Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 2 data and assumptions A1–A5

Definition
DKFAsymptotics_Assumptions

by burkh4rt · Sep 15, 2026 · Mathlib 0df444a (Lean v4.33.1)

gaussiankalman-filterprobabilitystatistics

For a fixed stationary Gaussian model, the problem data specify two probability-density sequences, two Gaussian approximating density sequences, a point bbb, and an arbitrary probability measure PPP. The exact normalized update must be well defined at every positive index. The exact and approximate updates are both defined by the same discriminative density-update formula. The assumptions are exactly A1–A5: weak convergence of the previous-state densities to PPP, L1L^1L1 approximation by their Gaussian sequence, weak convergence of the current-observation densities to δb\delta_bδb​, L1L^1L1 approximation by their Gaussian sequence, and weak convergence of the exact updated densities to δb\delta_bδb​. The approximate update's validity, Gaussianity and convergence are conclusions, not fields in these assumptions.

Definition code
import Definitions.Def_DKFAsymptotics_Filter

noncomputable section
open MeasureTheory Filter Matrix
open scoped Topology ENNReal

namespace DKFAsymptotics
variable {d : ℕ}

structure Problem (M : Model d) where
  s : ℕ+ → Density d
  u : ℕ+ → Density d
  s' : ℕ+ → Density d
  u' : ℕ+ → Density d
  b : State d
  P : ProbabilityMeasure (State d)
  s_pdf : ∀ n, IsPDF (s n)
  u_pdf : ∀ n, IsPDF (u n)
  s'_gaussian : ∀ n, IsGaussianPDF (s' n)
  u'_gaussian : ∀ n, IsGaussianPDF (u' n)
  original_defined : ∀ n, UpdateDefined M (s n) (u n)

def Problem.p (M : Model d) (D : Problem M) (n : ℕ+) : Density d :=
  updateDensity M (D.s n) (D.u n)

def Problem.p' (M : Model d) (D : Problem M) (n : ℕ+) : Density d :=
  updateDensity M (D.s' n) (D.u' n)

structure Assumptions (M : Model d) (D : Problem M) : Prop where
  A1 : Weakly D.s D.P
  A2 : Tendsto (fun n => l1Distance (D.s n) (D.s' n)) atTop (𝓝 0)
  A3 : Weakly D.u (Measure.dirac D.b)
  A4 : Tendsto (fun n => l1Distance (D.u n) (D.u' n)) atTop (𝓝 0)
  A5 : Weakly (D.p M) (Measure.dirac D.b)

end DKFAsymptotics
Source
M. C. Burkhart, D. M. Brandman, B. Franco, L. R. Hochberg and M. T. Harrison, The Discriminative Kalman Filter for Bayesian Filtering with Nonlinear and Nongaussian Observation Models, Neural Computation 32(5), 969–1017 (2020), doi:10.1162/neco_a_01275; https://burkh4rt.github.io/pubs/Burkhart-et-al-2020-Neural-Comput.pdf; Theorem 2, Appendix B, p. 1003, equation (B.1), assumptions A1–A5 and Remark 2.
Read-back

What the Lean code literally says, in plain math · GPT-6 (Codex independent sub-agent)

