On the approximation gain for a ​ b ​ c abc -triples Benne de Weger Abstract The concept of approximation gain was introduced recently by Müller and Taktikos for some a ​ b ​ c abc -triples related to convergents of surds, where there is a relatively large gap between min ⁡ { a , b , c } \min\{a,b,c\} and max ⁡ { a , b , c } \max\{a,b,c\} . This note proposes a generalization of the concept to all a ​ b ​ c abc -triples, with several variants. Extensive numerical computations are provided. 1 Introduction We start with common definitions. Definition 1 The radical \rad(n) of a positive integer n is defined as the product of its prime divisors. Definition 2 Let a,b,c be positive coprime integers such that a+b=c . (a) The quality of the triple (a,b,c) is defined as \qu(a,b,c)=\dfrac{\log c}{\log\rad(abc)}. (b) A triple (a,b,c) is called an abc -triple when \qu(a,b,c)>1 . The abc -conjecture says that for every \epsilon>0 there are only finitely many abc -triples with \qu(a,b,c)>1+\epsilon . The concept of approximation gain for abc -triples was introduced recently by Müller and Taktikos [ 1 ] . Their starting point is the surd k^{1/s} where k,s are positive integers, with convergents \dfrac{p}{q} , so that d=p^{s}-kq^{s} is relatively close to 0 . This gives a triple of quality at least close to 1 . To be a bit more precise, when e is the partial quotient occurring just after the convergent \dfrac{p}{q} , then by continued fraction theory \left|k^{1/s}-\dfrac{p}{q}\right|\approx\dfrac{1}{eq^{2}} , which implies |d|=\left|p^{s}-kq^{s}\right|=\left|p-k^{1/s}q\right|\displaystyle\sum_{i=0}^{s-1}p^{i}\left(k^{1/s}q\right)^{s-1-i}\approx\dfrac{sk^{1-1/s}}{e}q^{s-2}. For the proper permutation a,b,c of d,p^{s},kq^{s} this gives for the abc -quality in the worst case (where dkpq is squarefree) \qu(a,b,c)=\dfrac{\log\max\{|d|,p^{s},kq^{s}\}}{\log(|d|pkq)}\approx\dfrac{\log(kq^{s})}{\log\left(\dfrac{sk^{2}}{e}q^{s}\right)}\approx\dfrac{\log(kq^{s})}{\log(kq^{s})+\log(ks)-\log e}. The intuition is that the partial quotients have a tendency to be small often, and almost certainly e\ll q for growing q . This means that indeed the quality will tend towards 1 . When e is big enough it comes above 1 . For some k and s this actually does happen, but the abc -conjecture implies that big partial quotients e are rare and can’t be extremely big. To quantify this a bit more elegantly, Müller and Taktikos split up the abc -quality in a product of two terms: the approximation gain \dfrac{\log\max\{p^{s},kq^{s}\}}{\log(|d|pkq)} measuring the effect on the abc -quality of how well \dfrac{p}{q} approximates k^{1/s} , and the cofactor called power gain \dfrac{\log(|d|pkq)}{\log\rad(|d|pkq)} measuring the other effects. For the remainder of this note we assume without loss of generality that a\dfrac{1}{3} , and \rpg(a,b,c)<3\qu(a,b,c) . 2.3 Examples As an example we take Reyssat’s abc -triple 2+3^{10}\cdot 109=23^{5} , of record quality 1.62991\ldots , with d=5 . This gives b^{\prime}=109 , c^{\prime}=1 , \beta=3^{2} , \gamma=23 . We compute the continued fraction 109^{1/5}=[2,1,1,4,77\,733,2,\ldots] , with convergents \dfrac{2}{1} , \dfrac{3}{1} , \dfrac{5}{2} , \dfrac{23}{9} , \dfrac{1787864}{699599},\ldots , and the convergent \dfrac{23}{9} occurring just before the exceptionally large partial quotient 77733 gives us exactly \beta and \gamma . Data for the real approximation and power gains for