\documentclass[11pt]{amsart} \usepackage[a4paper,margin=27mm]{geometry} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage{amsmath,amssymb,mathtools} \usepackage{booktabs,array,longtable} \usepackage{microtype} \usepackage[hidelinks]{hyperref} \hypersetup{pdftitle={A note on Somos sequences},pdfauthor={Hernandez Juan}} \usepackage{enumitem} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} \allowdisplaybreaks[2] \setlength{\emergencystretch}{2em} \title{A note on Somos sequences} \author{Hernandez Juan} \date{November 18, 2026} \begin{document} \begin{abstract} We prove that the standard Somos-\(k\) sequence contains a noninteger for every \(k\geq8\), and give an explicit bound for its first occurrence. We classify the three-term Somos recurrences with independent coefficients and unit initial values whose terms are polynomials in the coefficients, and determine the possible first denominators. We also give exact prime-adic valuation formulas for the standard Somos-4 sequence. \end{abstract} \maketitle \section{Introduction} For an integer \(k\geq2\), the standard Somos-\(k\) sequence is defined by \begin{equation}\label{eq:intro-somos} a_0=\cdots=a_{k-1}=1,\qquad a_na_{n-k}=\sum_{j=1}^{\lfloor k/2\rfloor}a_{n-j}a_{n-k+j} \quad(n\geq k). \end{equation} Somos discovered the order-six example in 1982 during his study of addition formulas and presented its integrality problem in 1989 \cite{SomosHistory,SomosProblem}. Gale's accounts \cite{Gale,GaleUpdate} described the early computations and the related recurrences proposed by Raphael Robinson. Several independent proofs of integrality for orders four and five were found in 1990; Malouf published one in 1992 \cite{Malouf}. Propp's chronology \cite{Propp} records Hickerson's proof for order six and Lotto's proof for order seven, both obtained in April 1990. Fomin and Zelevinsky placed these examples in their theory of the Laurent phenomenon \cite{FZLaurent}. With independent initial variables, the terms in the relevant recurrences are Laurent polynomials over the coefficient ring. Their Theorems 1.4 and 1.8 apply respectively to the three-term and two-term right-hand sides of the Gale--Robinson recurrences. Setting the initial variables equal to one gives the integrality of Somos-4 through Somos-7. Subsequent combinatorial formulas express terms of the octahedron and cube recurrences through perfect matchings and groves \cite{Speyer,CarrollSpeyer}. Explicit perfect-matching models for the two-term right-hand-side Gale--Robinson recurrences were developed by Bousquet-M\'elou, Propp, and West \cite{BMPW}. \subsection{Higher orders} The first noninteger in Somos-8 is \[ a_{17}=\frac{420514}{7}. \] The all-orders question is whether every standard Somos-\(k\) sequence with \(k\geq8\) contains a noninteger. Stanley recorded this question in his list of bijective proof problems \cite[p.~68]{Stanley}. Somos's 2014 comment in OEIS A030127 also records the question of a published proof for all orders \cite{OEIS030127}. Let \(N(k)\) denote that position, with the indexing in~\eqref{eq:intro-somos}. Theorem~\ref{thm:p8} proves that \(N(k)\) exists for every \(k\geq8\), with \[ 2k\leq N(k)\leq 2k+\max\left\{1, \left\lfloor\frac{\lfloor k/2\rfloor-1}{4}\right\rfloor\right\}. \] For \(k=2,3\), every term is \(1\) by induction. Together with the known results for orders four through seven, this gives the exact classification: the standard sequence~\eqref{eq:intro-somos} is integral precisely for \(2\leq k\leq7\). Hanna's 2004 conjecture predicts the sharper location \(N(k)\in\{2k,2k+1\}\) \cite{OEIS030127,OEIS101591}. The proof for \(k\geq8192\) begins with an exact elimination of the first two blocks of the recurrence. If a specified interval of terms were integral, explicit quotients with cubic numerators would be integers. An integer difference operator annihilates sixteen terms of a geometric expansion. The remaining values have absolute value less than one and consequently vanish. Clearing denominators gives an exponential polynomial with more real zeros than Rolle's theorem permits. The finite interval \(8\leq k\leq8191\) is checked by exact integer remainders in the original recurrence; the complete verifier is included in Section~\ref{sec:p8}. \subsection{Prime powers in Somos-4} The divisibility properties of Somos-4 belong to the arithmetic of elliptic divisibility sequences. Ward studied the repetition of primes in these sequences \cite{WardMemoir,WardRepetition}. Robinson established periodicity properties for Somos sequences and recorded observations about the spacing of prime-power divisors \cite{Robinson}. The elliptic interpretation of Somos-4 was developed by Swart and Hone \cite{Swart,Hone}; Hone