Filter Algebra — The Category of Cognitive Filters¶

Jean-Paul Niko · Sole Author

1. The Five Species¶

Each filter is a smooth endomorphism \(F : \mathbb{R}^{n(e)}_{\geq 0} \to \mathbb{R}^{n(e)}_{\geq 0}\) acting on the intelligence vector \(\mathbf{I}\).

The canonical pipeline composes right to left:

2. The Category \(\mathsf{Filt}\)¶

2.1 Objects¶

The objects of \(\mathsf{Filt}\) are pipeline stages:

• Stage 0: raw capacity \(\mathbf{I}_{\text{raw}}\)
• Stage 1: ceiling-bounded
• Stage 2: developmentally shaped
• Stage 3: culturally modulated
• Stage 4: state-gated
• Stage 5: attention-projected = \(\mathbf{I}_{\text{eff}}\)

2.2 Morphisms¶

\(\text{Hom}(k, k+1) = \{F : \mathbb{R}^{n}_{\geq 0} \to \mathbb{R}^{n}_{\geq 0} \mid F \text{ is species } k+1\}\)

Composition: \(F \circ G \in \text{Hom}(j, k+1)\) whenever \(G \in \text{Hom}(j, k)\) and \(F \in \text{Hom}(k, k+1)\).

The category is skeletal (no non-trivial isomorphisms between stages) and thin in the forward direction (the canonical filter at each stage is essentially unique up to parameters).

2.3 The Monoidal Structure¶

\(\mathsf{Filt}\) is a monoidal category under composition:

• Tensor product: \(F \otimes G\) applies \(F\) and \(G\) to disjoint dimension subsets (parallel filtering)
• Unit: identity filter \(\text{Id}\)
• Associator: \((F \otimes G) \otimes H \cong F \otimes (G \otimes H)\) (trivial, direct sum decomposition)

3. Algebraic Properties of Each Species¶

3.1 Ceiling Filter \(F_{\text{ceil}}\)¶

Properties: - Diagonal (acts dimension-by-dimension) - Idempotent: \(F_{\text{ceil}}^2 = F_{\text{ceil}}\) - Non-expansive: \(\|F(\mathbf{I})\| \leq \|\mathbf{I}\|\) - The ceiling vector \(\mathbf{c}\) is species-dependent (humans: \(n=12\), current AI: \(n\) varies) - Not a linear map (min is nonlinear). Linearization: \(F_{\text{ceil}} \approx \text{diag}(\sigma(c_t - I_t))\) where \(\sigma\) is a sigmoid.

3.2 Developmental Filter \(F_{\text{dev}}\)¶

Properties: - Lower-triangular in the developmental partial order \(\prec\): dimension \(s\) only activates after prerequisite dimensions \(r \prec s\) pass threshold \(\tau_r\) - Monotonically increasing in developmental time \(t\) - Saturates: \(F_{\text{dev}} \to \text{Id}\) as \(t \to \infty\) - Non-commutative with cultural: developmental order constrains what cultural content can be absorbed. A child can't absorb abstract algebra before concrete operations.

3.3 Cultural Filter \(F_{\text{cult}}\)¶

Properties: - Full matrix — cross-type couplings. A culture that values linguistic intelligence (\(I_L\)) can suppress spatial (\(I_S\)) through resource competition. - \((M_{\text{cult}})_{st} = K_{st}^{\text{cult}}\): the cultural component of the K-matrix - Context-loadable: different \(M_{\text{cult}}\) for work, family, solitude (filter modularity) - Non-commutative with state: cultural norms about emotional expression modulate how state-gating works. Stoic culture: \(M_{\text{cult}}\) suppresses \(I_{IE}\)-dependent cross-terms.

3.4 State Filter \(F_{\text{state}}\)¶

Properties: - Diagonal (Hadamard/element-wise product) - \(\boldsymbol{\eta}\) depends on psychophysiological state vector \(\Psi\) (cortisol, sleep, arousal, etc.) - Invertible: \(F_{\text{state}}^{-1}(\mathbf{I}) = \mathbf{I} \oslash \boldsymbol{\eta}\) (element-wise division) - Only filter species that is generically invertible - Fatigue: \(\eta_t \to 0\) as sleep debt increases. Flow: \(\eta_t \to 1\) for task-relevant dimensions.

