diff --git a/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Inversion.lean b/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Inversion.lean
index 716413e3..4a865963 100644
--- a/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Inversion.lean
+++ b/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Inversion.lean
@@ -7,6 +7,7 @@ Authors: Samuel Schlesinger
 module
 public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Inversion.Collisions
 public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Inversion.Components
+public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Inversion.Features
 public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Inversion.Parameters
 public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Multioutput.LowerBound
 
@@ -19,6 +20,8 @@ size restriction is `n ≥ 3`; the permutation, nonprojection, and affine-restri
 properties are proved for inversion itself, not assumed as hardness hypotheses.
 The entropy and output-rank refinement has leading coefficient
 `(3 + 2c)/(2 + c)`, approximately 1.543112, where `c = 1 - H₂(1/4)`.
+The actual nonlinear-feature budget further gives leading coefficient
+`(3 + 4c)/(2 + 2c)`, approximately 1.579380, with its explicit additive penalty.
 -/
 
 @[expose] public section
diff --git a/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Inversion/Features.lean b/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Inversion/Features.lean
new file mode 100644
index 00000000..29053c1a
--- /dev/null
+++ b/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Inversion/Features.lean
@@ -0,0 +1,154 @@
+/-
+Copyright (c) 2026 Samuel Schlesinger. All rights reserved.
+Released under MIT license as described in the file Complexitylib/Algebraic/LICENSE.
+Authors: Samuel Schlesinger
+-/
+
+module
+public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Inversion.Collisions
+public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Inversion.Components
+public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Multioutput.Features
+public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Multioutput.LowerBound
+
+/-!
+# The actual nonlinear-feature budget for inversion circuits
+
+Actual conjunction gate outputs generate every output modulo affine primary
+functions. Their multiple-primary subfamily has quarter-biased marginals, so the
+average-collision estimate charges these features as well as the circuit's
+primary summaries. No independence, arity, depth, or reuse restriction is needed.
+-/
+
+@[expose] public section
+
+namespace Algebraic.Aggregate.Geometry.Inversion
+
+open scoped BigOperators Classical
+open Cutwidth.Aggregate.Geometry
+
+/-- An inversion circuit's actual conjunction outputs pay an entropy charge for every
+multiple-primary conjunction. -/
+theorem input_add_feature_bias_le_conjunctionCount {K : Type*} [Field K] [Fintype K]
+    [Algebra (ZMod 2) K] {n : ℕ} (e : (Fin n → ZMod 2) ≃ₗ[ZMod 2] K)
+    (c : Circuit signature n n) (computes : c.Computes interpretation (inverseFunction e)) :
+    (n : ℝ) + multiCount c.program * Entropy.bitSaving - Real.log 5 / Real.log 2 ≤
+      conjunctionCount c.program := by
+  let : CharP K 2 := charP_of_injective_algebraMap (algebraMap (ZMod 2) K).injective 2
+  let J := {i : Fin c.size // (c.program.lines i).op.isConjunction = true}
+  let key (x : K) (j : J) := c.program.gateFunction interpretation j ((encode e).symm x)
+  choose b L w representation using output_affine_conjunction_representation c
+  let A : K →ₗ[ZMod 2] K :=