Reyssat’s abc -triple can be found in Table 1 . \begin{array}[]{c|cc|cc|c|cc}d&\beta&b^{\prime}&\gamma&c^{\prime}&\rrad_{d}(a,b,c)&\rag_{d}(a,b,c)&\rpg_{d}(a,b,c)\\ \hline\cr 2&3^{5}&109&23^{2}&23&2\cdot 3^{5}\cdot 23^{3}\cdot 109&0.77289\ldots&2.10883\ldots\\ 3&3^{3}&3\cdot 109&23&23^{2}&2\cdot 3^{4}\cdot 23^{3}\cdot 109&0.81715\ldots&1.99461\ldots\\ 4&3^{2}&3^{2}\cdot 109&23&23&2\cdot 3^{4}\cdot 23^{2}\cdot 109&0.97679\ldots&1.66863\ldots\\ 5&3^{2}&109&23&1&2\cdot 3^{2}\cdot 23\cdot 109&1.46283\ldots&1.11421\ldots\\ 6&3&3^{4}\cdot 109&1&23^{5}&2\cdot 3^{5}\cdot 23^{5}\cdot 109&0.59037\ldots&2.76079\ldots\\ 7&3&3^{3}\cdot 109&1&23^{5}&2\cdot 3^{4}\cdot 23^{5}\cdot 109&0.61585\ldots&2.64658\ldots\\ 8&3&3^{2}\cdot 109&1&23^{5}&2\cdot 3^{3}\cdot 23^{5}\cdot 109&0.64363\ldots&2.53236\ldots\\ 9&3&3\cdot 109&1&23^{5}&2\cdot 3^{2}\cdot 23^{5}\cdot 109&0.67403\ldots&2.41814\ldots\\ 10&3&109&1&23^{5}&2\cdot 3\cdot 23^{5}\cdot 109&0.70744\ldots&2.30392\ldots\\ \geq 11&1&3^{10}\cdot 109&1&23^{5}&2\cdot 3^{10}\cdot 23^{5}\cdot 109&0.48918\ldots&3.33188\ldots\\ \end{array} Table 1: d ’th degree real approximation and power gains for Reyssat’s abc -triple. Clearly at degree d=5 the influence of the d th powers is most prominent, and this is where the convergent \dfrac{\gamma}{\beta}=\dfrac{23}{9} to 109^{1/5} shows up (note that this also happens at d=4 ). In this case \rag(a,b,c)=\rag_{5}(a,b,c)=1.46283\ldots . Just for fun we also give a very bad sequence of examples. Let p be a prime such that also p-6 and p+6 are prime. Then \rag(p-6,p+6,2p)=\qu(p-6,p+6,2p)=\dfrac{\log(2p)}{\log(2p(p^{2}-36))} and \rpg(p-6,p+6,2p)=1 . Conjecturally there are infinitely many such triples, and when p\to\infty then \rag(p-6,p+6,2p)\to\dfrac{1}{3} , so the lower bound \dfrac{1}{3} for the approximation gain can not be improved. 2.4 Computational results The real approximation gain was computed for the 14482065 abc -triples below 10^{18} , and for the known 9345651 abc -triples between 10^{18} and 2^{63} , from Bart de Smit’s tables [ 2 ] , containing a wealth of data. The top 10 real approximation gain abc -triples are given in Table 2 . The largest real approximation gain, 1.46283\ldots , occurs, not unexpectedly, for Reyssat’s abc -triple. \begin{array}[]{ccc|cc|c}a&b&c&\rag&\rpg&\qu\\ \hline\cr 2&3^{10}\cdot 109&23^{5}&1.46283\ldots&1.11421\ldots&1.62991\ldots\\ 1&2^{4}\cdot 5&3^{4}&1.29203\ldots&1&1.29203\ldots\\ 3\cdot 11&2^{7}\cdot 5^{6}&7^{6}\cdot 17&1.28721\ldots&1.06552\ldots&1.37156\ldots\\ 5&3^{11}&2^{10}\cdot 173&1.25190\ldots&1.12842\ldots&1.41268\ldots\\ 1&2^{6}\cdot 3\cdot 5\cdot 7\cdot 13^{4}\cdot 17&239^{4}&1.24408\ldots&1.08545\ldots&1.35040\ldots\\ 1&2\cdot 3^{4}\cdot 5\cdot 11^{4}&7\cdot 13\cdot 19^{4}&1.22892\ldots&1&1.22892\ldots\\ 5&2\cdot 3\cdot 257^{3}&467^{3}&1.22139\ldots&1&1.22139\ldots\\ 2^{2}&3^{10}\cdot 5\cdot 19\cdot 109\cdot 67751&23^{10}&1.20645\ldots&1.02740\ldots&1.23951\ldots\\ 2\cdot 61&3^{5}\cdot 7^{5}&11\cdot 13^{5}&1.18821\ldots&1&1.18821\ldots\\ 3\cdot 2713&2^{44}\cdot 23\cdot 107&5\cdot 7^{14}\cdot 17\cdot 751&1.15075\ldots&1.09085\ldots&1.25530\ldots\\ \end{array} Table 2: Top 10 real approximation gains known. The largest real power gain observed is \rpg(a,b,c)=2.94376\ldots at (a,b,c)=(2^{49},7^{7}\cdot 19^{3}\cdot 123499,3^{13}\cdot 5^{5}\cdot 503^{2}) , with \rag(a,b,c)=0.45020\ldots , \qu(a,b,c)=1.32528\ldots . Note that in Table 2 \rag is always close to \qu . Much more data are available on the website [ 6 ] for this paper. 