3.5 Attentional Filter \(F_{\text{att}}\)¶

Properties: - Simplex-constrained: \(\sum_t \alpha_t = 1\), \(\alpha_t \geq 0\) - Rank-deficient: projects onto a low-dimensional subspace of intelligence space - Dynamics: \(\dot{\alpha}_t = \alpha_t(u_t - \bar{u})\) (replicator equation on the attention simplex) - Fastest timescale: millisecond switching - Not invertible: once attention collapses, the unattended dimensions are gone

4. Non-Commutativity¶

4.1 Theorem: \(F_{\text{cult}} \circ F_{\text{dev}} \neq F_{\text{dev}} \circ F_{\text{cult}}\)¶

Proof: Let \(\mathbf{I} = (3, 1)\) in a 2D toy model. Let \(F_{\text{dev}}\) have threshold \(\tau_2 = 2\) (dimension 2 requires dimension 1 to exceed 2) and \(F_{\text{cult}}\) swap dimensions (\(M_{\text{cult}} = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}\)).

\(F_{\text{dev}} \circ F_{\text{cult}}(\mathbf{I}) = F_{\text{dev}}(1, 3)\). Now dim 1 has value 1, which is below \(\tau_2 = 2\), so dim 2 is gated: output \((1, 0)\).

\(F_{\text{cult}} \circ F_{\text{dev}}(\mathbf{I}) = F_{\text{cult}}(3, 3)\). Dim 1 was 3 > \(\tau_2\), so dim 2 passes. Then swap: output \((3, 3)\).

\((1, 0) \neq (3, 3)\). \(\square\)

Physical meaning: A culture that swaps what's valued (e.g., a society that prizes kinesthetic over linguistic) produces different developmental outcomes depending on whether cultural exposure happens before or after developmental gating.

4.2 Theorem: \(F_{\text{att}} \circ F_{\text{state}} \neq F_{\text{state}} \circ F_{\text{att}}\)¶

Proof: Attention is simplex-constrained. State-gating before attention changes which dimensions dominate the simplex allocation. Attention before state-gating locks in allocations that may become suboptimal after state changes. Concretely: coffee (\(F_{\text{state}}\): boost \(I_M\)) after focusing on writing (\(F_{\text{att}}\): \(\alpha_L \gg \alpha_M\)) doesn't retroactively reallocate attention. \(\square\)

4.3 The Non-Abelian Structure¶

The filter monoid \(\mathcal{F} = \langle F_{\text{ceil}}, F_{\text{dev}}, F_{\text{cult}}, F_{\text{state}}, F_{\text{att}} \rangle\) is:

• Non-abelian (§4.1, §4.2)
• Not a group (\(F_{\text{ceil}}, F_{\text{dev}}, F_{\text{att}}\) are non-invertible)
• A monoid with zero (the zero filter \(F_0 : \mathbf{I} \mapsto \mathbf{0}\) absorbs: \(F \circ F_0 = F_0\))
• Partially ordered by the pipeline order \(\text{ceil} \prec \text{dev} \prec \text{cult} \prec \text{state} \prec \text{att}\)

5. The Kernel Composition Lemma¶

5.1 Statement¶

For any composition \(G = F_k \circ \cdots \circ F_1\):

Information loss is monotonically non-decreasing through the pipeline. Each filter can only maintain or increase the kernel. No filter can recover information lost by a previous filter (except \(F_{\text{state}}^{-1}\), which only reverses its own gating).

5.2 Effective Dimension¶

Define the effective dimension of \(\mathbf{I}\) after filtering:

Then: \(d_{\text{eff}}(G(\mathbf{I})) \leq d_{\text{eff}}(F_1(\mathbf{I})) \leq n(e)\)

The pipeline is a dimension funnel. Raw capacity \(n(e) = 12\) for humans. Effective output dimension after all filters: typically 2–4 in any given moment (attention bottleneck).

6. Connection to P vs NP¶

6.1 Filter Complexity¶

Define the filter complexity of a computation as the minimum pipeline length needed:

6.2 The Verification/Generation Gap¶

Conjecture: For NP-complete problems, the filter chain for verification is polynomial:

but the chain for generation is exponential:

The gap arises because verification uses \(F_{\text{att}}\) (cheap simplex projection: check each clause), while generation requires \(F_{\text{cult}}^{-1}\) (exponential search through the non-invertible cultural filter's kernel to find the pre-image).

6.3 Non-Commutativity as Separation Witness¶

The commutator \([F_{\text{verify}}, F_{\text{generate}}] \neq 0\) witnesses the separation:

Verification after generation is polynomial (check the output). Generation after verification is meaningless (you can't reverse the funnel). The non-commutativity is essential — no rearrangement of the pipeline avoids the exponential bottleneck.

⚠ This is the weakest Millennium connection. It requires proving that the filter monoid's non-commutativity implies a complexity-theoretic lower bound, which is a substantial gap.