+    e.toLinearMap.comp ((LinearMap.pi L).comp e.symm.toLinearMap)
+  let a := e b
+  let decode (h : J → Bool) := e (fun j => ∑ i, w j i * bitValue (h i))
+  have represents (x : K) : x⁻¹ + A x + a = decode (key x) := by
+    have coordinates : e.symm x⁻¹ = b + (LinearMap.pi L) (e.symm x) +
+        (fun j => ∑ i, w j i * bitValue (key x i)) := by
+      funext j
+      have hj := representation j ((encode e).symm x)
+      rw [computes, bitVector_decode] at hj
+      simp only [inverseFunction, Equiv.apply_symm_apply] at hj
+      change bitEquiv ((encode e).symm x⁻¹ j) = _ at hj
+      rw [bitEquiv_decode] at hj
+      exact hj
+    have field := congrArg e coordinates
+    simp only [map_add, LinearEquiv.apply_symm_apply] at field
+    change x⁻¹ + e ((LinearMap.pi L) (e.symm x)) + e b =
+      e (fun j => ∑ i, w j i * bitValue (key x i))
+    rw [field]
+    have cancel : ∀ u v z : K, (u + v + z) + v + u = z := by
+      intro u v z
+      calc
+        (u + v + z) + v + u = (u + u) + (v + v) + z := by ac_rfl
+        _ = z := by simp only [CharTwo.add_self_eq_zero, zero_add]
+    exact cancel _ _ _
+  let B : Finset J := Finset.univ.filter fun j => multiPrimary (c.program.lines j.val) = true
+  have countJ : Fintype.card J = conjunctionCount c.program := Fintype.card_subtype _
+  have countB : B.card = multiCount c.program := by
+    rw [multiCount_eq_card_filter]
+    have image : B.image Subtype.val =
+        Finset.univ.filter fun j => multiPrimary (c.program.lines j) = true := by
+      ext j
+      constructor
+      · intro h
+        obtain ⟨i, hi, rfl⟩ := Finset.mem_image.mp h
+        exact Finset.mem_filter.mpr ⟨Finset.mem_univ _, (Finset.mem_filter.mp hi).2⟩
+      · intro h
+        have marked := (Finset.mem_filter.mp h).2
+        have conjunction := ((multiPrimary_iff_exists_pair _).mp marked).1
+        exact Finset.mem_image.mpr ⟨⟨j, conjunction⟩,
+          Finset.mem_filter.mpr ⟨Finset.mem_univ _, marked⟩, rfl⟩
+    rw [← image, Finset.card_image_of_injective _ Subtype.val_injective]
+  have quarter (j : J) : ∃ rare : Bool, multiPrimary (c.program.lines j.val) = true →
+      4 * (Finset.univ.filter fun x : K => key x j = rare).card ≤ Fintype.card K := by
+    by_cases marked : multiPrimary (c.program.lines j.val) = true
+    · obtain ⟨rare, small⟩ := multiPrimary_quarter c.program j.val marked
+      refine ⟨rare, fun _ => ?_⟩
+      have cards := Fintype.card_congr ((encode e).symm.subtypeEquivOfSubtype
+        (p := fun y => c.program.gateFunction interpretation j.val y = rare))
+      have same : (Finset.univ.filter fun x : K => key x j = rare).card =
+          Fintype.card {y : Fin n → Bool //
+            c.program.gateFunction interpretation j.val y = rare} := by
+        simpa only [Fintype.card_subtype] using cards
+      rw [same, card_field e]
+      exact small
+    · exact ⟨false, fun h => (marked h).elim⟩
+  choose rare bias using quarter
+  have bound := log_card_le_of_inverse_features_and_bias key decode A.toAddMonoidHom a
+    represents B rare (fun j hj => bias j (Finset.mem_filter.mp hj).2)
+  rw [countJ, countB, card_field e] at bound
+  simp only [Nat.cast_pow, Nat.cast_ofNat, Real.log_pow] at bound
+  have logpos : 0 < Real.log 2 := Real.log_pos (by norm_num)
+  apply (mul_le_mul_iff_left₀ logpos).mp
+  rw [sub_mul, add_mul, Entropy.bitSaving, mul_assoc,
+    div_mul_cancel₀ _ logpos.ne', div_mul_cancel₀ _ logpos.ne']
+  linarith
+
+/-- Charging actual nonlinear features and primary summaries strengthens the inversion
+bound, with leading coefficient `(3 + 4c) / (2 + 2c)` for `c = 1 - H₂(1/4)`. -/