3 The p -adic approximation gain 3.1 Introduction A potential disadvantage of the real approximation gain is that it gives interesting results only when a\ll b , and that is not typical for abc -triples. As an illustration we look at the example of a=2^{19}\cdot 13\cdot 103 , b=7^{11} , c=3^{11}\cdot 5^{3}\cdot 11^{2} , with a pretty good quality 1.45261\ldots , while a is relatively big, indeed a>\dfrac{1}{4}c . For the real approximation gain we get \rag(a,b,c)=\rag_{11}(a,b,c)=\dfrac{\log\left(3^{11}\cdot 5^{3}\cdot 11^{2}\right)}{\log\left(2^{19}\cdot 13\cdot 103\cdot 7\cdot 3\cdot 5^{3}\cdot 11^{2}\right)}=0.65708\ldots, which is quite low compared to the quality. This leads to the following variation on the theme of approximation gains. Instead of looking at a that is small in the real sense, we could look at any of a,b or c that is small in some other sense. Then the p -adic metric comes to mind, because ‘small’ then means ‘divisible by a high power of p ’. Indeed, for good abc -triples one expects high prime powers in all three of a,b,c , meaning that each of them can be considered small in a p -adic sense for some well chosen primes p , or even a combination of primes. 3.2 Examples We start with a number of examples to get a feeling for what we should use as definitions. For a p -adic integer x=\displaystyle\sum_{i=0}^{\infty}x_{i}p^{i} with x_{i}\in\{0,1,\ldots,p-1\} we write x=0.x_{0}x_{1}x_{2}\ldots , where we omit the trailing zeros if x\in\mathbb{Z} . 3.2.1 Example 1 – big gap First let’s take the example of the previous subsection, a=2^{19}\cdot 13\cdot 103 , b=7^{11} , c=3^{11}\cdot 5^{3}\cdot 11^{2} , with the pretty good quality 1.45261\ldots but a quite disappointing real approximation gain of 0.65708\ldots , due to a being big in the real sense. However, a is small in the p -adic sense for p=2 , and b and c both have a big exponent 11 at other primes. So the numbers \dfrac{3^{11}}{7^{11}} and (5^{3}11^{2})^{-1} are pretty close in the 2 -adic sense, indeed, \displaystyle\dfrac{3^{11}}{7^{11}} \displaystyle= \displaystyle 0.10111\,10000\,11110\,01011\,11000\,\ldots, \displaystyle(5^{3}11^{2})^{-1} \displaystyle= \displaystyle 0.10111\,10000\,11110\,01010\,11111\,\ldots, agreeing in exactly the first 19 binary digits, corresponding to the exponent 19 of the prime 2 in a . Turning to 11 th roots, we compute, using Hensel’s Lemma, \displaystyle\dfrac{3}{7} \displaystyle= \displaystyle 0.10100\,10010\,01001\,00100\,10010\,\ldots, \displaystyle(5^{3}11^{2})^{-1/11} \displaystyle= \displaystyle 0.10100\,10010\,01001\,00101\,11111\,\ldots, again agreeing in exactly the first 19 binary digits. Instead of working with the surd itself, we can work with an integer that is at 2 -adic distance at most 2^{19} , indeed, with (5^{3}11^{2})^{-1/11}\approx 0.10100\,10010\,01001\,0010=149797 . Note that, when looking for abc -triples starting from a surd, we don’t know a priori at which exponent of 2 we’re going to find something, so we need to go through a whole set of exponents, which corresponds to not knowing at which point we’ll find a big partial quotient in a continued fraction expansion of a real surd. But when we start out with an abc -triple, we exactly know the right exponent, in this case 19 . The set of all rational approximations \dfrac{x}{y} to a given p -adic integer \alpha of a given approximation order n , on identifying the fraction \dfrac{x}{y} with a vector (x,y)^{\top} , forms a twodimensional lattice \Lambda_{n}(\alpha)=\left\{\left.