Blind mathematical read-back: asymptotic problem and assumptions

DKFAsymptotics.Problem

For any natural number ddd, including 000, fix the state space E=RdE=\mathbb R^dE=Rd with its usual product topology and product Lebesgue measure λ\lambdaλ, and fix a model consisting of real d×dd\times dd×d matrices A,Γ,SA,\Gamma,SA,Γ,S, where Γ\GammaΓ and SSS are symmetric positive definite and S=ASAT+ΓS=ASA^\mathsf T+\GammaS=ASAT+Γ. Here positive definite means symmetric with xTCx>0x^\mathsf TCx>0xTCx>0 for every nonzero state vector xxx; no positive-definiteness or invertibility condition is separately placed on AAA. Write φm,C(z)=((2π)ddet⁡C)−1exp⁡(−12(z−m)TC−1(z−m))\varphi_{m,C}(z)=\bigl(\sqrt{(2\pi)^d\det C}\bigr)^{-1}\exp\bigl(-\tfrac12(z-m)^\mathsf TC^{-1}(z-m)\bigr)φm,C​(z)=((2π)ddetC​)−1exp(−21​(z−m)TC−1(z−m)), where m,z∈Em,z\in Em,z∈E and CCC is a real d×dd\times dd×d matrix. A problem for this model consists of four sequences sn,un,sn′,un′:E→Rs_n,u_n,s'_n,u'_n:E\to\mathbb Rsn​,un​,sn′​,un′​:E→R, indexed by every positive integer nnn, a vector b∈Eb\in Eb∈E, and a measure PPP on EEE of total mass one, together with the following requirements. For each nnn, both sns_nsn​ and unu_nun​ are nonnegative λ\lambdaλ-almost everywhere, Bochner integrable with respect to λ\lambdaλ, and have integral one. For each nnn, each of sn′s'_nsn′​ and un′u'_nun′​ has these same three properties and, separately for each function, there exist a mean m∈Em\in Em∈E and a symmetric positive-definite matrix CCC such that the function equals φm,C\varphi_{m,C}φm,C​ λ\lambdaλ-almost everywhere. These Gaussian witnesses may vary with nnn and with the sequence; equality and nonnegativity are not required at every point. Define the original update weight by wn(z)=un(z)(∫EφAy,Γ(z)sn(y) dλ(y))/φ0,S(z)w_n(z)=u_n(z)\bigl(\int_E\varphi_{Ay,\Gamma}(z)s_n(y)\,d\lambda(y)\bigr)/\varphi_{0,S}(z)wn​(z)=un​(z)(∫E​φAy,Γ​(z)sn​(y)dλ(y))/φ0,S​(z), and define its extended nonnegative absolute integral by Ln=∫E+∣wn(z)∣ dλ(z)∈[0,∞]L_n=\int_E^{+}|w_n(z)|\,d\lambda(z)\in[0,\infty]Ln​=∫E+​∣wn​(z)∣dλ(z)∈[0,∞]. For every nnn, the weight wnw_nwn​ must have a strongly measurable representative agreeing with it λ\lambdaλ-almost everywhere, and 0<Ln<∞0<L_n<\infty0<Ln​<∞. These requirements are imposed on the original weights only; there is no corresponding requirement here for weights formed from sn′,un′s'_n,u'_nsn′​,un′​. No convergence condition is part of this structure, and PPP is not required to have a density or any specified relationship to bbb. No index n=0n=0n=0 is supplied. Dimension zero is allowed: the state space then has one point and the positive-definiteness condition on nonzero vectors has no instances. The integrals used to form the weights are the total Bochner integrals, which return zero if their integrands are not integrable.