+theorem feature_entropy_mul_size_lower_bound {K : Type*} [Field K] [Fintype K]
+    [Algebra (ZMod 2) K] {n : ℕ} (e : (Fin n → ZMod 2) ≃ₗ[ZMod 2] K)
+    (dimension : 3 ≤ n) (c : Circuit signature n n)
+    (computes : c.Computes interpretation (inverseFunction e)) :
+    (3 + 4 * Entropy.bitSaving) * n - (Real.log 5 / Real.log 2 + 8 * Entropy.bitSaving) ≤
+      (2 + 2 * Entropy.bitSaving) * c.size := by
+  have out (i : Fin n) : c.outputFunction interpretation i = fun x => inverseFunction e x i := by
+    funext x
+    exact congrFun (computes x) i
+  have same : c.eval interpretation = inverseFunction e := funext computes
+  have bijective : Function.Bijective (c.eval interpretation) := by
+    rw [same]
+    exact inverseFunction_bijective e
+  have nonliteral : ∀ i j b, c.outputFunction interpretation i ≠ fun x => b ^^ x j := by
+    intro i j b
+    rw [out]
+    exact inverseFunction_nonliteral e dimension i j b
+  have outputs := input_add_conjunctionCount_le_size_add_outputConjunctionCount c bijective
+    (by intro i j; simpa using nonliteral i j false)
+  have information := input_le_size_sub_outputConjunctionCount_sub_bias c bijective nonliteral
+  have pairing := two_mul_input_le_size_add_multi_of_affine_restrictions (r := 2) c (by
+    intro S affine
+    exact card_flat_le_four e S (fun i => by simpa only [out i] using affine i))
+  have features := input_add_feature_bias_le_conjunctionCount e c computes
+  have counts : (n : ℝ) + conjunctionCount c.program ≤
+      c.size + outputConjunctionCount c := by exact_mod_cast outputs
+  have pairs : 2 * (n : ℝ) ≤ c.size + multiCount c.program + 4 := by
+    exact_mod_cast pairing
+  have positive := Entropy.bitSaving_pos
+  nlinarith [mul_nonneg positive.le (sub_nonneg.mpr pairs)]
+
+/-- The nonlinear-feature inequality in normalized gate-count form. -/
+theorem featureCoefficient_mul_sub_penalty_le_size {K : Type*} [Field K] [Fintype K]
+    [Algebra (ZMod 2) K] {n : ℕ} (e : (Fin n → ZMod 2) ≃ₗ[ZMod 2] K)
+    (dimension : 3 ≤ n) (c : Circuit signature n n)
+    (computes : c.Computes interpretation (inverseFunction e)) :
+    (3 + 4 * Entropy.bitSaving) / (2 + 2 * Entropy.bitSaving) * n -
+        (Real.log 5 / Real.log 2 + 8 * Entropy.bitSaving) / (2 + 2 * Entropy.bitSaving) ≤
+      c.size := by
+  have bound := feature_entropy_mul_size_lower_bound e dimension c computes
+  have positive : 0 < 2 + 2 * Entropy.bitSaving := by
+    linarith [Entropy.bitSaving_pos]
+  rw [div_mul_eq_mul_div, ← sub_div]
+  exact (div_le_iff₀ positive).mpr (by simpa only [mul_comm] using bound)
+
+end Algebraic.Aggregate.Geometry.Inversion
diff --git a/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Multioutput/Features.lean b/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Multioutput/Features.lean
new file mode 100644
index 00000000..5857f51b
--- /dev/null
+++ b/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Multioutput/Features.lean
@@ -0,0 +1,36 @@
+/-
+Copyright (c) 2026 Samuel Schlesinger. All rights reserved.
+Released under MIT license as described in the file Complexitylib/Algebraic/LICENSE.
+Authors: Samuel Schlesinger
+-/
+
+module
+public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Multioutput.Features.Internal
+
+/-!
+# Explicit affine-plus-conjunction representations of circuit outputs
+
+Every output coordinate is an affine function of the primary bits plus a linear
+combination of actual conjunction gate outputs. This representation allows arbitrary
+arity, depth, and nonlinear reuse, including signed and constant gates.