(x,y)^{\top}\in\mathbb{Z}^{2}\right||x-y\alpha|_{p}\leq p^{-n}\right\}, see [ 3 ] (note that we see lattice points as column vectors). A basis of \Lambda_{n}(\alpha) is given by the columns of the matrix \begin{pmatrix}p^{n}&\alpha_{p}\\ 0&1\end{pmatrix} , where \alpha_{p} is the unique integer in \{0,1,\ldots,p^{n}-1\} with \alpha_{p}\equiv\alpha\pmod{p^{n}} . Finding the best approximations therefore can be done efficiently by lattice basis reduction. Note that twodimensional lattice basis reduction is closely related to the continued fraction algorithm and Euclid’s algorithm. In our example we have the matrix \begin{pmatrix}2^{19}&149797\\ 0&1\end{pmatrix} . The lattice basis reduction algorithm reduces the longer column by the smaller one, meaning that one subtracts the smaller column k times from the longer one where k is chosen optimally so that the resulting vector gets the shortest length, and then repeating this until no more improvement is possible. In this case: \begin{pmatrix}2^{19}&149797\\ 0&1\end{pmatrix}\stackrel{{\scriptstyle k=3}}{{\to}}\begin{pmatrix}74897&149797\\ -3&1\end{pmatrix}\stackrel{{\scriptstyle k=2}}{{\to}}\begin{pmatrix}74897&3\\ -3&7\end{pmatrix}\stackrel{{\scriptstyle k=3874}}{{\to}}\begin{pmatrix}63275&3\\ -27121&7\end{pmatrix}. Clearly the factors k play the role of the partial quotients, and we indeed recover \dfrac{3}{7} as a very good rational approximation to the surd (5^{3}11^{2})^{-1/11} just before a huge k . One could also use a real continued fraction here, namely of \dfrac{149797}{2^{19}} , being [0,3,2,24965,1,2] , and the convergent found just before the big partial quotient 24965 is \dfrac{2}{7} . Its interpretation is that \dfrac{2}{7} is an excellent real approximation to \dfrac{149797}{2^{19}} , so that 7\cdot 149797-2\cdot 2^{19} is an extremely small integer, in fact it is 3 . Taken modulo 2^{19} we find 7\cdot(5^{3}11^{2})^{-1/11}\equiv 7\cdot 149797\equiv 3\pmod{2^{19}} , revealing the very good 2 -adic approximation \dfrac{3}{7} to (5^{3}11^{2})^{-1/11} . Going back to lattices, this precisely corresponds to (-2,7)^{\top} being the coordinate vector of the lattice point (3,7)^{\top} w.r.t. the original basis. How to define a p -adic approximation gain to take advantage of the fact that now we see a as small because it is so 2 -adically? It seems to make sense to take the already defined real approximation gain where in the denominator the a was present in full glory, and simply divide a by the relevant power of 2 , leaving one factor 2 . So in this case it becomes \dfrac{\log\left(3^{11}\cdot 5^{3}\cdot 11^{2}\right)}{\log\left(2\cdot 13\cdot 103\cdot 7\cdot 3\cdot 5^{3}\cdot 11^{2}\right)}=1.05580\ldots, indeed a much better approximation gain than the real one. 3.2.2 Example 2 – multiple primes Let’s also give an example of a case where it makes sense to use more primes at once as p -adic base primes. Let’s pick a=3^{2}\cdot 7^{3} , b=17^{3} , and c=2^{6}\cdot 5^{3} , with quality 1.09863\ldots . The original method of Karsten and Taktikos with d=3 would give b^{\prime}=c^{\prime}=1 , \beta=17 , \gamma=20 , and one gets \dfrac{20}{17} as a ‘good’ approximation to 1^{1/3} , which is a bit silly. It would give a real approximation gain of