DKFAsymptotics.Problem.p

For any natural number ddd, let E=RdE=\mathbb R^dE=Rd with product Lebesgue measure λ\lambdaλ, let A,Γ,SA,\Gamma,SA,Γ,S be real d×dd\times dd×d matrices with Γ,S\Gamma,SΓ,S symmetric positive definite and S=ASAT+ΓS=ASA^\mathsf T+\GammaS=ASAT+Γ, and let a problem for this model supply sequences sn,un,sn′,un′:E→Rs_n,u_n,s'_n,u'_n:E\to\mathbb Rsn​,un​,sn′​,un′​:E→R for positive integers nnn, a state bbb, and a probability measure PPP. The problem requires every sequence member to be nonnegative λ\lambdaλ-almost everywhere, Bochner integrable, and of integral one; in addition, every primed member equals λ\lambdaλ-almost everywhere some φm,C\varphi_{m,C}φm,C​ with symmetric positive-definite CCC, where φm,C(z)=((2π)ddet⁡C)−1exp⁡(−12(z−m)TC−1(z−m))\varphi_{m,C}(z)=\bigl(\sqrt{(2\pi)^d\det C}\bigr)^{-1}\exp\bigl(-\tfrac12(z-m)^\mathsf TC^{-1}(z-m)\bigr)φm,C​(z)=((2π)ddetC​)−1exp(−21​(z−m)TC−1(z−m)). For each positive integer nnn, set wn(z)=un(z)(∫EφAy,Γ(z)sn(y) dλ(y))/φ0,S(z)w_n(z)=u_n(z)\bigl(\int_E\varphi_{Ay,\Gamma}(z)s_n(y)\,d\lambda(y)\bigr)/\varphi_{0,S}(z)wn​(z)=un​(z)(∫E​φAy,Γ​(z)sn​(y)dλ(y))/φ0,S​(z) and Ln=∫E+∣wn(z)∣ dλ(z)∈[0,∞]L_n=\int_E^{+}|w_n(z)|\,d\lambda(z)\in[0,\infty]Ln​=∫E+​∣wn​(z)∣dλ(z)∈[0,∞]; the problem further requires wnw_nwn​ to agree almost everywhere with a strongly measurable function and requires 0<Ln<∞0<L_n<\infty0<Ln​<∞. The declared function pn:E→Rp_n:E\to\mathbb Rpn​:E→R is the pointwise formula pn(z)=wn(z)/real⁡(Ln)p_n(z)=w_n(z)/\operatorname{real}(L_n)pn​(z)=wn​(z)/real(Ln​), where real⁡\operatorname{real}real sends a finite extended nonnegative real to the same ordinary real and sends +∞+\infty+∞ to zero. Thus the denominator in this definition is strictly positive and finite by the problem's requirement. The formula uses the absolute integral of the weight for normalization. It is defined on the supplied real-valued representatives at every zzz, including exceptional sets where the original input densities might be negative. The integral defining the prediction is the total Bochner integral, taking value zero for a nonintegrable integrand. The additional data sn′,un′,b,Ps'_n,u'_n,b,Psn′​,un′​,b,P are required by the problem argument but do not enter this formula, and no asymptotic assumptions enter it. Dimension d=0d=0d=0 is admitted, in which case EEE has one point and positive definiteness tests no nonzero vectors; n=0n=0n=0 is excluded.

DKFAsymptotics.Problem.p'