and Swart subsequently studied stronger Laurent rings and arithmetic conditions for integrality \cite{HoneSwart}. Uchida described valuations of general Somos-4 sequences using canonical local heights \cite{Uchida}. For the standard Somos-4 sequence, the curve and point are \[ E:\ y^2+y=x^3-x,\qquad P=(0,0). \] The denominator correspondence appears in Everest--McLaren--Ward \cite[Examples 2.7--2.8]{EverestMcLarenWard}; Jones and Rouse give explicit coordinates for \([2n-3]P\) in terms of neighboring Somos terms \cite[proof of Corollary 1.2]{JonesRouse}. Together with the classical local denominator-growth formula \cite[Lemma 3.1]{EverestMcLarenWard}, these identities yield the following consequence. Write \(v_p\) for the \(p\)-adic valuation normalized by \(v_p(p)=1\). If an odd prime \(p\) divides \(a_n\), then \[ v_p\left(a_{\frac{(2n-3)p^j+3}{2}}\right)=v_p(a_n)+j \qquad(j\geq0). \] Theorem~\ref{thm:valuation} supplies a direct proof, including the coordinate calculation and the convergence range of the formal logarithm. It also gives the complete \(2\)-adic exception and the absence of the bad prime \(37\). Corollary~\ref{cor:all-valuations} describes all prime-power divisibility indices by the reduction order of \(P\). Van der Kamp proved arithmetic-progression results for prime-power divisibility indices and reviewed Robinson's observations \cite[Section 8]{Kamp}. The explicit formula above answers the arbitrary-height question there for every occurring odd prime in the standard Somos-4 sequence. Its proof uses the established elliptic correspondence and local multiplication law. \subsection{Three independent coefficients} Gale's 1991 account extended Robinson's unit-coefficient observations to recurrences with three coefficients \cite{Gale,SomosRefined}. On 11 April 1999, Somos formulated a refined polynomial version of the Gale--Robinson conjecture \cite{SomosRefined}. For integers \(k,p,q,r\) with \(00$ the polynomial $Mp_j/A^j$ has integer coefficients, as the displayed lists show. Define \begin{equation}\label{eq:p8-Pi} \Pi(r)=M\{J+\mathcal P(r,-1/A)\}=\widetilde J+f(r),\qquad f(0)=0. \end{equation} Explicitly, \begin{align} f(r)&=-144A(6174r^3+441r^2-6741r) &&(\varepsilon=1),\label{eq:p8-fodd}\\ f(r)&=-16\{55566Ar^3+161406Ar^2 +(86436A^2+25032A+333396)r\} &&(\varepsilon=0).\label{eq:p8-feven} \end{align} Thus the cubic coefficient of $\Pi$ is $-889056A$ in both cases. In particular, $\Pi$ is nonzero. Suppose now that $D_0,\ldots,D_R$ are integers. Then $MK_1$ and $M\mathcal P(1,2)$ are integers, so $MJ\in\mathbb Z$. For $r\ge1$, $b_r=1+A2^r$ is odd, and $K_r\in b_r\mathbb Z[1/2]$, where $\mathbb Z[1/2]=\{z/2^v:z\in\mathbb Z,\ v\in\mathbb Z_{\ge0}\}$. The coefficient conditions above give \[ M\{\mathcal P(r,2^r)-\mathcal P(r,-1/A)\} \in b_r\mathbb Z[1/2]. \] Indeed, the difference for a positive power $t^j$ is an integer polynomial multiple of $(At)^j-(-1)^j$, which is divisible by $1+At$; the difference for $t^{-1}$ is $t^{-1}+A=(1+At)/t$. The constant term makes no contribution. The same coefficient conditions show that $M\mathcal P(r,-1/A)$ is an integer. Equation \eqref{eq:p8-KP} therefore implies $\Pi(r)\in b_r\mathbb Z[1/2]\cap\mathbb Z=b_r\mathbb Z$, since $b_r$ is odd. Consequently \begin{equation}\label{eq:p8-integralquotients} \frac{\Pi(r)}{1+A2^r}\in\mathbb Z \quad(1\le r\le R). \end{equation} We record a bound on $\Pi$ which is independent of the integrality assumption. At $r=0,1$, \eqref{eq:p8-split} gives $b_0D_0=Q_0$ and $b_1D_1=D_0b_2+Q_1$. Hence \begin{align*} K_1 &=\frac{(A+1)Q_1-(4A+3)Q_0}{4}\\ &=-(A+1)Z+\frac{3A(A+2)}4W +\frac{(A+1)B_1-(4A+3)B_0}{4}. \end{align*} Inserting this expression in \eqref{eq:p8-Pi} gives \begin{equation}\label{eq:p8-Jtilde} \widetilde J=-M(A+1)Z+\frac{3MA(A+2)}4W+S(A,H). \end{equation} For odd $k$ the polynomial $S$ is \begin{align} S=18\bigl(&-1152A^4+57624A^3H^2-38416A^3 +192080A^2H^2\notag\\ &-115248A^2+153664AH^2-154816A\bigr). \label{eq:p8-Sodd} \end{align} For even $k$ it is \begin{align} S={}&-3600A^4+518616A^3H^2-264208A^3 +1728720A^2H^2\notag\\ &+1037232A^2H-6914880A^2+1382976AH^2 +1382976AH-5497088A+2540160. \label{eq:p8-Seven} \end{align} Both identities follow by substituting the displayed polynomials in \[ S=M\left\{ \frac{(A+1)B(1,2)-(4A+3)B(0,1)}4 -\mathcal P(1,2)+\mathcal P(0,-1/A)\right\}, \] and using $m=A+1$. Equations \eqref{eq:p8-fodd}--\eqref{eq:p8-Seven} also give $\Pi\in\mathbb Z[r]$ directly. The sums of the absolute coefficients of $S$ are $12834000$ and $21270456$, both less than $2^{25}$. Those of $f$ are $1923264$ and $10589376$, both less than $2^{24}$. Every monomial of $S$ has absolute value at most $(k+1)^4H^2$; for $0\le r\le R$, every monomial of $f$ has absolute value at most $(k+1)^4$. Using $M<2^{22}$, $H^2\le2^k$ and \eqref{eq:p8-firstheight} in \eqref{eq:p8-Jtilde} now gives \begin{align*} M(A+1)Z+\frac{3MA(A+2)}4W &<2^{k+23}(k+1)^3\mathcal B^d,\\ |S|&<2^{k+25}(k+1)^4\mathcal B^d,\\ \max_{0\le r\le R}|\Pi(r)| &<2^{k+27}(k+1)^4\mathcal B^d. \end{align*} Here and below the maxima range over integers $r$. Let $s=\log_2(k+1)$. For $m\ge4096$, \[ s\le\log_2(2m+2)\le\frac{\sqrt m}{4}. \] The last inequality holds at $m=4096$, and its difference has positive derivative thereafter, since $(m+1)\log2>8\sqrt m$. Since $\mathcal B\le(k+1)^2$ and $d\le s+1$, \begin{align*} \log_2\max_{0\le r\le R}|\Pi(r)| &0) \] gives \[ u_r=\sum_{h=1}^{16}(-1)^{h-1}A^{-h}P(r)2^{-hr}+\rho_r, \qquad |\rho_r|\le |P(r)|2^{-17r}. \] The operator $T$ kills the finite sum. Take $r_0=R-324$. For each of $r_0,r_0+1,\ldots,r_0+259$, every index used by $T$ lies between $1$ and $R-1$, because $r_0\ge699$ and $r_0+259+64=R-1$. By \eqref{eq:p8-Tnorm}, \[ |Tu_r|<2^{550+5m/2-17r_0} \le2^{6075-7m/4}<1. \] The middle inequality uses $R\ge m/4-1$, and the last uses $m\ge4096$. If all $u_r$ in the stated interval were integers, $Tu_r$ would therefore vanish at these $260$ distinct points. Define the ordinary polynomial in independent variables $r,x$ \[ G(r,x)=\sum_{j=0}^{64}t_jP(r+j) \prod_{\substack{0\le h\le64\\h\ne j}}(1+A2^hx). \] It has degree at most $64$ in $x$ and at most $3$ in $r$, and \[ G(r,2^r)=\left(\prod_{h=0}^{64}(1+A2^{r+h})\right)Tu_r. \] It is nonzero, since \[ G(r,-1/A)=P(r)\prod_{h=1}^{64}(1-2^h)\ne0 \] as a polynomial. Writing $G(r,x)=\sum_{j=0}^{64}g_j(r)x^j$, its substitution $G(r,2^r)$ is also nonzero: after division by the largest occurring exponential, the corresponding nonzero polynomial dominates all terms with smaller exponents as $r\to+\infty$. Consider a function \[ f(x)=\sum_{j=1}^s q_j(x)e^{\lambda_jx},\qquad \lambda_1<\cdots<\lambda_s,\qquad d_j=\deg q_j, \] where each $q_j$ is a nonzero real polynomial. It has at most $\sum_j(d_j+1)-1$ distinct real zeros. This follows by induction on $s$. For $s=1$, divide by the nonvanishing exponential and use the polynomial root bound. For $s>1$, divide by $e^{\lambda_1x}$ and differentiate $d_1+1$ times. The first summand vanishes. Each other summand retains a nonzero polynomial of the same degree, since the map $q\mapsto q'+(\lambda_j-\lambda_1)q$ is invertible on polynomials of bounded degree. Repeated Rolle's theorem shows that any finite set of $N$ zeros produces at least $N-d_1-1$ zeros of the resulting function. The induction hypothesis bounds that number and gives the assertion. Apply this bound to $G(r,2^r)=\sum_{j=0}^{64}g_j(r)e^{j(\log2)r}$, omitting zero coefficients. There are at most $65$ exponents and each polynomial has degree at most $3$. Thus it has at most $259$ real zeros, contrary to the $260$ zeros already obtained. \end{proof} If $D_0,\ldots,D_R$ were all integers, equations \eqref{eq:p8-integralquotients} and \eqref{eq:p8-Piheight} would contradict Lemma~\ref{lem:p8-quotients}, because $\Pi$ is a nonzero cubic. Hence $N(k)\le2k+R$ for $k\ge8192$. \subsection{The finite interval} For $8\le k\le8191$, the following exact finite computation gives $N(k)\in\{2k,2k+1\}$. In this subsection put $m=\lfloor k/2\rfloor$. The first block can be computed modulo any positive integer $q$, since all its divisors are $1$. The divisor at index $2k$ is $a_k=m$; compute its numerator modulo $m$. A nonzero remainder proves $N(k)=2k$. If the remainder is zero, $a_{2k}$ is an integer. The next divisor is $a_{k+1}=2m-1$. Modulo $2m-1$, the inverse of $m$ is $2$, so the residue of $a_{2k}$ is twice the residue of its numerator. Substituting this residue in the original sum computes the next remainder exactly. Its nonvanishing proves $N(k)=2k+1$. Here is a complete C++17 verifier. The integers in its products are residues less than $8191$, and each sum has at most $4095$ terms. Every sum is therefore less than $4095\cdot8190^2<2^{39}$, within the range of \texttt{uint64\_t}. The output is \texttt{8184 49 0}: $8184$ checked orders, $49$ with first failure at $2k+1$, and zero orders passing both tests. \begin{verbatim} #include #include #include using U = std::uint64_t; U numerator(const std::vector& a, int n, int k) { U v = 0; for (int j = 1; j <= k/2; ++j) v += a[n-j] * a[n-k+j]; return v; } std::vector block(int k, U q) { std::vector a(2*k+2, 1); for (int n = k; n < 2*k; ++n) a[n] = numerator(a,n,k) % q; return a; } int main() { int checked = 0, delayed = 0, escaped = 0; for (int k = 8; k <= 8191; ++k) { U m = k/2; auto a = block(k,m); U r = numerator(a,2*k,k) % m; if (r == 0) { U q = 2*m-1; a = block(k,q); a[2*k] = 2*(numerator(a,2*k,k) % q) % q; r = numerator(a,2*k+1,k) % q; if (r != 0) ++delayed; else ++escaped; } ++checked; } std::cout << checked << ' ' << delayed << ' ' << escaped << '\n'; return escaped != 0; } \end{verbatim} For every $k$, the terms preceding $a_{2k}$ are integers, since the only divisors used there are initial entries equal to $1$. The finite verification and the estimate for $k\ge8192$ prove both bounds in Theorem~\ref{thm:p8}. \clearpage \section{Prime powers in Somos-4}\label{sec:valuations} In this section, \[ a_0=a_1=a_2=a_3=1,\qquad a_{n+4}a_n=a_{n+3}a_{n+1}+a_{n+2}^2\qquad(n\geq0). \] For a prime \(p\), \(v_p\) denotes the valuation normalized by \(v_p(p)=1\). \begin{theorem}\label{thm:valuation} If \(p\) is an odd prime and \(p\mid a_n\), then, for every integer \(j\geq0\), \[ n_j=\frac{(2n-3)p^j+3}{2},\qquad v_p(a_{n_j})=v_p(a_n)+j. \] Moreover, \[ v_2(a_n)= \begin{cases}1,&n\equiv4\pmod5,\\0,&n\not\equiv4\pmod5,\end{cases} \qquad 37\nmid a_n . \] \end{theorem} We give the coordinate and local calculations explicitly. The elliptic curve correspondence is due to the earlier work discussed in the introduction; the coordinates below are those in \cite[proof of Corollary 1.2]{JonesRouse}. \subsection{The elliptic curve and its denominators} Set \[ F(a,b,c,d)=a^2d^2-4abcd+ac^3+b^3d+b^2c^2. \] With \(e=(bd+c^2)/a\), expansion gives \begin{equation}\label{eq:quartic-invariant} F(b,c,d,e)=\frac ea F(a,b,c,d). \end{equation} The recurrence first defines positive rational numbers. Since \(F(1,1,1,1)=0\), equation~\eqref{eq:quartic-invariant} proves that \(F\) vanishes on every consecutive quadruple. We now prove simultaneously that all terms are integers and that the terms in every consecutive quadruple are pairwise coprime. Suppose this holds for \(a,b,c,d\). Reduction of \(F=0\) modulo \(a\) gives \(b^2(bd+c^2)\equiv0\pmod a\). Since \(\gcd(a,b)=1\), the next term \(e=(bd+c^2)/a\) is an integer. A prime dividing \(e\) and \(b\), or \(e\) and \(d\), would divide \(c\). A prime dividing \(e\) and \(c\) would divide \(bd\). Each possibility contradicts pairwise coprimality of \(a,b,c,d\). Thus \(b,c,d,e\) are pairwise coprime, completing the induction. Consider \[ E:\ y^2+y=x^3-x,\qquad P=(0,0). \] A bar denotes reduction modulo the prime under consideration. The identity of the group law is denoted by \(O\), and \([m]P\) denotes the integer multiple of \(P\). The equation has discriminant \(37\), and \(2P=(1,0)\). It is minimal at every prime: discriminant valuations of isomorphic integral Weierstrass equations differ by a multiple of \(12\), whereas the present valuations are at most \(1\). For \(n\geq1\), define \begin{equation}\label{eq:somos-elliptic} X_n=\frac{a_n^2-a_{n-1}a_{n+1}}{a_n^2},\qquad Y_n=\frac{a_{n-1}^2a_{n+2}-2a_{n-1}a_na_{n+1}}{a_n^3}. \end{equation} We claim that \begin{equation}\label{eq:odd-orbit} (X_n,Y_n)=[2n-3]P. \end{equation} For its verification, write \[ X=\frac{b^2-ac}{b^2},\qquad Y=\frac{a^2d-2abc}{b^3}, \qquad (a,b,c,d)=(a_{n-1},a_n,a_{n+1},a_{n+2}). \] Then \[ Y^2+Y-X^3+X=\frac{a^2}{b^6}F(a,b,c,d)=0. \] Positivity gives \(X\neq1\). The line through \((X,Y)\) and \((1,0)\) has slope \(\lambda=(2bc-ad)/(bc)\). Hence their sum has coordinates \[ X_+=\lambda^2-X-1,\qquad Y_+=-Y-1+\lambda(X-X_+). \] Let \(X',Y'\) be the expressions for \(X,Y\) with \((a,b,c,d)\) replaced by \((b,c,d,e)\). The two required identities are \[ X_+-X'=\frac{F(a,b,c,d)}{b^2c^2},\qquad Y_+-Y'=\frac{a^2d-2abc-b^3}{ab^3c^3}F(a,b,c,d). \] They prove that \((X_{n+1},Y_{n+1})=(X_n,Y_n)+2P\). Since \((X_1,Y_1)=(0,-1)=-P\), equation~\eqref{eq:odd-orbit} follows. If a rational point has \(v_p(x)<0\), its equation forces \(v_p(y)<0\) and \(2v_p(y)=3v_p(x)\): on each side the term of lowest valuation is unique. Thus every negative valuation of \(x\) is even. For every multiple \([m]P\neq O\), write its reduced \(x\)-coordinate as \[ x([m]P)=\frac{A_m}{B_m^2},\qquad A_m\in\mathbb Z,\quad B_m>0,\quad \gcd(A_m,B_m)=1. \] Negation sends \((x,y)\) to \((x,-y-1)\), so \(B_{-m}=B_m\). Pairwise coprimality yields \[ \gcd(a_n^2-a_{n-1}a_{n+1},a_n)=1. \] Consequently, \begin{equation}\label{eq:denominator-correspondence} B_{|2n-3|}=a_n\qquad(n\geq0). \end{equation} At \(n=0\) this says \(B_3=1\), as follows from \(3P=(-1,-1)\). The construction already shows that each odd multiple appearing here is finite. The next calculation also shows that \(P\) has infinite order. \subsection{A local valuation calculation} We recall the formal group of an integral Weierstrass equation at \(O\); see \cite[Chapter IV]{Silverman}. Put \(t=-x/y\) and \(w=-1/y\). For the present equation, \[ w=t^3+(1-t)w^2. \] Successive comparison of coefficients gives \(w=t^3U(t)\), with \(U(t)\in1+t\mathbb Z[[t]]\). The