only \dfrac{\log\left(2^{6}\cdot 5^{3}\right)}{\log\left(3^{2}\cdot 7^{3}\cdot 17\cdot 2^{2}\cdot 5\right)}=0.64824\ldots. A second shot would be to look at the largest exponent, which occurs at the prime 2 , and therefore work 2 -adically. We would get -\dfrac{17}{7} as a rational approximation 2 -adically close to 3^{2/3} , and the 2 -adic approximation gain then is \dfrac{\log\left(2^{6}\cdot 5^{3}\right)}{\log\left(3^{2}\cdot 7\cdot 17\cdot 2\cdot 5^{3}\right)}=0.71910\ldots. The p -adic approach with, e.g., p=7 , does even better with an approximation gain of 0.90123\ldots . But we can even do better, when we look p -adically for the primes 2 and 5 simultaneously. As we have third powers in a and b we look for good rational approximations of -3^{2/3} . We take approximation degrees 2^{6} and 5^{3} . It’s not hard to see that -3^{2/3}\equiv 39\pmod{2^{6}} , -3^{2/3}\equiv 56\pmod{5^{3}} , and the Chinese Remainder Theorem then gives -3^{2/3}\equiv 3431\pmod{2^{6}\cdot 5^{3}} . So we compute the real continued fraction of \dfrac{3431}{2^{6}\cdot 5^{3}} , being [0,2,3,66,1,16] , the approximation corresponding to the biggish partial quotient of 66 is \dfrac{3}{7} , and we find the expected answer by 7\cdot 3431-3\cdot 8000=17 to be \dfrac{17}{7} . For computing the improved approximation gain we replace the powers 2^{6},5^{3} by the primes 2,5 only, and get \dfrac{\log\left(2^{6}\cdot 5^{3}\right)}{\log\left(3^{2}\cdot 7\cdot 17\cdot 2\cdot 5\right)}=0.96855\ldots which looks not too bad. See section 3.4 for a continuation of this example. Clearly with the p -adic case we can make any of the three of a,b,c small in some p -adic sense, and we can get three different p -adic approximation gains. It makes sense to take the biggest of them as the p -adic approximation gain. The exact definitions are written out in section 3.3 . 3.2.3 Example 3 – Reyssat Let’s finally see what happens with Reyssat’s abc -triple 2+3^{10}\cdot 109=23^{5} , with quality 1.62991\ldots , where we found an approximation gain of 1.46283\ldots . Computing with all four occurring primes p=2,3,23,109 we computed the best p -adic approximation gains, see Table 3 . \begin{array}[]{c|cc|c}p&\text{surd}&\text{approximation}&\text{approximation gain}\\ \hline\cr 2&109^{1/5}&\dfrac{23}{9}&1.46283\ldots\\[5.69054pt] 3&2^{-1/5}&\dfrac{1}{23}&1.62991\ldots\\ 23&-\left(\dfrac{2}{109}\right)^{1/10}&\dfrac{3}{1}&1.62991\ldots\\ 109&2^{-1/5}&\dfrac{1}{23}&0.80372\ldots\\ {3,109}&2^{-1/5}&\dfrac{1}{23}&1.62991\ldots\\ \end{array} Table 3: p -adic approximation and power gains for Reyssat’s triple. For p=2 the whole thing is almost trivial as there’s an exponent of only 1 , so we’re computing modulo 2 only, but this still gives the same 2 -adic approximation gain as in the real case. For p=3 and p=23 we reach even the best possible p -adic approximation gains, equal to the quality. For p=109 the approximation gain is not very good, which was to be expected as 109 has only exponent 1 . In general it makes sense to choose one of a,b,c , and take it as a whole as a product of prime powers that is multiple p -adically small. In this case in the enhanced radical the contribution of this number is set to be minimal, namely to only the product of its prime factors with multiplicity 1 , and the p -adic approximation gain so is maximized. 