For any natural number ddd, let E=RdE=\mathbb R^dE=Rd with product Lebesgue measure λ\lambdaλ, let A,Γ,SA,\Gamma,SA,Γ,S be real d×dd\times dd×d matrices with Γ,S\Gamma,SΓ,S symmetric positive definite and S=ASAT+ΓS=ASA^\mathsf T+\GammaS=ASAT+Γ, and let a problem for this model supply sequences sn,un,sn′,un′:E→Rs_n,u_n,s'_n,u'_n:E\to\mathbb Rsn​,un​,sn′​,un′​:E→R for positive integers nnn, a state bbb, and a probability measure PPP. Every sequence member is required to be nonnegative λ\lambdaλ-almost everywhere, Bochner integrable, and of integral one; each primed member must also agree λ\lambdaλ-almost everywhere with a function φm,C(z)=((2π)ddet⁡C)−1exp⁡(−12(z−m)TC−1(z−m))\varphi_{m,C}(z)=\bigl(\sqrt{(2\pi)^d\det C}\bigr)^{-1}\exp\bigl(-\tfrac12(z-m)^\mathsf TC^{-1}(z-m)\bigr)φm,C​(z)=((2π)ddetC​)−1exp(−21​(z−m)TC−1(z−m)) for some mean m∈Em\in Em∈E and symmetric positive-definite covariance CCC, with witnesses allowed to vary between members. The problem requires the original weights wn(z)=un(z)(∫EφAy,Γ(z)sn(y) dλ(y))/φ0,S(z)w_n(z)=u_n(z)\bigl(\int_E\varphi_{Ay,\Gamma}(z)s_n(y)\,d\lambda(y)\bigr)/\varphi_{0,S}(z)wn​(z)=un​(z)(∫E​φAy,Γ​(z)sn​(y)dλ(y))/φ0,S​(z) to have almost-everywhere strongly measurable representatives and strictly positive finite extended absolute integrals. For each positive integer nnn, however, the declared function pn′:E→Rp'_n:E\to\mathbb Rpn′​:E→R uses the primed weight wn′(z)=un′(z)(∫EφAy,Γ(z)sn′(y) dλ(y))/φ0,S(z)w'_n(z)=u'_n(z)\bigl(\int_E\varphi_{Ay,\Gamma}(z)s'_n(y)\,d\lambda(y)\bigr)/\varphi_{0,S}(z)wn′​(z)=un′​(z)(∫E​φAy,Γ​(z)sn′​(y)dλ(y))/φ0,S​(z): it is defined pointwise by pn′(z)=wn′(z)/real⁡(Ln′)p'_n(z)=w'_n(z)/\operatorname{real}(L'_n)pn′​(z)=wn′​(z)/real(Ln′​), where Ln′=∫E+∣wn′(z)∣ dλ(z)∈[0,∞]L'_n=\int_E^{+}|w'_n(z)|\,d\lambda(z)\in[0,\infty]Ln′​=∫E+​∣wn′​(z)∣dλ(z)∈[0,∞] and real⁡(a)=a\operatorname{real}(a)=areal(a)=a for finite aaa, while real⁡(+∞)=0\operatorname{real}(+\infty)=0real(+∞)=0. There is no explicit requirement in this declaration or in its problem argument that 0<Ln′<∞0<L'_n<\infty0<Ln′​<∞, or that wn′w'_nwn′​ have an almost-everywhere strongly measurable representative. Under the total real-division convention, if Ln′=0L'_n=0Ln′​=0 or Ln′=+∞L'_n=+\inftyLn′​=+∞, the displayed definition returns the identically zero function, since division by zero equals zero. A nonintegrable integrand in the prediction likewise has total Bochner integral zero. The declaration defines a real-valued function; it does not include a proof or requirement that this output integrates to one or is Gaussian. Its values depend on the supplied input representatives, whose Gaussian identities and nonnegativity are imposed only almost everywhere. The original sequences, bbb, and PPP occur among the required problem data but not in the formula for pn′p'_npn′​; no asymptotic assumptions are invoked. The dimension can be zero, giving a one-point space and vacuous positivity conditions on nonzero vectors, while the sequence index is always positive.

