(* Title: Exact logarithmic calibration of finite attention probes Author: Carles Marín (with Codex as formalization assistant) Copyright (c) 2026 Carles Marín. License: Apache-2.0 *)theory TAF_Probe imports TAF_Finite_Weightsbegin definition probe_null :: "'r set \<Rightarrow> ('k \<Rightarrow> 'r \<Rightarrow> real) \<Rightarrow> ('k \<Rightarrow> 'r \<Rightarrow> nat) \<Rightarrow> 'k \<Rightarrow> real" where "probe_null R a n k = (\<Sum>r\<in>R. a k r / real (n k r))" definition log_slope :: "'a set \<Rightarrow> ('a \<Rightarrow> real) \<Rightarrow> ('a \<Rightarrow> real) \<Rightarrow> ('a \<Rightarrow> real) \<Rightarrow> real" where "log_slope S w x A = - (\<Sum>i\<in>S. w i*(x i-fw_mean S w x)*ln (A i)) / fw_var S w x" lemma probe_uniform_average: assumes "\<And>r. r\<in>R \<Longrightarrow> A k r = 1 / real (n k r)" shows "(\<Sum>r\<in>R. a k r * A k r) = probe_null R a n k" using assms by (simp add: probe_null_def) lemma fixed_probe_null: assumes a: "\<And>k r. k\<in>K \<Longrightarrow> r\<in>R \<Longrightarrow> a k r = b r" and n: "\<And>k r. k\<in>K \<Longrightarrow> r\<in>R \<Longrightarrow> n k r = m r" and k: "k\<in>K" shows "probe_null R a n k = (\<Sum>r\<in>R. b r / real (m r))" unfolding probe_null_def using a n k by (intro sum.cong) auto lemma log_slope_ratio: assumes pos: "\<And>i. i\<in>S \<Longrightarrow> 0 < A i" and pos0: "\<And>i. i\<in>S \<Longrightarrow> 0 < A0 i" shows "log_slope S w x (\<lambda>i. A i/A0 i) = log_slope S w x A - log_slope S w x A0"proof - have logs: "ln (A i/A0 i) = ln (A i)-ln (A0 i)" if "i\<in>S" for i using pos[OF that] pos0[OF that] by (simp add: ln_div) have sums: "(\<Sum>i\<in>S. w i*(x i-fw_mean S w x)*ln (A i/A0 i)) = (\<Sum>i\<in>S. w i*(x i-fw_mean S w x)*ln (A i)) - (\<Sum>i\<in>S. w i*(x i-fw_mean S w x)*ln (A0 i))" proof - have point_eq: "w i*(x i-fw_mean S w x)*ln (A i/A0 i) = w i*(x i-fw_mean S w x)*ln (A i) - w i*(x i-fw_mean S w x)*ln (A0 i)" if "i\<in>S" for i by (simp only: logs[OF that] right_diff_distrib) show ?thesis using sum.cong[OF refl point_eq] by (simp only: sum_subtractf) qed show ?thesis unfolding log_slope_def sums by (simp add: diff_divide_distrib; linarith)qed context finite_probabilitybegin lemma centered_mean_zero: "(\<Sum>i\<in>S. p i*(x i-fw_mean S p x)) = 0" by (simp add: fw_mean_def algebra_simps sum_subtractf sum_distrib_right[symmetric] p_sum) lemma constant_log_slope: "log_slope S p x (\<lambda>_. c) = 0"proof - have "(\<Sum>i\<in>S. p i*(x i-fw_mean S p x)*ln c) = 0" using centered_mean_zero[of x] by (simp add: sum_distrib_right[symmetric]) then show ?thesis by (simp add: log_slope_def)qed lemma constant_profile_log_slope: assumes "\<And>i. i\<in>S \<Longrightarrow> A i = c" shows "log_slope S p x A = 0" using assms constant_log_slope[of x c] by (simp add: log_slope_def) lemma fixed_query_zero_null: assumes a: "\<And>k r. k\<in>S \<Longrightarrow> r\<in>R \<Longrightarrow> a k r = b r" and n: "\<And>k r. k\<in>S \<Longrightarrow> r\<in>R \<Longrightarrow> n k r = m r" shows "log_slope S p x (probe_null R a n) = 0" by (rule constant_profile_log_slope[where c="\<Sum>r\<in>R. b r / real (m r)"]) (rule fixed_probe_null[OF a n]) lemma covariance_centered: "fw_cov S p x y = (\<Sum>i\<in>S. p i*(x i-fw_mean S p x)*y i)" unfolding fw_cov_def fw_mean_def by (simp add: algebra_simps sum_subtractf sum_distrib_left) lemma slope_covariance: "log_slope S p x A = -fw_cov S p x (\<lambda>i. ln (A i)) / fw_var S p x" by (simp add: log_slope_def covariance_centered) end lemma variable_query_flat_counterexample: "(1/6::real) = ((1/4)+(1/12))/2" by simp end