OctoInformBeta

OpenAI preprint, September 2026 · §4. An exact coercive identity

Proposition 4.2: With constant 16615071521000000\tfrac{1661507}{1521000000} relative to v4v^4 under sign and flux conditions, and 11000\tfrac{1}{1000} for tilted Gaussian weights under shape conditions

Integrated coercivity of the quartic form z2−12vw2+6v4z^2-12vw^2+6v^4..

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

On an interval where the order parameter γ\gamma is constant, γ=c\gamma = c, write v=cΦγ,xxv = c\Phi_{\gamma,xx}, w=cΦγ,xxxw = c\Phi_{\gamma,xxx} and z=cΦγ,xxxxz = c\Phi_{\gamma,xxxx} for the scaled spatial derivatives of the value function Φγ\Phi_\gamma. The quartic form z2−12vw2+6v4z^2 - 12vw^2 + 6v^4(1) has a nonpositive middle term where v≥0v \ge 0, and two nonnegative ones. The quartic form (1) is coercive on average against a probability density ff on the line if 𝔼f[z2−12vw2+6v4]≥κ𝔼fv4\mathbb{E}_f[z^2 - 12vw^2 + 6v^4] \ge \kappa\, \mathbb{E}_f v^4 for some constant κ>0\kappa > 0.

The library states two bounds of this kind for real functions on the line.

The Polynomial F2F_2 and Its Flow Derivative

The first bound rests on a cubic polynomial F2F_2 in five real variables RR, KK, vv, bb and jj, whose coefficients are built from the rational constants A=113,C=5324,T0=7348,k=1916,g0=233110000,h=872500,x0=46635000,y=685310000,m=−1051400,n0=−1073500.\begin{gathered} A = \tfrac{1}{13},\quad C = \tfrac{53}{24},\quad T_0 = \tfrac{73}{48},\quad k = \tfrac{19}{16},\quad g_0 = \tfrac{2331}{10000},\\ h = \tfrac{87}{2500},\quad x_0 = \tfrac{4663}{5000},\quad y = \tfrac{6853}{10000},\quad m = -\tfrac{1051}{400},\quad n_0 = -\tfrac{1073}{500}. \end{gathered}(2)

From these constants, the polynomials QQ in R,K,v,bR, K, v, b and TT in R,v,jR, v, j are the quadratic forms Q=Avb+Cv2+kKv−2kRb+T0Rv,T=v(g0R+hj),Q = Avb + Cv^2 + kKv - 2kRb + T_0Rv, \qquad T = v\,(g_0R + hj),(3) and the polynomials F1F_1 and F2F_2 in R,K,v,b,jR, K, v, b, j are the cubic forms F1=b2(2kR−Av)−bv(2kK+g0R+2T0R+2Cv+hj),F2=v2(x0j+yv+mK+n0R)−F1.\begin{aligned} F_1 &= b^2(2kR - Av) - bv\,(2kK + g_0R + 2T_0R + 2Cv + hj),\\ F_2 &= v^2(x_0j + yv + mK + n_0R) - F_1. \end{aligned}(4)

In each variable, the partial derivative of F2F_2 is a quadratic form: ∂RF2=n0v2−2kb2+(g0+2T0)bv,∂KF2=mv2+2kbv,∂jF2=x0v2+hbv,∂vF2=2x0vj+3yv2+2mKv+2n0Rv+Ab2+b(2kK+g0R+2T0R+4Cv+hj),∂bF2=2Q+T.\begin{aligned} \partial_R F_2 &= n_0v^2 - 2kb^2 + (g_0 + 2T_0)\,bv, \qquad \partial_K F_2 = mv^2 + 2kbv, \qquad \partial_j F_2 = x_0v^2 + hbv,\\ \partial_v F_2 &= 2x_0vj + 3yv^2 + 2mKv + 2n_0Rv + Ab^2 + b\,(2kK + g_0R + 2T_0R + 4Cv + hj),\\ \partial_b F_2 &= 2Q + T. \end{aligned}(5)

The flow derivative DF2DF_2 at real numbers R,K,v,b,j,z,LR, K, v, b, j, z, L is the derivative of F2F_2 along the vector field (Ṙ,K̇,v̇,ḃ,j̇)=(2Rv,2vK+Rj,−Rb,−z−bv,Kb−jv−L)(\dot R, \dot K, \dot v, \dot b, \dot j) = (2Rv,\, 2vK + Rj,\, -Rb,\, -z - bv,\, Kb - jv - L), with each partial of (5) taken at (R,K,v,b,j)(R, K, v, b, j)✦, and the weighted derivative WF2WF_2 at the same point adds (v+R−K)F2(v + R - K)F_2 to it: DF2=2Rv∂RF2+(2vK+Rj)∂KF2−Rb∂vF2−(z+bv)∂bF2+(Kb−jv−L)∂jF2,WF2=DF2+(v+R−K)F2.\begin{aligned} DF_2 &= 2Rv\,\partial_R F_2 + (2vK + Rj)\,\partial_K F_2 - Rb\,\partial_v F_2 - (z + bv)\,\partial_b F_2 + (Kb - jv - L)\,\partial_j F_2,\\ WF_2 &= DF_2 + (v + R - K)\,F_2. \end{aligned}(6)✦

Tilted Gaussian Weights and the Backward Shape

A bounded smooth function is a function gg on ℝ\mathbb{R} of class C∞C^\infty whose derivatives of every order are bounded: for each n≥0n \ge 0 there is a constant Cn≥0C_n \ge 0 with |g(n)(x)|≤Cn|g^{(n)}(x)| \le C_n for all xx.✦ A regular datum is a function gg on ℝ\mathbb{R} of class C∞C^\infty whose derivative gxg_x is a bounded smooth function.✦

