## 1. Precise formulation Let \[ \phi(x)=(2\pi)^{-1/2}e^{-x^2/2}. \] The claim is the following existence statement on \(\mathbb R\): For every \(M>0\), there is a strictly positive probability density \(p\) such that: - \(p\) is real-analytic on all of \(\mathbb R\); - \(p\) is centered: \(\int xp(x)\,dx=0\); - writing \(V=-\log p\), there exists \(m>0\) such that \(V''(x)\ge m\) for every \(x\); - for every real \(q>0\), \[ \int_{\mathbb R}p(x)^q\,dx =\int_{\mathbb R}\phi(x)^q\,dx =(2\pi)^{(1-q)/2}q^{-1/2}; \] - its variance satisfies \[ 1\le \operatorname{Var}(p)<4-\frac8\pi =1.4535209105\ldots<1.454; \] - its location Fisher information is finite but exceeds \(M\): \[ I(p):=\int_{\mathbb R}\frac{p'(x)^2}{p(x)}\,dx>M. \] “Strongly log-concave” does **not** here mean \(V''\ge1\), the Gaussian’s curvature. In fact, however, the construction supplies the common lower bound \(V''\ge1/4\), stronger than the supporting text asserts. ## 2. Checking the argument ### Coordinate change and power integrals Put \[ c(t)=a\bigl(\sqrt{t^2+\delta^2}-\delta\bigr),\qquad F(t)=t+c(t). \] Then \[ F'(t)=1+\frac{at}{\sqrt{t^2+\delta^2}}\in(1-a,1+a). \] Consequently, \(F\) is an increasing bijection \(\mathbb R\to\mathbb R\), with a real-analytic inverse. Thus \[ p_0(F(t))=\phi(t) \] defines a positive real-analytic function. For every \(q>0\), \[ \int p_0(x)^q\,dx =\int\phi(t)^q(1+c'(t))\,dt =\int\phi(t)^q\,dt, \] because \(\phi^q\) is even and \(c'\) is odd. All these integrals converge. In particular, \(p_0\) is normalized. The absence of a Jacobian factor in the definition is intentional: \(p_0\) is **not** the ordinary pushforward density of a Gaussian under \(F\). ### Strong log-concavity Since \[ V(F(t))=\frac{t^2}{2}+\frac12\log(2\pi), \] differentiation gives \[ V'(F(t))=\frac{t}{1+c'(t)} \] and \[ V''(F(t)) =\frac{1+c'(t)-tc''(t)}{(1+c'(t))^3} =\frac{1+a t^3/(t^2+\delta^2)^{3/2}} {(1+a t/\sqrt{t^2+\delta^2})^3}. \] This matches the text. There are particularly simple sharp bounds. Set \[ u=\frac{t}{\sqrt{t^2+\delta^2}}\in(-1,1),\qquad g(u)=\frac{1+au^3}{(1+au)^3}. \] Then \[ g'(u)=\frac{3a(u^2-1)}{(1+au)^4}\le0, \] so \[ \boxed{\frac1{(1+a)^2}\le V''(x)\le\frac1{(1-a)^2}.} \] Thus every density is uniformly strongly log-concave. Moreover, all members of the family satisfy \(V''\ge1/4\). Only the upper curvature bound necessarily deteriorates. ### Fisher information The score is \[ (\log p_0)'(F(t))=-\frac{t}{1+c'(t)}. \] Therefore \[ I(p_0)=\int_{\mathbb R}\frac{t^2\phi(t)}{1+c'(t)}\,dt, \] as claimed. It is finite for each \(a<1\), since \[ I(p_0)\le \frac1{1-a}\int t^2\phi(t)\,dt=\frac1{1-a}. \] If \(\delta=1-a\to0\), the denominator tends uniformly to zero on any fixed negative interval bounded away from zero. Integrating over such an interval proves divergence. This step is valid. ### Variance The variance identity and Lipschitz estimate are also correct. Here is a derivation that independently verifies them. ### Independent probabilistic re-derivation The superlevel interval associated with radius \(r\) is \[ [F(-r),F(r)]=[c(r)-r,c(r)+r]. \] Using \[ \phi(t)=\int_{|t|}^{\infty}r\phi(r)\,dr, \] we obtain \[ p_0(x)=\int_0^\infty r\phi(r) \,\mathbf1_{\{|x-c(r)|\le r\}}\,dr. \] Equivalently, \[ X\stackrel d=c(R)+RU, \] where \(U\) is uniform on \([-1,1]\), independent of \(R\), and \[ f_R(r)=2r^2\phi(r),\qquad r>0. \] This is a probability density. Its moments are \[ ER=2\sqrt{\frac2\pi},\qquad ER^2=3,\qquad \operatorname{Var}(R)=3-\frac8\pi. \] Conditional expectation and variance now give \[ EX=Ec(R) \] and \[ \operatorname{Var}(X) =E\!\left[\frac{R^2}{3}\right]+\operatorname{Var}(c(R)) =1+\operatorname{Var}(c(R)). \] For independent copies \(R,R'\), the Lipschitz property yields \[ \operatorname{Var}(c(R)) =\frac12E(c(R)-c(R'))^2 \le\frac{a^2}{2}E(R-R')^2 =a^2\operatorname{Var}(R). \] Hence \[ 1\le\operatorname{Var}(X) \le1+a^2\left(3-\frac8\pi\right) <4-\frac8\pi. \] For the stated parameters, \(c(R)\) is nonconstant, so the lower bound is actually strict. There is also an independent, quantitative verification of Fisher-information divergence. Pairing the positive and negative halves of the integral gives, for \(Z\sim N(0,1)\), \[ I(p_0) =E\left[\frac{Z^2(Z^2+\delta^2)} {\delta^2+(1-a^2)Z^2}\right]. \] Writing \(b=1-a^2\), this becomes \[ I(p_0) =\frac1b-\frac{a^2\delta^2}{b} E\left[\frac{Z^2}{\delta^2+bZ^2}\right]. \] Since the expectation is at most \(1/b\), \[ \frac1b-\frac{a^2\delta^2}{b^2} \le