3.3 Definitions We now develop various definitions for p -adic approximation gains. Definition 5 Let (a,b,c) be an abc -triple, and let d\geq 2 be any integer. Let \left(a^{\ast},b^{\ast},c^{\ast}\right) be any permutation of (a,b,c) . (a) Let \beta , b^{\prime} , \gamma , c^{\prime} be positive integers with (for given d ) \beta , \gamma as large as possible such that b^{\ast}=\beta^{d}b^{\prime} , c^{\ast}=\gamma^{d}c^{\prime} . The multiple p -adic enhanced radicals \mrad_{a^{\ast},d}(a,b,c) are defined by \mrad_{a^{\ast},d}(a,b,c)=\rad(a^{\ast})\beta b^{\prime}\gamma c^{\prime}. (b) The d th degree multiple p -adic approximation gains \mpag_{a^{\ast},d}(a,b,c) and the d th degree multiple p -adic power gains \mpg_{a^{\ast},d}(a,b,c) are defined by \mpag_{a^{\ast},d}(a,b,c)=\dfrac{\log c}{\log\mrad_{a^{\ast},d}(a,b,c)},\quad\mpg_{a^{\ast},d}(a,b,c)=\dfrac{\log\mrad_{a^{\ast},d}(a,b,c)}{\log\rad(abc)}. Note that the symbol p in the naming of those approximation gains is not explicitly present at all in the definitions’ formulæ. The multiple p -adic approach uses all primes p in a^{\ast} to get the most out of the p -adic metrics. By taking \rad(a^{\ast}) instead of a^{\ast} itself in the enhanced radical, we take care of the fact that a^{\ast} is multiple p -adically small. Definition 6 For an abc -triple (a,b,c) , the multiple p -adic approximation gain \mpag(a,b,c) and the multiple p -adic power gain \mpg(a,b,c) are defined by \mpag(a,b,c)=\max_{d\geq 2,a^{\ast}\in\{a,b,c\}}\mpag_{a^{\ast},d}(a,b,c),\quad\mpg(a,b,c)=\min_{d\geq 2,a^{\ast}\in\{a,b,c\}}\mpg_{a^{\ast},d}(a,b,c). To compute the multiple p -adic approximation gain one even does not have to do any computation with p -adic numbers. It is totally straightforward to compute it directly from the multiple p -adic enhanced radicals as defined in Definition 5 . Only to understand the nature of the p -adic approximation and the underlying p -adic approximation lattice / continued fraction, one has to understand the p -adics of the situation. One could also be stricter and allow for the p -adic approximation gain only single prime powers. Definition 7 Let (a,b,c) be an abc -triple, and let d\geq 2 be any integer. Let \left(a^{\ast},b^{\ast},c^{\ast}\right) be any permutation of (a,b,c) . Let p be a prime with p\mid a^{\ast} , with exponent e_{p}(a) . (a) Let \beta , b^{\prime} , \gamma , c^{\prime} be positive integers with (for given d ) \beta , \gamma as large as possible such that b^{\ast}=\beta^{d}b^{\prime} , c^{\ast}=\gamma^{d}c^{\prime} . The enhanced radicals \srad_{p,d}(a,b,c) are defined by \srad_{p,d}(a,b,c)=\dfrac{a^{\ast}}{p^{e_{p}(a)-1}}\beta b^{\prime}\gamma c^{\prime}. (b) The d th degree single p -adic approximation gains \sag_{p,d}(a,b,c) and the d th degree single p -adic power gains \spg_{p,d}(a,b,c) are defined by \sag_{p,d}(a,b,c)=\dfrac{\log c}{\log\srad_{p,d}(a,b,c)},\quad\spg_{p,d}(a,b,c)=\dfrac{\log\srad_{p,d}(a,b,c)}{\log\rad(a,b,c)}. Definition 8 For an abc -triple (a,b,c) , the single p -adic approximation gain \sag(a,b,c) and the single p -adic power gain \spg(a,b,c) are defined by \sag(a,b,c)=\max_{d\geq 2,p|abc\textrm{\emph{ prime}}}\sag_{p,d}(a,b,c),\quad\spg(a,b,c)=\min_{d\geq 2,p|abc\textrm{\emph{ prime}}}\spg_{p,d}(a,b,c). 