The Burgers rate of a function rr of time tt and space xx is ℛ=12rxx+rrx\mathcal{R} = \tfrac12 r_{xx} + r\,r_x.✦ The backward Wronskian of rr is the Wronskian rℛx−rxℛr\,\mathcal{R}_x - r_x\,\mathcal{R} of rr and its Burgers rate in the space variable.✦ A function rr on ℝ\mathbb{R} has the backward shape if it is odd, has rx>0r_x > 0 everywhere, and at every x≥0x \ge 0 satisfies the three inequalities of (7). rxx(x)≤0,ℛ(x)≤0,r(x)ℛx(x)−rx(x)ℛ(x)≥0,r_{xx}(x) \le 0, \qquad \mathcal{R}(x) \le 0, \qquad r(x)\,\mathcal{R}_x(x) - r_x(x)\,\mathcal{R}(x) \ge 0,(7)✦

The coercive polynomial PrP_r of a real function rr on ℝ\mathbb{R} is the form (1) with vv, ww and zz replaced by the certificates of rr, its first three derivatives rxr_x, rxxr_{xx} and rxxxr_{xxx}: Pr=rxxx2−12rxrxx2+6rx4,P_r = r_{xxx}^2 - 12\,r_x\,r_{xx}^2 + 6\,r_x^4,(8)✦ and the quartic moment of rr is the function rx4r_x^4✦.

For a real number dd and a real function UU on ℝ\mathbb{R}, the tilted Gaussian density ρd,U\rho_{d,U} is the unnormalized weight on ℝ\mathbb{R} given by ρd,U(x)=e−dx2+U(x).\rho_{d,U}(x) = e^{-dx^2 + U(x)}.(9)✦

The Claim

Theorem 1. (a) Let ρ\rho be a nonnegative function and rr, RR, KK, vv, bb, jj, zz, LL real functions on ℝ\mathbb{R} with R,K,v,j,L≥0R, K, v, j, L \ge 0, b≥2vb \ge 2v and S=z+bv−2Rb≥0S = z + bv - 2Rb \ge 0 everywhere, and let F2F_2 and WF2WF_2 stand for (4) and (6) evaluated along (R,K,v,b,j)(R, K, v, b, j) and (R,K,v,b,j,z,L)(R, K, v, b, j, z, L). Suppose that the flux rF2ρrF_2\rho has derivative ρWF2\rho\,WF_2 at every point and tends to 00 at −∞-\infty and at +∞+\infty, and that ρv4\rho v^4, ρWF2\rho\,WF_2 and ρ(z2−12Rvb2+6v4)\rho\,(z^2 - 12Rvb^2 + 6v^4) are integrable on ℝ\mathbb{R}. Then 16615071521000000∫ℝρv4dx≤∫ℝρ(z2−12Rvb2+6v4)dx.\frac{1661507}{1521000000}\int_{\mathbb{R}} \rho\, v^4\, dx \le \int_{\mathbb{R}} \rho\,(z^2 - 12Rvb^2 + 6v^4)\, dx.(10)✦

(b) Let rr be a bounded smooth function with the backward shape, and let ss be an odd regular datum with sx≥0s_x \ge 0 everywhere and with sxx≤0s_{xx} \le 0 and rsx−srx≥0r s_x - s r_x \ge 0 on [0,∞)[0,\infty). Let d>0d > 0, and let UU be an even Lipschitz function on ℝ\mathbb{R} with Ux(x)=r(x)−s(x)+2dxU_x(x) = r(x) - s(x) + 2dx at every xx. Then 11000∫ℝρd,Urx4dx≤∫ℝρd,UPrdx.\frac{1}{1000}\int_{\mathbb{R}} \rho_{d,U}\, r_x^4\, dx \le \int_{\mathbb{R}} \rho_{d,U}\, P_r\, dx.(11)✦

In the paper, the result reads: if the variables are related as in Lemma 3.4, with its derivative rules, with R,K,v,j,L,S≥0R, K, v, j, L, S \ge 0, b≥2vb \ge 2v and Rb2=w2Rb^2 = w^2, and with their quotients extending across zero, and if ff is a probability density with fx/f=r−sf_x/f = r - s for which every term of the identity of Lemma 4.1 is integrable and rfF2→0rfF_2 \to 0 at ±∞\pm\infty, then 𝔼f[z2−12vw2+6v4]≥16615071521000000𝔼fv4≥11000𝔼fv4\mathbb{E}_f[z^2 - 12vw^2 + 6v^4] \ge \tfrac{1661507}{1521000000}\,\mathbb{E}_f v^4 \ge \tfrac{1}{1000}\,\mathbb{E}_f v^4.

What Is Checked

Part (a) of the theorem is stated over the constants (2), the polynomials QQ, TT, F1F_1 and F2F_2, the partial derivatives of F2F_2, its flow derivative DF2DF_2 and weighted derivative WF2WF_2, and the hypotheses on the weight ρ\rho and on r,R,K,v,b,j,z,Lr, R, K, v, b, j, z, L; part (b) is stated over bounded smooth functions and regular data, the backward shape with the Burgers rate and backward Wronskian it is defined by, the coercive polynomial of rr built from its certificates rxr_x, rxxr_{xx}, rxxxr_{xxx}, its quartic moment, and the tilted Gaussian density ρd,U\rho_{d,U}.