DKFAsymptotics.Assumptions

For any natural number ddd, including zero, let E=RdE=\mathbb R^dE=Rd with its product topology and product Lebesgue measure λ\lambdaλ. Fix real d×dd\times dd×d matrices A,Γ,SA,\Gamma,SA,Γ,S with Γ,S\Gamma,SΓ,S symmetric positive definite and S=ASAT+ΓS=ASA^\mathsf T+\GammaS=ASAT+Γ, and fix a problem supplying four sequences sn,un,sn′,un′:E→Rs_n,u_n,s'_n,u'_n:E\to\mathbb Rsn​,un​,sn′​,un′​:E→R indexed by positive integers, a vector b∈Eb\in Eb∈E, and a probability measure PPP. The problem requires all four sequences' members to be nonnegative almost everywhere, Bochner integrable, and of integral one; every primed member must equal almost everywhere some φm,C(z)=((2π)ddet⁡C)−1exp⁡(−12(z−m)TC−1(z−m))\varphi_{m,C}(z)=\bigl(\sqrt{(2\pi)^d\det C}\bigr)^{-1}\exp\bigl(-\tfrac12(z-m)^\mathsf TC^{-1}(z-m)\bigr)φm,C​(z)=((2π)ddetC​)−1exp(−21​(z−m)TC−1(z−m)) with a symmetric positive-definite CCC, where the mean and covariance may vary with the member. Define wn(z)=un(z)(∫EφAy,Γ(z)sn(y) dλ(y))/φ0,S(z)w_n(z)=u_n(z)\bigl(\int_E\varphi_{Ay,\Gamma}(z)s_n(y)\,d\lambda(y)\bigr)/\varphi_{0,S}(z)wn​(z)=un​(z)(∫E​φAy,Γ​(z)sn​(y)dλ(y))/φ0,S​(z), Ln=∫E+∣wn(z)∣ dλ(z)L_n=\int_E^{+}|w_n(z)|\,d\lambda(z)Ln​=∫E+​∣wn​(z)∣dλ(z), and pn(z)=wn(z)/real⁡(Ln)p_n(z)=w_n(z)/\operatorname{real}(L_n)pn​(z)=wn​(z)/real(Ln​). As part of the problem data, every wnw_nwn​ agrees almost everywhere with a strongly measurable function and satisfies 0<Ln<∞0<L_n<\infty0<Ln​<∞; consequently real⁡(Ln)\operatorname{real}(L_n)real(Ln​) is the usual positive finite real value. For any real-valued function fff, write μf=λ\mu_f=\lambdaμf​=λ weighted by the extended nonnegative function max⁡(f,0)\max(f,0)max(f,0): equivalently, for measurable BBB, μf(B)=∫B+max⁡(f(z),0) dλ(z)\mu_f(B)=\int_B^{+}\max(f(z),0)\,d\lambda(z)μf​(B)=∫B+​max(f(z),0)dλ(z), with the integrand understood as an extended nonnegative real. The proposition declared here is the conjunction of precisely five requirements as nnn tends to infinity through positive integers: (A1) for every bounded continuous real-valued function ggg on EEE, ∫Eg dμsn→∫Eg dP\int_E g\,d\mu_{s_n}\to\int_E g\,dP∫E​gdμsn​​→∫E​gdP; (A2) the extended nonnegative numbers ∫E+∣sn(z)−sn′(z)∣ dλ(z)\int_E^{+}|s_n(z)-s'_n(z)|\,d\lambda(z)∫E+​∣sn​(z)−sn′​(z)∣dλ(z) converge to zero in the usual topology on [0,∞][0,\infty][0,∞]; (A3) for every bounded continuous real-valued ggg on EEE, ∫Eg dμun→∫Eg dδb=g(b)\int_E g\,d\mu_{u_n}\to\int_E g\,d\delta_b=g(b)∫E​gdμun​​→∫E​gdδb​=g(b); (A4) the extended nonnegative numbers ∫E+∣un(z)−un′(z)∣ dλ(z)\int_E^{+}|u_n(z)-u'_n(z)|\,d\lambda(z)∫E+​∣un​(z)−un′​(z)∣dλ(z) converge to zero in the usual topology on [0,∞][0,\infty][0,∞]; and (A5) for every bounded continuous real-valued ggg on EEE, ∫Eg dμpn→∫Eg dδb=g(b)\int_E g\,d\mu_{p_n}\to\int_E g\,d\delta_b=g(b)∫E​gdμpn​​→∫E​gdδb​=g(b). Here δb\delta_bδb​ is the unit point mass at the same bbb in both (A3) and (A5), and each integral against a measure is the total Bochner integral, defined to be zero when its integrand is not integrable. These statements quantify over each fixed test function separately; they do not assert a common uniform convergence rate over all such functions. The two absolute-distance expressions contain no factor of one half. The density-to-measure operation discards negative pointwise values, although the input-density requirements already exclude such values outside null sets. The proposition assumes convergence of the original updated sequence pnp_npn​ in (A5); it does not state convergence, Gaussianity, positive finite normalization, or eventual validity for a primed updated sequence. There is no requirement that PPP have a density or finite moments. There is no n=0n=0n=0 condition. When d=0d=0d=0, the underlying space is a singleton and the positive-definiteness tests on nonzero vectors are vacuous; that dimension is not excluded by any binder or field.

Human review
  • Endorsed by Shuze Chen · Sep 16, 2026

  • Endorsed by burkh4rt · Sep 16, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me