3.4 Example 2 revisited We continue the example from section 3.2.2 . We have 8 choices, namely real (indicated by p=\infty ), the single p -adic cases p=2,3,5,7,17 , and the p -adic combinations \{2,5\},\{3,7\} . The data are in Table 4 , where always d=3 , and by r we denote the enhanced radical. \begin{array}[]{c|cc|c}p&\text{surd}&\text{approximation}&\text{gain}\\ \hline\cr\infty&\left(\dfrac{1}{1}\right)^{1/3}&\dfrac{20}{17}&\rag_{3}=0.64824\ldots\\ 2&\left(\dfrac{9}{1}\right)^{1/3}&\dfrac{17}{7}&\sag_{2,3}=0.71910\ldots\\ 3&\left(\dfrac{1}{1}\right)^{1/3}&\dfrac{20}{17}&\sag_{3,3}=0.70403\ldots\\ 5&\left(\dfrac{9}{1}\right)^{1/3}&\dfrac{17}{7}&\sag_{5,3}=0.70517\ldots\\ 7&\left(\dfrac{1}{1}\right)^{1/3}&\dfrac{20}{17}&\sag_{7,3}=0.90123\ldots\\ 17&-\left(\dfrac{9}{1}\right)^{1/3}&\dfrac{20}{7}&\sag_{17,3}=\mpag_{b,3}=0.90123\ldots\\ \{2,5\}&\left(\dfrac{9}{1}\right)^{1/3}&\dfrac{17}{7}&\mpag_{c,3}=0.96855\ldots\\ \{3,7\}&\left(\dfrac{1}{1}\right)^{1/3}&\dfrac{20}{17}&\mpag_{a,3}=1.01281\ldots\\ \end{array} Table 4: Real and p -adic approximation and power gains for (a,b,c)=(3^{2}\cdot 7^{3},17^{3},2^{6}\cdot 5^{3}) . It turns out that \rag=0.64824\ldots , \mpag=1.01281\ldots , and \sag=0.90123\ldots . 3.5 Computational results The p -adic approximation gains were computed for the 14482065 abc -triples below 10^{18} , and for the known 9345651 abc -triples between 10^{18} and 2^{63} , from Bart de Smit’s tables [ 2 ] . The top 10 multiple p -adic approximation gain abc -triples are given in Table 5 , the top 10 single ones in Table 6 . Both the largest multiple and single p -adic approximation gain, 1.62991\ldots , occurs, not unexpectedly, for Reyssat’s abc -triple. \begin{array}[]{ccc|cc}a&b&c&\mpag&\qu\\ \hline\cr 2&3^{10}\cdot 109&23^{5}&1.62991\ldots&1.62991\ldots\\ 1&2\cdot 3^{7}&5^{4}\cdot 7&1.56788\ldots&1.56788\ldots\\ 37&2^{15}&3^{8}\cdot 5&1.48291\ldots&1.48291\ldots\\ 1&3^{16}\cdot 7&2^{3}\cdot 11\cdot 23\cdot 53^{3}&1.47444\ldots&1.47444\ldots\\ 2^{7}\cdot 5^{2}&7^{6}\cdot 41&13^{6}&1.46192\ldots&1.46192\ldots\\ 1&2^{5}\cdot 3\cdot 5^{2}&7^{4}&1.45567\ldots&1.45567\ldots\\ 1&3\cdot 5^{5}\cdot 47^{2}&2^{18}\cdot 79&1.44965\ldots&1.44965\ldots\\ 89&7\cdot 11^{8}&2^{20}\cdot 3^{3}\cdot 53&1.44774\ldots&1.44774\ldots\\ 1&19\cdot 509^{3}&2^{19}\cdot 3^{4}\cdot 59&1.43835\ldots&1.43835\ldots\\ 11&7^{3}\cdot 167^{2}&2\cdot 3^{14}&1.42832\ldots&1.42832\ldots\\ \end{array} Table 5: Top 10 multiple p -adic approximation gains known. \begin{array}[]{ccc|cc}a&b&c&\sag&\qu\\ \hline\cr 2&3^{10}\cdot 109&23^{5}&1.62991\ldots&1.62991\ldots\\ 1&2\cdot 3^{7}&5^{4}\cdot 7&1.56788\ldots&1.56788\ldots\\ 37&2^{15}&3^{8}\cdot 5&1.48291\ldots&1.48291\ldots\\ 1&3^{16}\cdot 7&2^{3}\cdot 11\cdot 23\cdot 53^{3}&1.47444\ldots&1.47444\ldots\\ 3&5^{3}&2^{7}&1.42656\ldots&1.42656\ldots\\ 2^{6}\cdot 5\cdot 137&3^{14}&13^{6}&1.41369\ldots&1.41369\ldots\\ 5&3^{11}&2^{10}\cdot 173&1.41268\ldots&1.41268\ldots\\ 7^{3}&3^{10}&2^{11}\cdot 29&1.40956\ldots&1.54707\ldots\\ 7\cdot 13&2\cdot 3^{11}\cdot 5&11^{6}&1.39548\ldots&1.39548\ldots\\ 1&11^{4}\cdot 47&2^{15}\cdot 3\cdot 7&1.34609\ldots&1.34609\ldots\\ \end{array} Table 6: Top 10 single p -adic approximation gains known. The largest multiple p -adic power gain observed is \mpg(a,b,c)=1.69801\ldots at (a,b,c)=(29^{3}\cdot 79^{5},2^{31}\cdot 17^{3}\cdot 31^{2},3^{4}\cdot 5^{9}\cdot 7\cdot 3037^{2}) , with \rag(a,b,c)=0.50951\ldots , \qu(a,b,c)=1.34683\ldots . The largest single p -adic power gain observed is \spg(a,b,c)=2.15027\ldots at (a,b,c)=(7^{10}\cdot 19^{6}\cdot 37,2^{8}\cdot 17^{5}\cdot 271^{2}\cdot 349^{2},3^{16}\cdot 5^{9}\cdot 211^{2}) , with \rag(a,b,c)=0.49929\ldots , \qu(a,b,c)=1.35574\ldots . Note that in Tables 5 and 6 of the 20 cases 19 have power gain equal to 1 . Much more data are available on the website [ 6 ] for this paper. 