chord-and-tangent law at \(O\) therefore defines a one-dimensional commutative formal group \[ \mathcal F(T_1,T_2)\in\mathbb Z[[T_1,T_2]],\qquad \mathcal F(T_1,T_2)=T_1+T_2+\text{terms of total degree at least }2. \] For any prime \(p\), the points reducing to \(O\) form the group parametrized by \(t\in p\mathbb Z_p\), with \(t=0\) corresponding to \(O\). For \(t\neq0\), \begin{equation}\label{eq:parameter-denominator} x=t^{-2}U(t)^{-1},\qquad y=-t^{-3}U(t)^{-1}. \end{equation} Here \(U(t)\) is a unit. Conversely, \(v_p(x(R))<0\) gives \(v_p(y(R))=3v_p(x(R))/2\), so the projective point \((x(R):y(R):1)\) reduces to \((0:1:0)=O\). Thus, if \(x(R)=A/B^2\) is reduced, then \(p\mid B\) exactly when \(R\) reduces to \(O\), and in this case \(v_p(t(R))=v_p(B)\). This construction uses the smooth point \(O\) also when \(p=37\). Let \(R\) be a nonzero point in this formal neighborhood and put \(e=v_p(t(R))\geq1\). For an integer \(u\) prime to \(p\), the integral multiplication series \[ [u]_{\mathcal F}(T)=uT+O(T^2) \] shows that \(v_p(t([u]R))=e\). For odd \(p\), the precise formula for every nonzero integer \(u\) is \begin{equation}\label{eq:formal-lifting} v_p(t([u]R))=e+v_p(u). \end{equation} Here is the convergence and valuation argument. The invariant differential of the formal group has the form \[ \omega(T)=(1+c_1T+c_2T^2+\cdots)\,dT,\qquad c_i\in\mathbb Z_p. \] Its integral \[ \log_{\mathcal F}(T)=T+ \sum_{\nu\geq2}\frac{c_{\nu-1}}{\nu}T^\nu \] satisfies \(\log_{\mathcal F}(\mathcal F(T_1,T_2)) =\log_{\mathcal F}(T_1)+\log_{\mathcal F}(T_2)\). For \(e\geq1\), odd \(p\), and \(\nu\geq2\), \[ \nu e-v_p(\nu)\geq e+1. \] Indeed \(v_p(2)=0\); for \(\nu\geq3\), the inequality \(v_p(\nu)\leq \nu-2\) follows from \(p\geq3\). The estimates tend to infinity with \(\nu\), so the series converges and preserves the valuation of every nonzero element of \(p\mathbb Z_p\). Moreover, \[ \log_{\mathcal F}([u]_{\mathcal F}(t)) =u\log_{\mathcal F}(t). \] For \(u\neq0\) and \(t\neq0\), the right-hand side is nonzero. Consequently \([u]_{\mathcal F}(t)\neq0\), and taking valuations proves \eqref{eq:formal-lifting}. This also proves that the formal neighborhood has no nonzero torsion for odd \(p\). Equation~\eqref{eq:somos-elliptic} gives \(7P=(-5/9,8/27)\), whose parameter \(15/8\) has positive \(3\)-adic valuation. Hence \(7P\), and therefore \(P\), has infinite order. Combining~\eqref{eq:parameter-denominator} and~\eqref{eq:formal-lifting} gives \begin{equation}\label{eq:elliptic-valuation} v_p(B_{mu})=v_p(B_m)+v_p(u) \quad\text{if \(p\) is odd and \(p\mid B_m\)}. \end{equation} This is the elliptic-denominator multiplication law \cite[Lemma 3.1]{EverestMcLarenWard}, in the square-root denominator convention. \begin{proof}[Proof of Theorem~\ref{thm:valuation}] If an odd prime \(p\) divides \(a_n\), then \(n\geq4\). Set \(m=2n-3>0\). Equations~\eqref{eq:denominator-correspondence} and~\eqref{eq:elliptic-valuation}, with \(u=p^j\), give \[ v_p(B_{mp^j})=v_p(a_n)+j. \] The integer \(mp^j\) is positive and odd, so its corresponding Somos index is \(n_j=(mp^j+3)/2\). This proves the first assertion. At \(p=2\), the reduced curve has exactly five points: \[ O,\quad(0,0),\quad(0,1),\quad(1,0),\quad(1,1). \] Thus the reduction of \(P\) has order \(5\). Good reduction and the kernel of reduction give \(2\mid B_m\) if and only if \(5\mid m\). Since \(5P=(1/4,-5/8)\), one has \(v_2(B_5)=1\). For odd \(m=5u\), multiplication by the odd integer \(u\) preserves this valuation. Therefore \(v_2(B_m)=1\) for odd multiples of \(5\), and is zero at all other odd indices. The condition \(5\mid2n-3\) is \(n\equiv4\pmod5\). At \(37\), the unique singular affine point is \((5,18)\). The reduction of \(P\) lies in the smooth part. The addition law there gives \[ \begin{array}{c|rrrr} m&1&2&4&8\\ \hline [m]\overline P&(0,0)&(1,0)&(2,34)&(26,3)\\ \hline m&16&18&19&38\\ \hline [m]\overline P&(15,31)&(31,33)&(27,18)&O \end{array} \] The slopes for the first four doublings are \(36,2,20,20\), respectively, modulo \(37\). Adding \(2\overline P\) to \(16\overline P\) gives \(18\overline P\) with slope \(26\), and adding \(\overline P\) gives \(19\overline P\) with slope \(13\). The point \((27,18)\) is a smooth nonidentity point with \(2y+1=0\), so it has order \(2\). Since \(2\overline P\neq O\) and \(19\overline P\neq O\), the order of \(\overline P\) is \(38\). Points with smooth reduction form a subgroup, and reduction on this subgroup is a homomorphism whose kernel is the formal neighborhood of \(O\); see \cite[Proposition VII.2.1]{Silverman}. Every multiple of \(P\) belongs to this subgroup. Hence \(37\mid B_m\) if and only if \(38\mid m\). An odd index \(m=2n-3\) cannot satisfy this