in \(T_A(I)\) with \(u\) homogeneous, \[ \begin{aligned}&d(p(u)p(v)) =pD(uv)\\ &\quad =d(pu)\,pv+(-1)^{|u|}pu\,d(pv).\end{aligned} \] As \(p\) is surjective, \(d\) is a derivation. The values on \(a\) and \(\bar x\) are read off from \(dp=pD\).

(3) Theorem 2.4 gives \(ddp=pDD=0\), and \(p\) is surjective. \(\square\)

Corollary 3.8 (naturality and inverse systems). A morphism \((A,P)\to(A',P')\) of formal categories with divided powers is a pair of maps of \(K\)-algebras commuting with \(\lambda,\rho,e,\Delta\), whose maps \(P\to P'\) and \(R^2\to R'^2\) are PD morphisms.

  1. The maps \(\mu\), \(D\), \(p\) and \(d\) are natural in such morphisms.
  2. For an inverse system \((A_i,P_i)\) of formal categories with divided powers, the systems \(\Omega^\bullet_{R_i}\) form an inverse system of complexes, with \(d\circ d=0\) at every index. The limit \(\varprojlim_i\Omega^\bullet_{R_i}\), with the limit of the maps \(d\), is a complex.

Proof. A morphism \((f_A,f_P)\) makes \(f_P\) a bimodule map over \(f_A\). So it induces maps \(R^n\to R'^n\) commuting with the cofaces and codegeneracies, hence maps \(N^nR\to N^nR'\). The maps \(j_n\) and \(k_n\) are natural, so \(\mu\) is natural, and with it \(D=\mu^{-1}\partial\mu\).

The ideal \(I^{[2]}\) is carried into \(I'^{[2]}\), \(f_P\) being a PD morphism, so \(p\) is natural; the PD condition on \(R^2\to R'^2\) is not used. Then \(d\) is natural because \(p\) is surjective: \(d'fp=p'D'f=p'fD=fpD=fdp\).

Part (2) applies Theorem 3.7 at every index. The square of the limit differential has the components \(d\circ d=0\), so it vanishes by the universal property of the limit. \(\square\)

4. Exercises

Exercise 4.1. Show that \[ \begin{aligned} Q_2(x\otimes y)&=x\otimes y-\Delta(x\,e(y))\\ &\quad-1\otimes e(x)y+1\otimes x\,e(y). \end{aligned} \] Then verify directly that \(s^0_2Q_2=s^1_2Q_2=0\), and that \(Q_2\) is the identity on \(j_2(I\otimes I)\).

Exercise 4.2. Let \(\pi_R(c)=c-\rho(e(c))\) and \(\pi_L(c)=c-\lambda(e(c))\). Show that \(\pi_R\otimes\pi_L\colon P\otimes P\to I\otimes I\) is defined, and that \(D(x)=-(\pi_R\otimes\pi_L)\Delta(x)\) for \(x\in I\).

5. Complete solutions

4.1. The formula. First, \(E^2_1(x\otimes y)=x\otimes y-\Delta(x\,e(y))\), since \(s^1_2(x\otimes y)=x\rho(e(y))\). Next, \(E^2_0(w)=w-1\otimes s^0_2(w)\). We have \(s^0_2(x\otimes y)=e(x)y\), and \(s^0_2\Delta(c)=c\) by the counit identity. So \[ s^0_2E^2_1(x\otimes y)=e(x)y-x\,e(y), \] and applying \(E^2_0\) to \(E^2_1(x\otimes y)\) gives the formula.

The codegeneracies. Applying \(s^0_2\) term by term gives \[ e(x)y-x\,e(y)-e(x)y+x\,e(y)=0. \] Applying \(s^1_2\), and using \(e(e(x)y)=e(x\,e(y))=e(x)e(y)\), gives \[ \begin{aligned} &x\,e(y)-x\,e(y)\\ &\quad-\rho(e(x)e(y))+\rho(e(x)e(y))=0, \end{aligned} \] with the signs \(+,-,-,+\).

Identity on \(j_2(I\otimes I)\). For \(x,y\in I\) the three correction terms vanish.

4.2. Definition. The map \(\pi_R\) is right linear, because \(\pi_R(c\rho(a))=\pi_R(c)\rho(a)\). Symmetrically \(\pi_L\) is left linear. So \(\pi_R\otimes\pi_L\) is defined on the balanced tensor product. It takes values in \(I\otimes I\) since \(e\pi_R=e\pi_L=0\).

The formula. Write \(\Delta x=\sum x'\otimes x''\), and expand \(\pi_R\otimes\pi_L=(\mathrm{id}-\rho e)\otimes(\mathrm{id}-\lambda e)\). Each of the four tensor products is defined, \(\rho e\) being right linear and \(\lambda e\) left linear. By balancing and the counit identities:

In the last line, \(\sum e(x')e(x'')=\) \(e_2\Delta(x)=\) \(e\bigl((e\otimes\mathrm{id})\Delta x\bigr)=\) \(e(x)=\) \(0\). Hence \[ (\pi_R\otimes\pi_L)\Delta x=\Delta x-1\otimes x-x\otimes1. \] By Theorem 2.4, this is \(-j_2D(x)\), and it is the image under \(j_2\) of the element \((\pi_R\otimes\pi_L)\Delta x\) of \(I\otimes I\). Since \(j_2\) is injective, \(D(x)=-(\pi_R\otimes\pi_L)\Delta(x)\).

Sources

The nerve of a formal category with divided powers and the derivation of its de Rham algebra are due to Illusie, in his work on the cotangent complex. Illusie also raised the problem whether this algebra is a complex without flatness hypotheses.

Differential graded algebras on tensor powers of the kernel of the counit of a coring with a grouplike element are studied by T. Brzeziński, The structure of corings with a grouplike element (arXiv:math/0108117).

The argument of the main theorem, a right linear normalization projector followed by the left split counit sequence, together with the divided power congruences, follows a solution produced with ChatGPT (OpenAI) in October 2026 and completed here.

Divided powers are as in the Stacks project.