+-/
+
+@[expose] public section
+
+namespace Algebraic.Aggregate.Geometry
+
+/-- Every output coordinate has an affine-plus-conjunction-feature representation,
+using actual conjunction outputs without changing their Boolean polarity. -/
+theorem output_affine_conjunction_representation {n m : ℕ} (c : Circuit signature n m)
+    (j : Fin m) :
+    ∃ (b : ZMod 2) (L : (Fin n → ZMod 2) →ₗ[ZMod 2] ZMod 2)
+      (w : {i : Fin c.size // (c.program.lines i).op.isConjunction = true} → ZMod 2),
+      ∀ x, bitValue (c.eval interpretation x j) = b + L (bitVector x) +
+        ∑ i, w i * bitValue (c.program.gateFunction interpretation i x) := by
+  obtain ⟨f, ⟨b, L, rfl⟩, g, ⟨w, rfl⟩, h⟩ :=
+    Submodule.mem_sup.mp (output_mem_affine_sup_conjunctionRange c j)
+  refine ⟨b, L, w, fun x => ?_⟩
+  simpa only [Pi.add_apply, Fintype.linearCombination_apply, Finset.sum_apply,
+    Pi.smul_apply, smul_eq_mul] using (congrFun h x).symm
+
+end Algebraic.Aggregate.Geometry
diff --git a/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Multioutput/Features/Internal.lean b/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Multioutput/Features/Internal.lean
new file mode 100644
index 00000000..5c8ca52c
--- /dev/null
+++ b/Complexitylib/Algebraic/LowerBound/Cutwidth/Aggregate/Geometry/Multioutput/Features/Internal.lean
@@ -0,0 +1,49 @@
+/-
+Copyright (c) 2026 Samuel Schlesinger. All rights reserved.
+Released under MIT license as described in the file Complexitylib/Algebraic/LICENSE.
+Authors: Samuel Schlesinger
+-/
+
+module
+public import Complexitylib.Algebraic.LowerBound.Cutwidth.Aggregate.Geometry.Multioutput.Rank.Internal
+
+/-!
+# Affine functions and actual conjunction features span circuit outputs
+
+The submodule generated by affine primary functions and actual conjunction gate
+functions contains every gate, and hence every designated output wire.
+-/
+
+@[expose] public section
+
+namespace Algebraic.Aggregate.Geometry
+
+/-- Every output is an affine primary function plus a linear combination of the
+actual conjunction gate functions. -/
+theorem output_mem_affine_sup_conjunctionRange {n m : ℕ} (c : Circuit signature n m)
+    (j : Fin m) :
+    (fun x => bitValue (c.eval interpretation x j)) ∈ affineFunctions n ⊔
+      (Fintype.linearCombination (ZMod 2)
+        (fun i : {i : Fin c.size // (c.program.lines i).op.isConjunction = true} =>
+          fun x => bitValue (c.program.gateFunction interpretation i x))).range := by
+  classical
+  let combination := Fintype.linearCombination (ZMod 2)
+    (fun i : {i : Fin c.size // (c.program.lines i).op.isConjunction = true} =>
+      fun x => bitValue (c.program.gateFunction interpretation i x))
+  let S := affineFunctions n ⊔ combination.range
+  have constants (b : ZMod 2) : (fun _ : Fin n → Bool => b) ∈ S :=
+    Submodule.mem_sup_left ⟨b, 0, by ext; simp⟩
+  have inputs (i : Fin n) : (fun x => bitValue (x i)) ∈ S :=
+    Submodule.mem_sup_left ⟨0, LinearMap.proj i, by ext; simp [bitVector]⟩
+  have gates := gateValues_mem c.program S constants inputs (by
+    intro i hi
+    apply Submodule.mem_sup_right
+    refine ⟨Pi.single ⟨i, hi⟩ 1, ?_⟩
+    simp [combination, Fintype.linearCombination_apply_single])
+  change (fun x => bitValue
+    (c.program.wireFunction interpretation (c.outputs j) x)) ∈ S
+  cases c.outputs j with
+  | input i => exact inputs i
+  | gate i => exact gates i
+
+end Algebraic.Aggregate.Geometry
diff --git a/blueprint/src/chapters/lowerbounds.tex b/blueprint/src/chapters/lowerbounds.tex
index 867ffb92..29286e8f 100644
--- a/blueprint/src/chapters/lowerbounds.tex
+++ b/blueprint/src/chapters/lowerbounds.tex
@@ -10860,6 +10860,68 @@ \section{Results from the algebraic-circuits library}
   finite penalty $4r/D_{\mathrm i}$.