condition. \end{proof} The same argument describes every prime-power divisibility index. \begin{corollary}\label{cor:all-valuations} Let \(p\neq2,37\) be prime and let \(d_p\) be the order of \(\overline P\) in \(E(\mathbb F_p)\). If \(d_p\) is even, then \(p\nmid a_n\) for every \(n\). If \(d_p\) is odd, put \(w_p=v_p(B_{d_p})\). Then \[ v_p(a_n)= \begin{cases} 0,&d_p\nmid 2n-3,\\ w_p+v_p((2n-3)/d_p),&d_p\mid 2n-3. \end{cases} \] When \(d_p\) is odd, for every integer \(s\geq1\), \[ p^s\mid a_n \quad\Longleftrightarrow\quad d_p p^{\max(0,s-w_p)}\mid2n-3. \] \end{corollary} \begin{proof} Good reduction gives \(p\mid B_m\) exactly when \(d_p\mid m\). For \(m=d_pu\), apply~\eqref{eq:elliptic-valuation}; then use~\eqref{eq:denominator-correspondence}. Since \(d_p p^{\max(0,s-w_p)}\) is odd, the displayed divisibility condition describes one residue class of \(n\) modulo that integer. \end{proof} \section{Three-term polynomial recurrences}\label{sec:three-term} Let $x,y,z$ be algebraically independent. For integers \[ 00$. For $0\le tp+q+r$ remains to be considered outside the two families \eqref{p6:good}. \subsection{The region \texorpdfstring{$2p\le rp$. At $y=0$, the linear $p,2p$ sequence is constant on blocks of length $p$, with block values $P_0=0$, $P_1=-x$, $P_m=x(P_{m-1}-P_{m-2})$ for $m\ge2$. Their leading terms are $-x^m$ for $m\ge1$. If $e\ge2p$, equation \eqref{p6:K} at $t=0$ gives $K_0=-xC_{e-p}\ne0$ on this axis. If $e=p+a$ with $1\le a2d-p$&$z=0$, $t=d-a$&$K_t=x^2$\\ $ap$, and \eqref{p6:K} gives \[ K_{p-1}=-x\bigl(C_{2p+a+d-1}-C_{2p+a-d-1}\bigr). \] The second index lies in $[p,q)$, whereas the first is at least $q$; \eqref{p6:Ccoeff} says that their coefficients of $x$ are respectively $-1$ and $0$. All the indices used on the $z=0$ axis are less than $k-q=e+2p$, so that the linear recurrence applies. Now let $2p2$&$[x^{m+1}]J=2-m$\\ $M=m=2$, $u\ge b$&$J=x^2$\\ \bottomrule \end{tabular} \end{center} For the last identity the two polynomials are $C_n=-x^2$ and $C_{n+q-p}=-x^4-x^3+x^2$; all other rows follow by using $v=0$ or $1$ in \eqref{p6:Ccoeff} for the two largest possible degrees. Each row is nonzero. If $q<2p$, write $q=p+d$, $r=2p+\delta$ with $0<\delta= 2p}}\label{p6:kernel-section} We next prove the remaining $S$ obstruction when $r>p+q$ and $q\ge2p$. Here $e\ge f>p$. On the $z=0$ axis, all six entries in \eqref{p6:K} are values of the linear $p,q$ sequence. For $0\le t0$ and, by symmetry, $a\le b$. After subtracting the common degree $\alpha+\beta+\gamma$, the degrees of $D,EFH,AH,GF$ are respectively \[ \begin{split} &\left\lfloor\frac{a+b+c+t}{p}\right\rfloor,\quad \left\lfloor\frac{a+t}{p}\right\rfloor+ \left\lfloor\frac{b+t}{p}\right\rfloor+ \left\lfloor\frac{c+t}{p}\right\rfloor,\\ &\left\lfloor\frac{a+c+t}{p}\right\rfloor+ \left\lfloor\frac{b+t}{p}\right\rfloor,\quad \left\lfloor\frac{b+c+t}{p}\right\rfloor+ \left\lfloor\frac{a+t}{p}\right\rfloor. \end{split} \] Their leading coefficients are $-1,-1,1,1$. The term $-FH$ is of smaller degree than $EFH$. Let $w=(c+t)\bmod p$. Comparing the four floors, the sum of leading coefficients at their largest degree vanishes only for the following $w$: \begin{center} \begin{tabular}{ll} \toprule Range of $t$&Allowed $w$\\ \midrule $[0,p-b)$&$[0,p-a-b)\cup[p-b,p-a)$\\ $[p-b,p-a)$&$[0,p-a-b)\cup[p-b,p-a)\cup[2p-a-b,p)$\\ $[p-a,p)$&$[p-b,p-a)\cup[2p-a-b,p)$\\ \bottomrule \end{tabular} \end{center} Intervals are intersected with $[0,p)$. For verification, put $A_0=\mathbf1_{t\ge p-a}$ and $B_0=\mathbf1_{t\ge p-b}$. The four degrees, after removing their common $\lfloor(c+t)/p\rfloor$, are \[ \left\lfloor\frac{a+b+w}{p}\right\rfloor,\quad A_0+B_0,\quad \mathbf1_{w\ge p-a}+B_0,\quad\mathbf1_{w\ge p-b}+A_0; \] the table is obtained by taking $(A_0,B_0)=(0,0),(0,1),(1,1)$. If all $t$ gave cancellation, rotating the first interval by $c$ forces it into $[p-b,p-a)$: its other allowed component has smaller length. Similarly the last interval must rotate into the same middle component. The resulting inequalities are \[ p-b\le c\le b-a,\qquad p+a-b\le c\le p-a. \] They imply $a