I(p_0)\le\frac1b. \] Along \(\delta=1-a\), \[ \boxed{ \frac1{1-a^2}-\frac{a^2}{(1+a)^2} \le I(p_0)\le\frac1{1-a^2}. } \] Thus \(I(p_0)\sim(1-a^2)^{-1}\to\infty\). Finally, define \[ p(x)=p_0(x+\mu),\qquad \mu=EX. \] Translation centers the density and preserves its power integrals, variance, Fisher information, analyticity, and curvature bounds. This completes the proof. ## 3. Attempts to break the claim ### The limiting objects are not admissible—but every approximant is - At \(\delta=0\), analyticity generally fails at the mode. - At \(a=1\), the coordinate map need not map onto all of \(\mathbb R\). Neither is a defect: the construction uses only \(a<1\) and \(\delta>0\). ### The joint parameter limit matters Letting \(a\to1\) while holding \(\delta>0\) fixed does **not** make Fisher information diverge. The paired formula instead gives \[ I(p_0)\longrightarrow 1+\frac3{\delta^2}. \] The text correctly sends \(\delta\) to zero as well. ### No hidden infinite-information examples Every constructed density has finite Fisher information. The information becomes arbitrarily large across the family, not infinite for a single admissible member. ### The variance bound is genuinely tight for this family Along \(a\to1\), \(\delta=1-a\), we have \(c(R)\to R\) in \(L^2\). Consequently, \[ \operatorname{Var}(X)\longrightarrow 1+\operatorname{Var}(R)=4-\frac8\pi. \] Thus the constant is not an arithmetic accident. For a numerical check, \(a=0.99,\delta=0.01\) gives the rigorous bounds \[ I(p)>50.0037,\qquad \operatorname{Var}(p)<1.4445. \] ### A stronger Gaussian-curvature interpretation would fail The construction does not satisfy \(V''\ge1\). Indeed, if one additionally required that bound, the power-integral identities would force a Gaussian. To see this, the identities imply \(\sup p=\sup\phi\). If \(m\) is the mode and \(V''\ge1\), then \[ p(x)\le p(m)e^{-(x-m)^2/2} =\phi(x-m). \] Both sides integrate to one, forcing equality. So specifying what “strongly log-concave” means is important. **Result of the attempted falsification:** no gap or counterexample was found under the stated interpretation. ## 4. Novelty The underlying example has a classical antecedent: the **two-piece normal**, also called the **split normal** distribution. Formally setting \(\delta=0\) gives \[ p_a(x)= \begin{cases} \phi\!\left(x/(1-a)\right),&x<0,\\[2mm] \phi\!\left(x/(1+a)\right),&x\ge0. \end{cases} \] Its two scale parameters are \(1-a\) and \(1+a\). Direct integration yields \[ \int p_a^q=\int\phi^q,\qquad I(p_a)=\frac1{1-a^2}, \] and \[ \operatorname{Var}(p_a) =1+a^2\left(3-\frac8\pi\right). \] Thus the same power-integral identities, variance bound, and information blow-up already appear in a classical distribution family. Its defect for the present claim is the lack of analyticity at the joining point. The supporting construction removes that defect while preserving the identities exactly. More generally, preservation of all these power integrals comes from **equimeasurability**: corresponding superlevel sets have the same lengths as Gaussian superlevel sets. This is standard rearrangement-theoretic structure. The classical Pólya–Szegő inequality, applied to \(\sqrt p\), explains why the symmetric Gaussian rearrangement minimizes Fisher information; it does not provide an upper bound. Accordingly, I would classify the result as an **analytic variant of known examples and rearrangement principles**, rather than an entirely new phenomenon. I cannot establish whether this exact analytic formulation has appeared previously without a literature search. ## 5. Fairness of the one-sentence claim The sentence fairly summarizes what is proved: - centering is legitimate; - real-analyticity holds globally on \(\mathbb R\); - strong log-concavity holds; - every positive power integral agrees exactly with the Gaussian’s; - variance stays below the advertised constant; - Fisher information is unbounded across the family. The main interpretive restriction is that “strongly log-concave” does not mean curvature at least \(1\). Indeed, the text is unnecessarily cautious about uniformity: a common lower curvature bound of \(1/4\) is available. {"verdict":"correct","novelty":"special case or variant of known results","fair_summary":"yes","confidence":4,"one_line":"The construction rigorously proves the claim, even with a common curvature lower bound of 1/4, and is an analytic smoothing of the classical split-normal example."}