 \end{proof}
 
+\begin{lemma}[Inversion's actual nonlinear-feature budget]
+  \label{lem:lowerbounds-algebraic-geometry-inversion-features}
+  \lean{Algebraic.Aggregate.Geometry.output_affine_conjunction_representation,
+    Algebraic.Aggregate.Geometry.Inversion.input_add_feature_bias_le_conjunctionCount}
+  \leanok
+  \uses{def:lowerbounds-algebraic-geometry-nonlinear-rank,
+    def:lowerbounds-algebraic-geometry-inversion}
+  Every output of a signed unbounded AND/OR/XOR circuit is an affine function
+  of the primary bits plus a linear combination of its actual conjunction
+  gate outputs. Suppose the circuit computes all coordinates of inversion
+  in any supplied binary-field basis. Let $N$ count conjunction gates and
+  let $m$ count those reading at least two distinct primary coordinates.
+  With $c=1-H_2(1/4)$,
+  \[
+    N\ge n+cm-\log_2 5.
+  \]
+  No independence, fan-in, fanout, depth, or nonlinear-reuse restriction is imposed.
+\end{lemma}
+\begin{proof}
+  \leanok
+  Apply the gate-span induction to the sum of the affine subspace and the
+  range of the linear-combination map on actual conjunction functions.
+  Assemble the output representations in the field basis to obtain
+  $I(x)+Ax+a=D(h(x))$. Inversion's classical differential-uniformity bound
+  gives at most $q+4(q-1)$ ordered collisions for this affine residual,
+  with $q=2^n$. Equality of feature vectors implies equality of residuals,
+  so their average fibre size is at most five. The weighted logarithmic
+  collision inequality charges every quarter-biased feature. Each of the
+  $m$ marked actual conjunction outputs has a rare Boolean value occurring
+  on at most one quarter of the cube, even with repeated or internal wires.
+\end{proof}
+
+\begin{theorem}[Nonlinear features strengthen binary-field inversion]
+  \label{thm:lowerbounds-algebraic-geometry-inversion-feature-bound}
+  \lean{Algebraic.Aggregate.Geometry.Inversion.feature_entropy_mul_size_lower_bound,
+    Algebraic.Aggregate.Geometry.Inversion.featureCoefficient_mul_sub_penalty_le_size}
+  \leanok
+  \uses{lem:lowerbounds-algebraic-geometry-inversion-features,
+    thm:lowerbounds-algebraic-geometry-inversion-entropy}
+  Let $K$ be a finite field over $\mathbf F_2$ with a supplied linear
+  isomorphism $e:\mathbf F_2^n\longrightarrow K$, where $n\ge3$.
+  Every signed unbounded AND/OR/XOR circuit computing all inversion
+  coordinates satisfies
+  \[
+    (2+2c)g\ge(3+4c)n-\log_2 5-8c,
+    \qquad c=1-H_2(1/4).
+  \]
+  Equivalently, its gate count is at least
+  $1.5793801642\ldots n-1.6116903280\ldots$.
+  This gives a larger leading coefficient than the majority-fibre bound;
+  both finite inequalities remain available with their different penalties.
+\end{theorem}
+\begin{proof}
+  \leanok
+  Let $o$ count conjunction output gates. Distinct designated outputs give
+  $n+N\le g+o$, while their constant primary summaries and the biased
+  internal summaries give $n\le g-o-cm$. Hence
+  $2g\ge2n+N+cm\ge3n+2cm-\log_2 5$.
+  The affine-restriction inequality $2n\le g+m+4$ eliminates $m$.
+  Divide by $2+2c>0$ for the normalized gate bound.
+\end{proof}
+
 \begin{definition}[Full Hamming residue modulo three]
   \label{def:lowerbounds-algebraic-modthree-residue}
   \lean{Algebraic.BooleanCube.ModThree.bit,