0$. It remains to treat one zero residue. By symmetry take $e=\alpha p$ with $\alpha\ge2$. Initially put $t=p-1$, write $H=C_{hp+w}$ with $h\ge2$, $0\le w0$, degree equality forces the relevant residues below $c$, and the same values occur. For $\gamma=2,c=0$, they are $B_2+\alpha$ for $J_{q-p+t}$ and $B_2+\alpha+1$ for $J_{f+q-p+t}$; their combination cancels, but $-FH$ contributes $-1$ at exactly this degree. These values can also be read directly from the following expansion. For $n=mp+v$, $m\ge2$, $0\le v2p+r,\end{cases}\quad \eta=\begin{cases}s&k\le2p+q,\\1&k>2p+q,\end{cases}\quad \psi=\begin{cases}s&k\le p+q+r,\\1&k>p+q+r.\end{cases} \] The first-block recurrence gives \[ \begin{gathered} u_{2p}=z_0s\theta,\qquad [y]b_{2p}=s(\eta-\theta),\\ u_{p+r}=x\theta u_r,\qquad [y]b_{p+r}=x\theta[y]b_r,\\ u_{p+q}=z_0s\psi,\qquad u_{k+p-q}=xs u_{k-q},\\ b_{k+p-r}=xs b_{k-r}+ys b_h. \end{gathered} \] For instance, the three small factors in the last equation have indices $2p-r,p+q-r,p$; their values are $s,s,0$. In $R_0$ the $p$ summand vanishes, while the $q$ summand is divisible by $y^2$. Thus $R_0=zb_{k-r}b_r+O(y^2)$. Substituting the displayed equations in \eqref{p6:U}, every derivative of $b_r$ or $b_{k-r}$ has the factor $z_0+xs=0$. The remaining coefficient is \begin{equation}\label{p6:small-U} [y]U=xz_0s\bigl(u_r(\eta-\theta)u_{k-r} +s^2\psi u_{k-q}+s\theta u_ru_h\bigr). \end{equation} The base sequence $u_t$ is related to $l_t$ by \begin{equation}\label{p6:small-boundary} u_t=\begin{cases} l_t,&t2p+q$, the third row is accordingly replaced by \[ s^4\bigl((1-s)(1+s)^2-xs^2\bigr). \] Every displayed rational function is nonzero: the brackets with constant term $1$ retain that term at $x=0$, and the other cases are immediate for $\delta=0,1$. For $2p+rp+q+r$, the bracket is $s^2g_N$, where \begin{equation}\label{p6:gN} g_N=l_N+s l_{N-r+p}+s^2l_{N-r},\qquad N=k-q>p+r. \end{equation} We verify that $g_N\ne0$ for all these $N$. Let $\epsilon=x+1$ and write $N=jr+h$, $0\le hp+r$ forces $j\ge2$, and the last term of \eqref{p6:gN} has a unique pole of order $j+2$. If $h\ge r-p$, the coefficient at that order is $2(-1)^j$ for $hp+q$}{p < q < 2p, r > p+q}}\label{p6:last-wedge} This is the only remaining parameter region. Set \[ d=q-p,\qquad f=r-q>p,\qquad e=k-q-r\ge f, \] and retain the $T=0$ notation from Subsection~\ref{p6:T-prelim}. Since $r>p+q$, the first-block equations give \[ \begin{gathered} b_p=0,\quad b_q=y(s-1),\quad b_{2p}=z_0+y(s-1),\quad u_{p+q}=z_0,\\ b_{k+p-r}=xb_{k-r}+yb_{e+p},\qquad b_{p+r}=xb_r+yb_{p+f},\\ u_{k+p-q}=xs u_{k-q}+z_0u_{e+p}u_{p+f}. \end{gathered} \] As in \eqref{p6:small-U}, substitution in $U$ cancels the derivative terms containing $z_0+xs$. It gives \begin{equation}\label{p6:Q} [y]U=xz_0Q,\quad Q=(su_r-V)u_{k-r}+sV u_{e+p}+s^2u_{k-q},\quad V=u_r-su_{p+f}. \end{equation} For explicit verification of the cancellation, write $R_0=zb_{k-r}b_r+O(y^2)$ and insert the preceding two equations for $b_{k+p-r},b_{p+r}$ into \eqref{p6:U}; the coefficients multiplying $[y]b_r$ and $[y]b_{k-r}$ both vanish. The remaining terms factor as $xz_0$ times the displayed $Q$. Write $f=\beta p+b$, $0\le bf$, whereas $u_r=l_r+s^3$ if $e=f$. \subsubsection{The range \texorpdfstring{$e3d$ and $d\le t2d$, take $t=d$. The indices are $4p+a+2d$ and $3p+b+2d$. Before reducing modulo $p$, their excesses are less than $4d$ and $3d$; crossing a $p$ block only decreases the residue and increases the block number. Again both are nonzero. Finally, if $p\le2d$, then $a+b3d$. Thus $K_0=(\delta+2\eta)x^4\ne0$. This finishes $e= r}} Write $e=jr+h$, where $j\ge1$ and $0\le hn}}\binom{a+b}{a}x^a z_0^b. \end{equation} Indeed the unrestricted sum is $(1-x-z_0)^{-1}=1+x$, which transforms \eqref{p6:lgen} into \eqref{p6:tail}. Since $z_0=-x^2+O(x^3)$, a pair $(a,b)$ has smallest degree $a+2b$. For $n=jr+h$, the minimal pairs and resulting terms are \begin{equation}\label{p6:minimum} l_{jr+h}\sim\begin{cases} (-1)^j(j+1)x^{2j+1},&0\le ha$ the residual position after crossing one $r$ block is at least \[ r+(a-i)p+u\ge(a+1)p+b+d+u>(a+1)p. \] Including the factor $x^i$ puts every unlisted term strictly above the orders just used. The indices $i\le a$ follow directly from \eqref{p6:minimum}. Every coefficient in the table is nonzero for $j\ge1$. Finally take $r<3p$. Write $r=2p+\rho$, where $d<\rhop+q+r$, the regions $rp+q$ with $q\ge2p$, and $r>p+q$ with $q<2p$. In each case \eqref{p6:window} has a nonpolynomial term by index $2k+p$. The discussion following \eqref{p6:window} proves the denominator assertion and finishes Theorem~\ref{p6:classification}. \begin{thebibliography}{99} \raggedright \bibitem{BMPW} M. Bousquet-M\'elou, J. Propp, and J. West, \emph{Perfect matchings for the three-term Gale--Robinson sequences}, Electron. J. Combin. \textbf{16} (2009), no.~1, R125. \href{https://doi.org/10.37236/214}{doi:10.37236/214}. \bibitem{CarrollSpeyer} G. D. Carroll and D. E. Speyer, \emph{The cube recurrence}, Electron. J. Combin. \textbf{11} (2004), no.~1, R73. \href{https://arxiv.org/abs/math/0403417}{arXiv:math/0403417}. \bibitem{EverestMcLarenWard} G. Everest, G. McLaren, and T. Ward, \emph{Primitive divisors of elliptic divisibility sequences}, J. Number Theory \textbf{118} (2006), 71--89. \href{https://doi.org/10.1016/j.jnt.2005.08.002}{doi:10.1016/j.jnt.2005.08.002}. \bibitem{FZLaurent} S. Fomin and A. Zelevinsky, \emph{The Laurent phenomenon}, Adv. Appl. Math. \textbf{28} (2002), 119--144. \href{https://doi.org/10.1006/aama.2001.0770}{doi:10.1006/aama.2001.0770}. \bibitem{Gale} D. Gale, \emph{The strange and surprising saga of the Somos sequences}, Math. Intelligencer \textbf{13} (1991), no.