4 Combining the real and p -adic approximation gains The multiple p -adic approximation gain is always at least as large as the real approximation gain, even in the case of a being extremely small in the real sense. This is because the contribution of a to the enhanced radical is always the best in the p -adic case. So in this case the ‘combined’ real and multiple p -adic approximation gain will just be the multiple p -adic approximation gain. We think this observation is important enough to call it a theorem. Theorem 1 For all abc -triples \mpag(a,b,c)\geq\rag(a,b,c). Proof. Definition 6 shows that \mpag(a,b,c)\geq\mpag_{a^{\ast},d}(a,b,c) for all d and for those permutations (a^{\ast},b^{\ast},c^{\ast}) of (a,b,c) satisfying a^{\ast}=a . For the enhanced radicals we have, by Definitions 3 and 5 , \mrad_{a^{\ast},d}(a,b,c)=\rad(a^{\ast})\beta b^{\prime}\gamma c^{\prime}=\rad(a)\beta b^{\prime}\gamma c^{\prime}\leq a\beta b^{\prime}\gamma c^{\prime}=\rrad_{d}(a,b,c) , so \mpag_{a^{\ast},d}(a,b,c)=\dfrac{\log c}{\mrad_{a^{\ast},d}(a,b,c)}\geq\dfrac{\log c}{\rrad_{d}(a,b,c)} for all d . The result follows by Definition 4 . \square For the single prime case usually \sag(a,b,c)\geq\rag(a,b,c) , but it might happen that \sag(a,b,c)<\rag(a,b,c) , an example is (a,b,c)=(1,2^{19}\cdot 5^{3}\cdot 4909^{3},11^{5}\cdot 31^{4}\cdot 43\cdot 601\cdot 2017) with \sag(a,b,c)=0.90325\ldots , \rag(a,b,c)=0.92954\ldots , \mpag(a,b,c)=1.18190\ldots , \qu(a,b,c)=1.26428\ldots . So in such cases it does make sense to introduce a combined real and single p -adic approximation gain as follows. Definition 9 For an abc -triple (a,b,c) , the combined approximation gain \cag(a,b,c) and the combined power gain \cpg(a,b,c) are defined by \cag(a,b,c)=\max\{\rag(a,b,c),\sag(a,b,c)\},\quad\cpg(a,b,c)=\min\{\rpg(a,b,c),\spg(a,b,c)\}. 5 Computational results We used the data on abc -triples on Bart de Smit’s webpage [ 2 ] , containing all 14 482 065 abc -triples with c<10^{18} , all 9 345 651 abc -triples with 10^{18}\leq c<2^{63} , and 371 special abc -triples with c\geq 2^{63} , namely those of high quality, high merit, or unbeaten. Of all those triples we computed the real and p -adic approximation and power gains. Figure 1 shows the distributions of the approximation and power gains in bins of size 0.01 . Figure 1: Distributions of the approximation and power gains. On the webpage [ 6 ] for this paper we present tables with abc -triples of high gain, according to the selection criteria in Table 7 . \begin{array}[]{c|cccccc}c&\rag>&\text{or }\rpg>&\text{or }\mpag>&\text{or }\mpg>&\text{or }\sag>&\text{or }\spg>\\ \hline\cr<10^{18}&1.10&2.66&1.37&1.57&1.31&1.90\\ \geq 10^{18},<2^{63}&0.90&2.45&1.09&1.51&1.02&1.86\\ \geq 2^{63}&\lx@intercol\text{no additional criteria}\hfil\lx@intercol\end{array} Table 7: Selection criteria for published tables. This resulted in 1133 abc -triples with c<10^{18} , 13302 abc -triples with 10^{18}\geq c<2^{63} and 371 abc -triples with c\geq 2^{63} . Figure 2 shows some scatterplots of those data. The structures one might observe in there are probably mostly caused by our somewhat randomly chosen selection criteria. Figure 2: Scatterplots of observed quality, approximation and power gain data.