~1, 40--42. \href{https://doi.org/10.1007/BF03024070}{doi:10.1007/BF03024070}. \bibitem{GaleUpdate} D. Gale, \emph{Somos sequence update}, Math. Intelligencer \textbf{13} (1991), no.~4, 49--50. \href{https://doi.org/10.1007/BF03028343}{doi:10.1007/BF03028343}. \bibitem{Hone} A. N. W. Hone, \emph{Elliptic curves and quadratic recurrence sequences}, Bull. London Math. Soc. \textbf{37} (2005), 161--171. \href{https://doi.org/10.1112/S0024609304004163}{doi:10.1112/S0024609304004163}; corrigendum, \textbf{38} (2006), no.~5, 741--742. \href{https://doi.org/10.1112/S0024609306018844}{doi:10.1112/S0024609306018844}. \bibitem{HKW} A. N. W. Hone, T. E. Kouloukas, and C. Ward, \emph{On reductions of the Hirota--Miwa equation}, SIGMA \textbf{13} (2017), 057, 17 pp. \href{https://doi.org/10.3842/SIGMA.2017.057}{doi:10.3842/SIGMA.2017.057}. \bibitem{HoneSwart} A. N. W. Hone and C. Swart, \emph{Integrality and the Laurent phenomenon for Somos 4 and Somos 5 sequences}, Math. Proc. Cambridge Philos. Soc. \textbf{145} (2008), 65--85. \href{https://doi.org/10.1017/S030500410800114X}{doi:10.1017/S030500410800114X}. \bibitem{JonesRouse} R. Jones and J. Rouse, \emph{Galois theory of iterated endomorphisms}, Proc. London Math. Soc. (3) \textbf{100} (2010), 763--794. \href{https://doi.org/10.1112/plms/pdp051}{doi:10.1112/plms/pdp051}. \bibitem{Malouf} J. L. Malouf, \emph{An integer sequence from a rational recursion}, Discrete Math. \textbf{110} (1992), 257--261. \href{https://doi.org/10.1016/0012-365X(92)90714-Q}{doi:10.1016/0012-365X(92)90714-Q}. \bibitem{OEIS030127} OEIS Foundation, \emph{A030127: Term at which \(n\)-Somos sequence first becomes nonintegral}, comments by P. D. Hanna (2004) and M. Somos (2014), \url{https://oeis.org/A030127}, accessed 18 November 2026. \bibitem{OEIS101591} OEIS Foundation, \emph{A101591: Indices \(k\geq8\) designating the \(k\)-Somos sequences that first become nonintegral at odd-indexed terms}, P. D. Hanna (2004), \url{https://oeis.org/A101591}, accessed 18 November 2026. \bibitem{Propp} J. Propp, \emph{A bare-bones chronology of Somos sequences}, July 2006, \url{https://faculty.uml.edu/jpropp/somos/chronology.html}. \bibitem{Robinson} R. M. Robinson, \emph{Periodicity of Somos sequences}, Proc. Amer. Math. Soc. \textbf{116} (1992), 613--619. \href{https://doi.org/10.1090/S0002-9939-1992-1140672-5}{doi:10.1090/S0002-9939-1992-1140672-5}. \bibitem{Silverman} J. H. Silverman, \emph{The Arithmetic of Elliptic Curves}, second ed., Graduate Texts in Mathematics, vol.~106, Springer, Dordrecht, 2009. \bibitem{SomosProblem} M. Somos, \emph{Problem 1470}, Crux Math. \textbf{15} (1989), 208. \bibitem{SomosHistory} M. Somos, \emph{Brief history of the Somos sequence problem}, 7 May 1990, updated 30 October 1994, \url{https://faculty.uml.edu/jpropp/somos/history.txt}. \bibitem{SomosRefined} M. Somos, \emph{The refined Gale--Robinson polynomial sequence conjecture}, 11 April 1999, updated 7 September 2022, \url{https://grail.eecs.csuohio.edu/~somos/robinson.html}. \bibitem{Speyer} D. E. Speyer, \emph{Perfect matchings and the octahedron recurrence}, J. Algebraic Combin. \textbf{25} (2007), 309--348. \href{https://arxiv.org/abs/math/0402452}{arXiv:math/0402452}. \bibitem{Stanley} R. P. Stanley, \emph{Bijective proof problems}, 18 August 2009, \url{https://math.mit.edu/~rstan/bij.pdf}. \bibitem{Swart} C. S. Swart, \emph{Elliptic curves and related sequences}, Ph.D. thesis, Royal Holloway, University of London, 2003. \bibitem{Uchida} Y. Uchida, \emph{Valuations of Somos 4 sequences and canonical local heights on elliptic curves}, Math. Proc. Cambridge Philos. Soc. \textbf{150} (2011), 385--397. \href{https://doi.org/10.1017/S030500411000068X}{doi:10.1017/S030500411000068X}. \bibitem{Kamp} P. H. van der Kamp, \emph{Somos-4 and Somos-5 are arithmetic divisibility sequences}, J. Difference Equations Appl. \textbf{22} (2016), no.~4, 570--580, corrected version. \href{https://doi.org/10.1080/10236198.2015.1113272}{doi:10.1080/10236198.2015.1113272}; erratum \href{https://doi.org/10.1080/10236198.2016.1193956}{doi:10.1080/10236198.2016.1193956}. \bibitem{WardRepetition} M. Ward, \emph{The law of repetition of primes in an elliptic divisibility sequence}, Duke Math. J. \textbf{15} (1948), 941--946. \bibitem{WardMemoir} M. Ward, \emph{Memoir on elliptic divisibility sequences}, Amer. J. Math. \textbf{70} (1948), 31--74. \end{thebibliography} \end{document}